Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
built_at: string
source: string
selection_method: string
selection_rule: string
figure_count: int64
annotation_count: int64
paper_count: int64
annotator_count_distribution: struct<3: int64, 4: int64, 6: int64, 7: int64, 8: int64, 9: int64>
  child 0, 3: int64
  child 1, 4: int64
  child 2, 6: int64
  child 3, 7: int64
  child 4, 8: int64
  child 5, 9: int64
venue_distribution: struct<ACL: int64, EMNLP: int64, ICML: int64, NeurIPS: int64>
  child 0, ACL: int64
  child 1, EMNLP: int64
  child 2, ICML: int64
  child 3, NeurIPS: int64
year_distribution: struct<2020: int64, 2021: int64, 2022: int64, 2023: int64, 2024: int64, 2025: int64>
  child 0, 2020: int64
  child 1, 2021: int64
  child 2, 2022: int64
  child 3, 2023: int64
  child 4, 2024: int64
  child 5, 2025: int64
with_context_figures: int64
without_context_figures: int64
with_context_ratio: double
human_gold_standard: struct<join_key: list<item: string>, aggregation: string, annotator_filter: string>
  child 0, join_key: list<item: string>
      child 0, item: string
  child 1, aggregation: string
  child 2, annotator_filter: string
target_figures: int64
output_dir: string
venue: string
paper_id_norm: string
height: int64
annotations: list<item: struct<source_type: string, source_name: string, visual_clarity: double, structure_layout (... 343 chars omitted)
  child 0, item: struct<source_type: string, source_name: string, visual_clarity: double, structure_layout: double, c (... 331 chars omitted)
      child 0, source_type: string
      child 1, source_name: string
      child 2, visual_clarity: double
      child 3, structure_layout: double
      child 4, caption_consistency: double
      child 5, context_consistency: double
      child 6, misleading_risk: double
      child 7, overall_score: double
      child 8, summary: string
      child 9, suggestion: string
      child 10, visual_clarity_reason: string
      child 11, structure_layout_reason: string
      child 12, caption_consistency_reason: string
      child 13, context_consistency_reason: string
      child 14, misleading_risk_reason: string
      child 15, annotated_at: string
context_count: int64
title: string
context_texts: list<item: string>
  child 0, item: string
caption: string
fig_index: int64
domain_l1: string
annotator_count: int64
image_file: string
width: int64
year: int64
domain_l2: string
paper_id: string
section: string
to
{'paper_id': Value('string'), 'paper_id_norm': Value('string'), 'venue': Value('string'), 'year': Value('int64'), 'fig_index': Value('int64'), 'title': Value('string'), 'caption': Value('string'), 'image_file': Value('string'), 'domain_l1': Value('string'), 'domain_l2': Value('string'), 'width': Value('int64'), 'height': Value('int64'), 'section': Value('string'), 'context_texts': List(Value('string')), 'context_count': Value('int64'), 'annotator_count': Value('int64'), 'annotations': List({'source_type': Value('string'), 'source_name': Value('string'), 'visual_clarity': Value('float64'), 'structure_layout': Value('float64'), 'caption_consistency': Value('float64'), 'context_consistency': Value('float64'), 'misleading_risk': Value('float64'), 'overall_score': Value('float64'), 'summary': Value('string'), 'suggestion': Value('string'), 'visual_clarity_reason': Value('string'), 'structure_layout_reason': Value('string'), 'caption_consistency_reason': Value('string'), 'context_consistency_reason': Value('string'), 'misleading_risk_reason': Value('string'), 'annotated_at': Value('string')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1816, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              built_at: string
              source: string
              selection_method: string
              selection_rule: string
              figure_count: int64
              annotation_count: int64
              paper_count: int64
              annotator_count_distribution: struct<3: int64, 4: int64, 6: int64, 7: int64, 8: int64, 9: int64>
                child 0, 3: int64
                child 1, 4: int64
                child 2, 6: int64
                child 3, 7: int64
                child 4, 8: int64
                child 5, 9: int64
              venue_distribution: struct<ACL: int64, EMNLP: int64, ICML: int64, NeurIPS: int64>
                child 0, ACL: int64
                child 1, EMNLP: int64
                child 2, ICML: int64
                child 3, NeurIPS: int64
              year_distribution: struct<2020: int64, 2021: int64, 2022: int64, 2023: int64, 2024: int64, 2025: int64>
                child 0, 2020: int64
                child 1, 2021: int64
                child 2, 2022: int64
                child 3, 2023: int64
                child 4, 2024: int64
                child 5, 2025: int64
              with_context_figures: int64
              without_context_figures: int64
              with_context_ratio: double
              human_gold_standard: struct<join_key: list<item: string>, aggregation: string, annotator_filter: string>
                child 0, join_key: list<item: string>
                    child 0, item: string
                child 1, aggregation: string
                child 2, annotator_filter: string
              target_figures: int64
              output_dir: string
              venue: string
              paper_id_norm: string
              height: int64
              annotations: list<item: struct<source_type: string, source_name: string, visual_clarity: double, structure_layout (... 343 chars omitted)
                child 0, item: struct<source_type: string, source_name: string, visual_clarity: double, structure_layout: double, c (... 331 chars omitted)
                    child 0, source_type: string
                    child 1, source_name: string
                    child 2, visual_clarity: double
                    child 3, structure_layout: double
                    child 4, caption_consistency: double
                    child 5, context_consistency: double
                    child 6, misleading_risk: double
                    child 7, overall_score: double
                    child 8, summary: string
                    child 9, suggestion: string
                    child 10, visual_clarity_reason: string
                    child 11, structure_layout_reason: string
                    child 12, caption_consistency_reason: string
                    child 13, context_consistency_reason: string
                    child 14, misleading_risk_reason: string
                    child 15, annotated_at: string
              context_count: int64
              title: string
              context_texts: list<item: string>
                child 0, item: string
              caption: string
              fig_index: int64
              domain_l1: string
              annotator_count: int64
              image_file: string
              width: int64
              year: int64
              domain_l2: string
              paper_id: string
              section: string
              to
              {'paper_id': Value('string'), 'paper_id_norm': Value('string'), 'venue': Value('string'), 'year': Value('int64'), 'fig_index': Value('int64'), 'title': Value('string'), 'caption': Value('string'), 'image_file': Value('string'), 'domain_l1': Value('string'), 'domain_l2': Value('string'), 'width': Value('int64'), 'height': Value('int64'), 'section': Value('string'), 'context_texts': List(Value('string')), 'context_count': Value('int64'), 'annotator_count': Value('int64'), 'annotations': List({'source_type': Value('string'), 'source_name': Value('string'), 'visual_clarity': Value('float64'), 'structure_layout': Value('float64'), 'caption_consistency': Value('float64'), 'context_consistency': Value('float64'), 'misleading_risk': Value('float64'), 'overall_score': Value('float64'), 'summary': Value('string'), 'suggestion': Value('string'), 'visual_clarity_reason': Value('string'), 'structure_layout_reason': Value('string'), 'caption_consistency_reason': Value('string'), 'context_consistency_reason': Value('string'), 'misleading_risk_reason': Value('string'), 'annotated_at': Value('string')})}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

