A psycholinguistically plausible alternative to the standard pipeline in Natural Language Processing

Monday December 16, 2013 at 12:15 – 13:00 (HF – bygget: 301)

Research in Natural Language Processing has assumed that high level tasks such as Natural
Language Understanding can be solved by special purpose processes operating in a pipeline,
called the standard pipeline. Work is then done on building the individual components. However,
there appears to be no way of guaranteeing that such components will actually solve high
level NLP tasks when combined as suggested by the standard pipeline.

In my alternative model,  syntax- and meaning representations are incrementally constructed
on a word-by-word basis. Processing is serial in that the model commits to a unique syntactic
and semantic increment for each word. Incrementally constructed structures are stored in states
that both result from- and influence processing. Meaning representations do not influence further
syntactic increments, although meaning representations can cause the process to throw an
error. I propose a reanalysis algorithm by which the process can recover from such errors.

Magnus Bakken, Master student at LLE, presents his MA thesis.