{ "paper_id": "I11-1039", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:30:45.626131Z" }, "title": "Predicting Opinion Dependency Relations for Opinion Analysis", "authors": [ { "first": "Lun-Wei", "middle": [], "last": "Ku", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Yunlin University of Science and Technology Douliou", "location": { "postCode": "64002", "settlement": "Yunlin", "country": "Taiwan" } }, "email": "lwku@yuntech.edu.tw" }, { "first": "Ting-Hao", "middle": [ "Kenneth" ], "last": "Huang", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan University", "location": { "addrLine": "No.1, Sec.4, Roosevelt Rd", "settlement": "Taipei", "country": "Taiwan" } }, "email": "tinghaoh@andrew.cmu.edu" }, { "first": "Hsin-Hsi", "middle": [], "last": "Chen", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan University", "location": { "addrLine": "No.1, Sec.4, Roosevelt Rd", "settlement": "Taipei", "country": "Taiwan" } }, "email": "hhchen@ntu.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Syntactic structures have been good features for opinion analysis, but it is not easy to use them. To find these features by supervised learning methods, correct syntactic labels are indispensible. Two possible sources to acquire syntactic structures are parsing trees and dependency trees. For the annotation processing, parsing trees are more readable for annotators, while dependency trees are easier to use by programs. To use syntactic structures as features, this paper tried to annotate on human friendly materials and transform these annotations to the corresponding machine friendly materials. We annotated the gold answers of opinion syntactic structures on the parsing tree from Chinese Treebank, and then proposed methods to find their corresponding dependency relations on the dependency trees generated from the same sentence. With these relations, we could train a model to annotate opinion dependency relations automatically to provide an opinion dependency parser, which is language independent if language resources are incorporated. Experiment results show that the annotated syntactic structures and their corresponding dependency relations improve at least 8% of the performance of opinion analysis.", "pdf_parse": { "paper_id": "I11-1039", "_pdf_hash": "", "abstract": [ { "text": "Syntactic structures have been good features for opinion analysis, but it is not easy to use them. To find these features by supervised learning methods, correct syntactic labels are indispensible. Two possible sources to acquire syntactic structures are parsing trees and dependency trees. For the annotation processing, parsing trees are more readable for annotators, while dependency trees are easier to use by programs. To use syntactic structures as features, this paper tried to annotate on human friendly materials and transform these annotations to the corresponding machine friendly materials. We annotated the gold answers of opinion syntactic structures on the parsing tree from Chinese Treebank, and then proposed methods to find their corresponding dependency relations on the dependency trees generated from the same sentence. With these relations, we could train a model to annotate opinion dependency relations automatically to provide an opinion dependency parser, which is language independent if language resources are incorporated. Experiment results show that the annotated syntactic structures and their corresponding dependency relations improve at least 8% of the performance of opinion analysis.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Opinion analysis has drawn much attention in research communities of machine learning and natural language processing. In the early stages, words in documents were used as the main features (Pang et al., 2002) . Some opinion dictionaries were created for this demand . However, researchers soon realized that word features were not sufficient for acquiring good performances, so they started to include syntactic structures and semantic in-formation (Qiu et al., 2008) . Their researches showed that linguistic knowledge is helpful in determining opinions.", "cite_spans": [ { "start": 190, "end": 209, "text": "(Pang et al., 2002)", "ref_id": "BIBREF12" }, { "start": 450, "end": 468, "text": "(Qiu et al., 2008)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "For various applications related to opinions, syntactic structures have become powerful tools for extracting useful clues. To find opinions in product reviews, modification relations were used to identify the product and their features (Lu et al., 2009) , e.g., a good price (feature) of this camera (product). To find opinion holders and targets, templates and linguistic rules were adopted (Breck et al., 2007) . To find more opinion words, dependency relations were utilized (Qiu et al., 2011) . Even when applying the basic negation rule that flips opinion polarity over, we need to find its modified word first by syntactic clues. However, we will show that syntactic relations do not directly suggest opinions.", "cite_spans": [ { "start": 236, "end": 253, "text": "(Lu et al., 2009)", "ref_id": "BIBREF10" }, { "start": 392, "end": 412, "text": "(Breck et al., 2007)", "ref_id": "BIBREF2" }, { "start": 478, "end": 496, "text": "(Qiu et al., 2011)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Syntactic relations are obtained usually from all kinds of syntax trees. Parsing trees (phrase structured) and dependency trees (grammatical) are the most commonly seen ones. Parsing trees are in-order trees which keep the order of words in sentences, so they are more readable for people. Instead, nodes in dependency trees are displayed by the head-modifier relations, in which the sentence sequence probably is not remained. People could find the opinion passages if they can understand the whole sentence, i.e. from parsing trees. However, when the linguistic background is needed, it could be difficult for most people to reconstruct the whole sentence from the dependency trees in order to find the opinion passage. Therefore, if we want to find annotators to build a corpus which could be used to train an opinion relation recognizer, parsing trees are the better materials compared to dependency trees. However, compared to relations between words, complicated tree structures are more challenge to be utilized by algorithms (Doan et al., 2008) . This paper