Elamur Research⌕
☰
home / papers / 2019 / Natural Language Inference with Monotonicity

Natural Language Inference with Monotonicity

Paper · 2019 · Hai Hu, Qi Chen and Larry Moss

Summary

The paper presents a working natural language inference system built on natural logic. It reasons directly over polarity-annotated CCG parse trees and never converts sentences into a formal logical representation. The polarity marker from Hu and Moss (2018) tags every constituent as upward-monotone, downward-monotone or neither. The system then rewrites constituents using an ordering relation (≤) collected from the premises, from quantifier knowledge, and optionally from WordNet. Upward positions get broader terms and downward positions get narrower ones, which yields entailed sentences. A separate one-step operation swaps quantifiers or adds negation to produce contradictions. A shallow depth-first search checks whether any generated sentence equals the hypothesis string. A few natural logic rules that replacement cannot express, notably DET, extend coverage to multi-premise problems. On the 74 generalized-quantifier problems of FraCaS section 1, the system reached 88% accuracy with perfect precision. Apart from parsing, it is fully deterministic, and its derivations read almost like human-language proofs.

Key points

  • Evaluation covers FraCaS section 1 (74 problems, multi-premise included). The system scores 88% accuracy against reported results of 95% (D14), 93% (A16), 80% (T14), 77% (M15) and 62% (LS13).
  • Every wrong answer was "unknown": 7 entailment items and 2 contradiction items were missed, and no entail or contradict verdict was false.
  • The search is depth-first with a default depth of 2 and allows one replacement per step. Contradiction generation is applied only once per sentence, because negating twice just returns the original.
  • The parser is the main bottleneck. C&C could not parse four section-1 sentences. EasyCCG parsed all of them, but its trees still needed semi-automatic fixes for inconsistent quantifier supertags and for multi-word expressions.
  • The authors report that Stanford CoreNLP v3.9.1, which finds polarity by pattern-matching on dependency trees, mislabels cases such as "no man walks" and "I like most dogs". Their CCG-based marker handles these correctly and also assigns polarity to whole constituents, not just single words.

Interesting ideas

  • I1. Mark each phrase as upward, downward or non-monotone, then produce entailments by swapping in broader or narrower phrases from an ordering; this gives sound inference without any logical form.
  • I2. Build the ordering relation on the fly from the problem itself: universal premises contribute "X ≤ N", and any modified noun counts as narrower than the bare noun.
  • I3. Use an asymmetric decision order: answer "entail" only when a derivation is found, then try generated contradictions, and otherwise say "unknown", which keeps false positives at zero and leaves a readable proof.
  • I4. Generate contradictions with a single quantifier swap (no↔some) or by negating the main verb, and avoid swaps inside relative clauses.
  • I5. Before inference, normalize surface variants such as existential "there are NP who VP" into one canonical quantified form, so that string matching stays reliable.

For Mova

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

  • I1 (world): Polarity could be computed over UD subtrees and combined with WordNet hypernym chains, so the world model checks entailments by explicit, traceable substitutions.
  • I2 (world): When reading a text, Mova could add these local subset facts to the knowledge graph, giving per-text hyponymy that need not exist in the compiled ontology.
  • I3 (answer): Test-taking readers could adopt this conservative policy and print the substitution chain as the explanation behind each yes/no answer.
  • I4 (learn): The same cheap edits could produce minimally different negative controls for evaluation, revealing readers that ignore quantifiers or negation.
  • I5 (frames): A small normalization pass from constructions to canonical frames would shrink the variation that later frame matching and hypothesis comparison must cope with.

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

🎓Indiana UniversityBloomington · United States24 works · 2,995 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.Reasoning · 510 works

Cite

Cite as Hai Hu, Qi Chen and Larry Moss. 2019. Natural Language Inference with Monotonicity. Proceedings of the 13th International Conference on Computational Semantics - Short Papers.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/w19-0502

@inproceedings{hu2019natural,
  title = {{Natural Language Inference with Monotonicity}},
  author = {Hai Hu, Qi Chen and Larry Moss},
  booktitle = {Proceedings of the 13th International Conference on Computational Semantics - Short Papers},
  year = {2019},
  doi = {10.18653/v1/w19-0502},
  url = {https://doi.org/10.18653/v1/w19-0502}
}

