Technical Report Number
Current natural language processing systems have a wide coverage of English, but are unforgiving of errors and generate numerous alternative but irrelevant analyses. These problems are exacerbated when speech is the input medium rather than text, because speech recognizers generate their own errors and ambiguities. These errors, together with the user's informality of speech, can lead to extremely large search spaces of possible interpretations. To add to the complexity, only very small subsets of English can be analyzed semantically. For our domain, which is very limited, our approach is to avoid detailed syntactic and semantic processing wherever possible. This is less error prone and increases the level of robustness. Rather than perform complete compositional analysis of utterances, we are using patternmatching techniques to get an interpretation of utterances by generating information from significant seeming pieces.
Balentine, Jennifer; Johnstone, Anne; and Mathias, David, "The DIM system: An Empirically Derived NLP System" Report Number: WUCS-92-21 (1992). All Computer Science and Engineering Research.