paper_id
string
paper_id_norm
string
venue
string
year
int64
fig_index
int64
title
string
caption
string
image_file
string
domain_l1
string
domain_l2
string
width
int64
height
int64
section
string
context_texts
list
context_count
int64
annotator_count
int64
annotations
list
0001_2020.acl-main.1
2020.acl-main.1
ACL
2,020
1
Learning to Understand Child-directed and Adult-directed Speech
Figure 1: Validation performance in early training on natural speech
processed/ACL/2020/figures_clean/images/0001_2020.acl-main.1__fig0001.png
computer_science
natural_language_processing
576
373
6 Results
[]
0
3
[ { "source_type": "human", "source_name": "Xiao Zhang (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 5, "context_consistency": null, "misleading_risk": 9, "overall_score": 8, "summary": "The figure is visually clear and well-structured, but its captio...
0001_2020.acl-main.1
2020.acl-main.1
ACL
2,020
2
Learning to Understand Child-directed and Adult-directed Speech
Figure 2: Validation performance in early training on synthetic speech
processed/ACL/2020/figures_clean/images/0001_2020.acl-main.1__fig0002.png
computer_science
natural_language_processing
573
379
6 Results
[]
0
3
[ { "source_type": "human", "source_name": "Yutong Wang (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 5, "context_consistency": null, "misleading_risk": 9, "overall_score": 8, "summary": "The figure is visually clear and well-structured, but its capti...
0002_2020.acl-main.2
2020.acl-main.2
ACL
2,020
1
Predicting Depression in Screening Interviews from Latent Categorization of Interview Prompts
Figure 1: The architecture of our *JLPC* model with K = 3. For each prompt Pij in interview i, the *Category Inference layer* computes a *latent category membership* vector, hij . These are used as weights to form K separate *Category-Aware Response Aggregations*, which in turn are used by the *Decision Layer* to predi...
processed/ACL/2020/figures_clean/images/0002_2020.acl-main.2__fig0001.png
computer_science
natural_language_processing
1,190
509
2 Joint Latent Prompt Categorization
[ "sponse to that prompt. Together, (Pij, Rij) form the jth turn in ith interview. Each interview Xi is labeled with a ground-truth class Yi ∈{1, ..C}, where C is the number of possible labels. In our case, there are two possible labels: depressed or not depressed. Our model, shown in Figure 1, takes as input an inte...
3
3
[ { "source_type": "human", "source_name": "Wanqing Xu (Human)", "visual_clarity": 8, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 8, "misleading_risk": 9, "overall_score": 8.600000381469727, "summary": "This figure provides a clear and well-structured ar...
0002_2020.acl-main.2
2020.acl-main.2
ACL
2,020
2
Predicting Depression in Screening Interviews from Latent Categorization of Interview Prompts
Figure 2: Ablation study on validation set demonstrating the importance of prompt categorization and entropy regularization for our model.
processed/ACL/2020/figures_clean/images/0002_2020.acl-main.2__fig0002.png
computer_science
natural_language_processing
435
264
4 Experiments
[ "We analyzed how the prompt categorization and entropy regularization contribute to our model’s validation performance. The contributions of each component are visualized in Figure 2. Our analy- sis shows that while both components are impor- tant, latent prompt categorization yields the high- est contribution to t...
1
3
[ { "source_type": "human", "source_name": "yucen qian", "visual_clarity": 6, "structure_layout": 5, "caption_consistency": 5, "context_consistency": 6, "misleading_risk": 4, "overall_score": 5.199999809265137, "summary": "This ablation study figure suffers from a lack of clarity r...
0003_2020.acl-main.3
2020.acl-main.3
ACL
2,020
1
Coach: A Coarse-to-Fine Approach for Cross-domain Slot Filling
Figure 1: Cross-domain slot filling frameworks.
processed/ACL/2020/figures_clean/images/0003_2020.acl-main.3__fig0001.png
computer_science
natural_language_processing
560
413
Abstract
[ "classification models from adapting to the target domain without any target domain supervision sig- nals. Recently, Bapna et al. (2017) proposed a cross-domain slot filling framework, which enables zero-shot adaptation. As illustrated in Figure 1a, their model conducts slot filling individually for each slot type. It...
2
3
[ { "source_type": "human", "source_name": "standard", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 6, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.399999618530273, "summary": "The figure clearly presents two slot filling frameworks with ...
0003_2020.acl-main.3
2020.acl-main.3
ACL
2,020
2
Coach: A Coarse-to-Fine Approach for Cross-domain Slot Filling
Figure 2: Illustration of our framework, *Coach*, and the template regularization approach.