focuses on extracting opinionated dependency relations from relations generated by the Stanford parser. We design an annotation mechanism on the syntactic structures on the sentence from Chinese Treebank to create an annotation environment with a lower entry barrier so that sufficient annotations can be labeled. Then these annotations are aligned to the relations in the corresponding dependency trees generated by the same parser from the same sentence as the gold standard for training the automatic annotator of the opinion dependency relations. We conduct experiments on the annotated opinion syntactic structures in parsing trees, and on the opinion dependency relations corresponding to them. The proposed process demonstrates a feasible direction toward the development of an opinion dependency parser.", "cite_spans": [ { "start": 1033, "end": 1052, "text": "(Doan et al., 2008)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Given a set of non-collapsed dependencies parsed from a specific sentence by the Stanford dependency parser (de Marneffe and Manning, 2008; Chang et al., 2009) , each associated with a dependency relation between two words in this sentence, our goal is to identify which of them are with sentiment, i.e., those which reveal a part of opinions or the aroused emotions. For example, in the sentence \"\u6d3b\u52a8 \u53d6\u5f97 \u4e86 \u5706\u6ee1 \u6210\u529f (Activities scored le perfect success)\", the Stanford dependency parser gives three relations: nmod(\u6210\u529f , \u5706 \u6ee1 ), nsubj( \u53d6 \u5f97 , \u6d3b \u52a8 ), dobj(\u53d6\u5f97 , \u6210\u529f), and asp(\u53d6\u5f97 , \u4e86). The goal is to identify the former three may bear sentiment or opinions. The corresponding dependency tree is shown in Figure 1 . From Figure 1 we can also see that it is not easy to read the original sentence without the linguistic background. Formally, the collection of the noncollapsed dependency relations of a sentence S, generated by the Stanford dependency parser, is denoted by Rdep(S) = {r 1 , r 2 , \u2026}, where each ) (S Rdep r i \uf0ce is associate with an opinion judgment of op(r).", "cite_spans": [ { "start": 108, "end": 139, "text": "(de Marneffe and Manning, 2008;", "ref_id": "BIBREF11" }, { "start": 140, "end": 159, "text": "Chang et al., 2009)", "ref_id": "BIBREF3" } ], "ref_spans": [ { "start": 762, "end": 770, "text": "Figure 1", "ref_id": "FIGREF0" }, { "start": 778, "end": 786, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Problem Definition", "sec_num": "2" }, { "text": "Definition: Dependency Relation The dependency relation r, generated by the Stanford parser, is composed of the type of relation rel, the head word w h and the modifier word w m in the form of rel(w h , w m ). w h and w m are two individual words in S. For example, in one relation in Figure 1 , r = nmod(\u6210\u529f , \u5706\u6ee1 ), where rel = nmod, w h =\u6210\u529f , and w m =\u5706\u6ee1. A list of rel is available in Stanford Parser Manual (de Marneffe and Manning, 2008; Chang et al., 2009) .", "cite_spans": [ { "start": 451, "end": 478, "text": "Marneffe and Manning, 2008;", "ref_id": "BIBREF11" }, { "start": 479, "end": 498, "text": "Chang et al., 2009)", "ref_id": "BIBREF3" } ], "ref_spans": [ { "start": 285, "end": 293, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Problem Definition", "sec_num": "2" }, { "text": "Definition: Opinion Judgment The opinion judgment op(r), generated by the proposed system, indicates whether the corresponding dependency relation r is opinionated, and \uf07b \uf07d false true r op , ) ( \uf0ce . For example, when r = nmod(\u6210\u529f , \u5706\u6ee1), op(r) =", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Definition", "sec_num": "2" }, { "text": "Definition: Gold Opinion Judgment The gold opinion judgment, generated by mapping from manually annotated data, indicates whether the corresponding dependency relation r is opinionated, and", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "true.", "sec_num": null }, { "text": "\uf07b \uf07d false true r gop , ) ( \uf0ce .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "true.", "sec_num": null }, { "text": "The gold answers come from the annotations on Chinese Treebank 5.1. In a parsing tree T of the sentence S, generated by the Stanford parser, an in-ordered set of tree nodes O = {o 1 , o 2 , \u2026} is used to draw a parsing tree for the annotation process, and its corresponding order, i.e., its index, is used as the node ID to record the annotations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "true.", "sec_num": null }, { "text": "The way we annotate an opinion relation on a parsing tree is annotating an opinion trio (Ku et al., 2009 Note that because the annotation of trios is on nodes of parsing trees, which appear inorderly, o left will always appear before o right in a sentence, and keeping this in mind will help understand the meaning of each inter-word relation t.", "cite_spans": [ { "start": 88, "end": 104, "text": "(Ku et al., 2009", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "true.", "sec_num": null }, { "text": "Now for the sentence S, we have its parsing tree T, the annotated opinion trios Tri(S) on it, and its dependency relations Rdep(S). The next step is to mark the op(r) on Rdep(S) according to its corresponding Tri(S). For each trio tri, if any descendent of its left node o left and any descendent of its right node o right together build a relation ) (S Rdep r \uf0ce , the opinion judgment of gop(r) of the relation r is set to true. Otherwise, gop(r) is set to false. Now we have gop(r) for each r in Rdep(S), our goal is to find good methods to generate op(r) so that it can predict gop(r) as precisely as possible. We propose methods to achieve this goal in Section 3.