Base_JSON/prefixA/json/argmining/2021.argmining-1.2.json

{
    "paper_id": "2021",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T02:09:59.160478Z"
    },
    "title": "Multi-Task and Multi-Corpora Training Strategies to Enhance Argumentative Sentence Linking Performance",
    "authors": [
        {
            "first": "Jan",
            "middle": [],
            "last": "Wira",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Gotama",
            "middle": [],
            "last": "Putra",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tokyo Institute of Technology",
                "location": {
                    "country": "Japan"
                }
            },
            "email": ""
        },
        {
            "first": "Simone",
            "middle": [],
            "last": "Teufel",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tokyo Institute of Technology",
                "location": {
                    "country": "Japan"
                }
            },
            "email": "simone.teufel@cl.cam.ac.uk"
        },
        {
            "first": "Takenobu",
            "middle": [],
            "last": "Tokunaga",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tokyo Institute of Technology",
                "location": {
                    "country": "Japan"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Argumentative structure prediction aims to establish links between textual units and label the relationship between them, forming a structured representation for a given input text. The former task, linking, has been identified by earlier works as particularly challenging, as it requires finding the most appropriate structure out of a very large search space of possible link combinations. In this paper, we improve a state-of-the-art linking model by using multi-task and multi-corpora training strategies. Our auxiliary tasks help the model to learn the role of each sentence in the argumentative structure. Combining multi-corpora training with a selective sampling strategy increases the training data size while ensuring that the model still learns the desired target distribution well. Experiments on essays written by English-as-a-foreign-language learners show that both strategies significantly improve the model's performance; for instance, we observe a 15.8% increase in the F1-macro for individual link predictions.",
    "pdf_parse": {
        "paper_id": "2021",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Argumentative structure prediction aims to establish links between textual units and label the relationship between them, forming a structured representation for a given input text. The former task, linking, has been identified by earlier works as particularly challenging, as it requires finding the most appropriate structure out of a very large search space of possible link combinations. In this paper, we improve a state-of-the-art linking model by using multi-task and multi-corpora training strategies. Our auxiliary tasks help the model to learn the role of each sentence in the argumentative structure. Combining multi-corpora training with a selective sampling strategy increases the training data size while ensuring that the model still learns the desired target distribution well. Experiments on essays written by English-as-a-foreign-language learners show that both strategies significantly improve the model's performance; for instance, we observe a 15.8% increase in the F1-macro for individual link predictions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Argument mining (AM) is an emerging area that addresses the automatic analysis of argumentation. Many recent studies commonly try to tackle two major tasks (Lawrence and Reed, 2020) . The first of these is argumentative component identification, in which argumentative units (ACs) and nonargumentative components (non-ACs) including their boundaries are determined. ACs can be further classified according to their role in argumentation, e.g., major claim, 1 claim and premise . The second task is called argumentative structure prediction, which first establishes links from source to target ACs (this is called the linking task) and then labels the relationship between them, for instance using the support and attack relation labels (this is called the relation labelling task) .",
                "cite_spans": [
                    {
                        "start": 156,
                        "end": 181,
                        "text": "(Lawrence and Reed, 2020)",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(S10) I think it will reduce the very bad example for children.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(S12) Smoking is very bad especially for children and students.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(S16) So people shouldn't smoke in public places, not only in restaurants.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(S18) It is to avoid harming other people.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(S17) We should know this rule.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "...",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Smoking should be completely banned at all the restaurants in the country.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "... (S2)",
                "sec_num": null
            },
            {
                "text": "Figure 1: Illustration of linking task, using part of an essay discussing the topic \"Smoking should be banned at all restaurants in the country.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "... (S2)",
                "sec_num": null
            },
            {
                "text": "The linking task has been identified by earlier works as particularly challenging (Lippi and Torroni, 2016; Cabrio and Villata, 2018; Lawrence and Reed, 2020) . There are many possible combinations of links between textual units, and a linking model has to find the most proper structure out of a very large search space. Another typical challenge in AM is the size of annotated corpora (Schulz et al., 2018) . Corpus construction is a complex and time-consuming process; it also often requires a team of expert annotators. Existing corpora in AM are relatively \"small\" compared with more established fields, such as machine translation or document classification. This hinders training AM models when using a supervised machine learning framework.",
                "cite_spans": [
                    {
                        "start": 82,
                        "end": 107,
                        "text": "(Lippi and Torroni, 2016;",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 108,
                        "end": 133,
                        "text": "Cabrio and Villata, 2018;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 134,
                        "end": 158,
                        "text": "Lawrence and Reed, 2020)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 387,
                        "end": 408,
                        "text": "(Schulz et al., 2018)",
                        "ref_id": "BIBREF35"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "... (S2)",
                "sec_num": null
            },
            {
                "text": "In this paper, we perform the linking task for essays written by English-as-a-foreign-language (EFL) learners. Given an essay, we identify links between sentences, forming a tree-structured representation of argumentation in the text. Figure 1 illustrates the task. Our contributions are twofold. First, we propose two structural-modelling related auxiliary tasks to train our model in a multi-task learning (MTL) fashion. Past studies commonly executed several AM tasks or using discourse and rhetorical auxiliary tasks in the MTL setting (cf. Section 2). In comparison with other works, our auxiliary tasks are advantageous in that they do not require any additional annotation. Second, we combine multi-corpora training with a selective sampling strategy to increase the training data size. We explore the use of mixed training data of in-domain EFL and out-domain 2 student-essay corpora that were annotated using different schemes and of different quality, while ensuring that the model still learns the properties of the target in-domain data well.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 235,
                        "end": 243,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "... (S2)",
                "sec_num": null
            },
            {
                "text": "The choice of EFL texts in this study aims to contribute to a less attended area. In AM, it is common to use well-written texts by proficient authors (e.g., Ashley, 1990; Peldszus and Stede, 2016) . However, student texts often suffer from many problems because they are still in the learning phase. Even more, EFL texts are also less coherent and less lexically rich, and exhibit less natural lexical choices and collocations (Silva, 1993; Rabinovich et al., 2016) . There are more non-native English speakers (Fujiwara, 2018) , and yet, to the best of our knowledge, only one preceding study in AM concerned the EFL genre (Putra et al., 2021b) . The codes accompanying this paper are publicly available. 3",
                "cite_spans": [
                    {
                        "start": 157,
                        "end": 170,
                        "text": "Ashley, 1990;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 171,
                        "end": 196,
                        "text": "Peldszus and Stede, 2016)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 427,
                        "end": 440,
                        "text": "(Silva, 1993;",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 441,
                        "end": 465,
                        "text": "Rabinovich et al., 2016)",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 511,
                        "end": 527,
                        "text": "(Fujiwara, 2018)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 624,
                        "end": 645,
                        "text": "(Putra et al., 2021b)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "... (S2)",
                "sec_num": null
            },
            {
                "text": "A variety of formulations have been proposed for the linking task. Traditional approaches formulated it as a pairwise classification task, predicting whether an argumentative link exists between a given pair of ACs (Stab and Gurevych, 2014) . A further post-processing step can also be performed to combine the local predictions into an optimised global structure, e.g., using the minimum-spanningtree algorithm . Recent studies proposed a more global approach instead, considering the entire input context. For instance, Potash et al. (2017) formulated the linking task as a sequence prediction problem. They jointly performed AC classification and AC linking at once, 2 Past studies used the term domain in a broad context. It has at least five different senses: text genre, text quality, annotation scheme, dataset and topic (or prompt). In the rest of this paper, we use the specific meaning whenever possible.",
                "cite_spans": [
                    {
                        "start": 215,
                        "end": 240,
                        "text": "(Stab and Gurevych, 2014)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 522,
                        "end": 542,
                        "text": "Potash et al. (2017)",
                        "ref_id": "BIBREF29"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Argumentative Structure Prediction",
                "sec_num": "2.1"
            },
            {
                "text": "3 https://github.com/wiragotama/ ArgMin2021 assuming that the segmentation and AC vs non-AC categorisation have been pre-completed. They experimented on the microtext corpus (Peldszus and Stede, 2016) and the persuasive essay corpus (PEC, ). formulated argumentative structure parsing in three ways: as relation extraction, as sequence tagging and as dependency parsing tasks. They defined a BIO tagging scheme and performed end-to-end parsing at token-level, executing all subtasks (i.e., segmentation, unit type classification, linking and relation labelling) at once. Ye and Teufel (2021) also performed endto-end parsing at the token-level. They proposed a more efficient representation for the dependency structure of arguments, and achieved the state-ofthe-art performance for component and relation identifications on the PEC using a biaffine attention model (Dozat and Manning, 2017) . The biaffine attention model was originally designed to parse token-to-token dependency, but Morio et al. (2020) extended it to parse proposition (segment) level dependency. Their model dealt with graph-structured arguments in the Cornell eRulemaking corpus (Park and Cardie, 2018) . Using the same architecture, Putra et al. (2021b) parsed tree-structured EFL essays in the ICNALE-AS2R corpus (Ishikawa, 2013 (Ishikawa, , 2018 Putra et al., 2021a,b) . In tree-structured argumentation, it is common for groups of sentences about the same sub-topic to operate as a unit, forming a sub-tree (sub-argument). Putra et al. (2021b) found that their linking model has problems in constructing sub-trees, that is, it splits a group of sentences that should belong together into separate sub-arguments (sub-trees) or, conversely, groups together sentences that do not belong together into the same sub-trees.",
