Text Retrieval by Term Co-occurrences in a Query-based Vector Space

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2356–2365, Osaka, Japan, December 11-17 2016. Text Retrieval by Term Co-occurrences in a Query-based Vector Space Eriks Sneiders Department of Computer and Systems Sciences Stockholm University Postbox 7003, SE-164 07, Kista, Sweden eriks@dsv.su.se Abstract Term co-occurrence in a sentence or paragraph is a powerful and often overlooked feature for text matching in document retrieval. In our experiments with matching email-style query messages to webpages, such term co-occurrence helped greatly to filter and rank documents, compared to matching document-size bags-of-words. The paper presents the results of the experiments as well as a text- matching model where the query shapes the vector space, a document is modelled by two or three vec- tors in this vector space, and the query-document similarity score depends on the length of the vectors and the relationships between them. 1 Introduction Vector-space in text retrieval is old news. Cosine similarity by Salton and McGill (1986), along with the probabilistic retrieval model by Robertson and Spark Jones (1976), has dominated text retrieval for the last three decades. Much development has focused on improving the relevance scoring factors be- yond term frequency and inverted document frequency, e.g., co-occurrence and proximity of query terms in the document help improve the relevance score (see Section 6). This paper reports the results of an inductive research. We started with a practical task, then ana- lyzed the prototype and spotted a text retrieval model, then twisted the model in order to improve it. The practical task was to increase automation of the “ask us” function on a municipal website: before the user submits his or her message, the system checks whether some answer-relevant pages are pub- lished on the website; if yes, the system delivers the pages to the user. Reuse of previously published information would save the user’s time and the municipality’s resources. The municipality gave us 25 anonymized sample messages to experiment with, which was not enough for machine learning and text categorization methods. We knew that text-pattern matching yields superior results in automated email answering (Sneiders, 2016). Unfortunately, text patterns require learning, good text patterns are difficult to develop and maintain. We needed something new. Our research problem is matching email-style query messages to webpages by a technique that does not require learning or development of a knowledge base. Our solution is a text-matching technique that relies on term co-occurrence in a limited text chunk (a sentence or a paragraph), yet without the complexity of good text patterns. The novelty and contribution of this paper are: (i) demonstration of the advantage of such term co-occurrence in text matching over matching of document-size bags-of- words, and (ii) a text-matching model where the query shapes the vector space, a document is mod- elled by two or three vectors in this vector space, and the query-document similarity score depends on the length of the vectors and the relationships between them. The model is easy to visualize. Further in this …

Content truncated to 500 words