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "true.", "sec_num": null }, { "text": "As mentioned, our goal is to predict opinion dependency relations as precisely as possible. However, to use more readable materials, opinion trios are first annotated on Chinese Treebank 5.1, and then they are mapped to the corresponding dependency relations. Before the aligning process, we use the annotated trios for training to predict the opinion trios in Section 3.1. Using these predict trios for opinion analysis shall show the performance before the aligning process. After that, the aligned depen-dency relations, i.e., the gold opinion dependency relations are adopted for training to predict the opinion dependency relations in Section 3.2. Because the parsing tree and the dependency tree are generated by the same parser, we can always align them by the provided word ID numbers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methods", "sec_num": "3" }, { "text": "After prediction, the opinion dependency relations are available, and they can provide necessary information for many applications. However, we go one step further to test whether they benefit the opinion analysis. To fulfill this purpose, a basic method which uses the opinion dependency relations to extract opinionated sentences and determine their polarities is proposed in Section 3.3.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methods", "sec_num": "3" }, { "text": "We predict the opinion trios by the sequential labeling model Conditional Random Field (CRF, Lafferty et al., 2001) . In a parsing tree, the tag of the internal node is the syntactic structure of its sub-tree, and the tag of the leaf node contains its part of speech and the content word. For each node, tags of its first four children (the first level), first four children of them (the second level), and their three children are used as features of this node. Features of its siblings (the window size is five) are considered, too.", "cite_spans": [ { "start": 87, "end": 115, "text": "(CRF, Lafferty et al., 2001)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Predicting Opinion Trios", "sec_num": "3.1" }, { "text": "The labels l we would like the CRF to predict labels for each node, which are N or labels of the form t-C, where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting Opinion Trios", "sec_num": "3.1" }, { "text": "Rpt t \uf0ce , } , { R L C \uf03d , L", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting Opinion Trios", "sec_num": "3.1" }, { "text": "indicates that the current node is o left in some opinion trio and R indicates o right . The label N indicates that the current node does not belong to any Tri tri \uf0ce . The cardinality of the set Rpt is five, so that a total of 11 labels are used in CRF. CRF++ 1 is selected for experiments.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting Opinion Trios", "sec_num": "3.1" }, { "text": "After aligning the opinion trios to the dependency relations, we will have gop(r) for each one of them. In the previous research, usually only some relations were selected for opinion analysis. No statistical numbers showed the connection between the dependency relations and the opinions. We believe that it is because the opinion annotation on dependency relations is more difficult than on words, sentences, or documents. However, because of this alignment, we are able to see the distribution of different dependency relations in opinion sentences and opinion segments (opinion trios).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicting Opinion Dependency Relations", "sec_num": "3.2" }, { "text": "We then predict opinion dependency relations based on these distributions: the op(r) of the relation r is set to true when its corresponding gop(r) appears massively frequently to be true in opinion sentences. To make this method reasonable, the assumption that there are no opinion trios in non-opinionated sentences must hold. A similar assumption that there are no opinion segments in non-opinionated sentences was made when annotating the NTCIR MOAT corpus, too (Seki et al., 2008) . Under this assumption, the relation that is in most case opinionated in opinion sentences is also in most case opinionated in all sentences. We believe that this assumption holds because intuitively if there is an opinion segment in one sentence, this sentence should be opinionated.", "cite_spans": [ { "start": 466, "end": 485, "text": "(Seki et al., 2008)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Predicting Opinion Dependency Relations", "sec_num": "3.2" }, { "text": "In the previous research, relations were usually extracted automatically and then were used in various applications. As these relations are available after the prediction (or alignment) and as our purpose is to provide easy to use opinion dependency relations for further applications, we simply design rules for these relations in opinion analysis to show the baseline enhancement of using them.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Using Syntactic Information for Opinion Extraction", "sec_num": "3.3" }, { "text": "In the past, Ku et al. (2009) have conducted rule based experiments for opinion trios. They designed formula for trios of each", "cite_spans": [ { "start": 13, "end": 29, "text": "Ku et al. (2009)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Using opinion trios", "sec_num": "3.3.1" }, { "text": ". Therefore, we adopted their rules on our augmentative experiment materials. We define the opinion scoring function S(.), and its output opinion score varies with the input variables. These rules are shown by trio types as follows. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Rpt t \uf0ce", "sec_num": null }, { "text": "o S o S o o S o S SIGN o S SIGN o S o o S o S o S \uf02b \uf03d \uf0b4 \uf0b4 \uf03d \uf0b9 \uf0b9 (3)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Rpt t \uf0ce", "sec_num": null }, { "text": "\uf06c Verb-Complement Type: The scoring function for trios of this type is defined the same as that of a Subjective-Predicate type in Formula (2). The complement node is the deciding factor of the opinion score.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Rpt t \uf0ce", "sec_num": null }, { "text": "The usages of opinion dependency relations were seen in several researches (Bikel and Castelli, 2008) . In these researches, rules for a small number of major dependency relations were proposed in different papers but they were not listed together for a better utilization. Some rules were not ever mentioned in pervious researches. Instead, all relations are analyzed in this paper. For each relation r of which gop(r) equals true (when gold opinion relations are used for opinion analysis) or op(r) equals true (when predicted opinion relations are used for opinion analysis), we calculate its opinion score ops(r). Let RM(w) be a function to return the dependency relations of word w's modifiers one at a time, n is the total number of relations RM(w) returns, and S(.) is also the defined opinion scoring function then ops(r) is defined as in Formula (4).", "cite_spans": [ { "start": 75, "end": 101, "text": "(Bikel and Castelli, 2008)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Using opinion dependency relations", "sec_num": "3.3.2" }, { "text": "\uf0e5 \uf02b \uf03d )) ( ( 1 ) , , ( ) ( m m h w RM ops n w w rel S r ops (4)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Using opinion dependency relations", "sec_num": "3.3.2" }, { "text": "That is, the opinion score of a dependency relation is an average of the aggregate scores of its descendent dependency relations. In practice, we design different rules for calculating opinion scores by the current relation type rel in S(.). Here to simplify the problem, we adopted Formula (1) and treated w m as o left and w h as o right in it.