📄 A Logic-Based Semantic Approach to Recognizing Textual Entailment

Summary

The paper addresses Recognizing Textual Entailment (RTE): deciding whether a hypothesis can be inferred from a given text. The authors frame this as logical implication between meanings. Both fragments become a two-layer logic form. The syntactic layer has one predicate per content word, plus links for prepositions, conjunctions and noun compounds. The semantic layer adds relations such as AGENT, THEME, TIME and LOCATION, along with normalized temporal predicates.

These forms are passed to COGEX, a prover derived from OTTER. COGEX tries to refute the negated hypothesis, using axioms that are generated only when a proof needs them: eXtended WordNet lexical chains, world knowledge, and semantic and temporal axioms. If no refutation appears, the prover first loosens predicate arguments and then removes predicates. Each step costs points, and the normalized score is compared with a learned threshold.

Two prover variants, fed by constituency and by Minipar dependency parses, are mixed with a shallow lexical edit-distance aligner using per-task linear weights. The combination reaches 73.75% accuracy on RTE 2006 data. The work shows that deep, inspectable reasoning becomes robust through graded proofs and hybridization.

Key points

• COGEX loads the text plus the negated hypothesis as its set of support and the axioms as its usable list. Proof scores are normalized by the largest possible penalty, which is the case where every hypothesis predicate is dropped.
• Logic forms come from two sources: a Collins-style constituency parser, enriched by the authors' semantic parser, and Minipar dependencies. The authors note that parser accuracy of about 86% limits the approach.
• Lexical chains have several safeguards:
  • first WordNet senses are used unless source and target are synonyms;
  • chains are at most three relations long;
  • some relation orderings are banned, for example IS-A followed by hyponymy, or more than one hyponymy link;
  • overly general concepts are excluded via a generality weight based on hierarchy depth and BNC information content.
• Dates are normalized into begin/end time intervals. Learned event-ordering triples are mapped by hand-coded rules to SUMO-style predicates such as earlier and during.
• WordNet chains and NLP axioms are used most often in proofs, but semantic and temporal axioms give the biggest accuracy gains. The three-system combination reaches 73.75% on RTE 2006.

Interesting ideas

• I1. When a strict proof fails, the prover first relaxes arguments, then drops predicates, charging a penalty for each step and normalizing by sentence size; this turns brittle yes/no logic into a graded, explainable confidence.
• I2. Scoring depends on quantification: with existential statements, extra modifiers in the hypothesis are penalized, while universal statements reverse the subset direction.
• I3. Background axioms are created lazily from WordNet paths only when a proof needs a link, with length limits, banned relation sequences and a generality filter to block unsound inferences.
• I4. Negation from "not/never" or a "no" determiner is not kept as a separate word; it is moved onto the verb predicate as a negated event.
• I5. A cheap asymmetric edit distance is combined with the deep systems through per-task weights tuned by grid search. In this edit distance, deletions are free, substitutions are allowed only between synonyms, and insertion cost depends on part of speech.

For Mova

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

• I1 (answer): readers could report which hypothesis facts had to be weakened or dropped to match the text world, giving a scored and fully traceable answer instead of a bare failure.
• I2 (world): entailment checks over the text world should flip the direction of modifier subsumption when an entity group is universally quantified ("all", "every").
• I3 (world): the compiled ontology could expose on-demand inference links with the same filters, keeping each derived fact justified by a short, inspectable relation path.
• I4 (frames): frames could carry polarity on the event itself, so negated actions are never confused with asserted ones in later world-building.
• I5 (learn): this kind of lexical scorer can act as a baseline witness or negative control, showing where deep reasoning actually adds value over word overlap.