                "cite_spans": [
                    {
                        "start": 174,
                        "end": 200,
                        "text": "(Peldszus and Stede, 2016)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 571,
                        "end": 591,
                        "text": "Ye and Teufel (2021)",
                        "ref_id": "BIBREF41"
                    },
                    {
                        "start": 866,
                        "end": 891,
                        "text": "(Dozat and Manning, 2017)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 987,
                        "end": 1006,
                        "text": "Morio et al. (2020)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 1152,
                        "end": 1175,
                        "text": "(Park and Cardie, 2018)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 1288,
                        "end": 1303,
                        "text": "(Ishikawa, 2013",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1304,
                        "end": 1321,
                        "text": "(Ishikawa, , 2018",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 1322,
                        "end": 1344,
                        "text": "Putra et al., 2021a,b)",
                        "ref_id": null
                    },
                    {
                        "start": 1500,
                        "end": 1520,
                        "text": "Putra et al. (2021b)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Argumentative Structure Prediction",
                "sec_num": "2.1"
            },
            {
                "text": "Several approaches have been applied to alleviate the data sparsity problem in AM. Al-Khatib et al. Data augmentation can also be applied to mitigate the data sparsity problem. This aims to increase the amount of training data without directly collecting more data Feng et al., 2021) . A relatively straightforward strategy is to use multiple corpora when training models. For example, Chu et al. (2017) proposed a mixed fine-tuning approach for machine translation; they trained a model on an out-genre corpus and then fine-tune it on a dataset that is a mix of the targetgenre and out-genre corpora. However, the use of multiple corpora of different genres is challenging in AM because argumentation is often modelled differently across genres (Lippi and Torroni, 2016; Lawrence and Reed, 2020) . found that training a claim identification model with mixed-genre corpora only perform as good as training on each specific corpus. The use of data augmentation may cause the distributional shift problem as well, where the augmented data alter the target distribution that should be learned by the model (Feng et al., 2021). Putra et al. (2021b) provided some insights on the possible conditions for multi corpora training by experimenting using texts of different quality. They trained argumentative structure prediction models on EFL essays (in-domain) and parallel improved version of the texts (out-domain), and evaluated on the EFL texts. Their cross-domain system attained 94% accuracy of end-to-end indomain performance. This signals the potential to use corpora of different text quality altogether to train a parsing model, as long as the text genre stays the same. However, mixed-quality corpora training has yet to be tried in practice.",
                "cite_spans": [
                    {
                        "start": 265,
                        "end": 283,
                        "text": "Feng et al., 2021)",
                        "ref_id": null
                    },
                    {
                        "start": 386,
                        "end": 403,
                        "text": "Chu et al. (2017)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 746,
                        "end": 771,
                        "text": "(Lippi and Torroni, 2016;",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 772,
                        "end": 796,
                        "text": "Lawrence and Reed, 2020)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 1124,
                        "end": 1144,
                        "text": "Putra et al. (2021b)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Low-Resource and Cross-Domain Argument Mining",
                "sec_num": "2.2"
            },
            {
                "text": "In recent years, there has also been a growing interest towards more generic AM models. For example, Stab et al. (2018) proposed a simple annotation scheme for argument retrieval that is applicable in heterogeneous sources and can be performed by untrained annotators. Cocarascu et al. (2020) explored various deep learning architectures and features that work well across various datasets for the relation labelling task. They also provided a comparison of using contextualised and non-contextualised embeddings for the task. On the other hand, Ruiz-Dolz et al. (2021) compared the performance of transformer-based language mod-els on the cross-topic relation labelling task.",
                "cite_spans": [
                    {
                        "start": 101,
                        "end": 119,
                        "text": "Stab et al. (2018)",
                        "ref_id": "BIBREF40"
                    },
                    {
                        "start": 269,
                        "end": 292,
                        "text": "Cocarascu et al. (2020)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Low-Resource and Cross-Domain Argument Mining",
                "sec_num": "2.2"
            },
            {
                "text": "Our target texts are sourced from the ICNALE-AS2R, a corpus of 434 essays written by Asian college students with intermediate proficiency. 4 There are 6,021 sentences in total with 13.9 sentences on average per essay. To the best of our knowledge, this is the only currently publicly-available AM corpus focusing on EFL texts. The corpus is annotated at sentence-level, that is, a sentence corresponds to one argumentative unit. The corpus differentiates sentences as ACs (5,799 sents.) and non-ACs (222 sents.), without further classification of AC types. Links are established between ACs to form tree structure (avg. depth of 4.3, root at depth 0 in the corpus), where the major claim acts as the root. Four relation labels are employed to label the links: support, attack, detail and restatement.",
                "cite_spans": [
                    {
                        "start": 139,
                        "end": 140,
                        "text": "4",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": "3"
            },
            {
                "text": "In this work, we experiment on the linking task. Given an entire essay of N sentences as input, s 1 , ..., s N , a linking model outputs the distance d 1 , ..., d N between each sentence s i to its target. For instance, if a sentence is connected to its preceding sentence, the distance is d = \u22121. We consider those sentences that have no explicitly annotated outgoing links as linked to themselves (d = 0); this concerns major claims (roots) and non-ACs. We do not consider labelling the links with their relationships in this paper. Adjacent links predominate (50.4%). Short distance links (2 \u2264 |d| \u2264 4) make up 21.2% of the total. Backward long distance links at d \u2264 \u22125 are 16.6%, whereas forward long distance links are rare (1.0%). Self-loops make up 10.9% of the total. Following recent advances in AM, we approach the linking task as a sentence-to-sentence dependency parsing.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposed Method",
                "sec_num": "4"
            },
            {
                "text": "\u2264-5 -4 -3 -2 -1 0 1 2 3 4 \u22655",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposed Method",
                "sec_num": "4"
            },
            {
                "text": "We use Putra et al.'s (2021b) architecture as the state-of-the-art baseline (\"BIAF\" model, cf. Figure 3) . In this architecture, input sentences are first encoded into their respective sentence embeddings using sentence-BERT encoder (Reimers and Gurevych, 2019) . The resulting sentence embeddings are then passed into a dense layer for dimensionality reduction. The results are then fed into a bidirectional long-short-term memory network (BiLSTM) layer (#stack = 3) to produce contextual sentence representations. These representations are then passed into two different dense layers to encode the corresponding sentence when it acts as a source h (source) or target h (target) in a relation. Finally, a biaffine transformation (Dozat and Manning, 2017) is applied to all source and target representations to produce the final output matrix G \u2208 R N \u00d7N , where each cell g i,j represents the probability (or score) of the source sentence s i pointing to s j . Equation (1) and(2) show the biaffine transformation (f ), where U and W are weight matrices and b is bias.",
                "cite_spans": [
                    {
                        "start": 7,
                        "end": 29,
                        "text": "Putra et al.'s (2021b)",
                        "ref_id": null
                    },
                    {
                        "start": 233,
                        "end": 261,
                        "text": "(Reimers and Gurevych, 2019)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 650,
                        "end": 658,
                        "text": "(source)",
                        "ref_id": null
                    },
                    {
                        "start": 671,
                        "end": 679,
                        "text": "(target)",
                        "ref_id": null
                    },
                    {
                        "start": 730,
                        "end": 755,
                        "text": "(Dozat and Manning, 2017)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 95,
                        "end": 104,
                        "text": "Figure 3)",
                        "ref_id": "FIGREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Baseline",
                "sec_num": "4.1"
            },
            {
                "text": "f (x 1 , x 2 ) = x 1 T Ux 2 + W(x 1 \u2295 x 2 ) + b (1) g i,j = f h (source) i , h (target) j (2)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Baseline",
                "sec_num": "4.1"
            },
            {
                "text": "The Chu-Liu-Edmonds algorithm (Chu and Liu, 1965; Edmonds, 1967) is applied to create a minimum spanning tree out of the matrix G. The links in the tree are then converted into distance predictions between source and target sentences for evaluation purpose.",
                "cite_spans": [
                    {
                        "start": 30,
                        "end": 49,
                        "text": "(Chu and Liu, 1965;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 50,
                        "end": 64,
                        "text": "Edmonds, 1967)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Baseline",
                "sec_num": "4.1"
            },
            {
                "text": "We propose to extend the BIAF model in an MTL setup using two novel structural-modelling-related auxiliary tasks. The first auxiliary task is a quasi argumentative component type (QACT) prediction. ICNALE-AS2R corpus does not assign AC types per se, but we can compile the following four sentence types from the tree typology:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multi-Task Learning with Structural Signal",
                "sec_num": "4.2"
            },
            {
                "text": "\u2022 major claim (root): only incoming links, \u2022 AC (non-leaf): both outgoing and incoming links, \u2022 AC (leaf): only outgoing links and \u2022 non-AC: neither incoming nor outgoing links.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multi-Task Learning with Structural Signal",
                "sec_num": "4.2"
            },
            {
                "text": "For example, S2 is the major claim of Figure 1 , S10 is AC (non-leaf) and S17 is AC (leaf). The QACT prediction task should help the linking model to learn the role of sentences in the argumentative structure, as well as the property of links for each sentence. This auxiliary task improved a sequencetagger based linking model in Putra et al. (2021b) , but it has not been applied to the BIAF model yet.",
