📄 Using Phrasal Patterns to Identify Discourse Relations

Summary

The paper addresses automatic recognition of the discourse relation linking two adjacent Japanese sentences when no explicit connective is available. Following Marcu and Echihabi (2002), the authors build training data automatically. They collect sentence pairs whose second sentence opens with an unambiguous cue phrase and treat that cue as the relation label. The cue phrases come from the ChaSen dictionary and are grouped into classes adapted from Ichikawa. The data comes from roughly 20G of Web text and 38 years of newspapers. Two evidence sources are combined:

• Lexical: cross-sentence word pairs, restricted by part of speech and marked for negation.
• Phrasal patterns: combinations of bunsetsu fragments in which function words are required and content words are optional.

Frequency-based scores, some corrected for class priors, are merged by a simple decision rule. Phrasal patterns raise weighted accuracy from 53% to 65%. That result roughly matches human judges given the same two-sentence context. It shows that grammatical, function-word structure carries discourse signal beyond lexical content.

Key points

• Relation inventory: 167 cue phrases from ChaSen (IPADIC, Ver.2.7.0) are mapped to 7 classes. OTHER is dropped, leaving six: ELABORATION (43.0% of the corpus), CONTRAST (32.2%), CAUSE-EFFECT, EQUIVALENCE, CHANGE-TOPIC and EXAMPLE.
• Data: about 1,300,000 sentence pairs come from the Web and about 150,000 from newspapers. A balanced test set of 300 pairs (50 per relation) is held out.
• Pattern construction: every pattern must contain a minimum of three subphrases, and both sentences must contribute one or more of them. Patterns seen more than 1000 times are discarded as uninformative. A single example pair yields 348 distinct patterns.
• Results: unweighted accuracy rises from 57% to 64%, and weighted accuracy from 53% to 65%. The majority baseline is 43%, and Marcu and Echihabi reported 49.7% in a comparable setup. CONTRAST benefits most, rising from 62% to 86%.
• Human comparison: three annotators scored 63%, 54% and 48%, with 69% average inter-annotator agreement. System accuracy was 71% on items all three humans got right and 47% on items none of them did.