Named Entity Recognition of Spoken Documents Using Subword Units

Abstract

The output of a speech recognition system is a stream of text features that is overlayed by noise resulting from errors in the system's statistical classification of the audio input. Conditional Random Fields (CRFs), which have already proven themselves to be efficient, high-performance Named Entity Recognizers (NERs) for named entities from text, offer the promise to compensate part of these errors. In this paper we use CRFs to extract named entities from spoken audio documents. We consider a real-world audio information extraction scenario under which CRFs are trained to recognize named entities in unedited radio audio documents that have been converted into a stream of text features by a speech recognition system. The automatic speech recognition system (ASR) is able to produce word transcriptions as well as syllables. It uses general speaker-independent acoustic models and a domain-independent statistical language model, insuring that recognizer performance is not specific to the experimental domain. Using an additional syllable model increases the generality of the spoken document classification system, giving it the flexibility to handle words that are not present in the vocabulary. In this paper we apply for the first time CRFs to different features produced by German ASR. The experiments confirm that using transcribed syllables together with words can compensate for part of the NER errors caused by ASR transcription.

Factoid question answering for spoken documents

“…We report the standard NERC measures: precision, recall and the F β=1 harmonic mean of both. When adding the previously described phonetic features for the ASR transcripts, the overall F β=1 score improves significantly but no more than 2 points in data-sets asrA, asrB and asrC (Paaß et al [2009] reports an improvement of 1% for German broadcast news ASR transcripts when using syllables).…”

“…A joint ASR--NERC model can be applied to the graph, determining both where the named entities are and how they are transcripted, as shown in Favre et al [2005]. Classical text-based approaches to NERC usually enrich their recognition models with speech related information (e.g., representing named entities as syllables instead of words [Paaß et al, 2009]) or adapting rule-based NERCs to the new corpora [Brun and Ehrmann, 2009].…”