Introducing the Sequence Model for Text Retrieval

We propose and explore a novel approach, called the sequence model, to text retrieval. The model differs from classical ones in the extent of how positional information of term occurrences is used for relevance judgment. In the sequence model, documents and queries are viewed as sequences of term-position pairs and the relevance of a document to a query is judged by the similarity between their respective representative sequences. We suggest three primitive measures of sequence similarity, each capturing a distinct aspect of resemblance between two sequences. These similarity measures can be combined in various ways to suit different information needs. We have developed a prototype system with the sequence model as its core. Experimental results show that our sequence-based approach is often more effective than appearance-based approaches.