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Using opinion dependency relations", "sec_num": "3.3.2" }, { "text": "Though there were researches which predicted opinion dependency relations, they did not predict directly from the parsing results. Instead, they predicted from documents or sentences according to the context and a large quantity of training instances were needed. They did not predict on all dependency relations either. Therefore, there is no existing dataset containing correct opinion labels on dependency relations. In this section, we describe how to generate opinionated syntactic dataset on parsing trees, and align the annotated labels to dependency trees. After that, qualitative and quantitative analyses of opinion dependency relations are provided. At the end, we discuss the evaluation results of the proposed methods.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "4" }, { "text": "To use the Stanford parser as our tool to generate dependency tree for experimental sentences and to avoid errors as possible, we adopted Chinese Treebank 5.1 as experiment materials. Sentences in Chinese Treebank are already segmented and part of speech tagged, and its tagging set is the same with the one Stanford parser uses. Therefore, the Stanford parser can take the data from the Chinese Treebank to generate more accurate dependency trees.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Set and Preprocessing", "sec_num": "4.1" }, { "text": "The dataset Chinese Treebank 5.1 contains 507,222 words, 824,983 Hanzi, 18,782 sentences, and 890 data files. For the opinion analysis experiments, opinionated labels, i.e., opinionated, non-opinionated, positive, neutral, negative, were annotated on all sentences in Chinese Opinion Treebank. Afterward 57,706 trios were annotated on the parsing trees of gold opinion sentences, i.e., sentences which were annotated as opinionated. Methods for generating the gold opinion sentences proposed by were adopted.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Set and Preprocessing", "sec_num": "4.1" }, { "text": "Next, the Stanford parser took all sentences in Chinese Treebank as input to generate their dependency trees. A total of 416,581 dependency relations were generated, and 284,590 of them were in opinion sentences. Then the annotated trios were aligned to their corresponding dependency relations, and because trios were only annotated on opinionated sentences, the gop(r) of these aligned relations were set to true. At the end, a total of 54,753 relations gop(r) were set to true. Table 2 . Statistics of structural trios. Table 1 shows the distribution of the opinion and polarity labels. Table 2 shows the statistics of trios. Trios of the Substantive-Modifier and Verb-Object types are the majority in opinion sentences, while trios of the Verb-Complement type are few. Table 3 further shows the distribution of dependency relations. It shows that previously the most adopted dependency relations for opinion analysis, e.g., amod (adjective modifier) or advmod (adverb modifier), do not certainly bear opinions or appear in opinion sentences. In Section 4.3, we will further test the performance of finding the opinionated relations with the help of the opinion word dictionary, which was also widely adopted by previous work (Feng et al., 2009) .", "cite_spans": [ { "start": 1229, "end": 1248, "text": "(Feng et al., 2009)", "ref_id": "BIBREF5" } ], "ref_spans": [ { "start": 481, "end": 488, "text": "Table 2", "ref_id": null }, { "start": 523, "end": 530, "text": "Table 1", "ref_id": null }, { "start": 590, "end": 597, "text": "Table 2", "ref_id": null }, { "start": 773, "end": 780, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Data Set and Preprocessing", "sec_num": "4.1" }, { "text": "In this section, results of predicting opinion trios by CRF mentioned in Section 3.1 are shown. We first predicted the appearance of o left and o right in trios, and then predicted the trio type", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation of Opinion Trio Prediction", "sec_num": "4.2" }, { "text": "for each trio. The performance in Table 4 is not promising. Therefore, we consider the structure of trios, that is, o left and o right should appear as an ordered pair, and otherwise the label was viewed as illegal. The performance is shown in Table 5 . Table 5 shows that all predicted trios were opinionated, and this tells that some opinion trios are of certain structures, but not all of them. We observed that the precisions 1.00 came from the collocations of specific words and structures, while the low recalls were from other trios which were not identified. However, these results still confirmed that we can find opinion trios by phrase structures and they may benefit in the opinion analysis process. Table 3 . Distributions of dependency relations and opinion dependency relations.", "cite_spans": [], "ref_spans": [ { "start": 34, "end": 41, "text": "Table 4", "ref_id": null }, { "start": 244, "end": 252, "text": "Table 5", "ref_id": null }, { "start": 255, "end": 262, "text": "Table 5", "ref_id": null }, { "start": 713, "end": 720, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Rpt t \uf0ce", "sec_num": null }, { "text": "(A: type of dependency relations (rel); B: total occurrences in generated dependency trees; C: total occurrence in generated dependency trees of opinions; D: total occurrence in generated dependency trees of opinions when it bears opinions (gop(r) equals true); E: percentage that this relation appears in generated dependency trees of opinions; F: percentage that this relation appears in generated dependency trees of opinions when it bears opinions (gop(r) equals true).)