processed/ACL/2020/figures_clean/images/0003_2020.acl-main.3__fig0002.png
computer_science
natural_language_processing
1,113
530
1 Introduction
[ "As depicted in Figure 2, the slot filling process in our Coach framework consists of two steps. In the first step, we utilize a BiLSTM-CRF struc- ture (Lample et al., 2016) to learn the general pattern of slot entities by having our model pre- dict whether tokens are slot entities or not (i.e.,", "In many cases, s...
2
3
[ { "source_type": "human", "source_name": "Chuanzhi Xu (Human)", "visual_clarity": 7, "structure_layout": 8, "caption_consistency": 9, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.399999618530273, "summary": "This figure provides a detailed and well-structure...
0004_2020.acl-main.4
2020.acl-main.4
ACL
2,020
2
Designing Precise and Robust Dialogue Response Evaluators
Figure 2: Performance of the RoBERTa evaluator w.r.t amount of supervised training data (§[6.2\)](#page-3-3).
processed/ACL/2020/figures_clean/images/0004_2020.acl-main.4__fig0002.png
computer_science
natural_language_processing
555
267
7 Conclusion
[ "of training with even fewer data. Figure 2 shows that, with only around 100 samples, the RoBERTa evaluator can reach performance close to the result obtained using the entire 720 samples." ]
1
3
[ { "source_type": "human", "source_name": "Xiao Zhang (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 7, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.600000381469727, "summary": "The figure is visually clear, well-structured, and ...
0004_2020.acl-main.4
2020.acl-main.4
ACL
2,020
3
Designing Precise and Robust Dialogue Response Evaluators
Figure 3: Distributions of human annotations and model outputs on the test data (90 responses).
processed/ACL/2020/figures_clean/images/0004_2020.acl-main.4__fig0003.png
computer_science
natural_language_processing
1,160
336
7 Conclusion
[ "The distribution of human annotation scores on the 900 annotated responses has been given in Fig- ure 1(a). To analyze the distribution of model out- puts, we show the distributions of human anno- tation, ADEM’s outputs, RUBER’s outputs, and RoBERTa-eval’s outputs on the test data of 90 re- sponses in Figure 3. We...
1
3
[ { "source_type": "human", "source_name": "Yuyan Lin (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 10, "context_consistency": 10, "misleading_risk": 10, "overall_score": 9.600000381469727, "summary": "This multi-panel figure is exceptionally clear, w...
0005_2020.acl-main.5
2020.acl-main.5
ACL
2,020
1
Dialogue State Tracking with Explicit Slot Connection Modeling
Figure 1: DST-SC model architecture (best viewed in color). Three processing flows leading to Pgen, Pwc, Pvc are respectively generation (brown), copying from dialogue history (green), copying from last dialogue states (purple).
processed/ACL/2020/figures_clean/images/0005_2020.acl-main.5__fig0001.png
computer_science
natural_language_processing
1,729
417
2 Model
[ "In this section, we will describe DST-SC model in detail. DST-SC is an open vocabulary model based on the encoder-decoder architecture. As shown in Figure 1, there are three components that contribute to obtain the target slot value: (1) word generation from the vocabulary; (2) word copying from the dialogue histo...
1
3
[ { "source_type": "human", "source_name": "Siyuan Zhao (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 10, "misleading_risk": 9, "overall_score": 9.199999809265137, "summary": "The figure clearly illustrates the DST-SC model a...
0006_2020.acl-main.6
2020.acl-main.6
ACL
2,020
1
Generating Informative Conversational Response using Recurrent Knowledge-Interaction and Knowledge-Copy
Figure 1: The architecture of KIC. Here, U t d is calculated by decode-input and utterance context vector C t u at current step , C t k represents the knowledge context vector resulted from dynamic knowledge attention. ugen and kgen are two soft switches that control the copy pointer to utterance attention distribution...
processed/ACL/2020/figures_clean/images/0006_2020.acl-main.6__fig0001.png
computer_science
natural_language_processing
1,206
545
2 Model Description
[ "Given a dataset D = {(Xi, Yi, Ki)}N i=1, where N is the size of the dataset, a dialog response Y = {y1, y2, ..., yn} is produced by the conver- sation history utterance X = {x1, x2, ..., xm}, using also the relative knowledge set K = {k1, k2, ..., ks}. Here, m and n are the numbers of tokens in the conversation hi...
2
3
[ { "source_type": "human", "source_name": "Junjie Ma (Human)", "visual_clarity": 7, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.600000381469727, "summary": "This figure presents a clear and well-structured arc...
0006_2020.acl-main.6
2020.acl-main.6
ACL
2,020
2
Generating Informative Conversational Response using Recurrent Knowledge-Interaction and Knowledge-Copy
