📄 Open Text Semantic Parsing Using FrameNet and WordNet

Summary

The paper tackles a core weakness of syntax-based language understanding: one syntactic pattern can carry different meanings, and one meaning can surface in many syntactic forms. To get a representation that does not depend on syntax, the authors built a rule-based semantic parser for open text instead of a statistical role labeler.

The parser draws on two knowledge sources:

• FrameNet supplies sentence-level knowledge. Each annotated example is turned into an ordered rule that maps syntactic features to frame roles.
• WordNet supplies word-level features such as attribute links and adjective types.

Parsing runs in three steps:

1. A feature-augmented grammar analyzes the sentence and can reject ill-formed input.
2. Semantic roles are assigned by scoring candidate FrameNet rules.
3. Hand-written default rules cover constituents that FrameNet leaves unlabeled.

The output is a recursive, tree-like structure of entities, interactions, roles and modifiers. On 350 held-out FrameNet sentences, the parser assigned the correct role to 74.5% of correctly identified elements. Because it is rule-based, it also works for words with very few annotated examples.

Key points

• Each FrameNet-derived rule encodes, in sentence order: grammatical function (GF), phrase type (PT), position relative to the target word, voice, and the preposition for PP phrases.
• Role assignment scans candidate rules in one direction only, from left to right. Each exact feature match adds 1 point, and the rule with the highest score determines the frame and roles.
• About 100 manual default rules fill gaps that FrameNet does not cover. For example, an "on …" phrase is read as the location of an interaction.
• WordNet-derived lexicon entries record features such as category, degree, type and attribute. For example, slow is linked to the attribute speed.
• The parser also labels the sentence function as assertion, query, yn-query or command.
• Evaluation covered only role assignment, since FrameNet has no gold annotations for the other outputs. The 350 test sentences were removed from the corpus before the rules were learned, and role accuracy reached 74.5%.