Morpholexical and Discriminative Language Models for Turkish Automatic Speech Recognition

Abstract

This paper introduces two complementary language modeling approaches for morphologically rich languages aiming to alleviate out-of-vocabulary (OOV) word problem and to exploit morphology as a knowledge source. The first model, morpholexical language model, is a generative -gram model, where modeling units are lexical-grammatical morphemes instead of commonly used words or statistical sub-words. This paper also proposes a novel approach for integrating the morphology into an automatic speech recognition (ASR) system in the finite-state transducer framework as a knowledge source. We accomplish that by building a morpholexical search network obtained by the composition of lexical transducer of a computational lexicon with a morpholexical language model. The second model is a linear reranking model trained discriminatively with a variant of the perceptron algorithm using morpholexical features. This variant of the perceptron algorithm, WER-sensitive perceptron, is shown to perform better for reranking -best candidates obtained with the generative model. We apply the proposed models in Turkish broadcast news transcription task and give experimental results. The morpholexical model leads to an elegant morphology-integrated search network with unlimited vocabulary. Thus, it is highly effective in alleviating OOV problem and improves the word error rate (WER) over word and statistical sub-word models by 1.8% and 0.4% absolute, respectively. The discriminatively trained morpholexical model further improves the WER of the system by 0.8% absolute.

Subword-based modeling for handling OOV words inkeyword spotting

“…This contrast is similar to some of the methods explored in [2] for open vocabulary recognition, and we confirm their finding that unsupervised morphology learning gives similar results compared to the rule-based system. Further, unlike this and other prior work which uses morphology in Turkish broadcast news transcription [18,4,19,2], our study involves keyword spotting in conversational Turkish with minimal training resources (10 vs. roughly 200 hours).…”

“…ATWV on dev and eval-part1 test sets with dense lattices. Note: (2) and (3), as in all above experiments, use the word system (1) as a first stage.…”

“…Graphones (coupled phonetic and orthographic sequences) [1] and morphs are particularly well suited to open vocabulary recognition. While some work has based the vocabulary entirely on morphs (see [2,3,4] and references therein), other studies obtain better results using a combination of morphs and words in Arabic [5] and German [6]. However, a mixed word and syllable vocabulary outperformed a mixed word and morph vocabulary for Polish [7].…”

“…A mixed word and graphone vocabulary has also been explored for English [8]. Morphs have the potential advantage of introducing more powerful constraints in language modeling, and several studies have investigated novel language model structures that take advantage of morphological features in a variety of languages [9,6,4,7,10,2,11,12]. While these studies motivate our use of morphs in this work, only standard n-grams are used here since our focus will be primarily on the keyword search strategies that take advantage of a mixed word and morph vocabulary.…”

“…Further, unlike this and other prior work which uses morphology in Turkish broadcast news transcription [18,4,19,2], our study involves keyword spotting in conversational Turkish with minimal training resources (10 vs. roughly 200 hours).…”