Improve word embedding using both writing and pronunciation.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Speech Technology

Background:

• Traditional word embedding models primarily rely on written text, neglecting the semantic richness of spoken language.
• Spoken language is considered a more direct expression of meaning, with writing serving as its recorded form.
• Optimizing word embedding training often involves context, weight, dependency, and morphology, but lacks phonetic integration.

Purpose of the Study:

• To propose and evaluate a novel pronunciation-enhanced word embedding (PWE) model.
• To integrate speech information and pronunciation characteristics into the word embedding process.
• To leverage both spoken and written language modalities for improved semantic representation.

Main Methods:

• Developed several PWE models integrating word pronunciation features.
• Utilized Chinese, English, and Spanish languages for model development and testing.
• Compared PWE performance against baseline models without speech information.

Main Results:

• The PWE model demonstrated superior performance in word similarity tasks.
• PWE achieved better results in text classification experiments compared to baseline models.
• The integration of speech information significantly enhanced word embedding quality.

Conclusions:

• The proposed PWE model effectively utilizes both speech and writing for richer semantic representation.
• PWE offers a significant improvement over traditional word embedding techniques.
• The PWE approach is applicable to a wide range of languages, given language's nature as sound-images.