Searching for Entities: When Retrieval Meets Extraction

[摘要] Retrieving entities from inside of documents, instead of searching for documents or web pages themselves, has become an active topic in both commercial search systems and academic information retrieval research area. Taking into account information needs about entities represented as descriptions with targeted answer entity types, entity search tasks are to return ranked lists of answer entities from unstructured texts, such as news or web pages. Although it works in the same environment as document retrieval, entity retrieval tasks require finer-grained answers entities which need more syntactic and semantic analyses on germane documents than document retrieval. This work proposes a two-layer probability model for addressing this task, which integrates germane document identification and answer entity extraction. Germane document identification retrieves highly related germane documents containing answer entities, while answer entity extraction finds answer entities by utilizing syntactic or linguistic information from those documents. This work theoretically demonstrates the integration of germane document identification and answer entity extraction for the entity retrieval task with the probability model. Moreover, this probability approach helps to reduce the overall retrieval complexity while maintaining high accuracy in locating answer entities. Serial studies are conducted in this dissertation on both germane document identification and answer entity extraction. The learning to rank method is investigated for germane document identification. This method first constructs a model on the training data set using query features, document features, similarity features and rank features. Then the model estimates the probability of the germane documents on testing data sets with the learned model. The experiment indicates that the learning to rank method is significantly better than the baseline systems, which treat germane document identification as a conventional document retrieval problem.The answer entity extraction method aims to correctly extract the answer entities from the germane documents. The methods of answer entity extraction without contexts (such as named entity recognition tools for extraction and knowledge base for extraction) and answer entity extraction with contexts (such as tables/lists as contexts and subject-verb-object structures as contexts) are investigated. These methods individually, however, can extract only parts of answer entities. The method of treating the answer entity extraction problem as a classification problem with the features from the above extraction methods runs significantly better than any of the individual extraction methods.

[发布机构] the University of Pittsburgh