                "cite_spans": [
                    {
                        "start": 331,
                        "end": 351,
                        "text": "Putra et al. (2021b)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 38,
                        "end": 46,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Multi-Task Learning with Structural Signal",
                "sec_num": "4.2"
            },
            {
                "text": "The second auxiliary task concerns node depth (ND) prediction. There are six depth categories employed in this paper: depth 0 to depth 4, and depth 5+. The argumentative structure in ICNALE-AS2R corpus is hierarchical, and there is no relations between nodes (sentences) at the same depth. The ND prediction task should help the model to learn the placement of sentences in the hierarchy and guide where each sentence should point at, that is, sentences at depth X point at sentences at depth X \u2212 1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multi-Task Learning with Structural Signal",
                "sec_num": "4.2"
            },
            {
                "text": "We also propose to use sentence position (spos) embedding as an input feature because it has been proved to be useful in other studies (e.g., Song et al., 2020) . The sentence position encoding is calculated by dividing a sentence position by the essay length. When using this feature, sentence embeddings and spos encoding are concatenated before being passed to the first dense layer. Figure 4 illustrates our proposed architecture.",
                "cite_spans": [
                    {
                        "start": 142,
                        "end": 160,
                        "text": "Song et al., 2020)",
                        "ref_id": "BIBREF37"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 387,
                        "end": 395,
                        "text": "Figure 4",
                        "ref_id": "FIGREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Multi-Task Learning with Structural Signal",
                "sec_num": "4.2"
            },
            {
                "text": "s 1 ... s N Concat Sentence-BERT (SBERT) Encoder BiLSTM h 1 ( t a r g e t ) h 1 ( s o u r c e ) h 1 ( Q A C T ) h 1 ( N D ) ... h N ( t a r g e t ) h N ( s o u r c e ) h N ( Q A C T ) h N ( N D ) = G U c 1 c N . . . l 1 l N . . . Q u a s i",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multi-Task Learning with Structural Signal",
                "sec_num": "4.2"
            },
            {
                "text": "In this paper, we also consider the use of multicorpora training strategy, employing essays in the PEC as additional training data. Both ICNALE-AS2R and PEC provide argumentative essays written by students (in English) and represents argumentative structures as trees. However, there is no information on the essay authors proficiency (L1 or L2) nor the observed quality of the PEC. Hence, these two corpora might be of different quality. The writing prompts of ICNALE-AS2R and PEC essays are also different.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multi-Corpora Training",
                "sec_num": "4.3"
            },
            {
                "text": "There are two settings when training using multiple corpora. The first is to use the entire 402 essays in the PEC on top of ICNALE-AS2R train set (\"[P+I]\" setting). However, PEC and ICNALE-AS2R are different in terms of length and annotation scheme. The PEC essays have 18.2 sentences on average (15.1 ACs and 3.1 non-ACs), while ICNALE-AS2R essays have 13.9 sentences on average (13.4 ACs and 0.5 non-ACs). The difference in non-ACs proportion between these corpora is likely caused by the difference in the set of relation labels employed. The essays in PEC were annotated using two relation types: support and attack, while ICNALE-AS2R additionally employed detail and restatement. Hence, there is arguably more information in the ICNALE-AS2R corpus. Particularly, some sentences that should have been annotated as non-ACs (hence not linked to other sentences) in the PEC's scheme might have been annotated as ACs in the ICNALE-AS2R's scheme by using the additional relations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multi-Corpora Training",
                "sec_num": "4.3"
            },
            {
                "text": "Due to the differences in annotation scheme and statistical properties, there is a possibility that the model might not properly learn the distribution of ICNALE-AS2R in the [P+I] setting. Therefore, we also propose a second selective sampling (\"[SS]\" setting) strategy to account for the differences between these two corpora. In this setting, we only use PEC essays that are somewhat similar to those of ICNALE-AS2R considering the following two criteria. First, we only use PEC essays having 17 sentences at maximum (ICNALE-AS2R avg. 13.9 + 3.3 SD). Second, the employed PEC essays contain two non-ACs at maximum (ICNALE-AS2R avg. 0.5 + 0.9 SD). There are 110 remaining PEC essays after the selective sampling procedure.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multi-Corpora Training",
                "sec_num": "4.3"
            },
            {
                "text": "Note that ICNALE-AS2R was annotated at the sentence level while PEC was annotated at the segment level. Therefore, we convert the PEC annotation to the sentence level, following the strategy described by Song et al. (2020) . If a sentence contains only one AC, we use the whole sentence as an AC; if a sentence contains two or more ACs, we split it into multiple sentences while including the preceding tokens into each AC. Figure 5 illustrates the splitting procedure, where a sentence contain- ing three ACs is split into three sentences.",
                "cite_spans": [
                    {
                        "start": 204,
                        "end": 222,
                        "text": "Song et al. (2020)",
                        "ref_id": "BIBREF37"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 424,
                        "end": 432,
                        "text": "Figure 5",
                        "ref_id": "FIGREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Multi-Corpora Training",
                "sec_num": "4.3"
            },
            {
                "text": "In this paper, we do not evaluate the model performance only on individual link predictions but also analyse the structural properties of the outputs, giving more insights into the models' ability to learn different aspects of the argumentative structure. We also investigate whether our proposed strategies improve the model performance concerning the grouping of related sentences into the same sub-tree, a challenge to linking models previously (manually) observed in Putra et al. (2021b) (cf. Section 2.1). A metric is needed to quantify the improvement on this aspect, and we propose to use a novel structural metric called MAR dSet , that was developed by Putra et al. (2021a). 5 Given two structures A and B (i.e., predicted and gold structures), we can quantify their similarity based on the presence of common substructures. We first define a descendant set (dSet) of a node X as the set consisting the node X itself and its descendants. Figure 6 shows examples of dSets (brackets given below the node ID). 6 For example, the dSet of node-2 in annotation A of Figure 6 is {2, 3, 4, 5}. Two corresponding nodes in A and B are required to have identical dSets in order to score a value of 1. For example, the matching score for node-2 between annotation A and B in Figure 6 is 0, while the matching score for node-3 is 1. Non-AC nodes are counted as a match if they are deemed non-argumentative in both structures.",
                "cite_spans": [
                    {
                        "start": 684,
                        "end": 685,
                        "text": "5",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 947,
                        "end": 955,
                        "text": "Figure 6",
                        "ref_id": "FIGREF6"
                    },
                    {
                        "start": 1069,
                        "end": 1077,
                        "text": "Figure 6",
                        "ref_id": "FIGREF6"
                    },
                    {
                        "start": 1272,
                        "end": 1280,
                        "text": "Figure 6",
                        "ref_id": "FIGREF6"
                    }
                ],
                "eq_spans": [],
                "section": "Structural Metric",
                "sec_num": "5"
            },
            {
                "text": "With this formulation, we can get a vector v representing the matching scores for all nodes in the two structures. For example, v = [0, 0, 1, 1, 0] for Figure 6 . The similarity score between two structures are then calculated as 5 MAR stands for Mean Agreement in Recall. Descendant set (dSet) denotes the unit of analysis. 6 In the implementation, we use sentence position in the input text as node ID.",
                "cite_spans": [
                    {
                        "start": 230,
                        "end": 231,
                        "text": "5",
                        "ref_id": null
                    },
                    {
                        "start": 325,
                        "end": 326,
                        "text": "6",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 152,
                        "end": 160,
                        "text": "Figure 6",
                        "ref_id": "FIGREF6"
                    }
                ],
                "eq_spans": [],
                "section": "Structural Metric",
                "sec_num": "5"
            },
            {
                "text": "MAR dSet = N i=1 v i N ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Structural Metric",
                "sec_num": "5"
            },
            {
                "text": "where N denotes the number of nodes; The MAR dSet score for Figure 6 is 0.4. 6 Experiment and Discussion",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 60,
                        "end": 68,
                        "text": "Figure 6",
                        "ref_id": "FIGREF6"
                    }
                ],
                "eq_spans": [],
                "section": "Structural Metric",
                "sec_num": "5"
            },
            {
                "text": "In this experiment, the ICNALE-AS2R corpus is split into 80% train set (347 essays , 4,841 sents.) and 20% test set (87 essays, 1,180 sents.). We use the same splits as Putra et al. (2021b) . The number of training essays for [P+I] and [SS] settings are 749 (12,162 sents.) and 457 (6,418 sents.), respectively. We run the experiment for 20 times 7 and report the average performance. Statistical testing, whenever possible, is conducted using the permutation test (Noreen, 1989) on the performance scores of the 20 runs with a significance level of \u03b1 = .05. Here, we also perform an ablation study. Appendix A shows some implementation notes. Table 1 shows the experimental results on individual link predictions and MAR dSet . Training the BIAF model using the QACT task improved the performance over the baseline, particularly in terms of F1-macro (non significant difference). In addition, using both QACT+ND auxiliary tasks significantly improved the performance over the baseline in terms of F1-macro. This signals that the proposed MTL setting benefits model performance.",
                "cite_spans": [
                    {
                        "start": 169,
                        "end": 189,
                        "text": "Putra et al. (2021b)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 465,
                        "end": 479,
                        "text": "(Noreen, 1989)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 644,
                        "end": 651,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Experimental Setting",
                "sec_num": "6.1"
            },
            {
                "text": "We next look at how spos encoding affects the model performance. Introducing spos to the BIAF+QACT+ND model improves accuracy and MAR dSet . However, the difference is not significant. Similarly, the difference between Table 1 : Results of individual link predictions (Acc. and F1-macro) and MAR dSet . The best result is shown in bold-face. The \u2020 symbol indicates that the difference to the second-best result (underlined) is significant.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 219,
                        "end": 226,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Quantitative Analysis",
                "sec_num": "6.2"
            },
