«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年05期

页码:

188-196

栏目:

新型计算应用系统

出版日期:

2025-05-10

文章信息/Info

Title:

Prompt Learning Based Classification for Few-shot Cultural and Tourism Customer Service Questions

文章编号:

1673-629X(2025)05-0188-09

作者:

薛赟琨1; 王亮2; 朱欣娟1

1. 西安工程大学 计算机科学学院,陕西 西安 710600;

2. 秦始皇帝陵博物院,陕西 西安 710699

Author(s):

XUE Yun-kun1; WANG Liang2; ZHU Xin-juan1

1. School of Computer Science and Technology,Xi’an Polytechnic University,Xi’an 710600,China;

2. Emperor Qinshihuang’s Mausoleum Site Museum,Xi’an 710699,China

关键词:

提示学习; 问题分类; 小样本; 层次结构; 文旅客服问答

Keywords:

prompt learning;  question classification;  few - shot;  hierarchical structure;  question answering with cultural and tourism customer service

分类号:

TP391

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0399

摘要:

为解决传统分类方法需要大量的标签数据,且在样本量不足的情况下模型容易过拟合的问题,提出了一种基于提示学习的小样本文旅客服问题分类方法。 首先,收集整理了文旅客服领域数据,对问题类别进行了层次分类,构建了文旅客服问题数据集,并对其进行数据类别标注;其次,基于提示学习的方法,设计了八类不同的提示模板,分别在 Bert、RoBerta 等预训练语言模型上进行对比实验,选择了效果较优的提示模板和预训练语言模型,并对文本分类模型进行领域微调;最后,设计了基于提示学习的层次分类算法,将算法和微调后的分类模型应用于某博物馆文旅客服问答系统。 实验结果表明,在小样本数据条件下,文旅客服问题分类的准确度和 F1 值达到 92. 03% 和 96. 43% ,相较于 BERT 文本分类基线模型,准确度和 F1 值分别提升了 5. 77 百分点和 5. 78 百分点。

Abstract:

To solve the problem of traditional classification methods requiring a large amount of labeled data and model overfitting in the case of insufficient sample size,a few-shot cultural and tourism customer service question classification method based on prompt learning is proposed. Firstly,data in the field of cultural and tourism customer service was collected and organized,and question categories were hierarchically classified. A dataset of cultural and tourism customer service questions was constructed, and the data categories were annotated. Secondly,based on the method of prompt learning, eight different prompt templates were designed and compared on pre trained language models such as Bert and RoBerta. The better performing prompt templates and pretrained language models were selected,and domain fine-tuning was performed on the text classification model. Finally,a hierarchical classification algorithm based on prompt learning was designed,and the algorithm and the fine tuned classification model were applied to a museum cultural and tourism customer service Q&A system. The experimental results show that under few-shot data conditions, the accuracy and F1 score of cultural and tourism customer service question classification reach 92. 03% and 96. 43% . Compared with the BERT text classification baseline model,the accuracy and F1 score have increased by 5. 77 percentage points and 5. 78 percentage points.

常用功能

更新日期/Last Update: 2025-05-10