Elamur Research
☰
home / papers / 2006 / Computing relative polarity for textual inference

Computing relative polarity for textual inference

What it is about

The paper handles what a sentence commits its author to when complement clauses sit inside other clauses. Implicative verbs (manage to, forget to, refuse to, force to, attempt to) entail or deny their complement depending on the polarity of the environment, so "forget to" commits the author to the opposite of "forget that". The authors compile an implication signature for about 400 of 1,250 complement-taking verbs from the BNC, marking each with a sign for positive and negative environments. Polarity is relative to an ancestor context and is computed recursively down the chain of embedding verbs and negations, so a sentence that denies managing to remember to do something commits the author to the thing not being done. The output is a set of instantiable and uninstantiable event contexts, built inside the XLE system. Presuppositions are left out, and conversational implicatures such as "be able to" in a negative environment are flagged as cancellable rather than entailed.

Ideas

For Mova

The chapters are languages/en/27-negation.md, languages/en/23-complementation.md and languages/en/33-implicature-and-presupposition.md, plus languages/en/31-speech-acts.md for I4. The code is prag/src/infer.rs and the tables in prag/data. The paper's examples are for reading and writing our own tests; its lookup table is its own work and is not in the repo.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/W06-3907.pdf

Summary

The paper addresses a core problem in local textual inference: deciding whether a sentence commits its author to the truth or falsity of an embedded complement clause. The answer depends on the embedding verb (e.g. forget to versus forget that), the complement type, and every operator above it. The authors built a lexicon of "implication signatures". Working through complement-taking verbs ordered by frequency in the British National Corpus, several annotators judged each verb's entailments and presuppositions, checking real usage via Google and the Linguist's Search Engine. They then designed a recursive algorithm that propagates relative polarity through nested contexts. It runs inside the canonicalization stage of PARC's XLE-based pipeline, built for the AQUAINT project, as rewrite rules over packed representations. Each context collects the ancestor contexts relative to which it is positive or negative. From these sets, the system derives which event descriptions are instantiable or uninstantiable at the author's top level. The work matters because neither the signatures nor their stacked interactions can be found in WordNet, VerbNet or FrameNet.

Key points

  • The lexicon holds 1250 relevant verbs. On a first pass, 400 were classified, and about a third of those carried some implication. This could be an entailment (positive or negative), a presupposition of factivity, or a presupposition of counterfactivity.
  • Verbs are grouped into the following classes, each defined by its behaviour in positive and negative environments:
    • two-way implicatives (manage to, forget to)
    • one-way +implicatives (force to, refuse to)
    • one-way −implicatives (attempt to, hesitate to)
    • factives (forget that)
    • counterfactives (pretend that)
    • neutral verbs (want to)
  • Parsing uses the XLE parser with a hand-written broad-coverage English grammar. Logical forms are skolemized, flattened into contexts and canonicalized by packed term rewriting. Entailment and contradiction detection (ECD) then works on the packed output, so no disambiguation step is needed.
  • The AQUAINT evaluation uses three labels: true, false and unknown. The PASCAL challenge, by contrast, merges false and unknown into a single label.
  • The authors describe this as the first systematic implementation of inferences from the six implicative types and their interaction with factives. Remaining gaps include nominal collocations (e.g. take the trouble to), cancelable implicatures, and graded factivity.

Interesting ideas

  • Polarity as a set of ancestors. Instead of one global sign, each context stores the ancestors relative to which it is positive and those relative to which it is negative. This makes it possible to draw inferences inside a believer's context as well as the author's, for example inside "John believes…". The same design could be applied to nested modal or attitude scopes in other languages.
  • One lookup table for operators and verbs. Sentential negation is treated exactly like a polarity-inverting verb such as forget to. This keeps the propagation algorithm uniform and easy to extend with new operators.
  • Signatures with a "none" value. One-way implicatives give no entailment in one environment, so a one-way verb can simply drop its parent's negative or positive set. Recording the absence of an implication explicitly helps the system separate "unknown" from "false".
  • Separating entailment from implicature during annotation. Annotators tested whether a suggested inference could be cancelled without contradiction, as in refuse to or be able to. They flagged implicatures separately instead of encoding them as entailments, which is a sound protocol for building any inference lexicon.
  • Frequency-driven lexicon building with corpus checks. Verbs were processed in BNC frequency order and judged against real web usage. This approach would transfer well to building implication lexicons for other languages, such as Ukrainian complement-taking verbs.