            {
                "text": "BIAF+QACT+spos and BIAF+QACT is not significant. There are two possible explanations for this phenomenon. First, studies in contrastive rhetoric found that non-native speakers tend to structure and organise their texts differently from native speakers (Kaplan, 1966; Silva, 1993) . Some students might have used the writing customs and rhetorical strategies of their first language instead of using the common English writing patterns. For example, Asian students sometimes put a claim after its reason, which is not common in Anglo-Saxon cultures (Silva, 1993; Johns, 1986) . We observe that some texts in the ICNALE-AS2R corpus are written in the \"claim-support\" structure but some are written in the \"support-claim\" structure. Sentences on the same topic are sometimes distantly separated as well. These inconsistencies might have negated the effect of the spos encoding. Second, the output of the BIAF model is a graph G which considers the directed links between all pairs of sentences. The spos feature might not affect the biaffine transformation much in this context.",
                "cite_spans": [
                    {
                        "start": 252,
                        "end": 266,
                        "text": "(Kaplan, 1966;",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 267,
                        "end": 279,
                        "text": "Silva, 1993)",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 548,
                        "end": 561,
                        "text": "(Silva, 1993;",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 562,
                        "end": 574,
                        "text": "Johns, 1986)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Quantitative Analysis",
                "sec_num": "6.2"
            },
            {
                "text": "The models trained using multiple corpora attain the best performance for individual link and substructure (MAR) predictions. BIAF+QACT+ND [P+I] achieves the best performance of .455 in terms of MAR dSet , and BIAF+QACT+ND [SS] achieves the best performance of .489 and .374 in terms of accuracy and F1-macro. The [SS] model also achieves the second-best performance of .452 in terms of MAR dSet . These improvements are significant over the baseline and the BIAF+QACT+ND model. Note that when using the [SS] setting, the model performance is consis-tently improved with regard to all metrics, while the accuracy of the [P+I] model is lower than the baseline. In general, the [SS] model attains a better performance compared with the [P+I] model despite fewer training instances. This means that when training a model using multiple corpora, it is essential to consider training instances having the same properties as our goal. Simply having more training data does not guarantee improvement. To gain deeper insights into the individual link prediction improvement brought by the [SS] model over the baseline, Figure 7 shows F1 score per target linking distance. BIAF+QACT+ND [SS] is better than the baseline model, particularly at predicting short-distance links (2 \u2264 |d| \u2264 4, avg. F1 = .24 vs .17), which is the weakest range of the baseline model. However, this is still the weakest range even for the [SS] model. Figure 8 shows the performance across depths, that is, whether the model places each node at the proper depth in the predicted structure. BIAF+QACT+ND [SS] performs better than the baseline particularly in [0,3] and [6,8] ranges. This plot indicates that the model performance declines as one moves further down the tree.",
                "cite_spans": [
                    {
                        "start": 139,
                        "end": 144,
                        "text": "[P+I]",
                        "ref_id": null
                    },
                    {
                        "start": 223,
                        "end": 227,
                        "text": "[SS]",
                        "ref_id": null
                    },
                    {
                        "start": 1177,
                        "end": 1181,
                        "text": "[SS]",
                        "ref_id": null
                    },
                    {
                        "start": 1624,
                        "end": 1639,
                        "text": "[0,3] and [6,8]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 1111,
                        "end": 1119,
                        "text": "Figure 7",
                        "ref_id": null
                    },
                    {
                        "start": 1418,
                        "end": 1426,
                        "text": "Figure 8",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Quantitative Analysis",
                "sec_num": "6.2"
            },
            {
                "text": "We next look at the models' ability to predict the role of each node (QACT task) based on the predicted hierarchical structures. the results. The MTL models perform better compared with the baseline model. Both BIAF+QACT and BIAF+QACT+ND achieve a significant improvement over the baseline in terms of non-AC prediction and F1-macro. This reconfirms that both of our proposed MTL tasks are useful to improve linking performance. Similar to the previous result on individual links, the spos feature does not provide much improvement. We notice that the BIAF+QACT+ND [P+I] model performs worse than the baseline in some aspects, particularly in terms of AC (non-leaf) and non-AC predictions. On the other hand, BIAF+QACT+ ND [SS] consistently achieves better performance compared with the baseline model. It attains the highest score of .622 in F1-macro, which is a significant improvement over other configurations. This confirms our hypothesis that the difference between annotation schemes and statistical properties between PEC and ICNALE-AS2R affects the model in terms of the distribution learned. The proposed selective sampling strategy helps to alleviate this problem.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Quantitative Analysis",
                "sec_num": "6.2"
            },
            {
                "text": "We also analyse the overall shape of the predicted structures by all models, as shown in Table 3. The gold standard trees in ICNALE-AS2R have a particular shape, expressed as the average depth of 4.3 (SD = 1.4) and the leaf ratio of .439 (SD = 0.11). The baseline model tends to produce trees that are deeper and narrower than the ICNALE-AS2R gold trees. Our MTL auxiliary tasks help to improve the leaf ratio to become closer to the gold standard (significant difference between BIAF+QACT+ND and BIAF), while spos embedding does not provide additional improvement. When we introduce the multi-corpora training strategy, the predicted structures become shal- lower and wider compared with the baseline, particularly when using the [P+I] setting. We believe this is due to the shallower trees in the PEC compared with the ICNALE-AS2R (i.e., the distributional shift problem). However, this is less of an issue for the [SS] model as it produces the most similar structure to the ICNALE-AS2R essays. We conclude BIAF+QACT+ND [SS] as the best model in this experiment, achieving the new stateof-the-art performance for the linking task. It consistently performs better compared with other configurations across all evaluation aspects.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Quantitative Analysis",
                "sec_num": "6.2"
            },
            {
                "text": "We conducted a qualitative analysis on some random outputs of our best model, namely BIAF+QACT+ND [SS] . Some prediction errors are likely caused by the inability of the model to re-solve anaphora. Figure 9 shows an output example. The model successfully groups sentences S5-S7 as a single sub-argument, as the gold annotation does. However, it connects S7 to S6, while the gold annotation connects S7 to S5 (with the support label). The link from S7 to S5 is the correct interpretation because the statement in S7 can be broadly applied to \"skills of working and cooperating,\" not only to some specific instantiations mentioned in S6 (\"set the working agenda and get help from colleagues\"). A possible direction to alleviate this problem is to equip the model with a better ability to resolve anaphora. Another direction is to perform an anaphora resolution step before parsing the argumentative structure.",
                "cite_spans": [
                    {
                        "start": 98,
                        "end": 102,
                        "text": "[SS]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 198,
                        "end": 206,
                        "text": "Figure 9",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S3) The importance of a part-time job is recognized by a great number of people, including me.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S4) Most students go to the university to prepare themselves to blend into the society in the future.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S7) Without these skills, even if the student gets a good occupation upon his graduation, they will possible feel it difficult to get used to the job, and thus their performance will be unsatisfactory.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S5) Being successful in the society calls for skills of working and cooperating, which can be developed by working at part-time jobs (S6) For instance, one can learn how to set the working agenda and get help from colleagues. Figure 9 : An excerpt of prediction for essay \"W_HKG_PTJ0_029_B2_0_EDIT\". The essay was written in response to a prompt \"It is important for college students to have a part-time job.\" The dashed line illustrates an erroneous link, while the dotted line (grey) illustrates the corresponding gold link.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 227,
                        "end": 235,
                        "text": "Figure 9",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S7) Those who have a part-time job are more likely to get in touch with various people in society.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S6) To begin with, a part-time job will provide more social experience.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S10) In addition, a part-time job will show them the significance of cooperation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "(S5) Generally speaking, it's important for college students to have a part-time job.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "Figure 10: An excerpt of prediction for essay \"W_CHN_PTJ0_242_B2_0_EDIT\". The essay was written in response to a prompt \"It is important for college students to have a part-time job.\" Dashed lines and dotted lines as above.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "Another source of error is the difficulty in judging whether two statements argue at the same con-ceptual level (hence the same depth in the tree). Figure 10 shows an illustration. S7 and S10 are exemplifications of S6, and therefore, they should be placed lower than S6. However, the model recognises S6, S7 and S10 arguing at the same conceptual level. The exemplifications are probably not obvious for the model because S7 is not accompanied by a discourse marker (e.g., \"for example\"). In general, EFL essays are often challenging to process due to improper use (can be excessive or limited too) of discourse markers.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 148,
                        "end": 157,
                        "text": "Figure 10",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Qualitative Analysis",
                "sec_num": "6.3"
            },
            {
                "text": "In this paper, we presented a study on the argumentative linking task; given an essay as input, a linking model outputs the argumentative structure in a tree representation. We conducted experiments using an EFL corpus, namely ICNALE-AS2R. We extended the state-of-the-art biaffine attention model using a novel set of structural auxiliary tasks in the multi-task learning setup. Additionally, we also proposed a multi-corpora training strategy using the PEC to increase training instances. It has to be noted that simply increasing the training data size does not guarantee improvement. We need to ensure that the model still learns the desired target distribution as well. To this end, we filtered PEC essays using a selective sampling technique. These two strategies provided useful supervision signals to the biaffine model and significantly improved its performance. The F1-macro for individual link predictions was boosted to .374 from .323. Our strategies also improved the model performance on the structural aspects, achieving the MAR dset of .452 from .419, and the F1-macro of .622 from .601 for the QACT task.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "7"
