| Zaidan et al. |
2007 |
Using “Annotator Rationales” to Improve Machine Learning for Text Categorization |
Sentiment classification |
Reviews (IMDB) |
Authors |
Snippets |
E |
1800 with rationales + 200 without |
link |
To justify why a review is positive, highlight the most important words and phrases that would tell someone to see the movie |
ML learning |
|
| Yano et al. |
2010 |
Shedding (a Thousand Points of) Light on Biased Language |
Bias classification |
American political blog posts |
Crowdsourcing (MTurk) |
Snippets |
E |
1k sentences |
|
workers are asked to check the box to indicate the region which “give away” the bias |
Task insight |
5? |
| Abedin et al. |
2011 |
Learning Cause Identifiers from Annotator Rationales |
Identify cause of aviation incident |
Aviation safety reports |
Author and student worker |
Snippets |
E |
1233 docs with rationales + 1000 unlabeled |
|
the annotators are asked to “do their best to mark enough rationales to pro- vide convincing support for the class of interest”, but are not expected to “go out of their way to mark everything”. |
Reduce required data size, ML learning |
2 |
| MCDonnell et al. |
2016 |
Why is that relevant? collecting annotator rationales for relevance judgments |
Webpage relevance |
Webpages |
Crowdsourcing (MTurk) |
Sentences |
E + A |
10000+ |
link |
|
Improve data quality, transparency, verification |
|
| Chhatwall et al. |
2018 |
Explainable text classification in legal document review a case study of explainable predictive coding |
Determine responsiveness of legal documents |
Legal documents |
Legal domain experts |
Snippets |
E + A |
688,294 documents including email, Microsoft Office documents, PDFs, and other text-based documents. Only rationales for responsive documents |
- |
Justification of coding a document as responsive |
ML learning |
|
| Carton et al. |
2018 |
Extractive Adversarial Networks: High-Recall Explanations for Identifying Personal Attacks in Social Media Posts |
Personal attacks in comments |
Wikipedia revision comments |
Students |
Snippets |
E |
1089 |
|
to high- light sections of comments that they considered to constitute personal attacks |
Explanation verification |
40 |
| Bao et al. |
2018 |
Deriving Machine Attention from Human Rationales |
Sentiment classification |
Reviews (BeerAdvocate, Tripadvisor) |
Students |
Snippets |
E |
200 |
link |
to highlight rationales that are short and coherent, yet sufficient for supporting the label |
ML learning |
5 |
| Ramirez et al. |
2019 |
Understanding the impact of text highlighting in crowdsourcing tasks |
Topic classification |
Reviews (Amazon), Systematic Literature Review (SLR) |
Crowdsourcing (Figure Eight) |
Snippets + Sentences |
E + A |
400, 135 + 150 |
link |
Explain you decision. Tell us why you think the paper is relevant / If you were to select one or more sentences most useful for your decisions, which ones would you select? |
Enrich data / improve classification by humans |
449, 424 + 464 = 1337 |
| Wang et al. |
2019 |
Learning from Explanations with Neural Execution Tree |
Relation extraction + Sentiment classification |
Reviews (restaurant, laptop) |
Crowdsourcing (MTurk) |
Sentences |
A |
170 (TACRED), 203 (SemEval) 40 (Laptop), 45 (Restaurant) |
|
|
Augment labels |
|
| Sharma et al |
2021 |
A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support |
Empathy expression |
TalkLife, Reddit |
Crowdsourcing (Upwork) |
Snippets |
E |
approx. 10k |
link |
Along with the categorical annotations, crowdworkers were also asked to highlight portions of the response post that formed the rationale behind their annotation. |
Gold rationales |
8 |
| Sen et al. |
2020 |
Human Attention Maps for Text Classification: Do Humans and Neural Networks Focus on the Same Words? |
Sentiment classification |
Yelp |
Crowdsourcing (MTurk) |
Snippets |
E |
5000 reviews, each review 3, annotators |
link |
Participants are asked to complete two tasks: 1) Identify the sentiment of the review as positive, negative, or neither, and 2) Highlight (ALL) the words that are indicative of the chosen sentiment. |
Gold rationales |
3 |
| Kanchinadam. et al. |
2020 |
Rationale-based Human-in-the-Loop via Supervised Attention |
Sentiment classification |
Reviews (IMDB) |
Crowdsourcing (MTurk) |
Snippets |
E |
22k |
link |
|
ML learning |
|
| Kutlu et al. |
2020 |
Annotator rationales for labeling tasks in crowdsourcing |
Webpage relevance |
Webpages |
Crowdsourcing (MTurk) |
Sentences |
E + A |
10000+ |
|
Please copy and paste text 2-3 sentences from the web page which you believe support your decision / Please describe in words why you agree or disagree with Tom’s decision. |
Improve data quality, transparency, and verification |
|
| Socher et al. |
2013 |
Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank |
Sentiment classification |
Sentences from reviews (IMDB) |
Crowdsourcing (MTurk) |
Snippets + Sentences |
E |
215154 phrases |
|
- |
ML learning |
|
| Sap et al. |
2020 |
Social Bias Frames: Reasoning about Social and Power Implications of Language |
Implications in text classification |
Twitter, Reddit, Gap, Stormfront |
Crowdsourcing (MTurk) |
Sentences |
A |
34k implications |
link |
What aspect…/.. of this group is reference or implied in this post? Use simple phrases |
Gold rationales, ML learning |
|
| Vidgen et al. |
2021 |
Introducing CAD: the Contextual Abuse Dataset |
Abusive content |
Reddit |
experts/trained annotators |
Snippets |
E |
approx. 25k |
link |
For each entry they highlighted the part of the text which contains the abuse |
