基于大语言模型命名实体识别的AI智能问答优化  

AI-Enhanced Q&A Optimization via LLM-Based Named Entity Recognition

作　　者：施志雄 段该甲 马龙轩 吴婕 Shi Zhixiong;Duan Gaijia;Ma Longxuan;Wu Jie(China Unicom Guangdong Branch,Guangzhou 510627,China)

机构地区：[1]中国联通广东分公司,广东广州510627

出　　处：《邮电设计技术》2025年第3期80-84,共5页Designing Techniques of Posts and Telecommunications

摘　　要：为优化AI问答效果,提出基于大语言模型命名实体识别的优化方法。首先,通过在多种分割方式中选取最优方案,结合词语划分概率判断结果,对语料库文本进行分词。其次,在预训练的BERT模型顶部添加线性层,并通过标注数据对预测实体类别进行微调,将预测的同类标签词组合得到命名实体。最后,通过上下文构建整合用户输入与识别结果,将整合结果输入模型生成回答,并结合用户反馈优化输出。结果表明,所提方法生成结果与参考文本之间的语义相似度较高,具备较为理想的问答效果。To optimize the effectiveness of AI question answering,an optimization method based on large language model named entity recognition is proposed.Firstly,by selecting the optimal segmentation method from multiple options and combining it with the probability judgment results of word segmentation,the corpus text is segmented.Secondly,a linear layer is added at the top of the pre trained BERT model,and the predicted entity categories are fine tuned through annotated data.It combines the predicted same class label words to obtain named entities.Finally,it constructs and integrates user input and recognition results through context,and inputs the integrated results into the model to generate answers,and optimizes the output based on user feedback.The results indicate that the proposed method has a high semantic similarity between the generated results and the reference text,and has a relatively ideal question answering effect.

关 键 词：大语言模型 BERT 命名实体识别 智能问答 分词 

分 类 号：TN915[电子电信—通信与信息系统]