Elamur Research⌕
☰
home / papers / 2023 / Goal Driven Discovery of Distributional Differences via Language Descriptions

Goal Driven Discovery of Distributional Differences via Language Descriptions

Paper · 2023 · Ruiqi Zhong, Peter Zhang, Steve Li et al.

Summary

Comparing two large text collections by hand is slow and ad hoc. The authors frame this as an ML task called D5. A system receives a user's exploration goal and a pair of corpora. It must return natural-language predicates that describe how the corpora differ in ways that matter for that goal. For example, reports from patients on one drug might more often mention paranoia. Outputs are judged on two criteria: validity (the difference really holds) and relevance to the goal.

Open-ended discovery has no reference answers, so the authors build two evaluation resources. SynD5 contains synthetic corpus pairs with a known goal-relevant difference plus a more salient distractor, and it is scored automatically by an LM judge. OpenD5 collects 675 real problems from many fields, and its outputs are partly assessed by humans. A goal-conditioned version of an earlier system (Zhong et al., 2022) outperforms the goal-agnostic one on both benchmarks. It also surfaces findings the authors did not know before.

This work makes exploratory text analysis measurable and learnable. However, the discoveries reflect correlation, not causation, and could reinforce biases if misused.

Key points

  • OpenD5 contains 675 problems drawn from domains including health, machine learning, the humanities, the social sciences and business.
    • It totals about 4.4M text samples, with roughly 17K per corpus pair on average.
    • Each pair is split into an exploration half and a validation half.
  • SynD5 has 300 synthetic problems.
    • Corpora vary along two of three dimensions: topic, genre and language. One dimension is goal-relevant and the other is a distractor.
    • The share of samples showing the goal-relevant difference ranges from 0.6 to 1.
    • Claude-v1.3, prompted with few-shot examples, checks whether an output matches the reference.
  • Effect of conditioning on the goal:
    • 12% more correct answers on SynD5.
    • Relevant candidate discoveries 31% more often on OpenD5.
  • Example findings include insights from commercial reviews, demographic and temporal differences in discussion topics, political stances in speeches, lyric styles, and NLP model error patterns.
  • Self-supervised training: a proof-of-concept trains LMs on unlabeled problems to propose more valid hypotheses.
  • Released resources: code is at github.com/ruiqi-zhong/D5, and the OpenD5 download scripts are at github.com/petezh/OpenD5.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of California, BerkeleyBerkeley · United States209 works · 244,782 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSocial Sciences · 1,954 worksGeneral Social Sciences · 254 works

Cite

Cite as Ruiqi Zhong, Peter Zhang, Steve Li, Jinwoo Ahn, Dan Klein and Jacob Steinhardt. 2023. Goal Driven Discovery of Distributional Differences via Language Descriptions. arXiv:2302.14233.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2302.14233 · arXiv 2302.14233

@misc{zhong2023goal,
  title = {{Goal Driven Discovery of Distributional Differences via Language Descriptions}},
  author = {Ruiqi Zhong, Peter Zhang, Steve Li, Jinwoo Ahn, Dan Klein and Jacob Steinhardt},
  year = {2023},
  doi = {10.48550/arxiv.2302.14233},
  eprint = {2302.14233},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2302.14233}
}

