Scholarlink

<p>Commonsense question answering (CQA) task is an effective way to measure the ability of language models to understand commonsense knowledge. The previous research focus on combining pre-trained language models (PLMs) with knowledge graphs (KG) for knowledge reasoning. However, these methods still have some limitations, such as the lacking of entity background knowledge in KG, how to fuse heterogeneous information effectively and poor model robustness. To solve the above problems, we propose a novel commonsense reasoning method based on heterogeneous knowledge fusion and adversarial training (HKF-AT). A multi-source knowledge retrieval method is adopted to retrieve entity-related texts in a large corpus and extract the knowledge path from question entity to option entity in KG. Further, multi-source information fusion strategy makes the model utilize and learn rich external knowledge information directly. Specially, the new adversarial samples are generated for both the retrieved texts and triples, and adversarial training is carried out. The experimental results on public benchmark dataset indicate that the proposed method effectively improves the performance and robustness of the model compared with other excellent models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel commonsense reasoning method based on heterogeneous knowledge fusion and adversarial training

  • 期刊论文

  • 发表日期：

    2025年3月11日

查看全文
  • Yongping Du,
  • Qi Zhang,
  • Jingya Yan,
  • Ying Hou,
  • Honggui Han

摘要

Commonsense question answering (CQA) task is an effective way to measure the ability of language models to understand commonsense knowledge. The previous research focus on combining pre-trained language models (PLMs) with knowledge graphs (KG) for knowledge reasoning. However, these methods still have some limitations, such as the lacking of entity background knowledge in KG, how to fuse heterogeneous information effectively and poor model robustness. To solve the above problems, we propose a novel commonsense reasoning method based on heterogeneous knowledge fusion and adversarial training (HKF-AT). A multi-source knowledge retrieval method is adopted to retrieve entity-related texts in a large corpus and extract the knowledge path from question entity to option entity in KG. Further, multi-source information fusion strategy makes the model utilize and learn rich external knowledge information directly. Specially, the new adversarial samples are generated for both the retrieved texts and triples, and adversarial training is carried out. The experimental results on public benchmark dataset indicate that the proposed method effectively improves the performance and robustness of the model compared with other excellent models.