運用基於深度學習的標注系統 針對一對一與一對多的實體關係萃取方法

Abstract

實體關係萃取(Entity Relation Extraction, ERE)是自然語言處理(Natural Language Processing, NLP)領域中一項重要的任務，而其可以被拆分為實體萃取 (Entity Extraction, EE 及關係萃取(Relation Extraction, RE)兩個子任務，旨在抓取 句子中所有可以組成關係的實體。近年來，諸多實體關係萃取研究僅專注於一對 一的關係萃取，而忽略了一對多關係的實體關係萃取問題。 在本研究中，我們將透過聯合模型(joint model)將 ERE 視作實體標記任務， 藉以同時萃取出需要的實體以及關係。我們也使用了消融(ablation)實驗檢驗我們 在模型中所使用的組件是否提升我們所提出的模型之效能。實驗結果顯示，我們 提出的模型在一對一以及一對一加一對多的資料集中比基準模型還要好。最後我 們展示了幾個經過模型所產生的結果，藉此展示我們的模型的確比基準模型還要 有效。

Entity relation extraction is an important task in Natural Language Processing (NLP), which can be divided into Entity Extraction (EE) and Relation Extraction (RE), aiming to capture all types of relations triplets from unstructured texts. Most of prior studies on entity relation extraction concentrate on extracting one-to-one relations only; however, there do exist one-to-many relations in real world applications. In this research, we are going to treat entity-relation extraction as a sequence labeling problem using an end-to-end joint model to extract entities along with relations simultaneously. To see whether the components in our proposed model can improve the extraction effectiveness, we conduct ablation experiments in our evaluation. As the experiment result shows, our proposed model performs better than the state-of-the-art model on one-to-one and one-to-one + one-to-many datasets. Finally, we show some cases that illustrate the advantages of our model comparing to the benchmark model.