A Text Retrieval System Based on Distributed Representations

Abstract

Most text retrieval systems are essentially based on bag-of-words (BOW) text representations. Despite popularity of BOW, it ignores the internal semantic meanings of words since each word is treated as an atomic unit. Recently, distributed word and text representations become increasingly popular in NLP literatures. They embed syntactic and semantic information of words and texts into low-dimensional vectors, thus overcome the weaknesses of traditional BOW representations to some extent. In this paper, we implement a text retrieval system that are totally supported by distributed representations. Our new system no longer relies on the matchings of words in queries and texts, but uses semantic similarity to judge if a text is relevant to a query and to what extent, which provides better user experience compared with traditional text retrieval systems.