Figure 2: Bleu improvements on Wizard-of-Wikipedia.
processed/ACL/2020/figures_clean/images/0006_2020.acl-main.6__fig0002.png
computer_science
natural_language_processing
507
327
3 Experiments
[ "Bleu metrics are shown in Figure 2, we can find that the improvement of result increasing with the augment of Bleu’s grams, which means the dia- log response produced via model KIC is more in line with the real distribution of ground-truth re- sponse in the phrase level, and the better improve- ment on higher gram’...
1
3
[ { "source_type": "human", "source_name": "yucen qian", "visual_clarity": 9, "structure_layout": 8, "caption_consistency": 5, "context_consistency": 8, "misleading_risk": 6, "overall_score": 7.199999809265137, "summary": "The plot is visually clear and well-structured, supported b...
0007_2020.acl-main.7
2020.acl-main.7
ACL
2,020
1
Guiding Variational Response Generator to Exploit Persona
Figure 1: The architecture of the Persona-Aware Variational Response Generator (PAGenerator) described in this paper. ⊕ represents the concatenation of inputs and CE denotes the cross-entropy of predictions. The dotted arrow line indicates the connection is optional, and the default model named PAGenerator decodes with...
processed/ACL/2020/figures_clean/images/0007_2020.acl-main.7__fig0001.png
computer_science
natural_language_processing
1,206
311
1 Introduction
[ "Utilizing latent variables in response generation has become a widely accepted methodology in NRG due to their Bayesian essence. It helps to deal with external knowledge efficiently, e.g. Per- sona. Therefore, our proposed model is built based on the generation model with latent variables. The overall architecture ...
2
3
[ { "source_type": "human", "source_name": "Xiao Zhang (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 8, "misleading_risk": 9, "overall_score": 8.800000190734863, "summary": "The figure provides an exceptionally clear and well...
0008_2020.acl-main.8
2020.acl-main.8
ACL
2,020
1
Large Scale Multi-Actor Generative Dialog Modeling
Figure 1: Illustration of input representation for a conversation from three different speakers (A, B, C) composed of a sequence of four turns (denoted A1 , B, A2 , C in the "Tokens" row) when separately modeling two different target speakers (A and C for (i) and (ii) respectively). The model receives different referen...
processed/ACL/2020/figures_clean/images/0008_2020.acl-main.8__fig0001.png
computer_science
natural_language_processing
1,049
270
4 Model
[ "Due to supporting multi-actor conversations present in our dataset, special care is needed for presenting this information to the model. In gen- eral, this is accomplished by designating a speaker of interest to model in a conversation. As visualized in Figure 1, the designated speaker’s reference history tokens a...
2
3
[ { "source_type": "human", "source_name": "Yutong Wang (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 8, "misleading_risk": 9, "overall_score": 8.800000190734863, "summary": "This figure is exceptionally clear and well-struct...
0008_2020.acl-main.8
2020.acl-main.8
ACL
2,020
2
Large Scale Multi-Actor Generative Dialog Modeling
Figure 2: Test scores compared against the number of dialog turns given as context prior to generating samples for GCC-DEC (355M) and GCC-NRC (355M).
processed/ACL/2020/figures_clean/images/0008_2020.acl-main.8__fig0002.png
computer_science
natural_language_processing
463
489
5 Experiments
[ "Reference use From our qualitative study we can clearly see the benefit of using reference history as was alluded to in prior sections. In all four exper- iments the presence of references leads to better ground truth performance compared to GCC-NRC . In Figure 2 we delve deeper into the results of the ground truth...
1
3
[ { "source_type": "human", "source_name": "standard", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 5, "context_consistency": 8, "misleading_risk": 7, "overall_score": 7.599999904632568, "summary": "The figure is visually clear and well-structured but suffers ...
0008_2020.acl-main.8
2020.acl-main.8
ACL
2,020
6
Large Scale Multi-Actor Generative Dialog Modeling
Figure 6: Coherency Task Turk Layout.
processed/ACL/2020/figures_clean/images/0008_2020.acl-main.8__fig0006.png
computer_science
natural_language_processing
1,208
539
7 Conclusions
[]
0
3
[ { "source_type": "human", "source_name": "zihan Deng (Human)", "visual_clarity": 7, "structure_layout": 9, "caption_consistency": 5, "context_consistency": null, "misleading_risk": 7, "overall_score": 7, "summary": "The figure effectively shows a human evaluation task layout, but...
0009_2020.acl-main.9
2020.acl-main.9
ACL
2,020
1
PLATO: Pre-trained Dialogue Generation Model with Discrete Latent Variable
