An Investigation of Linguistic Information for Speech Recognition Error Detection

After several decades of effort, signiﬁcant progress has been made in the area of speech recognition technologies, and various speech-based applications have been developed. However, current speech recognition systems still generate erroneous output, which hinders the wide adoption of speech applications. Given that the goal of error-free output can not be realized in near future, mechanisms for automatically detecting and even correcting speech recognition errors may prove useful for amending imperfect speech recognition systems. This dissertation research focuses on the automatic detection of speech recognition errors for monologue applications, and in particular, dictation applications. <p> Due to computational complexity and efﬁciency concerns, limited linguistic information is embedded in speech recognition systems. Furthermore, when identifying speech recognition errors, humans always apply linguistic knowledge to complete the task. This dissertation therefore investigates the effect of linguistic information on automatic error detection by applying two levels of linguistic analysis, speciﬁcally syntactic analysis and semantic analysis, to the post processing of speech recognition output. Experiments are conducted on two dictation corpora which differ in both topic and style (daily ofﬁce communication by students and Wall Street Journal news by journalists). <p> To catch grammatical abnormalities possibly caused by speech recognition errors, two sets of syntactic features, linkage information and word associations based on syntactic dependency, are extracted for each word from the output of two lexicalized robust syntactic parsers respectively. Conﬁdence measures, which combine features using Support Vector Machines, are used to detect speech recognition errors. A conﬁdence measure that combines syntactic features with non-linguistic features yields consistent performance improvement in one or more aspects over those obtained by using non-linguistic features alone. <p> Semantic abnormalities possibly caused by speech recognition errors are caught by the analysis of semantic relatedness of a word to its context. Two different methods are used to integrate semantic analysis with syntactic analysis. One approach addresses the problem by extracting features for each word from its relations to other words. To this end, various WordNet-based measures and different context lengths are examined. The addition of semantic features in conﬁdence measures can further yield small but consistent improvement in error detection performance. The other approach applies lexical cohesion analysis by taking both reiteration and collocation relationships into consideration and by augmenting words with probability predicted from syntactic analysis. Two WordNet-based measures and one measure based on Latent Semantic Analysis are used to instantiate lexical cohesion relationships. Additionally, various word probability thresholds and cosine similarity thresholds are examined. The incorporation of lexical cohesion analysis is superior to the use of syntactic analysis alone. In summary, the use of linguistic information as described, including syntactic and semantic information, can provide positive impact on automatic detection of speech recognition errors.