tailieunhanh - Báo cáo khoa học: "Extending a Surface Realizer to Generate Coherent Discourse"

We present a discourse-level Tree Adjoining Grammar which tightly integrates syntax and discourse levels, including a representation for discourse entities. We show that this technique makes it possible to extend an optimisation algorithm used in natural language generation (polarity filtering) to the discourse level. We implemented the grammar in a surface realizer and show that this technique can be used to reduce the search space by filtering out referentially incoherent solutions. | Extending a Surface Realizer to Generate Coherent Discourse Eva Banik The Open University Milton Keynes UK Abstract We present a discourse-level Tree Adjoining Grammar which tightly integrates syntax and discourse levels including a representation for discourse entities. We show that this technique makes it possible to extend an optimisation algorithm used in natural language generation polarity filtering to the discourse level. We implemented the grammar in a surface realizer and show that this technique can be used to reduce the search space by filtering out referentially incoherent solutions. 1 Introduction A fundamental problem that microplanners and surface realizers face in natural language generation is how to restrict the search space of possible solutions. A traditional solution to this computational complexity problem is to divide the generation process into tractable sub-problems each represented as a module in a pipeline where every decision made by a module restricts the number of options available to others further down the line. Though such pipeline architectures are computationally efficient they severely restrict the flexibility of the system and the quality of the generated output. Most systems with pipeline architectures generate relatively simple domain-specific output. Systems that produce more complex linguistic constructions typically achieve this by adding more modules to the pipeline . a revision module Robin 1994 or aggregation Shaw 2002 . Since complex linguistic constructions often require interaction between modules adding them to the repertoire of pipelined NLG systems becomes an engineering and programming task. Integrated NLG systems have a simpler architecture because they do not need to model interactions between modules. However they still face the problem of computational complexity that was originally solved by the pipeline model. Strategies that have been introduced to reduce the search space in integrated