Elamur Research⌕
☰
home / papers / 2014 / “I Object!” Modeling Latent Pragmatic Effects in Courtroom Dialogues

“I Object!” Modeling Latent Pragmatic Effects in Courtroom Dialogues

Paper · 2014 · Dan Goldwasser and Hal Daumé III

Summary

The paper asks whether a system can learn which turns in a courtroom exchange provoke an objection, and why. Labeling the pragmatic rule behind each objection by hand would be costly and subjective. The authors instead take supervision directly from the transcripts. They search the O.J. Simpson murder trial record for judge rulings ("sustain"/"overrule"), which give a binary label, an objection type and the ruling. The few turns before each ruling form the input snippet.

A discriminative latent-variable model, with inference cast as an ILP, makes two kinds of hidden decisions:

  • which individual sentences matter for the decision;
  • which sentence pairs across turns are linked, such as a question and its answer.

These decisions are learned jointly with the objection classifier. Lexical features are enriched with lexicons for subjectivity, hedging and bias, plus named entities. Every feature is conditioned on the speaker's identity, role and side in the trial.

The learned dialogue structure clearly improves classification, and speaker information makes that structure more sensible. The broader point is that an outcome speech act can serve as cheap distant supervision for discourse structure.

Key points

  • Data and context window: objections are mined automatically from the O.J. Simpson trial transcripts by matching sustain/overrule lemmas in the judge's turns. The previous n=6 turns are treated as possible causes. Dialogues average 8.6 sentences.
  • Linguistic resources:
    • Stanford NER (Finkel et al., 2005)
    • the subjectivity/sentiment lexicon of Riloff and Wiebe (2003)
    • Hyland's (2005) hedges and boosters
    • the Wikipedia-derived bias lexicon of Recasens et al. (2013)
    • Patient Polarity Verbs (Goyal et al., 2010)
  • Sentence labels: each turn is split into sentences tagged FORMALITY, QUESTION, RESPONSE (POSITIVE/NEGATIVE) or STATEMENT.
  • Model: it follows the latent structured framework of Chang et al. (2010), with ILP inference (Roth and Yih, 2007). It has sentence-relevance variables and cross-turn edge variables.
  • Findings:
    • Most real objections are overruled, typically raised to interrupt the other side on sensitive topics.
    • Sustained objections are easier to detect than overruled ones.
    • The speaker-situated representation (x^Sit) usually improves predictions and changes which latent variables switch on.

Interesting ideas

  • I1. An action recorded in the text (a judge's ruling) can act as free binary supervision for hidden discourse structure, so no relations need to be annotated by hand.
  • I2. The model treats "which sentences matter" and "which cross-turn pairs are connected" as explicit latent decisions solved under ILP constraints, so its output shows exactly which evidence drove the prediction.
  • I3. Lexical cues are conditioned on who speaks and on their role and side, because the same words carry different pragmatic weight from a prosecutor than from a witness.
  • I4. Hedge/booster and bias lexicons separate evasive or overconfident answers from neutral ones, and loaded word choice ("death" vs. "murder") reveals framing.
  • I5. The evaluation splits data by outcome (sustained vs. overruled) and shows that justified cases are easier to detect, so models are judged on clean versus noisy positives.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Mova could mine naturally occurring outcome markers in texts (refusals, corrections, "that's not what I asked") as weak labels for training its speech-act and relation layers.
  • I2 (frames): crate prag could expose an explicit relevance mask plus question→answer links as part of its explainable output, rather than scoring the whole dialogue at once.
  • I3 (frames): implicature and irony detection in prag could key its cue features on the speaker entity and that entity's role in the world of the text.
  • I4 (world): marking each stated fact in the world model with the speaker's certainty and framing would let the reader separate asserted facts from slanted or tentative claims.
  • I5 (learn): evaluations could likewise report results separately on well-grounded and noisy silver examples, which acts as a sanity control on what the model really captures.

