Charting the Depths of Robust Speech Parsing

Charting the Depths of Robust Speech Parsing W. Kasper t, B. Kiefer t, H.-U. Kriegert, C. J. Rupp$, and K. L. Worm $ tGerman Research Center for Artificial Intelligence (DFKI) $Computational Linguistics Department, Universit~t des Saarlandes {kasper, kiefer, krieger}@dfki, de and {c j, worm}@coli, uni-sb, de Abstract We describe a novel method for coping with un- grammatical input based on the use of chart-like data structures, which permit anytime process- ing. Priority is given to deep syntactic anal- ysis. Should this fail, the best partial analy- ses are selected, according to a shortest-paths algorithm, and assembled in a robust process- ing phase. The method has been applied in a speech translation project with large HPSG grammars. 1 Introduction This paper describes a new method of deal- ing robustly with deficient speech or text in- put, which may be due to recognition errors, spontaneous speech phenomena, or ungrammat- ical constructions. Two key features of this ap- proach are: • the priority given to a deep and restrictive grammatical analysis, • and the use of chart-like data structures at each level of processing. The initial input is taken from a Word Hy- pothesis Graph, or WHG, (Oerder and Ney, 1993) from which the best ranked paths are successively selected until a result is found or a time limit proportional to the length of the utterance 1 is reached. Each path is parsed with an incremental chart parser that uses a Head- Driven Phrase Structure grammar (HPSG). The parser is adapted to input from WHGs and optimized to meet the needs of real-time speech processing. Since the goal of the parsing compo- nent is to process as many WHG paths as pos- sible, in order to find a grammatical utterance 1This is currently up to four times real time. and analyze it with highest accuracy, neither relaxation of the constraints imposed by the grammar nor repair rules are used at this stage. If the analysis of the current path is successful, the parsing process is complete. However, in most cases there is no spanning and syntacti- cally correct analysis. So a sequence of partial analyses is determined by incrementally eval- uating the passive edges in the parser's chart. These analyzed fragments are passed on to a robust semantic processing component for fur- ther treatment, while the next best WHG path is analyzed by the parser 2. Robust semantic processing similarly builds up a chart-like data structure including analyzed fragments and the results of applying robustness rules at the se- mantic level. After the first path of the WHG has been (unsuccessfully) analyzed, processing in both the restrictive parser and the robust- ness component proceeds in parallel, with the aid of a parallel virtual machine, until one of the following conditions is fulfilled: 1. a spanning grammatical analysis is found, 2. all the WHG paths have been explored, or 3. the time limit is reached. In the case of either of the latter two condi- tions, robust semantic processing is allowed a limited time …

Content truncated to 500 words