An Approach of Using the Web as a Live Corpus for Spoken Transliteration Name Access

An Approach of Using the Web as a Live Corpus for Spoken Transliteration Name Access Ming-Shun Lin, Chia-Ping Chen*, Hsin-Hsi Chen** Department of Computer Science and Information Engineering National Taiwan University, Taipei 106 TAIWAN {d91022 , **hhchen}@csie.ntu.edu.tw Department of Computer Science and Engineering National Sun Yat-Sen University 70, Lien-Hai Road, Kaohsiung, Taiwan 804 *cpchen@cse.nsysu.edu.tw Abstract Recognizing transliteration names is challenging due to their flexible formulation and coverage of a lexicon. This paper employs the Web as a huge-scale corpus. The patterns extracted from the Web are considered as a live dictionary to correct speech recognition errors. In our approach, the plausible character strings recognized by ASR (Automated Speech Recognition) are regarded as query terms and submitted to Google. The top n returned web page summaries are entered into PAT trees. The terms of the highest scores are selected. Total 100 Chinese transliteration names, including 50 person names and 50 location names, are used as test data. In the ideal case, we input the correct syllable sequences, convert them to text strings and test the recovery capability of using Web corpus. The results show that both the recall rate and the MRR (Mean Reciprocal Rank) are 0.94. That is, the correct answers appear in the top 1 position in 94 cases. When a complete transliteration name recognition system is evaluated, the experiments show that ASR model with a recovery mechanism can achieve 3.82% performance increases compared to ASR only model on character level. Besides, the recovery capability improves the average ranks of correct transliteration names from the 18th to the 3rd positions on word level. 1. Introduction Named entities [1], which denote persons, locations, organizations, etc., are common foci of searchers. Capturing named entities is challenging due to their flexible formulation and up-to-date use. The issues behind speech recognition make named entity recognition more challenging on spoken level than on written level. This paper emphasizes on a special kind of named entities, called transliteration names. They denote foreign people, places, etc. Spoken transliteration name recognition is useful for many applications. For example, cross language image retrieval via spoken query aims to employ spoken queries in one language to retrieve images with captions in another language [2]. In the past, Appelt and Martin [3] adapted TextPro system for processing text to processing transcripts generated by a speech recognizer. Miller et al [4] analyzed the effects of out-of-vocabulary errors and loss of punctuation in name finding of automatic speech recognition. Huang and Waibel [5] proposed an adaptive method of named entity extraction for meeting understanding. Chen [6] dealt with spoken cross-language access to image collection. The coverage of a lexicon is one of the major issues in spoken transliteration name access. Recently, researchers are interested in exploring the Web, which provides huge-collection of up-to-date data, as a corpus. Keller and Lapata [7] employed the Web to obtain frequencies for bigrams that are unseen in a given corpus. In this paper, we consider the Web as a live dictionary for recognizing spoken transliteration names, …

Content truncated to 500 words