SEEK AND FIND A retrieval approach to construction search

Abstract: Finding occurrences of specific grammatical constructions in running text is a central issue to constructionist approaches to linguistics and language processing. Of special concern are partially schematic constructions that cannot be distinguished from unrelated constructions by surfaceform alone. In order to detect instances of such complex constructions we consider using features that are intended to capture semantic restrictions of particular construction elements. We addressthis task as an information retrieval (IR) problem, and describe a simple interactive architecture for searching for constructions.The retrieval system is guided by the user who provides it with a number of positive seed examples (occurrences of the construction) and tailors a ranking function based on a combination of lexical semanticsimilarity features (lexicon-based or distributional). The system was evaluated using standard IR metrics on a new benchmark for construction retrievalin Swedish, and we observed that a lexical-semantic re-ranker can give significant improvement over a lemma-based baseline, but must be tailored for the construction at hand. The search system is effective even with a small number of positive seed examples, which proves the feasibility of our approach from a user perspective.