References

24 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abzianidze, L. (2014). Towards a wide-coverage tableau method for natural logic. In New Frontiers in Artificial Intelligence - JSAI-isAI 2014 Workshops, LENLS, JURISIN, and GABA, Kanagawa, Japan, October 27-28, 2014, Revised Selected Papers, pp. 66–82. find
  2. Abzianidze, L. (2015). A tableau prover for natural logic and language. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pp. 2492–2502. find
  3. Abzianidze, L. (2016). Natural solution to fracas entailment problems. In Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics, pp. 64–74. find
  4. Abzianidze, L. (2017). Langpro: Natural language theorem prover. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pp. 115–120. ↗
  5. Angeli, G. and C. Manning (2014). NaturalLI: Natural logic inference for common sense reasoning. In EMNLP, pp. 534–545.
  6. Angeli, G., N. Nayak, and C. Manning (2016). Combining natural logic and shallow reasoning for question answering. In ACL, pp. 442–452. find
  7. Bowman, S. R., C. Potts, and C. D. Manning (2015). Learning distributed word representations for natural logic reasoning. In AAAI Spring Symposium on Knowledge Representation and Reasoning, pp. 10–13. ↗
  8. Cases, I. and L. Karttunen (2017). Neural networks and textual inference: How did we get here and where do we go now? find
  9. Clark, S. and J. R. Curran (2007). Wide-coverage efficient statistical parsing with CCG and log-linear models. Computational Linguistics 33(4), 493–552. ↗
  10. Cooper, R., D. Crouch, J. Van Eijck, C. Fox, J. Van Genabith, J. Jaspars, H. Kamp, D. Milward, M. Pinkal, M. Poesio, et al. (1996). Using the framework. Technical report, Technical Report LRE 62-051 D-16, The FraCaS Consortium. find
  11. Dong, Y., R. Tian, and Y. Miyao (2014). Encoding generalized quantifiers in dependency-based compositional semantics. In Proceedings of the 28th Pacific Asia Conference on Language, Information and Computing. find
  12. Fyodorov, Y., Y. Winter, and N. Fyodorov (2003). Order-based inference in natural logic. Log. J. IGPL 11(4), 385–417. Inference in computational semantics: the Dagstuhl Workshop 2000. ↗
  13. Hu, H., T. F. Icard, and L. S. Moss (2018). Automated reasoning from polarized parse trees. In Proceedings of the Fifth Workshop on Natural Language and Computer Science. ↗
  14. Hu, H. and L. S. Moss (2018). Polarity computations in flexible categorial grammar. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pp. 124–129. find
  15. Lewis, M. and M. Steedman (2013). Combined distributional and logical semantics. Transactions of the Association of Computational Linguistics 1, 179–192. ↗
  16. Lewis, M. and M. Steedman (2014). A* CCG parsing with a supertag-factored model. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 990–1000.
  17. MacCartney, B. and C. D. Manning (2008). Modeling semantic containment and exclusion in natural language inference. In Proceedings of COLING, pp. 521–528. Association for Computational Linguistics. find
  18. MacCartney, B. and C. D. Manning (2009). An extended model of natural logic. In IWCS-8, Proceedings of the Eighth International Conference on Computational Semantics, pp. 140–156. find
  19. Marelli, M., S. Menini, M. Baroni, L. Bentivogli, R. Bernardi, and R. Zamparelli (2014). A SICK cure for the evaluation of compositional distributional semantic models.
  20. Miller, G. A. (1995). Wordnet: a lexical database for English. Communications of the ACM 38(11), 39–41.
  21. Mineshima, K., P. Martı́nez-Gómez, Y. Miyao, and D. Bekki (2015). Higher-order logical inference with compositional semantics. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 2055–2061. find
  22. Tian, R., Y. Miyao, and T. Matsuzaki (2014). Logical inference on dependency-based compositional semantics. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Volume 1, pp. 79–89. find
  23. van Benthem, J. (1986). Essays in Logical Semantics, Volume 29 of Studies in Linguistics and Philosophy. Dordrecht: D. Reidel Publishing Co. find
  24. Zamansky, A., N. Francez, and Y. Winter (2006). A ‘natural logic’ inference system using the Lambek calculus. Journal of Logic, Language, and Information 15(3), 273–295. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/hu-2019-natural-language-inference-monotonicity/