摘要
老挝语属于低资源语言,在有限的语料中获取更多的语义信息可以有效解决汉语和老挝语短文本相似度计算不准确的问题。多任务学习是有效获取语义信息的一种方法,该文对汉语和老挝语短文本特点进行研究后,提出一种融合词性位置特征的多任务汉老双语短文本相似度计算方法:首先,通过词性位置特征权重和TF-IDF权重加权表征双语短文本的同时,使用改进后的TextRank算法获取双语短文本的核心句;然后,通过带有自注意力机制的双向长短时记忆网络分别计算双语短文本的相似度与双语短文本对应核心句的相似度;最后,使用多任务学习方法,将双语短文本的核心句相似度计算作为辅助任务,获取更多的语义信息进行共享以提升汉老双语短文本相似度计算模型的性能。实验结果表明,该文提出的方法在有限的训练语料下取得了更好的效果,F_(1)值达76.16%。
To acquire more semantic information from short texts in both Chinese and Lao,this paper studies the characteristics of short texts in Chinese and Lao and proposed to employ part-of-speech and position information is characteristics in a multi-task short text similarity computation method.Firstly,part-of-speech and position weights are augmented to the TF-IDF weight presentation of short texts in both languages,and then the improved TextRank algorithm is applied to obtain the core sentences of short texts in both languages.Subsequently,bi-directional long-short-term memory network with self-attention mechanism is applied to compute the similarity of short texts in both languages and similarity of corresponding core sentences,respectively.Finally,the multi-task learning method is applied for the similarity computing model of short texts in both Chinese and Lao,with the core sentence similarity computing of short texts in both languages as the auxiliary task,.The test results indicated that the method pro-posed in this paper achieved good performances with limited training linguistic materials(76.16%).
作者
李炫达
周兰江
张建安
LI Xuanda;ZHOU Lanjiang;ZHANG Jian'an(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming,Yunnan 650500 China;Kunming Branch,No.3 College,PLA Information Engineering University,Kunming,Yunnan 650500,China)
出处
《中文信息学报》
CSCD
北大核心
2023年第4期18-27,33,共11页
Journal of Chinese Information Processing
基金
国家自然科学基金(61662040)。
关键词
汉语-老挝语
词性位置特征
多任务学习
短文本相似度计算
Chinese-Lao
part of speech and position characteristics
multi-task learning
short text similarity computation