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 Maryland, College ParkCollege Park · United States83 works · 23,784 citations

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 works

🧭 Topics 1Research subjects and languages the work is about.Dialogue systems · 401 works

Cite

Cite as Dan Goldwasser and Hal Daumé III. 2014. “I Object!” Modeling Latent Pragmatic Effects in Courtroom Dialogues. Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.3115/v1/e14-1069

@inproceedings{goldwasser2014i,
  title = {{“I Object!” Modeling Latent Pragmatic Effects in Courtroom Dialogues}},
  author = {Dan Goldwasser and Hal Daumé III},
  booktitle = {Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics},
  year = {2014},
  doi = {10.3115/v1/e14-1069},
  url = {https://doi.org/10.3115/v1/e14-1069}
}

References

26 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Adam Vogel and Christopher Potts and Dan Jurafsky. 2011. Implicatures and Nested Beliefs in Approximate Decentralized-POMDPs. In EMNLP.
  2. Yoav Artzi and Luke S. Zettlemoyer. 2011. Bootstrapping semantic parsers from conversations. In EMNLP. find
  3. Jason Baldridge and Alex Lascarides. 2005. Probabilistic head-driven parsing for discourse structure. In CoNLL. find
  4. B. E. Boser, I. M. Guyon, and V. N. Vapnik. 1992. A training algorithm for optimal margin classifiers. In Proc. 5th Annu. Workshop on Comput. Learning Theory, pages 144–152. find
  5. Nathanael Chambers and Dan Jurafsky. 2008. Unsupervised learning of narrative event chains. In Proceedings of ACL-08: HLT, June. find
  6. Nathanael Chambers and Dan Jurafsky. 2009. Unsupervised learning of narrative schemas and their participants. In ACL/IJCNLP, pages 602–610. find
  7. Ming-Wei Chang, Dan Goldwasser, Dan Roth, and Vivek Srikumar. 2010. Discriminative learning over constrained latent representations. In NAACL. find
  8. Pedro F. Felzenszwalb, Ross B. Girshick, David A. McAllester, and Deva Ramanan. 2010. Object detection with discriminatively trained part-based models. IEEE Trans. Pattern Anal. Mach. Intell. ↗
  9. Jenny Rose Finkel, Trond Grenager, and Christopher Manning. 2005. Incorporating non-local information into information extraction systems by gibbs sampling. In ACL.
  10. Dave Golland, Percy Liang, and Dan Klein. 2010. A game-theoretic approach to generating spatial descriptions. In EMNLP. find
  11. Amit Goyal, Ellen Riloff, and Hal Daumé III. 2010. Automatically producing plot unit representations for narrative text. In Empirical Methods in Natural Language Processing (EMNLP). find
  12. K. Hyland. 2005. Metadiscourse: Exploring interaction in writing. In Continuum, London and New York. find
  13. W. G. Lehnert. 1981. Plot units and narrative summarization. In Cognitive Science. ↗
  14. Daniel Marcu. 1997. The rhetorical parsing of natural language texts. In ACL. find
  15. R. Prasad, N. Dinesh, A. Lee, E. Miltsakaki, L Robaldo, A. Joshi, and B. Webber. 2008. The penn discourse treebank 2.0. In LREC.
  16. Ariadna Quattoni, Sybor Wang, L-P Morency, Michael Collins, and Trevor Darrell. 2007. Hidden conditional random fields. Pattern Analysis and Machine Intelligence, IEEE Transactions on. ↗
  17. Marta Recasens, Cristian Danescu-Niculescu-Mizil, and Dan Jurafsky. 2013. Linguistic models for analyzing and detecting biased language. In Proceedings of ACL.
  18. E. Riloff and J. Wiebe. 2003. Learning extraction patterns for subjective expressions. In NAACL.
  19. D. Roth and W. Yih. 2007. Global inference for entity and relation identification via a linear programming formulation. In Lise Getoor and Ben Taskar, editors, Introduction to Statistical Relational Learning. MIT Press. ↗
  20. Roger C. Schank and Robert P. Abelson. 1977. Scripts, plans, goals and understanding. In ACL/IJCNLP. find
  21. Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteer. 2000. Dialogue act modeling for automatic tagging and recognition of conversational speech. COMPUTATIONAL LINGUISTICS, 26:339–373. ↗
  22. Oscar Täckström and Ryan T. McDonald. 2011. Discovering fine-grained sentiment with latent variable structured prediction models. In ECIR. find
  23. Rakshit Trivedi and Jacob Eisenstein. 2013. Discourse connectors for latent subjectivity in sentiment analysis. classification. In NAACL. find
  24. Mengqiu Wang and Christopher D. Manning. 2010. Probabilistic tree-edit models with structured latent variables for textual entailment and question answering. In Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010). find
  25. Ainur Yessenalina, Yisong Yue, and Claire Cardie. 2010. Multi-level structured models for document-level sentiment classification. In EMNLP. find
  26. C. Yu and T. Joachims. 2009. Learning structural svms with latent variables. In Proc. of the International Conference on Machine Learning (ICML). find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/goldwasser-2014-i-object-modeling-latent-pragmatic/