Robust Unsupervised Topic-Based Language Model Adaptation

本論文的主要貢獻在於提出一個基於主題分析的語言模型調適法，這個方法主要是使用潛藏狄式配置（Latent Dirichlet Allocation, LDA）。我們使用機率式潛藏語意分析（Probabilistic Latent Semantic Analysis, PLSA）自動地把一個具有不同性質的文字語料加以聚成許多個潛藏主題，然後用這些結果當作我們LDA模型的初始化模型。我們用最後的LDA模型一句一句地建造主題式的文字語料，這些主題式語料則用來估計主題式的語言模型。當我們用語言模型調適進行N-best重新評分時，我們把這些主題式的語言模型以內插法跟一個背景（也就是非主題式的）語言模型結合在一起。本論文共提出幾個機制，可以讓主題推論的結果更強健，比較不會被辨識錯誤扭曲，我們也用詮釋資料做片段分割，進行節目層的語言模型調適。最後在多來源的美國國防部GALE計劃中文資料上的結果顯示比其他最新的語言模型調適方法更有效。We present a novel topic mixture-based language model adaptation approach that uses Latent Dirichlet Allocation (LDA). We use Probabilistic Latent Semantic Analysis (PLSA) to automatically cluster a heterogeneous training corpus, and then train an LDA model using the resultant topic-document assignments. Using this LDA model, we construct fine-grained topic-specific corpora at the utterance level, which we use to train topic language models. These topic LMs are interpolated with a background language model during language model adaptation under an N-best rescoring framework. We describe several techniques for hardening LDA topic inference to first-pass recognition errors, and demonstrate the effectiveness of metadata-based segmentation when combined with show-level language model adaptation. Good improvements over state-of-the-art schemes were obtained in experiments on multi-genre GALE Project data in Mandarin Chinese.1 Introduction 1 1.1 An Automatic Speech Recognition System . . . . . . . . . . . . . . . . . 2 1.2 Training an ASR System . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.3 This Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 2 Language Model Adaptation 5 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.1.1 Automatic Speech Recognition . . . . . . . . . . . . . . . . . . . 5 2.1.2 n-gram Language Models . . . . . . . . . . . . . . . . . . . . . 6 2.1.3 Language Model Adaptation . . . . . . . . . . . . . . . . . . . . 9 2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.3 Adaptation Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.3.1 Topic Corpus Construction and LM Training . . . . . . . . . . . 11 2.3.2 Language Model Adaptation . . . . . . . . . . . . . . . . . . . . 12 2.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.4.1 Topic LM Mixture Size Experiments . . . . . . . . . . . . . . . 13 2.4.2 Interpolation Weight Experiments . . . . . . . . . . . . . . . . . 14 2.4.3 Speech Recognition Experiments . . . . . . . . . . . . . . . . . 15 2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 3 Robust Language Model Adaptation 19 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.2 LDA Topic Inference Algorithm . . . . . . . . . . . . . . . . . . . . . . 19 3.2.1 Custom Prior Mixtures . . . . . . . . . . . . . . . . . . . . . . . 19 3.2.2 Custom Word Weights . . . . . . . . . . . . . . . . . . . . . . . 20 3.3 Adaptation Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 3.3.1 Topic LM Training . . . . . . . . . . . . . . . . . . . . . . . . . 21 3.3.2 Segmentation for Topic Inference . . . . . . . . . . . . . . . . . 23 3.3.3 Language Model Adaptation . . . . . . . . . . . . . . . . . . . . 25 3.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 3.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 25 3.4.2 Training Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 3.4.3 Topic Inference: tmtw . . . . . . . . . . . . . . . . . . . . . . . 27 3.4.4 Adaptation Windows . . . . . . . . . . . . . . . . . . . . . . . . 27 3.4.5 Show Adaptation: Metadata timegap . . . . . . . . . . . . . . . . 33 3.4.6 Weighted N-Best Topic Inference . . . . . . . . . . . . . . . . . 33 3.4.7 Soft vs. Hard Topic Language Models . . . . . . . . . . . . . . . 33 3.4.8 Topic LMs and Topic Inference . . . . . . . . . . . . . . . . . . 36 3.4.9 Optimal lB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4 Conclusion 39 4.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 4.1.1 Topic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 4.1.2 Topic Language Model Training . . . . . . . . . . . . . . . . . . 40 4.1.3 Adaptation Framework . . . . . . . . . . . . . . . . . . . . . . . 40 4.1.4 Topical Context Windows: Content-Based Segmentation . . . . . 40 4.1.5 Interpolation Weights: Dynamic lB . . . . . . . . . . . . . . . . 40 4.1.6 Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 4.2 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 A Latent Dirichlet Allocation 43 A.1 LDA Topic Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4