Kevin W. Humphreys

Reusing a statistical language model for generation

A relatively self-contained subtask of natural language generation is sentence realization: the process of generating a grammatically correct sentence from an abstract semantic / logical representation. We propose a method where sentence realization is carried out using a simplified (context free) version of a large analysis grammar, combined with a statistical language model from the full (context sensitive) version of the same grammar. The statistical model provides a measure of the probability of syntactic substructures, derived from the analysis of a corpus with the full grammar, and is used to guide both subsequent analysis and generation.

New Directions in Question Answering

It addresses all its goals well and in general is very much forward-looking. Apart from an interesting paper on the commercial deployment of question answering applications, there is little historical review of the development of the field, and the extent to which it has been driven by the TREC-QA track evaluations, which have played a significant role in defining the field to date. The editor’s introduction and several of the papers make clear the extent to which question answering (QA) is a proper subset of artificial intelligence (AI). The editor summarises the results of a panel session at the symposium to establish a roadmap for QA research, which raises several significant issues as requirements, including entire AI research areas such as user modelling and inferencing. The discussion of the roadmap is very brief and would perhaps have deserved its own chapter in the book, to better situate the symposium papers. The introduction also contains a useful table of the areas addressed by each paper in the collection. However the table does act to highlight the fact that none of the papers address multilingual issues, which are a significant area for future research. The editor’s introductions to each section are extremely thorough and well written, providing very useful abstracts of all the papers’ main results and contributions. Several of the papers have rather tangential relations to QA, but their relevance and contribution are made clear by the editor’s abstracts. Section 1: Foundation and History situates the book well, giving a broad view of the long term goals for QA as an AI research area, together with a review of practical progress so far and the current status. The first chapter by Nyberg, Burger, Mardis and Ferrucci follows on from the discussion of the roadmap to give a high-level view of the criteria required for QA in general. It makes clear that fully general QA is a very ambitious goal, requiring solutions to various significant AI problems, but it does highlight real systems which are addressing some of the criteria. The next chapter by Ulicny gives a thorough