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Rpt t \uf0ce", "sec_num": null }, { "text": "To predict which dependency relations are opinionated, we start with analyzing the distribution of them. Table 3 presented the distribution of dependency relations. The percentage of a relation appearing in dependency trees of opinion sentences when bearing opinions (gop(r) equals true), i.e., the value in F column, is taken as the support value. The support value indicates that in what degree this relation bears opinions. If the support value is high, it is confident to say that the relation is opinionated; otherwise, considering the content words is necessary. This idea conforms to the previous observation in Section 4.2: some of the opinions are structural, but not all of them. According to the support value, dependency relations were divided into four categories. The Chinese opinion word dictionary NTUSD is involved to help identify opinion dependency relations when the support value is not high. The selecting criteria are listed as follows. \uf06c Very supportive: with the support value above 0.8, e.g., dvpmod. Relations in this category are viewed as opinionated and their gop(r) are automatically set to true. \uf06c Supportive: with the support value above 0.35 but lower than 0. 8, e.g., pass, dobj, npsubj, ba, top, nsubj, neg, amod, rcmod Table 6 . Performance of predicting opinion dependency relations.", "cite_spans": [ { "start": 1194, "end": 1255, "text": "8, e.g., pass, dobj, npsubj, ba, top, nsubj, neg, amod, rcmod", "ref_id": null } ], "ref_spans": [ { "start": 105, "end": 112, "text": "Table 3", "ref_id": null }, { "start": 1256, "end": 1263, "text": "Table 6", "ref_id": null } ], "eq_spans": [], "section": "Evaluation of Opinion Dependency Relation Prediction", "sec_num": "4.3" }, { "text": "The results of two experiment settings are listed: prediction performed on all dependency relations and on only modification-related dependency relations (in the form of lex-mod, e.g., amod, rcmod, etc.) The later are the relations adopted in many previous researches. Table 6 shows the performance of predicting opinion dependency relations. It indicates that if only modification related relations were considered, the f-score dropped nearly half because more than half of the opinion dependency relations were expelled in this case. In other word, results show that predicting on all relations instead of taking only modificationrelated dependency relations as clues can capture more opinion relations, and hence the prediction of opinion relations is necessary.", "cite_spans": [], "ref_spans": [ { "start": 269, "end": 276, "text": "Table 6", "ref_id": null } ], "eq_spans": [], "section": "Evaluation of Opinion Dependency Relation Prediction", "sec_num": "4.3" }, { "text": "In this section, predicted opinion trios and predicted opinion dependency relations were utilized in an opinion extraction system. In order to make use of these structural cues, opinion analysis methods proposed by were selected. Their methods calculated opinion scores of sentences from characters and words accumulatively, so syntactic cues can be added in and function jointly. Five settings for opinion analysis were experimented: \uf06c C+W+N: characters, words, and negations were used as cues for calculating opinion scores. It was the original method proposed by Ku et al. \uf06c C+W+N+goldTrio: annotated opinion trios were utilized additionally. \uf06c C+W+N+Trio: predicted opinion trios were utilized additionally. \uf06c C+W+N+goldDep: opinion dependency relations aligned from the annotated trios were utilized additionally. \uf06c C+W+N+Dep: predicted opinion dependency relations were utilized additionally. The results were shown in Table 7 . The performance of the opinion extraction improves 10.40% (0.7162->0.7993) when utilizing opinion trios and 8.66% (0.7162->0.7782) when utilizing opinion dependency relations. These results clearly indicate that the syntactic information benefit opinion analysis. Because of the possible information loss in the automatic alignment process, that the performance of using trios is a little better than using dependency relations matches our expectation. Table 7 . Performance of using syntactic information for opinion analysis.", "cite_spans": [], "ref_spans": [ { "start": 925, "end": 932, "text": "Table 7", "ref_id": null }, { "start": 1388, "end": 1395, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Evaluation of Opinion Extraction Using Predicted Opinion Trios and Dependency Relations", "sec_num": "4.4" }, { "text": "For all we know, no previous work has annotated opinion information on all dependency relations, or mapped annotated opinionated structures to dependency relations on a large quantity of documents or sentences. Therefore, to the best of our knowledge, no statistically analysis of opinion dependency relations involving manually annotations has been conducted. Researchers designed ruled or extracted dependency relations as features for opinion analysis based on their linguistic knowledge (Qiu et al., 2011 ). Yet there are still several lines of related work, including (i) opinion analysis (ii) opinion corpora (iii) syntactic information. Several dozen papers have been published on the topic of opinion analysis. Two general approaches have been proposed previously. They are machine learning approaches and heuristicrule approaches. For both approaches, syntactic structures could be utilized. For the former, they can be used as features (Abbasi et al., 2008) ; for the later, rules can be designed according to them (Ku et al., 2009) . We can see from the previous work that syntactic structures can help to enhance the performance.", "cite_spans": [ { "start": 491, "end": 508, "text": "(Qiu et al., 2011", "ref_id": "BIBREF14" }, { "start": 946, "end": 967, "text": "(Abbasi et al., 2008)", "ref_id": "BIBREF0" }, { "start": 1025, "end": 1042, "text": "(Ku et al., 2009)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "As to the experimental corpora, some researchers managed to generate annotated materials and gold standards under constraints. Somasundaran (2007) annotated discourse information from meeting dialogs to train a sentiment model. MPQA annotated opinions and their sources (Wiebe et al., 