Konferensartikel

We present our work towards developing a system that should find, in a large text corpus, contiguous phrases expressing similar meaning as a query phrase of arbitrary length. Depending on the use case, this task can be seen as a form of (phrase-level) query rewriting. The suggested approach works in a generative manner, is unsupervised and uses a combination of a semantic word n-gram model, a statistical language model and a document search engine. A central component is a distributional semantic model containing word n-grams vectors (or embeddings) which models semantic similarities between n-grams of different order. As data we use a large corpus of PubMed abstracts. The presented experiment is based on manual evaluation of extracted phrases for arbitrary queries provided by a group of evaluators. The results indicate that the proposed approach is promising and that the use of distributional semantic models trained with uni-, bi- and trigrams seems to work better than a more traditional unigram model.

Unsupervised query rewriting Distributional semantics N-gram embeddings