Lattice Parsing to Integrate Speech Recognition and Rule-Based Machine Translation

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 469–477, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Lattice Parsing to Integrate Speech Recognition and Rule-Based Machine Translation Selçuk Köprü AppTek, Inc. METU Technopolis Ankara, Turkey skopru@apptek.com Adnan Yazıcı Department of Computer Engineering Middle East Technical University Ankara, Turkey yazici@metu.edu.tr Abstract In this paper, we present a novel approach to integrate speech recognition and rule- based machine translation by lattice pars- ing. The presented approach is hybrid in two senses. First, it combines struc- tural and statistical methods for language modeling task. Second, it employs a chart parser which utilizes manually cre- ated syntax rules in addition to scores ob- tained after statistical processing during speech recognition. The employed chart parser is a unification-based active chart parser. It can parse word graphs by using a mixed strategy instead of being bottom-up or top-down only. The results are reported based on word error rate on the NIST HUB-1 word-lattices. The presented ap- proach is implemented and compared with other syntactic language modeling tech- niques. 1 Introduction The integration of speech and language technolo- gies plays an important role in speech to text translation. This paper describes a unification- based active chart parser and how it is utilized for language modeling in speech recognition or speech translation. The fundamental idea behind the proposed solution is to combine the strengths of unification-based chart parsing and statistical language modeling. In the solution, all sentence hypotheses, which are represented in word-lattice format at the end of automatic speech recognition (ASR), are parsed simultaneously. The chart is initialized with the lattice and it is processed un- til the first sentence hypothesis is selected by the parser. The parser also utilizes the scores assigned to words during the speech recognition process. This leads to a hybrid solution. An important benefit of this approach is that it allows one to make use of the available grammars and parsers for language modeling task. So as to be used for this task, syntactic analyzer compo- nents developed for a rule-based machine trans- lation (RBMT) system are modified. In speech translation (ST), this approach leads to a perfect integration of the ASR and RBMT components. Language modeling effort in ASR and syntac- tic analysis effort in RBMT are overlapped and merged into a single task. Its advantages are twofold. First, this allows us to avoid unnecessary duplication of similar jobs. Secondly, by using the available components, we avoid the difficulty of building a syntactic language model all from the beginning. There are two basic methods that are being used to integrate ASR and rule-based MT systems: First-best method and the N-best list method. Both techniques are motivated from a software engi- neering perspective. In the First-best approach (Figure 1.a), the ASR module sends a single rec- ognized text to the MT component to translate. Any ambiguity existing in the recognition process is resolved inside the ASR. In contrast to the First- best …

Content truncated to 500 words