2002) . NTCIR annotated opinions, polarities, sources, and targets for its multilingual opinion analysis task (MOAT, Seki et al., 2008) . However, none of them were annotated on materials with syntac-tic structures, and it caused the lack of analysis of opinion syntactic structures.", "cite_spans": [ { "start": 127, "end": 146, "text": "Somasundaran (2007)", "ref_id": "BIBREF16" }, { "start": 270, "end": 290, "text": "(Wiebe et al., 2002)", "ref_id": "BIBREF17" }, { "start": 395, "end": 420, "text": "(MOAT, Seki et al., 2008)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "Researchers have acquired syntactic structures (Zhou, 2008) , but few of them have tried to associate syntactic structures with opinions. The most similar previous work to ours was proposed by Ku et al. (2009) . Compared to it, the proposed process made the development of opinion dependency parser feasible. As dependency relations and the predicted opinion dependency relations are of the same form, no extra knowledge or integration is needed for the use of them.", "cite_spans": [ { "start": 47, "end": 59, "text": "(Zhou, 2008)", "ref_id": "BIBREF18" }, { "start": 193, "end": 209, "text": "Ku et al. (2009)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "The proposed new process is the main contribution of this paper. It annotated opinion syntactic structures on phrase structure trees, which are more readable for annotators, and aligned these structures to grammatical structures, which facilitates their usage. Chinese Treebank was selected as the source of phrase structure trees, and dependency relations as the grammatical structures. They are both widely used in natural language processing.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" }, { "text": "Though the experiments were implemented on Chinese materials, this process is language independent. It can be applied to materials in different languages without modifications.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" }, { "text": "By predicting opinion dependency relations, we can say that a basic opinion dependency parser has been developed. Experiments have shown that the predicted opinion dependency relations are beneficial for opinion extraction. Although we still need a parser to generate syntactic structures, parsing is relatively a mature technique in natural language processing. For a comparably new research problem like opinion analysis, it is common that tools are not handy. The best of the proposed method is that it can function in a multilingual environment by incorporating a domain or language specific resources (here, NTUSD for Chinese).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" }, { "text": "Through the alignment, we made a large quantity of opinion dependency relations available. According to their distributions shown in this paper, researchers can select suitable relations to use according to their diverse needs, such as extracting evaluative features in product reviews or comments, opinions or their polarities.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" }, { "text": "http://crfpp.sourceforge.net/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Sentiment analysis in multiple languages: Feature selection for opinion classification in Web forums", "authors": [ { "first": "A", "middle": [], "last": "Abbasi", "suffix": "" }, { "first": "H", "middle": [], "last": "Chen", "suffix": "" }, { "first": "A", "middle": [], "last": "Salem", "suffix": "" } ], "year": 2008, "venue": "ACM Trans. Inf. Syst", "volume": "26", "issue": "", "pages": "1--34", "other_ids": { "DOI": [ "10.1145/1361684" ] }, "num": null, "urls": [], "raw_text": "Abbasi, A., Chen, H., and Salem, A. 2008. Senti- ment analysis in multiple languages: Feature se- lection for opinion classification in Web forums. ACM Trans. Inf. Syst. 26, 3 (Jun. 2008), 1-34. DOI= http://doi.acm.org/10.1145/1361684.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Event Matching Using the Transitive Closure of Dependency Relations", "authors": [ { "first": "D", "middle": [ "M" ], "last": "Bikel", "suffix": "" }, { "first": "V", "middle": [], "last": "Castelli", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 46th Annual Meeting on Association for Computational Linguistics", "volume": "", "issue": "", "pages": "145--148", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bikel, D. M. and Castelli, V. 2008. Event Matching Using the Transitive Closure of Dependency Re- lations. Proceedings of the 46th Annual Meeting on Association for Computational Linguistics, pages 145-148.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Identifying Expressions of Opinion in Context", "authors": [ { "first": "E", "middle": [], "last": "Breck", "suffix": "" }, { "first": "Y", "middle": [], "last": "Choi", "suffix": "" }, { "first": "C", "middle": [], "last": "Cardie", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the 20th International Joint Conferences on Artificial Intelligence", "volume": "", "issue": "", "pages": "2683--2688", "other_ids": {}, "num": null, "urls": [], "raw_text": "Breck, E., Choi, Y. and Cardie, C. 2007. Identify- ing Expressions of Opinion in Context. Proceed- ings of the 20th International Joint Conferences on Artificial Intelligence, pages 2683-2688.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Discriminative Reordering with Chinese Grammatical Relations Features", "authors": [ { "first": "P.-C", "middle": [], "last": "Chang", "suffix": "" }, { "first": "H", "middle": [], "last": "Tseng", "suffix": "" }, { "first": "D", "middle": [], "last": "Jurafsky", "suffix": "" }, { "first": "C", "middle": [ "D" ], "last": "Manning", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Third Workshop on Syntax and Structure in Statistical Translation", "volume": "", "issue": "", "pages": "51--59", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chang, P.-C., Tseng, H., Jurafsky, D. and Manning C. D. 2009. Discriminative Reordering with Chinese Grammatical Relations Features. Pro- ceedings of the Third Workshop on Syntax and Structure in Statistical Translation, pages 51-59.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Searching Question by Identifying Question Topic and Question Focus", "authors": [ { "first": "H", "middle": [], "last": "Doan", "suffix": "" }, { "first": "Y", "middle": [], "last": "Cao", "suffix": "" }, { "first": "C.