Genchi

Inspiration

The inspiration for this project came from the structural complexity of the Chinese language. Unlike English, Chinese relies heavily on specific grammatical markers (like 的, 了, or 在) to define relationships between words. I wanted to create a generative model that doesn't just predict the next likely word but truly understands the syntactic skeleton of a sentence, making the transition from a simple prompt like "I go..." (我北京...) to a grammatically perfect completion.

What it does

This is a context-aware Chinese text completion engine. It takes a short initial input and autonomously generates a coherent continuation.

Challenges we ran into

the project involves a custom training logic to improve grammatical coherence in Chinese. I implemented a mechanism that forces the model to prioritize essential structural particles like 的 (de) and 了 (le). Since these are critical for sentence structure, focusing on them prevents the model from generating 'garbled' or unnatural text.

Built With

• accelerator
• gpt-2
• pytorch