📄 Presupposed Content and Entailments in Natural Language Inference

Summary

The paper addresses a gap in the NatLog natural logic system for NLI. NatLog runs negation, monotonicity and implicatures through one projection mechanism, but presuppositions behave differently. They survive negation, yet some operators block them and conditionals can cancel them. The authors add a separate presupposition channel to NatLog's pipeline.

Lexical factivity triggers, meaning factive verbs and nouns such as "know" and "realize" and antifactive ones such as "pretend", are found with regular-expression patterns over lemmatized parse trees. Their effects are then projected up the tree using Karttunen's (1973) holes/plugs/filters theory. For each edit from premise to hypothesis, the system computes an atomic presupposition relation, composes these across the edit sequence, and combines the result with NatLog's ordinary entailment relation.

This repairs errors the original system made, for instance concluding that Gore won the election from a premise stating that Bush was unaware of Gore's election victory. On a PARC 3-way entailment test set, accuracy rose from 53.95% to 60.53%. The approach suggests that combining specialized reasoners is a workable path toward open-domain NLI.

Key points

• NatLog has 5 stages: linguistic analysis, alignment, lexical entailment classification, entailment projection, and entailment composition. Presupposition handling is added at stages 1, 4 and 5.
• Triggers are encoded as tree patterns with a lemma, a scope (SBAR/S complement) and a factivity label (FACT or ANTI). For example, "know" is matched as a VB-headed VP.
• Effective factivity is computed bottom-up by pairing the dominated and dominating labels:
  • ANTI over anything gives ANTI.
  • FACT over ANTI gives ANTI.
  • FACT over FACT gives FACT.
• Atomic presupposition relations depend on the edit type and the factivity of the edited operator:
  • Deleting a FACT operator gives Forward; inserting one gives Reverse.
  • Deleting or inserting an ANTI operator gives Alternation.
• Evaluation used the 3-way NLI problems from PARC (the Nairn et al. 2006 group): Augmented NatLog reached 60.53% versus 53.95% for the original. Other trigger types (aspectual verbs, definite descriptions, iteratives) are left for future work.

Interesting ideas

• I1. Keep presupposed content in its own relation channel, separate from asserted content, because the two project through different operators (negation flips assertions but leaves presuppositions intact).
• I2. Tag every constituent with an "effective factivity" by recursively composing the labels of all dominating operators, using a small lookup table.
• I3. Classify operators by how they treat embedded presuppositions: holes let them through, plugs block them, and filters decide based on the relation between antecedent and consequent.
• I4. Describe lexical triggers declaratively as small records: lemma pattern, structural scope, and factivity label.
• I5. Build a test set of minimal pairs that stress one phenomenon, such as negated factives and antifactive embeddings, and compare the system with and without the new module.

For Mova

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

• I1 (world): Store presupposed facts in the world of the text with their own provenance tag, so that negating a clause does not erase what its factive verb takes for granted.
• I2 (frames): A walk up the UD tree from each ccomp/xcomp to the root could assign every embedded frame a veridicality label (fact, counterfact, unknown), and the trace stays fully explainable.
• I3 (frames): Add hole/plug/filter features to verb lexicon entries (realize = hole, pretend = plug, if = filter) so that projection becomes a data-driven rule, not ad hoc code.
• I4 (tree): Keep a trigger table keyed on lemma plus UD relation of the complement, so new factives can be added without touching the parser or the inference code.
• I5 (answer): Add negated-factive and pretend-style items to the QA test suites, with the baseline run kept as a control, to check that the presupposition logic actually changes answers.