Edinburgh Linguistic Circle seminar

Marcus Tomalin gave a seminar, In Search of `Natural' Speech: Grammaticality, Acceptability, and Speech Technology, at the Edinburgh University Linguistic Circle on 2 March 2012.

Abstract Although state-of-the-art large vocabulary Automatic Speech Recognition (ASR) and Statistical Machine Translation (SMT) systems often achieve impressive Word Error Rates (WERs) and BLEU scores respectively, end-users frequently consider the word sequences output by such systems to be `unnatural'. The perceived `unnaturalness' usually results from the accumulation of many small linguistic errors (e.g., lack of subject-verb agreement, partially scrambled syntax, homophonic substitution). Consequently, in recent years there has been a renewed interest in improving the `naturalness' of ASR and SMT output, even in systems that produce good WER and BLEU scores.

In this talk, the perceived `naturalness' of ASR and SMT transcriptions will be considered in the context of on-going debates about grammaticality and acceptability. An experimental framework for exploring these aspects of ASR/SMT transcriptions is described, and a methodology for improving the `naturalness' of such outputs is presented. The simplest ways of modifying an input word sequence are insertion, permutation, deletion, and substitution, and the approach adopted in this work makes use of a Combinatory Categorial Grammar (CCG) text generation system which enables input word sequences to be modified so as to improve their `naturalness'. It is shown that the output produced by the CCG-based system is considerably improved if the N-best generated hypotheses are rescored and reranked using Ngram-based techniques.