Guji-TAI: A Task-Aware Interpretability Framework for Large Language Models in Ancient Chinese Text Processing

Guji-TAI is an interpretability analysis framework for large language models specifically designed for ancient Chinese text processing. Its core innovation lies in the introduction of a task-aware mechanism, enabling systematic interpretive analysis of heterogeneous tasks and helping researchers understand the decision-making process of models in ancient Chinese text processing. The project is positioned to combine in-depth exploration of a vertical domain (ancient Chinese texts) with methodological innovation (task-specific interpretation strategies).

With the outstanding performance of LLMs in NLP tasks, ancient Chinese text digitization processing has ushered in new opportunities. However, the differences between ancient and modern Chinese pose challenges to general-purpose models. The field of humanities research has high requirements for model interpretability (needing to know "why" rather than just "what"), while traditional interpretation methods take a "one-size-fits-all" approach and ignore the differences in interpretation needs across different tasks (e.g., the needs for named entity recognition vs. text classification are different).

Guji-TAI (Task-aware Interpretability for Guji) is an LLM interpretability tool tailored for ancient Chinese text scenarios, with a task-aware mechanism at its core. The technical architecture is built around heterogeneous tasks, supporting ancient Chinese text tasks such as character recognition and proofreading, named entity recognition, relation extraction, text classification, automatic punctuation and sentence segmentation, translation and paraphrasing, etc. Through the task-aware layer, it identifies task types and invokes corresponding interpretation strategies, enabling multi-task support under a unified architecture.

Guji-TAI's interpretation mechanism covers three levels:

1. Local Interpretation Layer: Through attention weight visualization, gradient attribution, etc., it shows the model's response to individual input features (such as Chinese characters/phrases);
2. Global Interpretation Layer: Aggregates sample interpretation results to identify the model's systematic preferences or biases in processing certain types of ancient Chinese texts;
3. Comparative Interpretation Layer: Supports cross-task and cross-model comparisons to analyze differences in focus between different tasks/models.

Guji-TAI's application scenarios include:

• Ancient Chinese Text Collators: Verify the reliability of automatic proofreading systems and assist manual review through interpretations;
• Linguistics Researchers: Observe the model's mechanism for understanding ancient Chinese and explore syntactic rules;
• Model Developers: Diagnose weak points of the model (e.g., poor handling of documents from specific dynasties) and optimize them targetedly.

Methodologically, the concept of task-aware interpretability can be extended to multi-task LLM application scenarios. Its academic value lies in promoting interdisciplinary research between computer science and the humanities, lowering the threshold for humanities scholars to use AI. Future directions include: enriching interpretation methods (integrating conceptual interpretation, causal inference), expanding multi-modal support (images/structured data), and optimizing user interaction (visual interfaces for non-technical users).

Guji-TAI is a beneficial attempt to apply cutting-edge AI to traditional cultural research. While pursuing performance, it pays attention to the model's "black box" mechanism, reflecting respect for cultural heritage. We look forward to the project's continuous development and community contributions, providing more robust and user-friendly interpretability tools for digital humanities.