-Y", "middle": [], "last": "Lin", "suffix": "" }, { "first": "Y", "middle": [], "last": "Yu", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 46th Annual Meeting on Association for Computational Linguistics", "volume": "", "issue": "", "pages": "156--164", "other_ids": {}, "num": null, "urls": [], "raw_text": "Doan, H., Cao, Y., Lin, C.-Y. and Yu, Y. 2008. Searching Question by Identifying Question Topic and Question Focus. Proceedings of the 46th Annual Meeting on Association for Compu- tational Linguistics, pages 156-164.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Chinese Blog Clustering by Hidden Sentiment Factors. ADMA", "authors": [ { "first": "S", "middle": [], "last": "Feng", "suffix": "" }, { "first": "D", "middle": [], "last": "Wang", "suffix": "" }, { "first": "G", "middle": [], "last": "Yu", "suffix": "" }, { "first": "C", "middle": [], "last": "Yang", "suffix": "" }, { "first": "N", "middle": [], "last": "Yang", "suffix": "" } ], "year": 2009, "venue": "", "volume": "5678", "issue": "", "pages": "140--151", "other_ids": {}, "num": null, "urls": [], "raw_text": "Feng, S., Wang, D., Yu, G., Yang, C. and Yang, N. 2009. Chinese Blog Clustering by Hidden Sen- timent Factors. ADMA, Vol. 5678, Springer (2009), pages 140-151.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Mining Opinions from the Web: Beyond Relevance Retrieval", "authors": [ { "first": "L.-W", "middle": [], "last": "Ku", "suffix": "" }, { "first": "H.-H", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2007, "venue": "Journal of American Society for Information Science and Technology, Special Issue on Mining Web Resources for Enhancing Information Retrieval", "volume": "58", "issue": "12", "pages": "1838--1850", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ku, L.-W. and Chen, H.-H. 2007. Mining Opinions from the Web: Beyond Relevance Retrieval. Journal of American Society for Information Science and Technology, Special Issue on Min- ing Web Resources for Enhancing Information Retrieval, 58(12), 1838-1850.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Using Morphological and Syntactic Structures for Chinese Opinion Analysis", "authors": [ { "first": "L.-W", "middle": [], "last": "Ku", "suffix": "" }, { "first": "T.-H", "middle": [], "last": "Huang", "suffix": "" }, { "first": "H.-H", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2009, "venue": "Proceedings of Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1260--1269", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ku, L.-W., Huang, T.-H. and Chen, H.-H. 2009. Using Morphological and Syntactic Structures for Chinese Opinion Analysis. Proceedings of Conference on Empirical Methods in Natural Language Processing, pages 1260-1269.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Test Collection Selection and Gold Standard Generation for a Multiply-Annotated Opinion Corpus", "authors": [ { "first": "L.-W", "middle": [], "last": "Ku", "suffix": "" }, { "first": "Y.-S", "middle": [], "last": "Lo", "suffix": "" }, { "first": "Chen H.-H", "middle": [], "last": "", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the 45th Annual Meeting on Association for Computational Linguistics", "volume": "", "issue": "", "pages": "89--92", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ku, L.-W., Lo, Y.-S. and Chen H.-H. 2007. Test Collection Selection and Gold Standard Genera- tion for a Multiply-Annotated Opinion Corpus. Proceedings of the 45th Annual Meeting on As- sociation for Computational Linguistics, pages 89-92.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data", "authors": [ { "first": "J", "middle": [], "last": "Lafferty", "suffix": "" }, { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" }, { "first": "F", "middle": [], "last": "Pereira", "suffix": "" } ], "year": 2001, "venue": "Proceedings of International Conference on Machine Learning", "volume": "", "issue": "", "pages": "282--289", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lafferty, J., McCallum, A. and Pereira, F. 2001. Conditional Random Fields: Probabilistic Mod- els for Segmenting and Labeling Sequence Data. Proceedings of International Conference on Ma- chine Learning, pages 282-289.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Rated Aspect Summarization of Short Comments", "authors": [ { "first": "Y", "middle": [], "last": "Lu", "suffix": "" }, { "first": "C", "middle": [ "X" ], "last": "Zhai", "suffix": "" }, { "first": "N", "middle": [], "last": "Sundaresan", "suffix": "" } ], "year": 2009, "venue": "Proceedings of 18 th International World Wide Web Conference", "volume": "", "issue": "", "pages": "131--140", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lu, Y., Zhai, C.X. and Sundaresan, N. 2009. Rated Aspect Summarization of Short Comments. Pro- ceedings of 18 th International World Wide Web Conference, pages 131-140.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Stanford typed dependencies manual", "authors": [ { "first": ", M.-C", "middle": [], "last": "De Marneffe", "suffix": "" }, { "first": "C", "middle": [ "D" ], "last": "Manning", "suffix": "" } ], "year": 2008, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "de Marneffe, M.-C. and Manning, C. D. 2008. Stan- ford typed dependencies manual. Technichal re- port. http://nlp.stanford.edu/software/depend encies_manual.pdf", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Thumbs up? Sentiment classification using machine learning techniques", "authors": [ { "first": "B", "middle": [], "last": "Pang", "suffix": "" }, { "first": "L", "middle": [], "last": "Lee", "suffix": "" }, { "first": "S", "middle": [], "last": "Vaithyanathan", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "79--86", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pang, B., Lee, L. and Vaithyanathan, S. 2002. Thumbs up? Sentiment classification using ma- chine learning techniques. Proceedings of the 2002 Conference on Empirical Methods in Natu- ral Language Processing, pages 79-86.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Incorporate the Syntactic. Knowledge in Opinion Mining in User-generated Content", "authors": [ { "first": "G", "middle": [], "last": "Qiu", "suffix": "" }, { "first": "C", "middle": [], "last": "Wang", "suffix": "" }, { "first": "J", "middle": [], "last": "Bu", "suffix": "" }, { "first": "K", "middle": [], "last": "Liu", "suffix": "" }, { "first": "C", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2008, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Qiu, G., Wang, C., Bu, J., Liu, K. and Chen, C. 2008. Incorporate the Syntactic. Knowledge in Opinion Mining in User-generated Content. Pro- ceedings of NLPIX'08.