| 引用本文: | 胡松华,唐家伟,马汝虎.一种基于提示学习的话题讽刺识别方法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2026，43（3）：108-115 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 面向话题的讽刺识别旨在通过引入话题,将话题视作对象来判断评论是否为讽刺表达,但是现有基于
提示学习的方法存在模板依赖人工构造的问题,存在复杂性与不稳定性。 方法 针对此问题,根据提示学习的模板
工程与映射器工程提出了新的设计方法,将连续与离散相结合的混合模板与映射器进行交互,通过对模板进行预
训练减少人工选取模板引起的复杂性与不稳定性问题,再将设计好的模板与原始输入拼接输入到预训练模型
BERT-base-Chinese 中对标记的数据进行分类;在此基础上,提出面向话题讽刺识别任务的模型 MTPrompt。 结果
该模型在 ToScarcasm 数据集上的实验表明,在全批量训练数据场景下,与现有手工模板与手工映射器交互模型
TOSPrompt 相比较,准确率提升了 2. 78%、精确率提升了 1. 89%、召回率提升了 5. 26%、F1 值提升了 3. 54%,同时在
小批量训练数据场景下,准确率与 F1 值均高于其他基线模型。 结论 在话题讽刺识别应用场景中,混合模板的方法
可以有效优化人工构造模板方法的复杂性与不稳定,并且达到更佳的识别性能。 |
| 关键词: 自然语言处理 提示学习 面向话题的讽刺识别 情感分析 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| A Topic-Oriented Sarcasm Detection Method Based on Prompt Learning |
|
HU Songhua TANG Jiawei MA Ruhu
|
|
School of Artificial Intelligence and Big Data Hefei University Hefei 230601 China
|
| Abstract: |
| Objective Topic-oriented sarcasm detection aims to judge whether a comment is a sarcastic expression by
introducing topics and regarding them as objects. Nevertheless the existing methods based on prompt learning have the
problem of relying on manually constructed templates which brings complexity and instability. Methods To tackle this
problem a novel design approach was put forward based on template engineering and mapper engineering of prompt
learning. A hybrid template integrating continuous and discrete elements interacted with the mapper. By pre-training the
template the complexity and instability resulting from manual template selection were mitigated. Subsequently the welldesigned template was concatenated with the original input and fed into the pre-trained model BERT-base-Chinese for
classifying the labeled data. On this basis a model named MTPrompt for topic-oriented sarcasm detection was proposed.
Results Experiments on the ToScarcasm dataset show that under the scenario of full-batch training data when compared
with the existing model TOSPrompt that involves the interaction between manual templates and manual mappers the
proposed model MTPrompt achieves significant improvements. Specifically its accuracy was increased by 2. 78%
precision by 1. 89% recall by 5. 26% and the F1
score by 3. 54%. Meanwhile under the scenario of small-batch training data both the accuracy and F1
score of this model were higher than those of other baseline models.
Conclusion In the application scenario of topic-oriented sarcasm detection the hybrid template method can effectively
alleviate the complexity and instability of manually constructed templates and attain better detection performance |
| Key words: natural language processing prompt learning topic-oriented sarcasm detection sentiment analysis |
