KEML: A Knowledge-Enriched Meta-Learning Framework for Lexical Relation Classification

Chengyu Wang, Minghui Qiu, Jun Huang, Xiaofeng He

Introduction

Lexical Relation Classification (LRC) is a fundamental task in Natural Language Processing (NLP) that involves identifying the semantic relationship between a pair of words. Common relations include synonymy, antonymy, hypernymy (is-a), and meronymy (part-of). Despite its importance for tasks like taxonomy construction and question answering, LRC remains challenging due to three primary factors:

  1. Pattern Sparsity: Explicit mentions of lexical relations are rare in natural text (e.g., "a wheel is a part of a car" is seldom written).
  2. Distributional Limitations: Context-based word embeddings often fail to distinguish between synonyms and antonyms as they appear in similar linguistic environments.
  3. Data Imbalance: Real-world datasets are often dominated by "random" word pairs, making it difficult for models to learn the nuances of structured relations.

To address these challenges, we introduce KEML (Knowledge-Enriched Meta-Learning), a framework that integrates external lexical knowledge with meta-learning to improve relation classification.

Methodology

The KEML framework consists of three primary components designed to leverage structured knowledge and robust learning strategies.

The KEML framework architecture, illustrating the Knowledge Encoder (LKB-BERT), Auxiliary Task Generator, and Relation Learner.

1. Knowledge Encoder (LKB-BERT)

Standard BERT models are fine-tuned into LKB-BERT using distant supervision from Lexical Knowledge Bases (LKBs) like WordNet. This process enriches the model's representations with explicit relational information, moving beyond simple distributional context.

2. Auxiliary Task Generator

To combat data imbalance, KEML employs a meta-learning approach. The recognition of each specific relation is treated as an auxiliary task. These tasks are sampled according to a probabilistic distribution, ensuring the model effectively learns rare relations while remaining robust against the noise of random word pairs.

3. Relation Learner (SRR Cells)

The final classification layer utilizes Single Relation Recognition (SRR) Cells. These cells compare the vector difference of a word pair against a "relation prototype"—a learned average representation of a specific relation derived from a knowledge base. This architecture allows for a "divide-and-conquer" approach, where the model evaluates a pair against multiple potential relations simultaneously.

Experimental Results

KEML was evaluated on several benchmark datasets, including K&H+N, BLESS, ROOT09, EVALution, and the challenging CogALex-V.

Accuracy trends across iterations for the K&H+N, BLESS, and ROOT09 datasets, showing consistent performance gains for various relation types.

The framework achieved state-of-the-art (SOTA) performance across all benchmarks. Notably, on the CogALex-V dataset—designed to eliminate "lexical memorization" shortcuts—KEML reached an F1-score of 50.0%, surpassing the previous best of 47.1%. On EVALution, it achieved a 66.0% F1-score, a 4% improvement over prior models like SphereRE.

Ablation studies confirmed that both the knowledge enrichment (LKB-BERT) and the meta-learning stage are critical. Removing either component led to a significant degradation in classification accuracy.

Conclusion

KEML demonstrates that combining deep contextual models with structured external knowledge and meta-learning provides a powerful solution for Lexical Relation Classification. By treating relation recognition as a set of meta-tasks and using knowledge-guided prototypes, the model overcomes traditional hurdles of data sparsity and imbalance. Future work will explore integrating broader commonsense knowledge and applying KEML to automated taxonomy generation.

Summary updated

Abstract

Lexical relations describe how concepts are semantically related, in the form of relation triples. The accurate prediction of lexical relations between concepts is challenging, due to the sparsity of patterns indicating the existence of such relations. We propose the Knowledge-Enriched Meta-Learning (KEML) framework to address lexical relation classification. In KEML, the LKB-BERT (Lexical Knowledge Base-BERT) model is first presented to learn concept representations from text corpora, with rich lexical knowledge injected by distant supervision. A probabilistic distribution of auxiliary tasks is defined to increase the model's ability to recognize different types of lexical relations. We further propose a neural classifier integrated with special relation recognition cells, in order to combine meta-learning over the auxiliary task distribution and supervised learning for LRC. Experiments over multiple datasets show KEML outperforms state-of-the-art methods.
BibTeX
@inproceedings{wang2021keml,
  author = {Chengyu Wang and Minghui Qiu and Jun Huang and Xiaofeng He},
  title = {{KEML: A Knowledge-Enriched Meta-Learning Framework for Lexical Relation Classification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i15.17640},
}
Figure