Figure 1: Graphical illustration of response generation (gray lines) and latent act recognition (dashed blue lines).
processed/ACL/2020/figures_clean/images/0009_2020.acl-main.9__fig0001.png
computer_science
natural_language_processing
523
188
1 Introduction
[ "The probabilistic relationships among these el- ements are elaborated with the graphical model in Figure 1. Given a context c, there are multi- ple latent speech acts which can be taken as re- sponse intents (represented by the latent variable z). Conditioned on the context and one selected latent speech act, the ...
1
3
[ { "source_type": "human", "source_name": "Yuyan Lin (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 10, "misleading_risk": 9, "overall_score": 9.199999809265137, "summary": "The figure provides an exceptionally clear and well...
0009_2020.acl-main.9
2020.acl-main.9
ACL
2,020
2
PLATO: Pre-trained Dialogue Generation Model with Discrete Latent Variable
Figure 2: Architecture of dialogue generation with discrete latent variable. In self-attention visualization, red and blue lines denote bi-directional attention, and dashed orange lines denote uni-directional attention.
processed/ACL/2020/figures_clean/images/0009_2020.acl-main.9__fig0002.png
computer_science
natural_language_processing
1,208
475
2 Dialogue Generation Pre-training
[ "The probabilistic relationships among these el- ements are elaborated with the graphical model in Figure 1. Given a context c, there are multi- ple latent speech acts which can be taken as re- sponse intents (represented by the latent variable z). Conditioned on the context and one selected latent speech act, the ...
3
3
[ { "source_type": "human", "source_name": "Chuanzhi Xu (Human)", "visual_clarity": 8, "structure_layout": 7, "caption_consistency": 5, "context_consistency": 7, "misleading_risk": 8, "overall_score": 7, "summary": "The figure provides a detailed architectural overview with good vi...
0010_2020.acl-main.10
2020.acl-main.10
ACL
2,020
1
Slot-consistent NLG for Task-oriented Dialogue Systems with Iterative Rectification Network
Figure 1: IRN consists of two modules: an experience replay buffer and a pointer rewriter. The experience replay buffer collects mistaken cases from NLG baseline, template and IRN itself (the red dashed arrow) whereas the pointer network outputs templates with improved performance metrics. In each epoch of rectificatio...
processed/ACL/2020/figures_clean/images/0010_2020.acl-main.10__fig0001.png
computer_science
reinforcement_learning
1,207
569
3 Architeture
[ "Figure 1 shows the architecture of Iterative Recti- fication Network. It consists of two components: a pointer rewriter to produce templates with im- proved performance metrics and an experience re- play buffer to gather and sample training data. The improvements on slot consistency are ob- tained via an iterative ...
1
3
[ { "source_type": "human", "source_name": "zihan Deng (Human)", "visual_clarity": 7, "structure_layout": 8, "caption_consistency": 8, "context_consistency": 9, "misleading_risk": 8, "overall_score": 8, "summary": "This figure provides a clear, well-structured architectural overvie...
0010_2020.acl-main.10
2020.acl-main.10
ACL
2,020
2
Slot-consistent NLG for Task-oriented Dialogue Systems with Iterative Rectification Network
Figure 2: Correcting a candidate given a reference template. d c , d l , and d π are inferred by simple rules.
processed/ACL/2020/figures_clean/images/0010_2020.acl-main.10__fig0002.png
computer_science
natural_language_processing
1,216
303
3 Architeture
[ "One key idea behind the proposed IRN model is to conduct distant supervision on the actions of tem- plate copy and generation. We diagram its motiva- tion in Figure 2. During training, only candidate y and its reference z are given. The exact actions that convert template y to z have to be inferred from the two te...
2
3
[ { "source_type": "human", "source_name": "Xiaotong Wu (Human)", "visual_clarity": 10, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 7, "misleading_risk": 9, "overall_score": 8.800000190734863, "summary": "This figure clearly illustrates the process of co...
0011_2020.acl-main.11
2020.acl-main.11
ACL
2,020
1
Span-ConveRT: Few-shot Span Extraction for Dialog with Pretrained Conversational Representations
Figure 1: Turn-based span extraction with the new RESTAURANTS-8K data set. Note how the requested slot feature is needed to differentiate time or party size in short utterances like "7". The single-turn examples are extracted from different conversations.
processed/ACL/2020/figures_clean/images/0011_2020.acl-main.11__fig0001.png
computer_science
natural_language_processing
596
375
1 Introduction
