Question-Answering Systems: Answer Search and Retrieval

• How are candidate answers found, ranked, and presented to the user?
  • Some systems use modifications of the vector-space model familiar from IR, but there are many more sophisticated strategies.
    • Mollá Aliod, et al. use a semantically based approach, in which they convert queries and sentences in the documents into statements in first-order logic, and use resolution theorem-proving to try to "prove" the query using the logical assertions derived from a document.
    • Cardie, et al. add syntactic and semantic components to the vector-space approach. (Note that their system, designed for the "50-byte answer" track of TREC, doesn't handle how/why questions.)
      • They first divide the documents into small chunks, and generate vector-space representations of each chunk. The most similar chunks of documents are retained for further consideration (these are the "query-dependent summary extracts").
      • Because answers to the kinds of questions the system is designed for are usually NPs, the system looks for NPs in the summary extracts as possible answers (this is a very simple filter, but improved performace somewhat).
      • Semantic type checking, by analyzing the question and positing restrictions on the kinds of entities that could be answers, also improved performance.
      • Finally, Cardie, et al. used a syntactic similarity measure (how many "base NPs" appear in a summary abstract) to reorder the document chunks returned and ranked by the vector-space similarity measure. Performace improved slightly
  • Typically, Q & A systems either present an answer extracted from a document, or highlight a passage containing the answer in a displayed document.

(3) (5)

(return to syllabus)