Certain approaches towards robust parsing take as a starting point the `inactive items' that were found by the parser. Such an approach requires that the parser discovers all categories anywhere in the input in a bottom-up fashion. It is therefore natural to consider a bottom-up parser in such a context, since otherwise the set of inactive items the parser constructs in case of failure will not be complete.
Head-corner parsing is a mixture of bottom-up and top-down parsing. Therefore it cannot be guaranteed that the parser finds all possible categories. In this paper I will argue that the use of underspecification can be exploited to obtain a version of the head-corner parser which is guaranteed to discover all relevant categories anywhere in the input. The grammar explicitly defines what categories are relevant in this sense. It is shown how this property is used in the design of the robustness component of the OVIS system.
An important issue in the context of robustness concerns evaluation. In the context of the OVIS system a new and simple evaluation criterion is available. In OVIS the input for the parser consists of the output of the speech recognizer. Given that the speech recognizer often is uncertain about what is actually said, it produces a word-graph of possible sentences [4]. This makes it possible to measure the success of the parser by comparing the path in the word-graph on which the parser bases its best analysis with the actual utterance. Different comparison operators will be discussed.
This paper is organised as follows. I briefly introduce the head-corner parser and some of the techniques to make the parser more efficient in the following section. A more detailed discussion can be found in [8]. Section 3 discusses the use of underspecification to make sure that all relevant categories anywhere in the input are found. In the fourth section I discuss the internal evaluation method that is being used. Section 5 presents some preliminary results.