A Noisy Channel Approach to Error Correction in Spoken Referring Expressions

International Joint Conference on Natural Language Processing, pages 234–242, Nagoya, Japan, 14-18 October 2013. A Noisy Channel Approach to Error Correction in Spoken Referring Expressions Su Nam Kim, Ingrid Zukerman, Thomas Kleinbauer and Farshid Zavareh Faculty of Information Technology, Monash University Clayton, Victoria 3800, Australia Abstract We offer a noisy channel approach for rec- ognizing and correcting erroneous words in referring expressions. Our mechanism handles three types of errors: it removes noisy input, inserts missing prepositions, and replaces mis-heard words (at present, they are replaced by generic words). Our mechanism was evaluated on a corpus of 295 spoken referring expressions, improv- ing interpretation performance. 1 Introduction One of the main stumbling blocks for Spoken Di- alogue Systems (SDSs) is the lack of reliability of Automatic Speech Recognizers (ASRs) (Pelle- grini and Trancoso, 2010). Recent research proto- types of ASRs yield Word Error Rates (WERs) be- tween 15.6% (Pellegrini and Trancoso, 2010) and 18.7% (Sainath et al., 2011) for broadcast news. However, the commercial ASR employed in this research had a WER of 30% and a Sentence Er- ror Rate (SER) (proportion of sentences for which no correct textual transcription was produced) of 65.3% for descriptions of household objects. In addition to mis-recognized entities or actions, ASR errors often yield ungrammatical sentences that cannot be processed by subsequent interpreta- tion modules of an SDS, e.g., “the blue plate” be- ing mis-heard as “to build played”, and hesitations (e.g., “ah”s) being mis-heard as “and” or “on” — all of which happened in our trials. In this paper, we offer a general framework for error detection and correction in spoken utterances that is based on the noisy channel model, and present a first-stage implementation of this frame- work that performs simple corrections of referring expressions. Our model is implemented as a pre- processing step for the Scusi? spoken language in- terpretation system (Zukerman et al., 2008; Zuk- erman et al., 2009). Table 1: Spoken, heard and labeled descriptions. Spoken: the stool to the left of the table Heard: the storm the left of the table Labels: Object Prep Specifier Landmark Spoken: the plate in the microwave Heard: to play it in the microwave Labels: Object Noise Prep Landmark The idea of the noisy channel model is that a message is sent through a channel that intro- duces errors, and the receiver endeavours to re- construct the original message by taking into ac- count the characteristics of the noisy channel and of the transmitted information (Ringger and Allen, 1996; Brill and Moore, 2000; Zwarts et al., 2010). The system described in this paper handles three types of errors: noise (which is removed), missing prepositions (which are inserted), and mis-heard words (which are replaced). Table 1 shows two descriptions that illustrate these errors. The first row for each description displays what was spo- ken, the second row displays what was heard by the ASR, and the third row shows the semantic la- bels assigned to each segment in the description by a shallow semantic parser …

Content truncated to 500 words