Analysis of Morph-Based Speech Recognition and the Modeling of Out-of-Vocabulary Words Across Languages

Proceedings of NAACL HLT 2007, pages 380–387, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Analysis of Morph-Based Speech Recognition and the Modeling of Out-of-Vocabulary Words Across Languages Mathias Creutz∗, Teemu Hirsim¨aki∗, Mikko Kurimo∗, Antti Puurula∗, Janne Pylkk¨onen∗, Vesa Siivola∗, Matti Varjokallio∗, Ebru Arısoy†, Murat Sarac¸lar†, and Andreas Stolcke‡ ∗ Helsinki University of Technology, <firstname>.<lastname>@tkk.fi, † Boˇgazic¸i University, arisoyeb@boun.edu.tr, murat.saraclar@boun.edu.tr, ‡ SRI International / International Computer Science Institute, stolcke@speech.sri.com Abstract We analyze subword-based language models (LMs) in large-vocabulary continuous speech recognition across four “morphologically rich” languages: Finnish, Estonian, Turkish, and Egyptian Colloquial Arabic. By estimating n-gram LMs over sequences of morphs instead of words, better vocabulary coverage and reduced data sparsity is obtained. Standard word LMs suffer from high out-of-vocabulary (OOV) rates, whereas the morph LMs can recognize previously unseen word forms by concatenating morphs. We show that the morph LMs generally outperform the word LMs and that they perform fairly well on OOVs without compromising the accuracy obtained for in-vocabulary words. 1 Introduction As automatic speech recognition systems are being developed for an increasing number of languages, there is growing interest in language modeling ap- proaches that are suitable for so-called “morpholog- ically rich” languages. In these languages, the num- ber of possible word forms is very large because of many productive morphological processes; words are formed through extensive use of, e.g., inflection, derivation and compounding (such as the English words ‘rooms’, ‘roomy’, ‘bedroom’, which all stem from the noun ‘room’). For some languages, language modeling based on surface forms of words has proven successful, or at least satisfactory. The most studied language, En- glish, is not characterized by a multitude of word forms. Thus, the recognition vocabulary can sim- ply consist of a list of words observed in the training text, and n-gram language models (LMs) are esti- mated over word sequences. The applicability of the word-based approach to morphologically richer lan- guages has been questioned. In highly compounding languages, such as the Germanic languages German, Dutch and Swedish, decomposition of compound words can be carried out to reduce the vocabulary size. Highly inflecting languages are found, e.g., among the Slavic, Romance, Turkic, and Semitic language families. LMs incorporating morphologi- cal knowledge about these languages can be applied. A further challenging category comprises languages that are both highly inflecting and compounding, such as the Finno-Ugric languages Finnish and Es- tonian. Morphology modeling aims to reduce the out- of-vocabulary (OOV) rate as well as data sparsity, thereby producing more effective language mod- els. However, obtaining considerable improvements in speech recognition accuracy seems hard, as is demonstrated by the fairly meager improvements (1–4 % relative) over standard word-based models accomplished by, e.g., Berton et al. (1996), Ordel- man et al. (2003), Kirchhoff et al. (2006), Whit- taker and Woodland (2000), Kwon and Park (2003), and Shafran and Hall (2006) for Dutch, Arabic, En- glish, Korean, and Czech, or even the worse perfor- mance reported by Larson et al. (2000) for German and Byrne et al. (2001) for Czech. Nevertheless, …

Content truncated to 500 words