Topics: Implicature

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{nairn2006computing,
  title = {{Computing relative polarity for textual inference}},
  author = {Nairn, Rowan and Condoravdi, Cleo and Karttunen, Lauri},
  booktitle = {Proceedings of the Fifth International Workshop on Inference in Computational Semantics (ICoS-5)},
  year = {2006},
  url = {https://aclanthology.org/W06-3907/}
}

References

12 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bobrow, D., C. Condoravdi, R. Crouch, R. Kaplan, L. Karttunen, T. King, V. de Paiva and A. Zaenen, A basic logic for textual inference, in: Proceedings of the AAAI Workshop on Inference for Textual Question Answering, Pittsburgh, PA, 2005, http://www2.parc.com/istl/groups/nltt/papers/textual-inference.pdf. ↗
  2. Condoravdi, C., R. Crouch, R. Stolle, V. de Paiva and D. Bobrow, Entailment, intensionality and text understanding, in: Proceedings of the Workshop on Text Meaning, Human Language Technology Conference (HLT-NAACL-2003), Edmonton, Canada, 2003, http://www2.parc.com/spl/members/stolle/Papers/condoravdi-textmeaning.pdf. ↗
  3. Crouch, R., Packed rewriting for mapping semantics to KR, in: Proceedings of the Sixth International Workshop on Computational Semantics, Tilburg, the Netherlands, 2005, http://www2.parc.com/istl/groups/nltt/papers/iwcs05_crouch.pdf. ↗
  4. Crouch, R., R. Sauri and A. Fowler, AQUAINT pilot knowledge-based evaluation: Annotation guidelines (2005), http://www2.parc.com/istl/groups/nltt/papers/aquaint_kb_pilot_evaluation_guide.pdf. ↗
  5. Geis, M. and A. Zwicky, On invited inferences, Linguistic Inquiry 2 (1971), pp. 561–565. find
  6. Grice, H. P., Logic and conversation, in: P. Cole and J. L. Morgan, editors, Speech Acts, Academic Press, New York, NY, 1989 pp. 41–58. find
  7. Ido Dagan, O. G. and B. Magnini, The PASCAL recognising textual entailment challenge, in: Proceedings of the PASCAL Challenges Workshop on Recognising Textual Entailment, Southampton, U.K., 2005, http://www.cs.biu.ac.il/~glikmao/rte05/dagan_et_al.pdf. ↗
  8. Karttunen, L., Implicative verbs, Language 47 (1971), pp. 340–358. find
  9. Karttunen, L., The logic of English predicate complement constructions (1971), distributed by the Indiana University Linguistics Club. http://www2.parc.com/istl/members/karttune/publications/english_predicate.pdf. ↗
  10. Karttunen, L. and A. Zaenen, Veridicity, in: G. Katz, J. Pustejovsky and F. Schilder, editors, Annotating, Extracting and Reasoning about Time and Events, number 05151 in Dagstuhl Seminar Proceedings (2005), http://drops.dagstuhl.de/opus/volltexte/2005/314. ↗
  11. Kiparsky, P. and C. Kiparsky, Fact, in: D. Steinberg and L. Jakobovits, editors, Semantics. An Inderdisciplinary Reader, Cambridge University Press, Cambridge, England, 1971 . find
  12. Zaenen, A., L. Karttunen and R. Crouch, Local textual inference: can it be defined or circumscribed?, in: Workshop on the Empirical Modeling of Semantic Equivalence and Entailment, Ann Arbor, MI, 2005, http://www2.parc.com/istl/members/karttune/publications/acl2005workshop.pdf. ↗