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Opinion Word Expansion and Target Extraction through Double Propagation", "authors": [ { "first": "G", "middle": [], "last": "Qiu", "suffix": "" }, { "first": "B", "middle": [], "last": "Liu", "suffix": "" }, { "first": "J", "middle": [], "last": "Bu", "suffix": "" }, { "first": "C", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2011, "venue": "Computational Linguistics", "volume": "37", "issue": "1", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Qiu, G., Liu, B., Bu, J. and Chen, C. 2011. Opinion Word Expansion and Target Extraction through Double Propagation. Computational Linguistics, March 2011, Vol. 37, No. 1: 9.27", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Overview of Multilingual Opinion Analysis Task at NTCIR-7", "authors": [ { "first": "Y", "middle": [], "last": "Seki", "suffix": "" }, { "first": "D", "middle": [ "K" ], "last": "Evans", "suffix": "" }, { "first": "L.-W", "middle": [], "last": "Ku", "suffix": "" }, { "first": "L", "middle": [], "last": "Sun", "suffix": "" }, { "first": "H.-H", "middle": [], "last": "Chen", "suffix": "" }, { "first": "N", "middle": [], "last": "Kando", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 7th NTCIR Workshop Meeting on Evaluation of Information Access Technologies:Information Retrieval, Question Answering, and Cross-Lingual Information Access", "volume": "", "issue": "", "pages": "185--203", "other_ids": {}, "num": null, "urls": [], "raw_text": "Seki, Y., D. K. Evans, L.-W. Ku, L. Sun, H.-H. Chen and N. Kando. 2008. Overview of Multi- lingual Opinion Analysis Task at NTCIR-7. Proceedings of the 7th NTCIR Workshop Meet- ing on Evaluation of Information Access Tech- nologies:Information Retrieval, Question Ans- wering, and Cross-Lingual Information Access, pages 185-203.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Detecting arguing and sentiment in meetings", "authors": [ { "first": "S", "middle": [], "last": "Somasundaran", "suffix": "" }, { "first": "J", "middle": [], "last": "Ruppenhofer", "suffix": "" }, { "first": "J", "middle": [], "last": "Wiebe", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the SIGdial Workshop on Discourse and Dialogue", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Somasundaran, S., J. Ruppenhofer and J. Wiebe. 2007. Detecting arguing and sentiment in meet- ings. Proceedings of the SIGdial Workshop on Discourse and Dialogue 2007.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "NRRC summer workshop on multi-perspective question answering, final report. ARDA NRRC Summer", "authors": [ { "first": "J", "middle": [], "last": "Wiebe", "suffix": "" }, { "first": "E", "middle": [], "last": "Breck", "suffix": "" }, { "first": "C", "middle": [], "last": "Buckly", "suffix": "" }, { "first": "C", "middle": [], "last": "Cardie", "suffix": "" }, { "first": "P", "middle": [], "last": "Davis", "suffix": "" }, { "first": "B", "middle": [], "last": "Fraser", "suffix": "" }, { "first": "D", "middle": [], "last": "Litman", "suffix": "" }, { "first": "D", "middle": [], "last": "Pierce", "suffix": "" }, { "first": "E", "middle": [], "last": "Riloff", "suffix": "" }, { "first": "T", "middle": [], "last": "Wilson", "suffix": "" } ], "year": 2002, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Wiebe, J., E. Breck, C. Buckly, C. Cardie, P. Davis, B. Fraser, D. Litman, D. Pierce, E. Riloff and T. Wilson. 2002. NRRC summer workshop on mul- ti-perspective question answering, final report. ARDA NRRC Summer 2002 Workshop.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Automatic rule acquisition for Chinese intra-chunk relations", "authors": [ { "first": "Q", "middle": [], "last": "Zhou", "suffix": "" } ], "year": 2008, "venue": "Proceedings of International Joint Conference of Natural Language Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhou, Q. 2008. Automatic rule acquisition for Chi- nese intra-chunk relations. Proceedings of Inter- national Joint Conference of Natural Language Processing (IJCNLP-2008).", "links": null } }, "ref_entries": { "FIGREF0": { "text": "A sample dependency tree with three aligned opinion dependency relations.", "uris": null, "type_str": "figure", "num": null }, "FIGREF1": { "text": "A sample parsing tree and opinion trios within the sentence inFigure 1are shown inFigure 2. The literal output of opinion trios are shown in", "uris": null, "type_str": "figure", "num": null }, "FIGREF2": { "text": "In the trio tri = (3, NP-OBJ, \u5706\u6ee1, \u6210\u529f, Substantive-Modifier), triID = 3, o parent = NP-OBJ, o left =\u5706\u6ee1 (perfect), o right =\u6210\u529f(success), and t = Substantive-Modifier.", "uris": null, "type_str": "figure", "num": null }, "FIGREF3": { "text": "A sample parsing tree with trios. 1, IP, \u6d3b\u52a8, VP, Subjective-Predicate 2, VP, \u53d6\u5f97, NP-OBJ, Verb-Object 3, NP-OBJ, \u5706\u6ee1, \u6210\u529f, Substantive-Modifier", "uris": null, "type_str": "figure", "num": null }, "FIGREF4": { "text": "Figure 3. Opinion trios", "uris": null, "type_str": "figure", "num": null }, "FIGREF5": { "text": "\uf06c", "uris": null, "type_str": "figure", "num": null }, "TABREF1": { "type_str": "table", "html": null, "num": null, "text": "o left of this trio type modifies o right , so that the trio's opinion weight comes from the absolute opinion score of o left , while the opinion polarity is determined by the occurrence of negative o left or o right . If at least one of them is negative, the trio is negative, else it is positive. Predicate Type: o left of this trio type is a subject and o right is the action it performs, so that the action decides the opinion score of the trio. If the action is not an opinion or it is neutral, the subject determines the opinion score of this trio.", "content": "
if(S(left o)\uf0b90andS(oright)\uf0b90)then
( o left left ) o S ( S else ( if0 right o \uf03e-and ) \uf03d1 S \uf0b4 (( right left ) o S o) \uf03e0)thenS(left oright o)\uf03dS(left o)(1)
elseS(left oright o)\uf03dS(left o)\uf02bS(oright)
else ( if S S (( o o right left) o) ) 0 right \uf0b9( then S \uf03d( left S o) oleftoright)\uf03dS(oright)(2)
if((left)0and(right)0)
then(leftright)(left)((left))((right))
else(leftright)(left)(right)
" } } } }