References

91 references, 22 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. The hewlett foundation: Automated essay scoring, 2012. competitions/asap-aes. find
  2. URL https://kaggle.com/The hewlett foundation: Short answer scoring, 2013. URL https://kaggle.com/competitions/ asap-sas. ↗
  3. Ad observer. https://adobserver.org/, 2021. Accessed: 2022-12-30. ↗
  4. Armen Aghajanyan, Lili Yu, Alexis Conneau, Wei-Ning Hsu, Karen Hambardzumyan, Susan Zhang, Stephen Roller, Naman Goyal, Omer Levy, and Luke Zettlemoyer. Scaling laws for generative mixed-modal language models. arXiv preprint arXiv:2301.03728, 2023.
  5. Roee Aharoni and Yoav Goldberg. Unsupervised domain clusters in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 7747–7763, Online, July 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020. acl-main.692. UR…
  6. Mohammad Alali, Shaayan Syed, Mohammed Alsayed, Smit Patel, and Hemanth Bodala. Justice: A benchmark dataset for supreme court’s judgment prediction. arXiv preprint arXiv:2112.03414, 2021. find
  7. Akari Asai, Sara Evensen, Behzad Golshan, Alon Halevy, Vivian Li, Andrei Lopatenko, Daniela Stepanov, Yoshihiko Suhara, Wang-Chiew Tan, and Yinzhan Xu. Happydb: A corpus of 100,000 crowdsourced happy moments. arXiv preprint arXiv:1801.07746, 2018. find
  8. Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022a. find
  9. Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022b.
  10. Miriam Barnum and James Lo. Is the npt unraveling? evidence from text analysis of review conference statements. Journal of Peace Research, 57(6):740–751, 2020. find
  11. Alexander Baturo, Niheer Dasandi, and Slava J Mikhaylov. Understanding state preferences with text as data: Introducing the un general debate corpus. Research & Politics, 4(2):2053168017712821, 2017. find
  12. Akshay Bhalotia. Yc company scraper. https://github.com/akshaybhalotia/yc_company_ scraper, 2022. ↗
  13. Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit. " O’Reilly Media, Inc.", 2009. find
  14. David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993–1022, 2003.
  15. Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. The snli corpus. 2015. find
  16. Divy Bramhecha. Poetry Foundation Poems, 2019. URL https://www.kaggle.com/datasets/ tgdivy/poetry-foundation-poems. ↗
  17. Dallas Card, Serina Chang, Chris Becker, Julia Mendelsohn, Rob Voigt, Leah Boustan, Ran Abramitzky, and Dan Jurafsky. Replication code and data for “Computational analysis of 140 years of US political speeches reveals more positive but increasingly polarized framing of immigration” [dataset]. https:… find
  18. Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. Neural legal judgment prediction in english. arXiv preprint arXiv:1906.02059, 2019.
  19. Sihao Chen, Daniel Khashabi, Wenpeng Yin, Chris Callison-Burch, and Dan Roth. Seeing Things from a Different Angle: Discovering Diverse Perspectives about Claims. In Proc. of the Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2019. URL http:…
  20. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. find
  21. Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019.
  22. Norman R Draper and Harry Smith. Applied regression analysis, volume 326. John Wiley & Sons, 1998. find
  23. Sabri Eyuboglu, Maya Varma, Khaled Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Ré. Domino: Discovering systematic errors with cross-modal embeddings. arXiv preprint arXiv:2203.14960, 2022. — [arXiv:2203.14960](https://arxiv.org/abs/2203.14960) find
  24. Yujia Gao, Jinu Jang, and Diyi Yang. Understanding the usage of online media for parenting from infancy to preschool at scale. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pp. 1–12, 2021. find
