Query Segmentation and Resource Disambiguation Leveraging Background Knowledge

Accessing the wealth of structured data available on the Data Web is still a key challenge for lay users. Keyword search is the most convenient way for users to access information (e.g., from data repositories). In this paper we introduce a novel approach for determining the correct resources for user-supplied keyword queries based on a hidden Markov model. In our approach the user-supplied query is modeled as the observed data and the background knowledge is used for parameter estimation. Instead of learning parameter estimation from training data, we leverage the semantic relationships between data items for computing the parameter estimations. In order to maximize accuracy and usability, query segmentation and resource disambiguation are mutually tightly interwoven. First, an initial set of potential segmentations is obtained leveraging the underlying knowledge base; then the final correct set of segments is determined after the most likely resource mapping was computed using a scoring function. While linguistic methods like named entity, multi-word unit recognition and POS-tagging fail in the case of an incomplete sentences (e.g. for keyword-based queries), we will show that our statistical approach is robust with regard to query expression variance. Our experimental results when employing the hidden Markov model for resource identification in keyword queries reveal very promising results.