📄 Weakly Supervised Learning of Presupposition Relations between Verbs

Summary

The paper targets presupposition between verbs, a lexical relation that WordNet and VerbOcean barely cover or merge with entailment. Examples are find presupposing seek and manage presupposing try. The two relations matter separately because a presupposition still holds when the sentence is negated, while an entailment does not. Tremper builds a weakly supervised, discriminative classifier that sorts verb pairs into five classes: presupposition, entailment, temporal inclusion, antonymy and other/no. Candidate pairs come from DIRT and are filtered with ukWAC frequencies and PMI. A few manually chosen seed pairs per class yield a training corpus that is partly hand-labelled, partly labelled by heuristics and partly built from WordNet synonyms. Sentences are parsed with XLE to extract shallow and deep features. One J48 binary classifier is trained per relation, and their votes are combined with confidence filters. Only the first iteration of the bootstrapping loop is evaluated. Type-level accuracy is about 36% overall, well above a random baseline, while token-level accuracy is lower. This is an early attempt to learn presupposition from corpora.

Key points

• Candidate pairs: DIRT provides 808,764 distinct pairs drawn from 5,604 distinct verbs. Filtering by ukWAC (parts 1–3) frequency and a PMI threshold of 2.0 leaves 199,393 candidates.
• Seeds and training data:
  • Each relation gets three seed pairs, except temporal inclusion, which gets six.
  • The first training set has 15,717 sentences: 5,032 hand-labelled, 9,918 heuristically labelled and 757 from WordNet synonym pairs.
  • The data are rebalanced because temporal inclusion is about 2% of sentences and entailment about 39%.
• Features:
  • verb distance and order, verb form, and negation
  • polarity verbs from the Nairn et al. lexicon
  • conjunctions, adverbials and PoS windows
  • length of the grammatical-function path between the verbs, from XLE parses
  • whether the verbs share a subject or object
• Ensemble and filtering:
  • A classifier vote below 0.75 confidence is dropped, and ties go to the most confident classifier.
  • At type level, a verb pair can receive up to three labels if each covers more than 10% of its instances.
  • Bootstrapping keeps instances whose pattern-based reliability exceeds 0.3, following Pantel and Pennacchiotti (2006).
• Results:
  • Antonymy scores best at type level (72% Majority, 42% Without NONE, against a 12% baseline).
  • At token level, the four real relations reach only 13–29%, mostly because single contexts are not informative enough.

Interesting ideas

• I1. Negation works as a diagnostic: presuppositions survive under negation and entailments do not, so the polarity of contexts can separate two relations that otherwise look identical.
• I2. Rather than mining candidates with patterns, the method starts from a precomputed list of associated verbs and prunes it with frequency and PMI before any classification.
• I3. A small hand-made stoplist of contexts (e.g. look followed by forward) acts as cheap disambiguation when spreading seed labels to new sentences.
• I4. One binary classifier per relation is combined with confidence cut-offs, and type-level labels come from instance proportions with a cap on how many labels a pair may receive.
• I5. Deep features such as the dependency path length between two verbs and whether they share a subject or object add structural evidence beyond word windows.

For Mova

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

• I1 (frames): The implicature layer can tag verb-licensed inferences as presupposed or entailed, and keep only the presupposed ones when the frame sits under negation.
• I2 (world): Verb-relation entries in compiled knowledge could be built by filtering a DIRT-style association list first and classifying only the survivors, which keeps cost and noise low.
• I3 (learn): When silver data is expanded from seeds, an explicit, inspectable blocklist of misleading contexts can stop systematic label leaks and stay explainable.
• I4 (learn): The same aggregation (discard weak votes, allow at most k labels, require a minimum share) can turn noisy per-sentence predictions into knowledge-graph facts with a stated support level.
• I5 (tree): The UD trees Mova already produces can feed these path and argument-sharing features directly into any verb-pair relation learner, with no bag of words.