[ "time and number of guests with correct values given by the user (e.g. tomorrow, 8pm, 3 people) in or- der to proceed with a booking. A particular chal- lenge is to deploy slot-filling systems in low-data regimes (i.e., few-shot learning setups), which is needed to enable quick and wide portability of con- versation...
2
3
[ { "source_type": "human", "source_name": "Chuanzhi Xu (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 6, "misleading_risk": 9, "overall_score": 8.399999618530273, "summary": "This figure clearly and effectively demonstrates t...
0011_2020.acl-main.11
2020.acl-main.11
ACL
2,020
2
Span-ConveRT: Few-shot Span Extraction for Dialog with Pretrained Conversational Representations
Figure 2: Span-ConveRT model architecture. Contextual subword embeddings, computed by ConveRT, are augmented with token features, and fed through a CNN. The outputs of the CNN parameterise a CRF sequence model, defining a distribution over sequence tag labellings, using the *before*, *begin*, *inside*, *after* scheme. ...
processed/ACL/2020/figures_clean/images/0011_2020.acl-main.11__fig0002.png
computer_science
natural_language_processing
503
755
2 Methodology: Span-ConveRT
[ "Span ConveRT: Final Model. We now describe our model architecture, illustrated in Figure 2. Our approach builds on established sequence tagging models using Conditional Random Fields (CRFs) (Ma and Hovy, 2016; Lample et al., 2016). We propose to replace the LSTM part of the model with fixed ConveRT embeddings.4 We ...
1
3
[ { "source_type": "human", "source_name": "Yuyan Lin (Human)", "visual_clarity": 8, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.800000190734863, "summary": "This figure provides a clear, well-structured, and t...
0011_2020.acl-main.11
2020.acl-main.11
ACL
2,020
3
Span-ConveRT: Few-shot Span Extraction for Dialog with Pretrained Conversational Representations
Figure 3: Breakdown of errors made on the test set of RESTAURANTS-8K after training on the entire train set.
processed/ACL/2020/figures_clean/images/0011_2020.acl-main.11__fig0003.png
computer_science
natural_language_processing
1,161
572
5 Conclusion and Future Work
[ "When training on the full training set (Figure 3), there is little difference in error breakdown between Span-ConveRT and V-CNN-CRF. This suggests the behavior of these models is similar when trained in a high-data setting, but improvements made by Span-ConveRT are on all fronts. When trained on a 16th of the data...
1
3
[ { "source_type": "human", "source_name": "Haoran Sun (Human)", "visual_clarity": 7, "structure_layout": 9, "caption_consistency": 6, "context_consistency": 9, "misleading_risk": 8, "overall_score": 7.800000190734863, "summary": "The figure is well-structured and visually clear, a...
0011_2020.acl-main.11
2020.acl-main.11
ACL
2,020
4
Span-ConveRT: Few-shot Span Extraction for Dialog with Pretrained Conversational Representations
Figure 4: Breakdown of errors made on the test set of RESTAURANTS-8K after training on a 16th of the train set.
processed/ACL/2020/figures_clean/images/0011_2020.acl-main.11__fig0004.png
computer_science
natural_language_processing
1,177
579
5 Conclusion and Future Work
[ "When training on the full training set (Figure 3), there is little difference in error breakdown between Span-ConveRT and V-CNN-CRF. This suggests the behavior of these models is similar when trained in a high-data setting, but improvements made by Span-ConveRT are on all fronts. When trained on a 16th of the data...
1
3
[ { "source_type": "human", "source_name": "Yutong Wang (Human)", "visual_clarity": 7, "structure_layout": 9, "caption_consistency": 6, "context_consistency": 9, "misleading_risk": 8, "overall_score": 7.800000190734863, "summary": "This figure is well-structured and visually clear ...
0012_2020.acl-main.12
2020.acl-main.12
ACL
2,020
3
Zero-Shot Transfer Learning with Synthesized Data for Multi-Domain Dialogue State Tracking
Figure 3: Breakdown of accuracy by turn number and number of slots of the TRADE model on the "Restaurant" domain. "Zero-shot" results are trained by withholding in-domain data, and "Zero-shot (DM)" is our data synthesis based on the Dialogue Model. "Full dataset" refers to training with all domains.
processed/ACL/2020/figures_clean/images/0012_2020.acl-main.12__fig0003.png
computer_science
natural_language_processing
554
262
5 Experiments
[ "To analyze the errors, we break down the result according to the turn number and number of slots in the dialogues in the test set, as shown in Fig. 3. We perform this analysis using the TRADE model on the “Restaurant” domain, which is the largest domain in MultiWOZ. We observe that the base- line model achieves 10...
1
3
[ { "source_type": "human", "source_name": "Xiao Zhang (Human)", "visual_clarity": 8, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.800000190734863, "summary": "The figure is visually clear and well-structured, w...
