Pragmatic and Logical Inferences in NLI Systems: The Case of Conjunction Buttressing

What it is about

"John repaired the engine and the car started" is read as "after John repaired the engine, the car started". This temporal succession, or a causal reading, is a pragmatic enrichment of "and" (conjunction buttressing). Logic treats "and" as commutative. The authors build an NLI dataset of 3,470 premise-hypothesis pairs from UkWaC sentences in which a "before", "after", "because" or "while" clause was rewritten with "and". Each pair has two labels: a logical label, where only commutativity counts as entailment, and a pragmatic label, where succession or causation counts as entailment. Negative pairs cover precedence and synchrony. Three Transformer NLI models trained on MNLI (BERT, RoBERTa, DeBERTa) were tested without training on the data. DeBERTa is the best, with pragmatic accuracy close to 1 on succession and causal pairs. The systems assign all tested relations as entailments at once, and their precedence answers split: DeBERTa entails precedence in 42% of cases and contradicts it in 51%. Eight participants per item were asked about 40 causal pairs. In 31 of the 40 items at least 80% agreed with the pragmatic label. A check of MNLI found 26 of 500 examples where a pragmatic "and" explains the gold label.

Ideas

• I1. world/src/events.rs: when two coordinated clauses both have dynamic (action) predicates, add a succession relation between their events, stored as an implicature with a flag, not as a fact. When the predicates are static, keep the commutative order. Test: static pairs must give the same answers under swap, dynamic pairs must not (negative control).
• I2. Two-label test set for the school: each item carries a logical and a pragmatic answer key, and the report shows which one Mova's answers match. This makes the choice visible, which fits the transparency rule.
• I3. Generator in Rust, not a neural one: take parsed clauses with the connectives "before", "after", "because" (UD advcl with mark), rewrite them with "and", and label them by rule. The paper's MNLI check shows the same pattern can be counted in a corpus, which is a cheap measure of how often Mova's own text needs the enrichment.
• I4. Dynamic versus static verbs from en/src/dict.rs: a lexicon flag, checked against the paper's split (dynamic conjuncts vs. "the food is super" type stative ones).

For Mova

The chapters are 24. Coordination and subordination (conjunction "and") and 33. Implicature and presupposition (generalized implicatures). The code is world/src/events.rs for event order and world/src/narr.rs for narrative sequence, with the verb flag in en/src/dict.rs. Measure: agreement with the two label sets on Mova's own generated pairs, run twice, plus the swap control on static pairs. The paper's dataset is not in this card; the UkWaC source is a separate licence check.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2022.unimplicit-1.2

@inproceedings{pedinotti2022pragmatic,
  title = {{Pragmatic and Logical Inferences in NLI Systems: The Case of Conjunction Buttressing}},
  author = {Pedinotti, Paolo and Chersoni, Emmanuele and Santus, Enrico and Lenci, Alessandro},
  booktitle = {Proceedings of the Second Workshop on Understanding Implicit and Underspecified Language},
  year = {2022},
  doi = {10.18653/v1/2022.unimplicit-1.2},
  url = {https://doi.org/10.18653/v1/2022.unimplicit-1.2}
}

References

7 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Clifton Poth, Jonas Pfeiffer, Andreas Rücklé, and Iryna Gurevych. 2021. What to Pre-Train on? Efficient Intermediate Task Selection. In Proceedings of EMNLP. find
2. Robyn Carston. 1988. Implicature, Explicature, and Truth-Theoretic Semantics. Mental Representations: The Interface between Language and Reality, pages 155–181. find
3. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt-sakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008. The Penn Discourse TreeBank 2.0. In Proceedings of LREC. find our card
4. Nathanael Chambers and Dan Jurafsky. 2008. Unsupervised Learning of Narrative Event Chains. In Proceedings of ACL-HLT. find
5. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference. In Proceedings of NAACL-HLT. find
6. 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
7. A Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of NAACL-HLT. Details about Data Collection find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/pedinotti-2022-pragmatic-logical-inferences-nli-systems/