一种面向事件事实性预测的平行语料库构建方法研究

Parallel Corpus Annotated for Event Factuality Prediction

事件事实性预测(Event Factuality Prediction,EFP)是将事实性评价(Factuality Assessment)问题建模为句子级别回归任务,判定句子中事件提及(event mention)的事实程度.EFP是自然语言处理中重要且具有挑战性的任务.与英文事件事实性语料库的资源丰富不同,目前中文领域事件事实性语料库十分缺乏,这明显阻碍了对中文事实性评价问题的进一步研究.针对此问题,本文探索并提出了基于机器翻译的半自动事件事实性平行语料库构建方法.实验结果表明,利用本文构建的事件事实性中英平行语料库Parallel FactBank配合DLEF语料库进行多任务学习可以有效提升中文EFP任务中模型的泛化能力,并使模型在各数据集上性能优于单任务学习模型.

Event factuality prediction(EFP)models the problem of factuality assessment as a sentence level regression task and determines the degree to which the event mention in a sentence corresponds to a fact in the world.EFP is an important and challenging task in natural language processing.Different from the rich resources of English event factuality corpus,there is a lack of event factuality corpus in Chinese field at present,which obviously hinders the further study of Chinese factuality assessment.To solve this problem,this paper explores and proposes a semi-automatic event factuality parallel corpus construction method based on machine translation.The experimental results show that using the event factuality Chinese-English parallel corpus Parallel FactBank constructed in this paper and the DLEF corpus for multi-task learning can effectively improve the generalization ability of the Chinese EFP task model,and make the performance of the model on each dataset better than that of the single-task learning model.