  25. Anthony G Greenwald and Mahzarin R Banaji. Implicit social cognition: attitudes, self-esteem, and stereotypes. Psychological review, 102(1):4, 1995. find
  26. Ivan Habernal and Iryna Gurevych. Which argument is more convincing? Analyzing and predicting convincingness of Web arguments using bidirectional LSTM. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1589–1599, Berlin, Germany,…
  27. Kevin Hartman. Advertisement Transcripts from Various Industries, 2019. URL https://tinyurl.com/5w36dwdx. ↗
  28. He He, Derek Chen, Anusha Balakrishnan, and Percy Liang. Decoupling strategy and generation in negotiation dialogues, 2018.
  29. Jibo He. Big Data Set from RateMyProfessor.com for Professors’ Teaching Evaluation, 2020. URL https://data.mendeley.com/datasets/fvtfjyvw7d/2. ↗
  30. Samuel He. Goodbye world: using natural language processing to identify suicidal posts, 2021. URL https://github.com/hesamuel/goodbye_world. ↗
  31. James J Heckman and Burton Singer. Abducting economics. American Economic Review, 107(5): 298–302, 2017. find
  32. Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. find
  33. Or Honovich, Uri Shaham, Samuel R Bowman, and Omer Levy. Instruction induction: From few examples to natural language task descriptions. arXiv preprint arXiv:2205.10782, 2022.
  34. Nabil Hossain, John Krumm, and Michael Gamon. " president vows to cut\< taxes\> hair": Dataset and analysis of creative text editing for humorous headlines. arXiv preprint arXiv:1906.00274, 2019. — [arXiv:1906.00274](https://arxiv.org/abs/1906.00274) find
  35. Kaggle. TMDB 5000 Movie Dataset, 2018. URL https://www.kaggle.com/datasets/tmdb/ tmdb-movie-metadata. ↗
  36. Rohit Kulkarni. A Million News Headlines, 2018. URL https://doi.org/10.7910/DVN/ SYBGZL. — [doi:10.7910/dvn/](https://doi.org/10.7910/dvn/) find
  37. Rohit Kulkarni. The Examiner - Spam Clickbait Catalog, 2020a. URL https://www.kaggle.com/ datasets/therohk/examine-the-examiner. ↗
  38. Rohit Kulkarni. Urban Dictionary Words And Definitions, 2020b. URL https://www.kaggle.com/datasets/therohk/urban-dictionary-words-dataset. ↗
  39. Rohit Kulkarni. India News Headlines Dataset, 2022. URL https://www.kaggle.com/datasets/ therohk/india-headlines-news-dataset. ↗
  40. Benjamin Letham, Cynthia Rudin, Tyler H. McCormick, and David Madigan. Interpretable classifiers using rules and bayesian analysis: Building a better stroke prediction model. The Annals of Applied Statistics, 9(3), Sep 2015. ISSN 1932-6157. doi: 10.1214/15-aoas848. URL http://dx.doi.org/10.1214/15-… ↗
  41. Derek Lim and Austin R Benson. Expertise and dynamics within crowdsourced musical knowledge curation: A case study of the genius platform. In ICWSM, pp. 373–384, 2021. find
  42. Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. Wanli: Worker and ai collaboration for natural language inference dataset creation, January 2022. URL https://arxiv.org/pdf/ 2201.05955. ↗
  43. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.
  44. Zhi Liu. Reuter_50_50 Data Set, 2011. URL https://archive.ics.uci.edu/ml/datasets/ Reuter_50_50. ↗
  45. Jens Ludwig and Sendhil Mullainathan. Algorithmic behavioral science: Machine learning as a tool for scientific discovery. Chicago Booth Research Paper, (22-15), 2022. find
  46. Christopher Manning and Hinrich Schutze. Foundations of statistical natural language processing. MIT press, 1999. find
  47. Ken McGarry. A survey of interestingness measures for knowledge discovery. The knowledge engineering review, 20(1):39–61, 2005. find
  48. Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models, 2016.
  49. Natan Mish. Federal Reserve Governors Speeches 1996 - 2020, 2020. URL https://tinyurl.com/3j2e79a6. ↗