Gold rationales |
2 |
| Mohseni et al. |
2021 |
Quantitative Evaluation of Machine Learning Explanations: A Human-Grounded Benchmark |
Sentiment and topic classification |
20news, Reviews (IMDB) |
Crowdsourcing (MTurk) |
Snippets |
E |
100IMDB, 100 20news |
link |
“select words and phrases which explain the positive (or negative) sentiment of the movie review” -> label is given |
Gold rationales |
200 |
| Arous et al. |
2021 |
Marta: Leveraging human rationales for explainable text classification |
Topic relevance classification |
Wikipedia articles |
Crowdsourcing (MTurk) |
Snippets |
E |
1413 |
link |
Then, workers were asked to annotate the article and provide a snippet from the text as a justification. |
ML learning, explainability |
58 |
| Chalkidis et al. |
2021 |
Paragraph-level Rationale Extraction through Regularization: A case study on European Court of Human Rights Cases |
alleged violation prediction |
European Court of Human Rights cases |
Legal domain experts |
Paragraphs |
E |
11k silver, 50 gold rationales |
link |
to identify the relevant facts (paragraphs) |
Gold rationales |
1 for gold rats |
| Hayati et al. |
2021 |
Does BERT Learn as Humans Perceive? Understanding Linguistic Styles through Lexica |
Style classification |
StanfordPoliteness, Standford Sentiment Treebank, tweet dataset for offensiveness, dataset for emotion classification |
Crowdsourcing (Prolific2) |
Words |
E |
500 texts |
link |
Each worker was asked what styles they perceive each of the texts to exhibit. If they think the text has certain styles, workers then highlight the words in the text which they believe make them think the text has those styles |
Feature importance verification |
622 |
| Mathew et al. |
2021 |
HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection |
Hate speech detection |
Twitter and Gap |
Crowdsourcing (MTurk) |
Snippets |
E |
9,055 Twitter + 11,093 Gab |
link |
if the text is considered as hate speech, or offensive by majority of the annotators, we further ask the annotators to annotate parts of the text, which are words or phrases that could be a potential reason for the given annotation. |
Explanability and reduce bias for ML models |
253 |
| Malik et al. |
2021 |
ILDC for CJPE: Indian Legal Documents Corpus for Court Judgment Prediction and Explanation |
Predict outcome of legal cases |
Indian Supreme Court cases |
Legal domain experts |
Sentences |
E |
56 |
link |
1. predict the judgement, and 2. mark the sentences that they think are explanation for the judgement |
Gold rationales |
5 |
| Wang et al. |
2022 |
Ranking-Constrained Learning with Rationales for Text Classification |
Topic classification AIvsCR |
Arxiv (scientific articles) |
Authors |
Snippets |
E |
394 annotated with rationales + 2k docs |
|
To justify why .. is .., highlight the most important words and phrases, but not all |
Enrich data with rationales |
2 |
| Guzman et al. |
2022 |
RaFoLa: A Rationale-Annotated Corpus for Detecting Indicators of Forced Labour |
Forced Labour Indicators |
news articles |
two annotators (30+ with masters degree) |
Snippets |
E |
989 |
link |
…we are asking you to identify the risks of forced labour in news articles…. tag what phrases or sentences led you to decide the presence of that indicator … and highlighting the phrases/sentences that support your decision. |
Enrich data with rationales |
2 |
| Jayaram et al. |
2021 |
Human Rationales as Attribution Priors for Explainable Stance Detection |
Stance detection (pro/con) |
VAST: comments for The New York Times |
Crowdsourcing (MTurk) |
Words |
E |
775 |
link |
workers are asked to (1) classify the stance of an argument with respect to a topic and (2) select the k most important words in the argument (for each example, we provide an acceptable range of values for k). A word is considered to be important if masking it would make (1) more difficult. |
Improve ML model in data-scarce setting |
3 |
| Meldo et al. |
2020 |
The natural language explanation algorithms for the lung cancer computer-aided diagnosis system |
Lung cancer image classification |
LUNA16 (lung photos) |
Medical domain experts (doctors) |
Sentences |
A |
240 |
- |
- |
Gold rationales |
|
| Zini et al. |
2022 |
On the Evaluation of the Plausibility and Faithfulness of Sentiment Analysis Explanations |
Sentiment classification |
Reviews (Rotten Tomatoes) |
Data scientists |
Words |
E |
1973 rationale sentences |
link |
|
Gold rationales |
10 |
| Srivastava et al. |
2020 |
Robustness to Spurious Correlations via Human Annotations |
Medical diagnosis, handwriting, police domain |
Multiple |
Crowdsourcing (MTurk) |
Sentences |
A |
3 datasets |
link |
What transformation do you think happened to the image?, What factors do you think led to the individual being stopped and [arrested/not arrested]? |
reduce spurious correlations |
3 per annotation |
| Lu et al. |
2022 |
A Rationale-Centric Framework for Human-in-the-loop Machine Learning |
Sentiment classification |
Reviews (IMDB) |
Crowdsourcing |
Snippets |
E |
5073 rationales in 855 movie reviews |
link |
|
eliminating the effect of spurious patterns by leveraging human knowledge |
|
| Beckh et al. |
2024 |
Limitations of Feature Attribution in Long Text Classification of Standard] |
Assessing the AI readiness of standards and specifications |
Technical documents |
Human experts |
Snippets |
E |
1000 documents |
|
Annotators were instructed to find and annotate evidence that is enough to justify a label. |
Task insight, gold rationales |
1 |