            },
            {
                "text": "A possible future direction is to evaluate whether our proposed methods are beneficial for more complex discourses, such as scientific articles. It is also necessary to provide an even richer supervision signal to improve the linking performance. One possibility is to pre-train the model on the anaphora resolution task.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "7"
            },
            {
                "text": "The major claim is the statement expressing the writer's view on the discussion topic; also called main stance or main claim.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://www.gsk.or.jp/en/catalog/ gsk2021-a. The approximated CEFR level of the essay authors are A2 (94 essays), B1 (253) and B2 (87).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Twenty experiments were repeated on the same dataset splits to account for random initialisation in neural networks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "This work was partially supported by Tokyo Tech World Research Hub Initiative (WRHI), JSPS KAKENHI grant number 20J13239 and Support Centre for Advanced Telecommunication Technology Research. We would like to thank anonymous reviewers for their useful and detailed feedback.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgements",
                "sec_num": null
            },
            {
                "text": "Hidden Units and Model Implementation 512 units for the first dense layer, 256 LSTM unit, and 256 units for the second dense layer after the BiLSTM layer. Dropout is applied between all layers. For a fair comparison, we follow Putra et al. (2021b) in using the sentence-BERT encoder fine-tuned on the NLI dataset (\"bert-base-nli-meantokens\", https://github.com/UKPLab/ sentence-transformers). All models are trained using Adam optimiser (Kingma and Ba, 2015), and implemented in PyTorch (Paszke et al., 2019) and AllenNLP (Gardner et al., 2018) .Loss The MTL loss is defined aswhere the loss L t of each task t is dynamically weighted, controlled by a learnable parameter \u03c3 t . The loss for the main task (sentence linking) is computed using max-margin criterion, while losses for auxiliary tasks are computed using cross-entropy.",
                "cite_spans": [
                    {
                        "start": 227,
                        "end": 247,
                        "text": "Putra et al. (2021b)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 487,
                        "end": 508,
                        "text": "(Paszke et al., 2019)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 522,
                        "end": 544,
                        "text": "(Gardner et al., 2018)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Implementation Notes",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Transferring knowledge from discourse to arguments: A case study with scientific abstracts",
                "authors": [
                    {
                        "first": "Pablo",
                        "middle": [],
                        "last": "Accuosto",
                        "suffix": ""
                    },
                    {
                        "first": "Horacio",
                        "middle": [],
                        "last": "Saggion",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 6th Workshop on Argument Mining",
                "volume": "",
                "issue": "",
                "pages": "41--51",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-4505"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Pablo Accuosto and Horacio Saggion. 2019. Trans- ferring knowledge from discourse to arguments: A case study with scientific abstracts. In Proceedings of the 6th Workshop on Argument Mining, pages 41- 51, Florence, Italy. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Crossdomain mining of argumentative text through distant supervision",
                "authors": [
                    {
                        "first": "Khalid",
                        "middle": [],
                        "last": "Al-Khatib",
                        "suffix": ""
                    },
                    {
                        "first": "Henning",
                        "middle": [],
                        "last": "Wachsmuth",
                        "suffix": ""
                    },
                    {
                        "first": "Matthias",
                        "middle": [],
                        "last": "Hagen",
                        "suffix": ""
                    },
                    {
                        "first": "Jonas",
                        "middle": [],
                        "last": "K\u00f6hler",
                        "suffix": ""
                    },
                    {
                        "first": "Benno",
                        "middle": [],
                        "last": "Stein",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "",
                "issue": "",
                "pages": "1395--1404",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N16-1165"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Khalid Al-Khatib, Henning Wachsmuth, Matthias Ha- gen, Jonas K\u00f6hler, and Benno Stein. 2016. Cross- domain mining of argumentative text through distant supervision. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1395-1404, San Diego, Califor- nia. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Modeling legal argument -reasoning with cases and hypotheticals. Artificial intelligence and legal reasoning",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Kevin",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ashley",
                        "suffix": ""
                    }
                ],
                "year": 1990,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kevin D. Ashley. 1990. Modeling legal argument -rea- soning with cases and hypotheticals. Artificial intel- ligence and legal reasoning. MIT Press.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Five years of argument mining: A data-driven analysis",
                "authors": [
                    {
                        "first": "Elena",
                        "middle": [],
                        "last": "Cabrio",
                        "suffix": ""
                    },
                    {
                        "first": "Serena",
                        "middle": [],
                        "last": "Villata",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI'18",
                "volume": "",
                "issue": "",
                "pages": "5427--5433",
                "other_ids": {
                    "DOI": [
                        "10.24963/ijcai.2018/766"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Elena Cabrio and Serena Villata. 2018. Five years of ar- gument mining: A data-driven analysis. In Proceed- ings of the 27th International Joint Conference on Artificial Intelligence, IJCAI'18, page 5427-5433. AAAI Press.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "An empirical comparison of domain adaptation methods for neural machine translation",
                "authors": [
                    {
                        "first": "Chenhui",
                        "middle": [],
                        "last": "Chu",
                        "suffix": ""
                    },
                    {
                        "first": "Raj",
                        "middle": [],
                        "last": "Dabre",
                        "suffix": ""
                    },
                    {
                        "first": "Sadao",
                        "middle": [],
                        "last": "Kurohashi",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics",
                "volume": "2",
                "issue": "",
                "pages": "385--391",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P17-2061"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Chenhui Chu, Raj Dabre, and Sadao Kurohashi. 2017. An empirical comparison of domain adaptation methods for neural machine translation. In Proceed- ings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Pa- pers), pages 385-391, Vancouver, Canada. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "On the shortest arborescence of a directed graph",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Chu",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 1965,
                "venue": "Science Sinica",
                "volume": "",
                "issue": "",
                "pages": "1396--1400",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Y. Chu and T. Liu. 1965. On the shortest arbores- cence of a directed graph. Science Sinica, page 1396-1400.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "A dataset independent set of baselines for relation prediction in argument mining",
                "authors": [
                    {
                        "first": "Oana",
                        "middle": [],
                        "last": "Cocarascu",
                        "suffix": ""
                    },
                    {
                        "first": "Elena",
                        "middle": [],
                        "last": "Cabrio",
                        "suffix": ""
                    },
                    {
                        "first": "Serena",
                        "middle": [],
                        "last": "Villata",
                        "suffix": ""
                    },
                    {
                        "first": "Francesca",
                        "middle": [],
                        "last": "Toni",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of Computational Models of Argument (COMMA)",
                "volume": "",
                "issue": "",
                "pages": "45--52",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Oana Cocarascu, Elena Cabrio, Serena Villata, and Francesca Toni. 2020. A dataset independent set of baselines for relation prediction in argument mining. In Proceedings of Computational Models of Argu- ment (COMMA), pages 45-52.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "What is the essence of a claim? cross-domain claim identification",
                "authors": [
                    {
                        "first": "Johannes",
                        "middle": [],
                        "last": "Daxenberger",
                        "suffix": ""
                    },
                    {
                        "first": "Steffen",
                        "middle": [],
                        "last": "Eger",
                        "suffix": ""
                    },
                    {
                        "first": "Ivan",
                        "middle": [],
                        "last": "Habernal",
                        "suffix": ""
                    },
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Stab",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "2055--2066",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D17-1218"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Johannes Daxenberger, Steffen Eger, Ivan Habernal, Christian Stab, and Iryna Gurevych. 2017. What is the essence of a claim? cross-domain claim identi- fication. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2055-2066, Copenhagen, Denmark. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Deep biaffine attention for neural dependency parsing",
                "authors": [
                    {
                        "first": "Timothy",
                        "middle": [],
                        "last": "Dozat",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the International Conference on Learning Representations (ICLR",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency pars- ing. In Proceedings of the International Conference on Learning Representations (ICLR).",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Optimum branchings",
                "authors": [],
                "year": 1967,
                "venue": "Journal of Research of the National Bureau of Standards -B. Mathematics and Mathematical Physics",
                "volume": "71",
                "issue": "",
                "pages": "233--240",
                "other_ids": {
                    "DOI": [
                        "10.6028/jres.071B.032"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jack Edmonds. 1967. Optimum branchings. Journal of Research of the National Bureau of Standards -B. Mathematics and Mathematical Physics, 71B:233- 240.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Neural end-to-end learning for computational argumentation mining",
                "authors": [
                    {
                        "first": "Steffen",
                        "middle": [],
                        "last": "Eger",
                        "suffix": ""
                    },