  50. Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. In ACL, 2022. find
  51. Rishabh Misra and Prahal Arora. Sarcasm detection using hybrid neural network. arXiv preprint arXiv:1908.07414, 2019. — [arXiv:1908.07414](https://arxiv.org/abs/1908.07414) find
  52. Rishabh Misra and Jigyasa Grover. Sculpting Data for ML: The first act of Machine Learning. 01 2021. ISBN 9798585463570. find
  53. Nuno Moniz and Luâ C™is Torgo. Multi-source social feedback of online news feeds. CoRR, [Web Link], 2018. find
  54. Mickaël Mouillé. Kickstarter Projects, 2017. URL https://www.kaggle.com/datasets/ kemical/kickstarter-projects?select=ks-projects-201612.csv. ↗
  55. Dong Nguyen, Maria Liakata, Simon DeDeo, Jacob Eisenstein, David Mimno, Rebekah Tromble, and Jane Winters. How we do things with words: Analyzing text as social and cultural data. Frontiers in Artificial Intelligence, 3:62, 2020. find
  56. Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-…
  57. Elle O’Brien. iterative/aita_dataset: Praw rescrape of entire dataset, February 2020. URL https://doi.org/10.5281/zenodo.3677563. — [doi:10.5281/zenodo.3677563](https://doi.org/10.5281/zenodo.3677563) ↗
  58. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022.
  59. Verónica Pérez-Rosas and Rada Mihalcea. Experiments in open domain deception detection. In Proceedings of the 2015 conference on empirical methods in natural language processing, pp. 1120–1125, 2015. find
  60. Verónica Pérez-Rosas, Mohamed Abouelenien, Rada Mihalcea, and Mihai Burzo. Deception detection using real-life trial data. In Proceedings of the 2015 ACM on International Conference on Multimodal Interaction, pp. 59–66, 2015. find
  61. Verónica Pérez-Rosas, Bennett Kleinberg, Alexandra Lefevre, and Rada Mihalcea. Automatic detection of fake news. arXiv preprint arXiv:1708.07104, 2017.
  62. Denis Peskov, Benny Cheng, Ahmed Elgohary, Joe Barrow, Cristian Danescu-Niculescu-Mizil, and Jordan Boyd-Graber. It takes two to lie: One to lie and one to listen. In Association for Computational Linguistics, 2020. find
  63. John P. Pestian, Chris Brew, Pawel Matykiewicz, DJ Hovermale, Neil Johnson, K. Bretonnel Cohen, and Wlodzislaw Duch. A shared task involving multi-label classification of clinical free text. In Biological, translational, and clinical language processing, pp. 97–104, Prague, Czech Republic, June 2007… find
  64. Michael Polanyi, John Ziman, and Steve Fuller. The republic of science: its political and economic theory minerva, i (1)(1962), 54-73. Minerva, 38(1):1–32, 2000. find
  65. Ilan Price, Jordan Gifford-Moore, Jory Fleming, Saul Musker, Maayan Roichman, Guillaume Sylvain, Nithum Thain, Lucas Dixon, and Jeffrey Sorensen. Six attributes of unhealthy conversation. arXiv preprint arXiv:2010.07410, 2020. — [arXiv:2010.07410](https://arxiv.org/abs/2010.07410) find
  66. Demand Progress. Statements of Administration Policy, 2022. URL https://github.com/unitedstates/statements-of-administration-policy\# statements-of-administration-policy. ↗
  67. PromptCloud. U.S. Technology Jobs on Dice.com, 2017. URL https://www.kaggle.com/ datasets/PromptCloudHQ/us-technology-jobs-on-dicecom. ↗
  68. WVO Quine. Naturalistic epistemology. Ontological relativity and other essays, pp. 69–90, 1969. find
  69. Quora. Quora Question quora-question-pairs. Pairs, 2017. find
  70. URL — Know What You Don't Know: Unanswerable Questions for SQuAD https://www.kaggle.com/c/Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don’t know: Unanswerable questions for squad. arXiv preprint arXiv:1806.03822, 2018. ↗
  71. Justin Robischon. Wikipedia Movie Plots, 2019. URL https://www.kaggle.com/datasets/ jrobischon/wikipedia-movie-plots. ↗
