Abstract
Extracting hierarchical "is-a" relationships is a cornerstone of natural language understanding. While well-established in English, Chinese hypernym-hyponym extraction faces unique challenges due to linguistic flexibility and a lack of diverse datasets. We introduce KEPL (Knowledge Enhanced Prompt Learning), a framework that integrates linguistic Hearst patterns with dynamic prompt-based learning. By utilizing a Dynamic Adaptor to weight linguistic cues, KEPL achieves state-of-the-art performance on a newly introduced multi-domain Chinese dataset.
Introduction
In NLP, a hypernym represents a general category (e.g., "fruit"), while a hyponym is a specific instance (e.g., "apple"). Identifying these pairs is essential for taxonomy construction, search optimization, and question-answering systems. In the Chinese context, researchers face two primary hurdles: the rigidity of traditional rule-based methods and the "data desert"—a lack of high-quality, multi-domain benchmarks. Modern deep learning pipelines often decouple entity extraction from relation recognition, losing semantic nuance.
The KEPL Framework
The KEPL framework addresses these issues by treating linguistic patterns as "hints" rather than strict rules. It consists of three main pillars:
- Prompt Learning with Hearst Patterns: The model uses natural language templates (e.g., "对象是一种分类" / "The object is a type of classification") to guide the extraction process.
- Dynamic Knowledge Adaptor: Recognizing that not all patterns fit every context, KEPL employs a Dynamic Adaptor. This component calculates a scoring matrix to select and weight the most relevant prompt patterns based on the input text's specific semantics.
- Unified Representation and Focal Loss: An attention mechanism blends prompt embeddings with unstructured text. To handle the imbalance where most words in a sentence are not part of a target pair, the model utilizes focal loss for precise span selection.
The CHR Dataset
To evaluate KEPL, the researchers introduced the Chinese Hypernym Recognition (CHR) Dataset. This benchmark contains 21,376 training contexts and over 12,000 hypernym pairs across three distinct domains:
- Baike: Structured encyclopedic data from sources like Baidu Encyclopedia.
- News: Formal journalistic reports from Xinhua and Tencent News.
- We-media: Informal social media content from public accounts.
Experimental Results
KEPL was compared against several state-of-the-art baselines, including UIE (Universal Information Extraction) and Ernie3. The KEPL-Ernie3 variant achieved an overall F1 score of 0.8701, a 2.3% improvement over UIE.
As shown in the results, the model excels in formal contexts like Baike (0.9334 F1) but faces greater difficulty with the informal language of We-media (0.7295 F1). Ablation studies highlighted the Dynamic Adaptor as the most critical component; replacing it with a random matrix caused a 6.73% drop in F1 score, proving that context-aware pattern selection is vital.
Conclusion
KEPL demonstrates that integrating traditional linguistic knowledge with modern prompt learning significantly enhances hypernym-hyponym extraction in Chinese. Future work will focus on improving performance in informal social media contexts and adapting the model for specialized domains like medicine or law.
Summary updated