📄 Automatic Detection of Causal Relations for Question Answering

Summary

Girju addresses the automatic extraction of explicit causal relations within a sentence. Earlier systems relied on hand-written, domain-bound patterns. Here the focus is on noun–verb–noun constructions (NP1 verb NP2) where the verb is a simple causative that is often ambiguous, such as produce or generate.

The method has two stages:

1. Pattern discovery. Causally linked noun pairs are harvested from WordNet (its CAUSE-TO relation and noun glosses). Sentences containing those pairs are then searched to collect 60 candidate causal verbs and verbal expressions.
2. Constraint learning. Matched instances from the LA TIMES section of TREC 9 are parsed with the Charniak parser and labelled by hand as causal or not. A C4.5 decision tree then learns lexical and semantic constraints on the verb and its noun arguments.

The author also proposes a taxonomy of cause-effect questions by ambiguity: explicit, ambiguous and implicit. Adding the causation module to an existing QA system raised precision on causal questions from 36% to 61%. This shows that relation-level semantics can help answer a question type that QA systems of the time handled poorly.

Key points

• WordNet 1.7 glosses yielded 429 noun–noun causal links. About 58.28% of them come from the medical domain.
• Pattern discovery produced 60 ambiguous causal verbs and expressions (Table 1), for example trigger, lead to, stem from and give rise to.
• For each verb, 120 sentences were taken from LA TIMES / TREC 9. The resulting split was a 6,000-sentence training corpus and a 1,200-sentence test corpus.
• The training data held 6,523 pattern instances, of which 2,101 were causal and 4,422 were not. The classifier used 19 lexical and semantic features with C4.5.
• On 50 causation questions (explicit and ambiguous) run on the Harabagiu et al. QA system, precision rose from 36% to 61% with the causal module.

Interesting ideas

• I1. Lexical-resource seeding. Seed pairs come from a lexical resource's definitions, by pairing a headword with a noun in its own gloss. Mining these for a relation gives cheap, high-confidence seeds for Hearst-style pattern bootstrapping, and would work for any relation that glosses describe.
• I2. Ambiguity is resolved separately from pattern finding. Generic patterns are harvested first, and a classifier then decides when an instance truly expresses the relation. This lets recall-oriented discovery coexist with precision-oriented filtering instead of demanding unambiguous hand-written patterns.
• I3. Argument normalisation. Each noun phrase is trimmed to the longest span that exists as a concept in the lexicon, such as "earthquake" from "a 7.1 magnitude earthquake". This maps noisy NPs onto ontology nodes so that semantic-class features can be computed.
• I4. Ontology-driven feature design. Following Kim's philosophical argument, the method assumes only certain entity types can be causes or effects: events, states, processes, phenomena and the like. Using such typed constraints on arguments is a principled way to choose features for other semantic relations.
• I5. Measuring only the component's contribution. The QA evaluation counted only answers delivered through the causal pattern, even though other correct answers existed. This isolates exactly what the new component adds, a useful design for ablation-style evaluations of pipeline modules.