Tackling Spuriousness with Similarity

The goal of Semantic Parsing is to translate natural language utterances into logical forms that are executed against a target knowledge base. Ideally one would train the parser on the utterances and their annotations in formal language. However, annotating logical forms is a costly operation so work in recent years has suggested training the parser on examples labeled only with their correct output, and not with the correct formula that produced it (Clarke et al., 2010; Liang et al., 2011). This weak supervision facilitates the training process but comes at the price of learning from spurious formulas, logical forms which execute to the correct output but do not capture the meaning of the original utterance. Consider Table 1 and the utterance ”When did China host the Olympics?”, its correct logical form should be ”value of Year where Country is China”, although the logical form ”first value of Year after 2004” will also yield the correct output, 2008. Both logical forms are consistent (execute to the correct output) but the latter is clearly spurious.