Predicting scalar inferences from "or" to "not both" using neural sentence encoders

What it is about

A short paper extending the «some → not all» strength studies (Degen 2015; Schuster et al. 2020) to «or → not both». 1,244 Switchboard sentences with or were rated, in context, for how close they are to the same sentence with «but not both» added, on a slider and on a 1–7 scale. Strength varies widely and individual items often have bimodal ratings: people disagree on whether exclusivity was meant. Two features matter. either strengthens the inference («You're either liberal or conservative», 6.6 of 7). A downward-entailing context such as negation kills it («I don't believe in drugs or alcohol», 1.1). Together the features explain much less variance than for some (marginal R² 0.07). The authors read this as evidence that or often has other jobs — definitional, metalinguistic, listing — where «and» is not a salient alternative. A BERT+BiLSTM model predicting a discrete rating distribution reaches only r = 0.47.

Ideas

I1. Keep «or → not both» in prag/src/infer.rs (the or < and scale) but gate it more strictly than «some». Fire only with either … or, or with two clausal or eventive disjuncts in an assertion. Never fire under negation or another downward-entailing operator (no, never, without, if-clause, question). Add tests: «I don't believe in drugs or alcohol» → no inference; «either X or Y» → strong.
I2. Recognise the non-exclusive uses of or before inferring: definitional or appositive or («the gnu, or wildebeest») and approximate numbers («two or three days»). Mark them as or-not-scalar in the explanation. Both are frequent in school lessons and tales, and both would otherwise yield false «not both» claims.
I3. Because people themselves split on many or items, do not train a strength classifier for or on teacher labels. Treat the rule's output as a hint the explanation states, and record the measured disagreement as a limit in chapter 33 §4.2.

For Mova

33. Implicature and presupposition §4.1–4.2; prag/src/infer.rs (implicatures, SCALES); measure: unit tests with negative controls, and a count of or inferences on lesson texts before and after the gate (with a checked sample).

Cite

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

@inproceedings{li2021predicting,
  title = {{Predicting scalar inferences from "or" to "not both" using neural sentence encoders}},
  author = {Li, Elissa and Schuster, Sebastian and Degen, Judith},
  booktitle = {Proceedings of the Society for Computation in Linguistics 2021},
  year = {2021},
  url = {https://aclanthology.org/2021.scil-1.54/}
}

References

14 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Mira Ariel. 2020. Or constructions, argumentative direction and disappearing ‘alternativity’. Language Sciences, 81:101195. Ad hoc categorization and language: the construction of categories in discourse. find
Mira Ariel and Caterina Mauri. 2018. Why use or? Linguistics, 56(5):939–993. find
Gennaro Chierchia et al. 2004. Scalar implicatures, polarity phenomena, and the syntax/pragmatics interface. Structures and beyond, 3:39–103. find
Judith Degen. 2015. Investigating the distribution of ‘some‘ (but not ‘all‘ ) implicatures using corpora and web-based methods. Semantics & Pragmatics, 8(11):1–55. find our card
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL 2019), pages 4171–4186. John J. Godfrey, Edward C. Holliman, and Jane. find
Mc-Daniel. 1992. SWITCHBOARD: Telephone speech corpus for research and development. In Proceedings of the 1992 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 92). find
Laurence Horn. 1985. Metalinguistic negation and pragmatic ambiguity. Language, 61(1):121–174. find
Masoud Jasbi. 2018. Learning Disjunction. Ph.D. thesis, Stanford University. find
Paloma Jeretic, Alex Warstadt, Suvrat Bhooshan, and Adina Williams. 2020. Are natural language inference models IMPPRESsive? learning IMPlicature and PRESupposition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020). find our card
Nanjiang Jiang and Marie-Catherine de Marneffe. 2019. Do you know that florence is packed with visitors? evaluating state-of-the-art models of speaker commitment. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL 2019). find
Christopher Potts and Roger Levy. 2015. Negotiating lexical uncertainty and speaker expertise with disjunction. In Proceedings of the 41st Annual Meeting of the Berkeley Linguistics Society. find
Uli Sauerland. 2012. The Computation of Scalar Implicatures: Pragmatic, Lexical or Grammatical? Language and Linguistics Compass, 6(1):36–49. find
Sebastian Schuster, Yuxing Chen, and Judith Degen. 2020. Harnessing the linguistic signal to predict scalar inferences. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020). find our card
Isabel Gomez Txurruka and Nicholas Asher. 2008. A discourse-based approach to natural language disjunction (revisited). In Language, Representation and Reasoning. University of the Basque country Press. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/li-2021-predicting-scalar-inferences-or-not/