0012_2020.acl-main.12
2020.acl-main.12
ACL
2,020
4
Zero-Shot Transfer Learning with Synthesized Data for Multi-Domain Dialogue State Tracking
Figure 4: Accuracy plots for the few-shot MultiWOZ experiments. X axis indicates the percentage of real target domain data included in training. Y axis indicates joint accuracy.
processed/ACL/2020/figures_clean/images/0012_2020.acl-main.12__fig0004.png
computer_science
natural_language_processing
1,236
222
5 Experiments
[ "Following Wu et al. (2019), we also evaluate the effect of mixing a small percentage of real train- ing data in our augmented training sets. We use a naive few-shot training strategy, where we di- rectly add a portion of the original training data in the domain of interest to the training set. Fig. 4 plots the joi...
1
3
[ { "source_type": "human", "source_name": "Xiao Zhang (Human)", "visual_clarity": 8, "structure_layout": 9, "caption_consistency": 6, "context_consistency": 8, "misleading_risk": 8, "overall_score": 7.800000190734863, "summary": "The figure is visually clear and well-structured, e...
0013_2020.acl-main.13
2020.acl-main.13
ACL
2,020
1
A Complete Shift-Reduce Chinese Discourse Parser with Robust Dynamic Oracle
Figure 1: Overview of our Chinese discourse parser.
processed/ACL/2020/figures_clean/images/0013_2020.acl-main.13__fig0001.png
computer_science
natural_language_processing
568
681
2 Methodology
[ "Figure 1 gives an overview of our parser. Five stages are performed to transform a raw document into a parse tree: EDU segmentation, tree structure construction, rhetorical relation and nuclearity clas- sification, binary tree conversion, and beam search." ]
1
3
[ { "source_type": "human", "source_name": "Xiao Zhang (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 6, "context_consistency": 8, "misleading_risk": 9, "overall_score": 8.199999809265137, "summary": "This figure provides a clear and well-structured ov...
0014_2020.acl-main.14
2020.acl-main.14
ACL
2,020
1
TransS-Driven Joint Learning Architecture for Implicit Discourse Relation Recognition
Figure 1: TransS-driven joint learning architecture of our proposed model.
processed/ACL/2020/figures_clean/images/0014_2020.acl-main.14__fig0001.png
computer_science
natural_language_processing
1,084
421
2 The Proposed Model
[ "The implicit discourse relation recognition task is usually formalized as a classification problem. In this section, we give an overview of the TransS- driven joint learning framework, which consists of four parts: embedding layer, multi-level encoder, latent geometric structure learning, and semantic feature learn...
1
3
[ { "source_type": "human", "source_name": "yucen qian", "visual_clarity": 8, "structure_layout": 9, "caption_consistency": 8, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.600000381469727, "summary": "This figure provides a clear, well-structured overview of t...
0014_2020.acl-main.14
2020.acl-main.14
ACL
2,020
2
TransS-Driven Joint Learning Architecture for Implicit Discourse Relation Recognition
Figure 2: The illustration of multi-level encoder.
processed/ACL/2020/figures_clean/images/0014_2020.acl-main.14__fig0002.png
computer_science
natural_language_processing
552
338
2 The Proposed Model
[ "To enrich the discourse argument representations, we exploit multi-level encoder shown in Figure 2 to learn the argument representations at the different levels. Particularly, the higher-level states of multi- level encoder could capture context-dependent as- pects of words while the lower-level states could model...
1
3
[ { "source_type": "human", "source_name": "Yutong Wang (Human)", "visual_clarity": 9, "structure_layout": 9, "caption_consistency": 9, "context_consistency": 9, "misleading_risk": 9, "overall_score": 9, "summary": "This figure is visually clear and well-structured, providing an ex...
0014_2020.acl-main.14
2020.acl-main.14
ACL
2,020
4
TransS-Driven Joint Learning Architecture for Implicit Discourse Relation Recognition
Figure 4: The effect of encoder layers' number.
processed/ACL/2020/figures_clean/images/0014_2020.acl-main.14__fig0004.png
computer_science
natural_language_processing
361
278
3 Experiments
[ "as comparison experiments on the 4-way classi- fication. Figure 4 shows that the F1 scores are increasing until three encoder layers. And when the size of the encoder layer is four or five, the performance of our model is decreasing obviously. With the increasing of the number of encoder lay- ers, the model could ca...
1
3
[ { "source_type": "human", "source_name": "lambert", "visual_clarity": 8, "structure_layout": 8, "caption_consistency": 7, "context_consistency": 9, "misleading_risk": 9, "overall_score": 8.199999809265137, "summary": "The figure clearly illustrates the effect of encoder layers on...
End of preview.

