TL;DR
This paper tackles the dual challenge of word segmentation and homograph disambiguation in Thai text-to-speech (TTS) synthesis by proposing a unified framework. They leverage the Winnow machine learning algorithm to automatically extract and combine local and long-distance contextual features, achieving state-of-the-art results that outperform previous approaches.
What Problem Does This Paper Address?
Accurate Text-to-Speech (TTS) synthesis relies heavily on correctly processing text. For languages like Thai, two significant hurdles exist:
- Word Boundary Ambiguity (WBA): Unlike languages with clear spaces between words, Thai often lacks explicit delimiters. This means a string of characters can be segmented into words in multiple ways, leading to different meanings and pronunciations. For instance, "ตากลม" can mean "round eye" ([ta:0 klom0]) or "expose wind" ([ta:k1 lom0]).
- Homograph Ambiguity: Even when words are correctly segmented, some words have the same spelling but different pronunciations and meanings, depending on the context. For example, "เพลา" can be pronounced as [phe:0 la:0] (time) or [phlaw0] (axe).
Traditional Thai TTS systems often address these issues separately, but this paper argues for a unified approach to improve overall pronunciation accuracy.
Methods
The authors propose a novel framework that jointly tackles both word boundary ambiguity and homograph ambiguity. They frame this as a pronunciation disambiguation task, where the goal is to select the correct pronunciation from a set of possibilities for a given string.
The core of their approach is the Winnow machine learning algorithm. Winnow is a type of neural network where "specialist" nodes focus on specific attribute-value pairs (features). These specialists vote for the correct pronunciation, and their weights are updated based on whether their predictions are correct. This allows Winnow to learn and combine evidence from various features effectively.
The features used are derived from the context surrounding the target word:
- Context Words: The presence of specific words within a window of +/- 10 words from the target.
- Collocations: Patterns of up to 2 contiguous words and/or their part-of-speech tags around the target word.
This approach allows the system to capture both local (nearby words) and long-distance (further away words) contextual information, which is crucial for disambiguation.
Datasets
The authors used a 25K-word Thai corpus for their experiments. This corpus was manually annotated with word boundaries, parts of speech, and correct pronunciations by linguists. The corpus was split into two parts: approximately 80% for training the Winnow model and the remaining 20% for testing.
Key Insights and Ideas
- Unified Framework: Treating word boundary ambiguity and homograph ambiguity as a single pronunciation disambiguation problem is more effective than separate approaches.
- Context is King: Both local and long-distance contextual features are essential for accurate disambiguation in Thai.
- Winnow's Adaptability: The Winnow algorithm proves adept at learning and combining diverse contextual features, outperforming traditional methods.
- Feature Engineering: The careful selection of context words and collocations as features is critical for Winnow's success.
Key Results
The paper presents a comparative evaluation of their Winnow-based approach against two other methods: a POS trigram model and a Bayesian hybrid approach. The results, summarized in Table 1, show significant improvements:
- CDSA (Context Dependent Segmentation Ambiguity): Winnow achieved 95.67% accuracy, significantly higher than Trigram (73.02%) and Bayesian hybrid (93.18%).
- CISA (Context Independent Segmentation Ambiguity): Winnow reached 99.70% accuracy, surpassing Trigram (98.25%) and Bayesian hybrid (99.67%).
- Homograph Ambiguity: Winnow achieved 96.45% accuracy, a substantial improvement over Trigram (52.46%) and Bayesian hybrid (94.25%).
These results demonstrate that the proposed Winnow-based system is highly effective in disambiguating both word boundaries and homographs in Thai.
Limitations and Future Work
The authors acknowledge that while Winnow performs well, they plan to investigate other machine learning techniques such as SNoW and SVM in the future. They also suggest that further exploration of feature extraction and combination strategies could lead to even greater accuracy. The current work focuses on Thai, and extending this framework to other languages with similar ambiguity challenges would be a valuable next step.
Summary updated