融合历史答案特征的多粒度语义交互答案排序方法
Multi-granularity semantic interactive answers ranking method based on historical answers
崔伟琪 1严馨 1刘艳超 2邓忠莹 3徐广义4
作者信息
- 1. 昆明理工大学信息工程与自动化学院,云南昆明 650504;昆明理工大学云南省人工智能重点实验室,云南 昆明 650504
- 2. 湖北工程学院信息技术中心,湖北孝感 432000
- 3. 昆明理工大学信息工程与自动化学院,云南昆明 650504
- 4. 云南南天电子信息产业股份有限公司昆明南天电脑信息系统有限公司,云南昆明 650040
- 折叠
摘要
为解决只根据单一特征判断答案质量的问题,提出一种结合历史答案特征及多粒度语义交互判断答案质量的排序方法.通过指针网络提取历史答案特征,用动态注意力剔除掉问答对及历史答案的弱相关部分,采用比较聚合池化提取局部语义特征向量,用池化归纳问答对及历史答案句子信息,通过加权求和提取全局语义特征向量.将问答对及历史答案的局部和全局语义特征向量融合,输入到分类器进行打分,按照得分对候选答案排名.实验结果表明,所提方法有效提升了答案选择的正确率.
Abstract
To solve the problem that the answer quality is judged only by a single feature,a ranking method combining historical answer features and multi granularity semantic interaction was proposed to judge the answer quality.The pointer network was used to extract the features of historical answers,and the dynamic attention was used to remove the weak correlation parts of question answer pairs and historical answers.The comparison aggregation pooling was used to extract local semantic feature vec-tors,and the pooling was used to summarize the sentence information of question answer pairs and historical answers.The weighted sum was used to extract global semantic feature vectors.The local and global semantic feature vectors of question an-swer pairs and historical answers were fused and inputted to the classifier for scoring,and the candidate answers were ranked according to the score.Experimental results show that the proposed method can effectively improve the accuracy of answer selec-tion.
关键词
答案排序/多粒度语义交互/注意力机制/指针神经网络/预训练模型/长短期记忆网络/深度学习Key words
answers ranking/multi granularity semantic interaction/attention mechanism/pointer network/pre-trained model/LSTM/deep learning引用本文复制引用
基金项目
国家自然科学基金项目(61562049)
国家自然科学基金项目(61462055)
出版年
2024
NETL
NSTL
万方数据