                    {
                        "first": "Johannes",
                        "middle": [],
                        "last": "Daxenberger",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "11--22",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P17-1002"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Steffen Eger, Johannes Daxenberger, and Iryna Gurevych. 2017. Neural end-to-end learning for computational argumentation mining. In Proceed- ings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pages 11-22, Vancouver, Canada. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Teruko Mitamura, and Eduard H. Hovy. 2021. A survey of data augmentation approaches for NLP",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Steven",
                        "suffix": ""
                    },
                    {
                        "first": "Varun",
                        "middle": [],
                        "last": "Feng",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Gangal",
                        "suffix": ""
                    },
                    {
                        "first": "Sarath",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    },
                    {
                        "first": "Soroush",
                        "middle": [],
                        "last": "Chandar",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Vosoughi",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Steven Y. Feng, Varun Gangal, Jason Wei, Sarath Chan- dar, Soroush Vosoughi, Teruko Mitamura, and Ed- uard H. Hovy. 2021. A survey of data augmentation approaches for NLP. CoRR, abs/2105.03075.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The role of grammar instruction in japanese efl context: Towards communicative language teaching",
                "authors": [
                    {
                        "first": "Yumi",
                        "middle": [],
                        "last": "Fujiwara",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Journal of the Academic Society for Quality of Life",
                "volume": "4",
                "issue": "4",
                "pages": "1--11",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yumi Fujiwara. 2018. The role of grammar instruction in japanese efl context: Towards communicative lan- guage teaching. Journal of the Academic Society for Quality of Life (JAS4QoL), 4(4):1-11.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Allennlp: A deep semantic natural language processing platform",
                "authors": [
                    {
                        "first": "Matt",
                        "middle": [],
                        "last": "Gardner",
                        "suffix": ""
                    },
                    {
                        "first": "Joel",
                        "middle": [],
                        "last": "Grus",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Neumann",
                        "suffix": ""
                    },
                    {
                        "first": "Oyvind",
                        "middle": [],
                        "last": "Tafjord",
                        "suffix": ""
                    },
                    {
                        "first": "Pradeep",
                        "middle": [],
                        "last": "Dasigi",
                        "suffix": ""
                    },
                    {
                        "first": "Nelson",
                        "middle": [
                            "F"
                        ],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Matthew",
                        "middle": [
                            "E"
                        ],
                        "last": "Peters",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Schmitz",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew E. Peters, Michael Schmitz, and Luke Zettlemoyer. 2018. Allennlp: A deep semantic natural language processing platform. CoRR, abs/1803.07640.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "The icnale and sophisticated contrastive interlanguage analysis of asian learners of english",
                "authors": [
                    {
                        "first": "Shinichiro",
                        "middle": [],
                        "last": "Ishikawa",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Learner Corpus Studies in Asia and the World",
                "volume": "1",
                "issue": "",
                "pages": "91--118",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shinichiro Ishikawa. 2013. The icnale and sophis- ticated contrastive interlanguage analysis of asian learners of english. Learner Corpus Studies in Asia and the World, 1:91-118.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "The icnale edited essays: A dataset for analysis of l2 english learner essays based on a new integrative viewpoint",
                "authors": [
                    {
                        "first": "Shinichiro",
                        "middle": [],
                        "last": "Ishikawa",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "English Corpus Linguistics",
                "volume": "25",
                "issue": "",
                "pages": "1--14",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shinichiro Ishikawa. 2018. The icnale edited essays: A dataset for analysis of l2 english learner essays based on a new integrative viewpoint. English Corpus Lin- guistics, 25:1-14.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "The esl student and the revision process: Some insights from scheme theory",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Ann",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Johns",
                        "suffix": ""
                    }
                ],
                "year": 1986,
                "venue": "Journal of Basic Writing",
                "volume": "5",
                "issue": "2",
                "pages": "70--80",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ann M. Johns. 1986. The esl student and the revision process: Some insights from scheme theory. Jour- nal of Basic Writing, 5(2):70 -80.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Cultural thought patterns in inter-cultural education",
                "authors": [
                    {
                        "first": "Robert",
                        "middle": [
                            "B"
                        ],
                        "last": "Kaplan",
                        "suffix": ""
                    }
                ],
                "year": 1966,
                "venue": "Language Learning",
                "volume": "16",
                "issue": "1-2",
                "pages": "1--20",
                "other_ids": {
                    "DOI": [
                        "10.1111/j.1467-1770.1966.tb00804.x"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Robert B. Kaplan. 1966. Cultural thought patterns in inter-cultural education. Language Learning, 16(1- 2):1-20.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Adam: a method for stochastic optimization",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Diederik",
                        "suffix": ""
                    },
                    {
                        "first": "Jimmy",
                        "middle": [
                            "Lei"
                        ],
                        "last": "Kingma",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of International Conference on Learning Representations",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Diederik P. Kingma and Jimmy Lei Ba. 2015. Adam: a method for stochastic optimization. In Proceedings of International Conference on Learning Represen- tations (ICLR).",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Investigating the role of argumentation in the rhetorical analysis of scientific publications with neural multi-task learning models",
                "authors": [
                    {
                        "first": "Anne",
                        "middle": [],
                        "last": "Lauscher",
                        "suffix": ""
                    },
                    {
                        "first": "Goran",
                        "middle": [],
                        "last": "Glava\u0161",
                        "suffix": ""
                    },
                    {
                        "first": "Simone",
                        "middle": [
                            "Paolo"
                        ],
                        "last": "Ponzetto",
                        "suffix": ""
                    },
                    {
                        "first": "Kai",
                        "middle": [],
                        "last": "Eckert",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "3326--3338",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D18-1370"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Anne Lauscher, Goran Glava\u0161, Simone Paolo Ponzetto, and Kai Eckert. 2018. Investigating the role of ar- gumentation in the rhetorical analysis of scientific publications with neural multi-task learning mod- els. In Proceedings of the 2018 Conference on Em- pirical Methods in Natural Language Processing, pages 3326-3338, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Argument Mining: A Survey",
                "authors": [
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Lawrence",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Reed",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Computational Linguistics",
                "volume": "45",
                "issue": "4",
                "pages": "765--818",
                "other_ids": {
                    "DOI": [
                        "10.1162/coli_a_00364"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "John Lawrence and Chris Reed. 2020. Argument Mining: A Survey. Computational Linguistics, 45(4):765-818.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Argumentation mining: State of the art and emerging trends",
                "authors": [
                    {
                        "first": "Marco",
                        "middle": [],
                        "last": "Lippi",
                        "suffix": ""
                    },
                    {
                        "first": "Paolo",
                        "middle": [],
                        "last": "Torroni",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "ACM Trans. Internet Technol",
                "volume": "16",
                "issue": "2",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.1145/2850417"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Marco Lippi and Paolo Torroni. 2016. Argumentation mining: State of the art and emerging trends. ACM Trans. Internet Technol., 16(2).",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "A survey of text data augmentation",
                "authors": [
                    {
                        "first": "Pei",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Xuemin",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Chao",
                        "middle": [],
                        "last": "Xiang",
                        "suffix": ""
                    },
                    {
                        "first": "Weiye",
                        "middle": [],
                        "last": "Meng",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "2020 International Conference on Computer Communication and Network Security (CCNS)",
                "volume": "",
                "issue": "",
                "pages": "191--195",
                "other_ids": {
                    "DOI": [
                        "10.1109/CCNS50731.2020.00049"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Pei Liu, Xuemin Wang, Chao Xiang, and Weiye Meng. 2020. A survey of text data augmentation. In 2020 International Conference on Computer Communica- tion and Network Security (CCNS), pages 191-195.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Towards better non-tree argument mining: Proposition-level biaffine parsing with task-specific parameterization",
                "authors": [
                    {
                        "first": "Gaku",
                        "middle": [],
                        "last": "Morio",
                        "suffix": ""
                    },
                    {
                        "first": "Hiroaki",
                        "middle": [],
                        "last": "Ozaki",
                        "suffix": ""
                    },
                    {
                        "first": "Terufumi",
                        "middle": [],
                        "last": "Morishita",
                        "suffix": ""
                    },
                    {
                        "first": "Yuta",
                        "middle": [],