SciFigQual-Bench

Full-manuscript-context benchmark for multidimensional quality assessment of scientific paper figures from ACL, EMNLP, ICML, and NeurIPS (2020–2025).

Dataset summary

Item Count
Curated figures 7,609 (corpus)
Gold-labeled instances (this release) 6,308
Raw annotation records (multi-rater) 20,166
Public eval split eval1200 1,200

Each instance is a tuple (I, c, T, m): figure image, caption, citing paragraphs, and metadata. Human gold scores aggregate multiple expert ratings per figure (mean over annotators).

Files

  • figures.jsonl — 6,308 instances with caption, context, raw annotations[]
  • human_means.csv — aggregated gold means (VC, SL, CC, CTX, MR, overall)
  • images/ — 6,308 PNG crops
  • splits/eval1200.jsonl — fixed 1,200-figure public test manifest
  • corpus_stats.json — construction funnel statistics

Join keys

figure_id or (paper_id, fig_index)

Five-dimensional rubric (1–10)

Visual Clarity (VC), Structure & Layout (SL), Caption Consistency (CC), Context Consistency (CTX), Misleading Risk (MR).

Model predictions

See companion release 3_supporting_eval_results/ (29 protocol–backend runs on eval1200).

Citation

@inproceedings{scifigqual2027,
  title={SciFigQual-Bench: A Benchmark for Scientific Figure Quality Assessment with Full-Manuscript Context},
  year={2027}
}
Downloads last month
52