Abstract:

Entity and Relation Extraction （ERE） is typically handled in a pipeline manner. However， such an approach relies only on the output of the preceding task， resulting in limited information interaction between named entity recognition and relation extraction， and is susceptible to error propagation. To address these challenges， a Memory-Enhanced model for Entity and Relation Extraction （MEERE） was proposed. This model introduced a memory-like mechanism， allowing each task not only to utilize the output of the preceding task， but also to influence it in reverse， thereby capturing complex interactions between entities and relations. To further mitigate error propagation， an entity span screening mechanism was incorporated. This mechanism dynamically screened and verified entity spans in the joint module， ensuring that only high-quality entities were used for relation extraction， thus enhancing the robustness and accuracy of the model. A table decoding method was finally employed to handle relation overlap. Experimental results on three widely used benchmark datasets （ACE05， SciERC， and CoNLL04） demonstrated significant advantages of MEERE in ERE tasks. In specific， on the CoNLL04 dataset， MEERE outperformed Tab-Seq model in both named entity recognition and relation extraction tasks with a 0.5 percentage point increase in named entity recognition F1-score and a 3.0 percentage point improvement in relation strict evaluation F1-score； compared to PURE-F model， MEERE achieved no less than a ninefold acceleration effect with improved relation extraction performance. These findings confirm the effectiveness of the proposed memory enhancement model in exploring interactions between entities and relations.

Key words: joint extraction of entities and relations, memory enhancement, span screening, Pre-Trained Language Model (PLM), cross-sentence context

摘要：

实体和关系抽取（ERE）通常采用流水线的方式进行处理，但这种流水线方法仅依赖于前一个任务的输出，导致命名实体识别和关系抽取之间出现信息交互问题，且容易引发误差传播问题。针对以上问题，提出一种面向实体和关系抽取的记忆增强模型（MEERE）。该模型引入类似记忆的机制，使每个任务不仅能利用前一任务的输出，还能反向影响前一任务，从而捕获实体和关系间的复杂交互。为进一步减轻误差传播，同时引入实体跨度筛选机制。该机制通过在联合模块中动态地筛选和验证实体跨度，确保只有高质量的实体被用于关系抽取，从而提升模型的鲁棒性和准确性。最后利用表格解码方式处理关系重叠问题。在3个广泛使用的基准数据集（ACE05、SciERC和CoNLL04）上的实验结果表明，MEERE在ERE任务上表现出了显著的优势。与Tab-Seq在CoNLL04数据集上相比，MEERE在命名实体识别和关系抽取上的性能都有显著提升，命名实体识别的F1值提升了0.5个百分点，关系严格评估的F1值提升了3.0个百分点；相较于PURE-F模型，MEERE实现了不少于9倍的加速效果，并且关系抽取性能更佳。这些结果验证了所提出的记忆增强模型在探索实体和关系交互作用方面的有效性。

关键词: 实体和关系联合抽取, 记忆增强, 跨度筛选, 预训练语言模型, 跨句子上下文