                        "last": "Koreeda",
                        "suffix": ""
                    },
                    {
                        "first": "Kohsuke",
                        "middle": [],
                        "last": "Yanai",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "3259--3266",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.acl-main.298"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Gaku Morio, Hiroaki Ozaki, Terufumi Morishita, Yuta Koreeda, and Kohsuke Yanai. 2020. Towards bet- ter non-tree argument mining: Proposition-level bi- affine parsing with task-specific parameterization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3259-3266, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Computer Intensive Methods for Testing Hypotheses: An Introduction",
                "authors": [
                    {
                        "first": "Eric",
                        "middle": [
                            "W"
                        ],
                        "last": "Noreen",
                        "suffix": ""
                    }
                ],
                "year": 1989,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Eric W. Noreen. 1989. Computer Intensive Methods for Testing Hypotheses: An Introduction. Wiley, New York.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "A corpus of eRulemaking user comments for measuring evaluability of arguments",
                "authors": [
                    {
                        "first": "Joonsuk",
                        "middle": [],
                        "last": "Park",
                        "suffix": ""
                    },
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Cardie",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Joonsuk Park and Claire Cardie. 2018. A corpus of eRulemaking user comments for measuring evalua- bility of arguments. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. Eu- ropean Language Resources Association (ELRA).",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Pytorch: An imperative style, high-performance deep learning library",
                "authors": [
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Paszke",
                        "suffix": ""
                    },
                    {
                        "first": "Sam",
                        "middle": [],
                        "last": "Gross",
                        "suffix": ""
                    },
                    {
                        "first": "Francisco",
                        "middle": [],
                        "last": "Massa",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Lerer",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Bradbury",
                        "suffix": ""
                    },
                    {
                        "first": "Gregory",
                        "middle": [],
                        "last": "Chanan",
                        "suffix": ""
                    },
                    {
                        "first": "Trevor",
                        "middle": [],
                        "last": "Killeen",
                        "suffix": ""
                    },
                    {
                        "first": "Zeming",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "Natalia",
                        "middle": [],
                        "last": "Gimelshein",
                        "suffix": ""
                    },
                    {
                        "first": "Luca",
                        "middle": [],
                        "last": "Antiga",
                        "suffix": ""
                    },
                    {
                        "first": "Alban",
                        "middle": [],
                        "last": "Desmaison",
                        "suffix": ""
                    },
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "Kopf",
                        "suffix": ""
                    },
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Zachary",
                        "middle": [],
                        "last": "Devito",
                        "suffix": ""
                    },
                    {
                        "first": "Martin",
                        "middle": [],
                        "last": "Raison",
                        "suffix": ""
                    },
                    {
                        "first": "Alykhan",
                        "middle": [],
                        "last": "Tejani",
                        "suffix": ""
                    },
                    {
                        "first": "Sasank",
                        "middle": [],
                        "last": "Chilamkurthy",
                        "suffix": ""
                    },
                    {
                        "first": "Benoit",
                        "middle": [],
                        "last": "Steiner",
                        "suffix": ""
                    },
                    {
                        "first": "Lu",
                        "middle": [],
                        "last": "Fang",
                        "suffix": ""
                    },
                    {
                        "first": "Junjie",
                        "middle": [],
                        "last": "Bai",
                        "suffix": ""
                    },
                    {
                        "first": "Soumith",
                        "middle": [],
                        "last": "Chintala",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "32",
                "issue": "",
                "pages": "8024--8035",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Te- jani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Py- torch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch\u00e9-Buc, E. Fox, and R. Gar- nett, editors, Advances in Neural Information Pro- cessing Systems 32, pages 8024-8035. Curran Asso- ciates, Inc.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Joint prediction in MST-style discourse parsing for argumentation mining",
                "authors": [
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "Peldszus",
                        "suffix": ""
                    },
                    {
                        "first": "Manfred",
                        "middle": [],
                        "last": "Stede",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "938--948",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D15-1110"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Andreas Peldszus and Manfred Stede. 2015. Joint pre- diction in MST-style discourse parsing for argumen- tation mining. In Proceedings of the 2015 Confer- ence on Empirical Methods in Natural Language Processing, pages 938-948, Lisbon, Portugal. Asso- ciation for Computational Linguistics.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "An annotated corpus of argumentative microtexts",
                "authors": [
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "Peldszus",
                        "suffix": ""
                    },
                    {
                        "first": "Manfred",
                        "middle": [],
                        "last": "Stede",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Argumentation and Reasoned Action -Proceedings of the 1st European Conference on Argumentation",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andreas Peldszus and Manfred Stede. 2016. An anno- tated corpus of argumentative microtexts. In Argu- mentation and Reasoned Action -Proceedings of the 1st European Conference on Argumentation, Lisbon, 2015.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Here's my point: Joint pointer architecture for argument mining",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Potash",
                        "suffix": ""
                    },
                    {
                        "first": "Alexey",
                        "middle": [],
                        "last": "Romanov",
                        "suffix": ""
                    },
                    {
                        "first": "Anna",
                        "middle": [],
                        "last": "Rumshisky",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1364--1373",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D17-1143"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Peter Potash, Alexey Romanov, and Anna Rumshisky. 2017. Here's my point: Joint pointer architecture for argument mining. In Proceedings of the 2017 Conference on Empirical Methods in Natural Lan- guage Processing, pages 1364-1373, Copenhagen, Denmark. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Annotating argumentative structure in english-as-foreign-language learner essays",
                "authors": [
                    {
                        "first": "Jan Wira Gotama",
                        "middle": [],
                        "last": "Putra",
                        "suffix": ""
                    },
                    {
                        "first": "Simone",
                        "middle": [],
                        "last": "Teufel",
                        "suffix": ""
                    },
                    {
                        "first": "Takenobu",
                        "middle": [],
                        "last": "Tokunaga",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Natural Language Engineering",
                "volume": "",
                "issue": "",
                "pages": "1--37",
                "other_ids": {
                    "DOI": [
                        "10.1017/S1351324921000218"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jan Wira Gotama Putra, Simone Teufel, and Takenobu Tokunaga. 2021a. Annotating argumentative struc- ture in english-as-foreign-language learner essays. Natural Language Engineering, pages 1-37.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Parsing argumentative structure in English-as-foreign-language essays",
                "authors": [
                    {
                        "first": "Jan Wira Gotama",
                        "middle": [],
                        "last": "Putra",
                        "suffix": ""
                    },
                    {
                        "first": "Simone",
                        "middle": [],
                        "last": "Teufel",
                        "suffix": ""
                    },
                    {
                        "first": "Takenobu",
                        "middle": [],
                        "last": "Tokunaga",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 16th Workshop on Innovative Use of NLP for Building Educational Applications",
                "volume": "",
                "issue": "",
                "pages": "97--109",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jan Wira Gotama Putra, Simone Teufel, and Takenobu Tokunaga. 2021b. Parsing argumentative structure in English-as-foreign-language essays. In Proceed- ings of the 16th Workshop on Innovative Use of NLP for Building Educational Applications, pages 97- 109, Online. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "On the similarities between native, non-native and translated texts",
                "authors": [
                    {
                        "first": "Ella",
                        "middle": [],
                        "last": "Rabinovich",
                        "suffix": ""
                    },
                    {
                        "first": "Sergiu",
                        "middle": [],
                        "last": "Nisioi",
                        "suffix": ""
                    },
                    {
                        "first": "Noam",
                        "middle": [],
                        "last": "Ordan",
                        "suffix": ""
                    },
                    {
                        "first": "Shuly",
                        "middle": [],
                        "last": "Wintner",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "1870--1881",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P16-1176"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ella Rabinovich, Sergiu Nisioi, Noam Ordan, and Shuly Wintner. 2016. On the similarities between native, non-native and translated texts. In Proceed- ings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pages 1870-1881, Berlin, Germany. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Sentence-BERT: Sentence embeddings using Siamese BERTnetworks",
                "authors": [
                    {
                        "first": "Nils",
                        "middle": [],
