Fusion Makes Perfection: An Efficient Multi-Grained Matching Approach for Zero-Shot Relation Extraction

TL;DR

Zero-shot relation extraction aims to identify unseen relations from text using only seen relations during training. EMMA blends a coarse-grained recall stage with a fine-grained classification stage and introduces virtual entity matching to avoid manual annotation, trained with a contrastive objective to align $x^{vec}$ with $d^{vec}$. On FewRel and Wiki-ZSL, EMMA achieves state-of-the-art $F_1$ on unseen relations, with sizable gains at larger unseen sets (e.g., $m=15$). The approach reduces labeling effort and maintains fast inference, offering a practical solution for scalable zero-shot relation extraction.

Abstract

Predicting unseen relations that cannot be observed during the training phase is a challenging task in relation extraction. Previous works have made progress by matching the semantics between input instances and label descriptions. However, fine-grained matching often requires laborious manual annotation, and rich interactions between instances and label descriptions come with significant computational overhead. In this work, we propose an efficient multi-grained matching approach that uses virtual entity matching to reduce manual annotation cost, and fuses coarse-grained recall and fine-grained classification for rich interactions with guaranteed inference speed. Experimental results show that our approach outperforms the previous State Of The Art (SOTA) methods, and achieves a balance between inference efficiency and prediction accuracy in zero-shot relation extraction tasks. Our code is available at https://github.com/longls777/EMMA.

Figures (5)

• Figure 1: The overall process of our method. The coarse-grained recall refers to the rough and rapid screening of several possible results, while the fine-grained classification denotes the detailed discrimination of these possible results.
• Figure 2: The overall architecture of EMMA. (a) The recall model swiftly matches to retrieve the top $k$ most probable relations. (b) The classification model further distinguishes among these similar relations.
• Figure 3: Comparison in terms of runtime(Bars) and matching F1 (Dotted lines).
• Figure 4: This is an example showcasing the role of the classification model. During the recall stage, the correct relation description wasn't ranked first, yet through the fine-grained classification model's correction, the accurate result was eventually obtained.
• Figure 5: The F1 scores of EMMA across different values of $k$.