  72. Allen Roush and Arvind Balaji. DebateSum: A large-scale argument mining and summarization dataset. In Proceedings of the 7th Workshop on Argument Mining, pp. 1–7, Online, December 2020. Association for Computational Linguistics. URL https://aclanthology.org/2020. argmining-1.1. ↗
  73. Dudley Shapere. The structure of scientific revolutions. The Philosophical Review, 73(3):383–394, 1964. find
  74. Chandan Singh, John X Morris, Jyoti Aneja, Alexander M Rush, and Jianfeng Gao. Explaining patterns in data with language models via interpretable autoprompting. arXiv preprint arXiv:2210.01848, 2022. — [arXiv:2210.01848](https://arxiv.org/abs/2210.01848) find
  75. Richard Socher, Danqi Chen, Christopher D Manning, and Andrew Ng. Reasoning with neural tensor networks for knowledge base completion. Advances in neural information processing systems, 26, 2013.
  76. Alane Suhr, Clara Vania, Nikita Nangia, Maarten Sap, Mark Yatskar, Samuel Bowman, and Yoav Artzi. Crowdsourcing beyond annotation: Case studies in benchmark data collection. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: Tutorial Abstracts, pp. 1–6, 2021. find
  77. J Sun. Daily News for Stock Market Prediction, 2017. datasets/aaron7sun/stocknews. find
  78. URL https://www.kaggle.com/Jesse Thomason, Jivko Sinapov, Maxwell Svetlik, Peter Stone, and Raymond J Mooney. Learning multi-modal grounded linguistic semantics by playing" i spy". In IJCAI, pp. 3477–3483, 2016. ↗
  79. Andrew Thompson. All the News 1.0, 2019. URL https://components.one/datasets/ all-the-news-articles-dataset. ↗
  80. Elsbeth Turcan and Kathleen McKeown. Dreaddit: A reddit dataset for stress analysis in social media. arXiv preprint arXiv:1911.00133, 2019. find
  81. Udacity. Armenian Online Job Postings, 2017. URL https://www.kaggle.com/datasets/ udacity/armenian-online-job-postings. ↗
  82. Veniamin Veselovsky, Manoel Horta Ribeiro, and Robert West. Artificial artificial artificial intelligence: Crowd workers widely use large language models for text production tasks. arXiv preprint arXiv:2306.07899, 2023.
  83. Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. URL https://arxiv. org/abs/2204.… find
  84. Zihan Wang, Jingbo Shang, and Ruiqi Zhong. Goal-driven explainable clustering via language descriptions. arXiv preprint arXiv:2305.13749, 2023. find
  85. Orion Weller and Kevin Seppi. The rJokes dataset: a large scale humor collection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pp. 6136–6141, Marseille, France, May 2020. European Language Resources Association. ISBN 979-10-95546-34-4. URL https://aclanthology.org/202… ↗
  86. Adina Williams, Nikita Nangia, and Samuel R Bowman. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426, 2017.
  87. Zonglin Yang, Li Dong, Xinya Du, Hao Cheng, Erik Cambria, Xiaodong Liu, Jianfeng Gao, and Furu Wei. Language models as inductive reasoners. arXiv preprint arXiv:2212.10923, 2022.
  88. Seonghyeon Ye, Doyoung Kim, Joel Jang, Joongbo Shin, and Minjoon Seo. Guess the instruction! making language models stronger zero-shot learners. arXiv preprint arXiv:2210.02969, 2022. — [arXiv:2210.02969](https://arxiv.org/abs/2210.02969) find
  89. Ruiqi Zhong, Charlie Snell, Dan Klein, and Jacob Steinhardt. Describing differences between text distributions with natural language. In International Conference on Machine Learning, pp. 27099–27116. PMLR, 2022. find
  90. Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910, 2022.
  91. Zhiying Zhu, Weixin Liang, and James Zou. Gsclip: A framework for explaining distribution shifts in natural language. arXiv preprint arXiv:2206.15007, 2022. — [arXiv:2206.15007](https://arxiv.org/abs/2206.15007) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhong-2023-goal-driven-discovery-distributional-differences/