                        "last": "Reimers",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "3982--3992",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D19-1410"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Nils Reimers and Iryna Gurevych. 2019. Sentence- BERT: Sentence embeddings using Siamese BERT- networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing (EMNLP-IJCNLP), pages 3982-3992.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Transformer-based models for automatic identification of argument relations: A cross-domain evaluation",
                "authors": [
                    {
                        "first": "Ramon",
                        "middle": [],
                        "last": "Ruiz-Dolz",
                        "suffix": ""
                    },
                    {
                        "first": "Jose",
                        "middle": [],
                        "last": "Alemany",
                        "suffix": ""
                    },
                    {
                        "first": "Stella",
                        "middle": [],
                        "last": "Heras",
                        "suffix": ""
                    },
                    {
                        "first": "Ana",
                        "middle": [],
                        "last": "Garcia-Fornes",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "IEEE Intelligent Systems",
                "volume": "",
                "issue": "",
                "pages": "1--1",
                "other_ids": {
                    "DOI": [
                        "10.1109/MIS.2021.3073993"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ramon Ruiz-Dolz, Jose Alemany, Stella Heras, and Ana Garcia-Fornes. 2021. Transformer-based mod- els for automatic identification of argument rela- tions: A cross-domain evaluation. IEEE Intelligent Systems, pages 1-1.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Multi-task learning for argumentation mining in low-resource settings",
                "authors": [
                    {
                        "first": "Claudia",
                        "middle": [],
                        "last": "Schulz",
                        "suffix": ""
                    },
                    {
                        "first": "Steffen",
                        "middle": [],
                        "last": "Eger",
                        "suffix": ""
                    },
                    {
                        "first": "Johannes",
                        "middle": [],
                        "last": "Daxenberger",
                        "suffix": ""
                    },
                    {
                        "first": "Tobias",
                        "middle": [],
                        "last": "Kahse",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "2",
                "issue": "",
                "pages": "35--41",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N18-2006"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Claudia Schulz, Steffen Eger, Johannes Daxenberger, Tobias Kahse, and Iryna Gurevych. 2018. Multi-task learning for argumentation mining in low-resource settings. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 35-41, New Orleans, Louisiana. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Toward an understanding of the distinct nature of l2 writing: The esl research and its implications",
                "authors": [
                    {
                        "first": "Tony",
                        "middle": [],
                        "last": "Silva",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "TESOL Quarterly",
                "volume": "27",
                "issue": "4",
                "pages": "657--677",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tony Silva. 1993. Toward an understanding of the dis- tinct nature of l2 writing: The esl research and its implications. TESOL Quarterly, 27(4):657-677.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "Discourse selfattention for discourse element identification in argumentative student essays",
                "authors": [
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Song",
                        "suffix": ""
                    },
                    {
                        "first": "Ziyao",
                        "middle": [],
                        "last": "Song",
                        "suffix": ""
                    },
                    {
                        "first": "Ruiji",
                        "middle": [],
                        "last": "Fu",
                        "suffix": ""
                    },
                    {
                        "first": "Lizhen",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Miaomiao",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Ting",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "2820--2830",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.emnlp-main.225"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Wei Song, Ziyao Song, Ruiji Fu, Lizhen Liu, Miao- miao Cheng, and Ting Liu. 2020. Discourse self- attention for discourse element identification in ar- gumentative student essays. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2820-2830, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Identifying argumentative discourse structures in persuasive essays",
                "authors": [
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Stab",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "46--56",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/D14-1006"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Christian Stab and Iryna Gurevych. 2014. Identifying argumentative discourse structures in persuasive es- says. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 46-56, Doha, Qatar. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "Parsing argumentation structures in persuasive essays",
                "authors": [
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Stab",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Computational Linguistics",
                "volume": "43",
                "issue": "3",
                "pages": "619--659",
                "other_ids": {
                    "DOI": [
                        "10.1162/COLI_a_00295"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Christian Stab and Iryna Gurevych. 2017. Parsing ar- gumentation structures in persuasive essays. Com- putational Linguistics, 43(3):619-659.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "Crosstopic argument mining from heterogeneous sources",
                "authors": [
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Stab",
                        "suffix": ""
                    },
                    {
                        "first": "Tristan",
                        "middle": [],
                        "last": "Miller",
                        "suffix": ""
                    },
                    {
                        "first": "Benjamin",
                        "middle": [],
                        "last": "Schiller",
                        "suffix": ""
                    },
                    {
                        "first": "Pranav",
                        "middle": [],
                        "last": "Rai",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "3664--3674",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D18-1402"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Christian Stab, Tristan Miller, Benjamin Schiller, Pranav Rai, and Iryna Gurevych. 2018. Cross- topic argument mining from heterogeneous sources. In Proceedings of the 2018 Conference on Em- pirical Methods in Natural Language Processing, pages 3664-3674, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF41": {
                "ref_id": "b41",
                "title": "End-to-end argument mining as biaffine dependency parsing",
                "authors": [
                    {
                        "first": "Yuxiao",
                        "middle": [],
                        "last": "Ye",
                        "suffix": ""
                    },
                    {
                        "first": "Simone",
                        "middle": [],
                        "last": "Teufel",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume",
                "volume": "",
                "issue": "",
                "pages": "669--678",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yuxiao Ye and Simone Teufel. 2021. End-to-end ar- gument mining as biaffine dependency parsing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Lin- guistics: Main Volume, pages 669-678, Online. As- sociation for Computational Linguistics.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF1": {
                "text": "Distribution of distance (in percent) between related sentences in the ICNALE-AS2R corpus.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF2": {
                "text": "shows the distance distribution between the source and target sentences in the ..., +15].",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF3": {
                "text": "Biaffine attention model (BIAF).",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF4": {
                "text": "Our proposed extension for the BIAF model. Newly added modules are coloured.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF5": {
                "text": "S1) [To conclude, art could play an active role in improving the quality of people's lives,] (S2) [but I think that governments should attach heavier weight to other social issues such as education and housing needs] (S3) [because those are the most essential ways enable to make people a decent life.] Illustration of conversion of PEC's segmentlevel annotation to the sentence-level (essay075). Annotated AC segments are written in bold.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF6": {
                "text": "Example of descendant set matching between structure A (left) and B (right). Matching scores are written in red. Grey nodes represent non-AC.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF7": {
                "text": "",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF9": {
                "text": "Model performance across distances. Model performance across depths.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "TABREF1": {
                "num": null,
                "text": "Acc. F1-macro MAR dset",
                "html": null,
                "type_str": "table",
                "content": "<table><tr><td>Baseline</td><td/><td/><td/></tr><tr><td>BIAF</td><td>.471</td><td>.323</td><td>.419</td></tr><tr><td>MTL</td><td/><td/><td/></tr><tr><td>BIAF+QACT</td><td>.473</td><td>.333</td><td>422</td></tr><tr><td>BIAF+QACT+ND</td><td>.472</td><td>.338</td><td>423</td></tr><tr><td>Spos</td><td/><td/><td/></tr><tr><td>BIAF+QACT+spos</td><td>.472</td><td>.327</td><td>.421</td></tr><tr><td colspan=\"2\">BIAF+QACT+ND+spos .475</td><td>.336</td><td>426</td></tr><tr><td>Multi-corpora Training</td><td/><td/><td/></tr><tr><td colspan=\"2\">BIAF+QACT+ND [P+I] .468</td><td>.360</td><td>.455</td></tr><tr><td colspan=\"2\">BIAF+QACT+ND [SS] .489  \u2020</td><td>.374  \u2020</td><td>.452</td></tr></table>"
            },
            "TABREF2": {
                "num": null,
                "text": "",
                "html": null,
                "type_str": "table",
                "content": "<table><tr><td>Model</td><td colspan=\"5\">Major Claim AC (non-leaf) AC (leaf) Non-AC F1-macro</td></tr><tr><td>Baseline</td><td/><td/><td/><td/><td/></tr><tr><td>BIAF</td><td>.730</td><td>.639</td><td>.599</td><td>.437</td><td>.601</td></tr><tr><td>MTL</td><td/><td/><td/><td/><td/></tr><tr><td>BIAF+QACT</td><td>.739</td><td>.639</td><td>.601</td><td>.453</td><td>.608</td></tr><tr><td>BIAF+QACT+ND</td><td>.734</td><td>.636</td><td>.601</td><td>.454</td><td>.606</td></tr><tr><td>Spos</td><td/><td/><td/><td/><td/></tr><tr><td>BIAF+QACT+spos</td><td>.725</td><td>.641</td><td>.602</td><td>.438</td><td>.602</td></tr><tr><td>BIAF+QACT+ND+spos</td><td>.738</td><td>.638</td><td>.603</td><td>.460</td><td>.610</td></tr><tr><td>Multi-corpora Training</td><td/><td/><td/><td/><td/></tr><tr><td>BIAF+QACT+ND [P+I]</td><td>748</td><td>.606</td><td>.634  \u2020</td><td>.420</td><td>.602</td></tr><tr><td>BIAF+QACT+ND [SS]</td><td>.767  \u2020</td><td>.633</td><td>.628</td><td>.462</td><td>.622  \u2020</td></tr></table>"
            },
            "TABREF3": {
                "num": null,
                "text": "Results for the QACT prediction. Node labels are automatically identified from the predicted topology in the main task. This table shows F1 score per node label and F1-macro. Bold-face, \u2020 and underline as above.",
                "html": null,
                "type_str": "table",
                "content": "<table/>"
            },
            "TABREF5": {
                "num": null,
                "text": "Structural-output qualities of linking models. Bold-face, \u2020 and underline as above.",
                "html": null,
                "type_str": "table",
                "content": "<table/>"
            }
        }
    }
}