We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
MSF-Net: Multi-level Semantic Feature Network Extractor for Paraphrase Identification.
- Authors
Wenrui Xue; Yujun Zhang; Xinyu Wang; Shuting Ge
- Abstract
To address inaccurate semantic representation and challenges in interpreting rare words within deep learningbased paraphrase identification tasks, this paper introduces a multi-level semantic feature network extractor (MSF-Net). The MSF-Net model represents an end-to-end dual-stage, multilevel semantic information learning architecture. Specifically, a topic-level semantic feature extraction module is incorporated to discern the topic distribution of the text. Initially, this module synergizes the text's hidden state, acquired from the Bi-GRU module, with the topic extractor for joint learning of local-global semantic information in the text. MSF-Net employs a multi-attention module to proficiently capture word relationships and semantic details by modeling the complete text sequence, informed by learned topic and context information, thereby aiding the model in paraphrase identification. Comparative and ablation experiments on the extensive LCQMC text dataset are presented in this paper. The MSF-Net model achieves precision, recall, F1 score, and accuracy rates of 78.69, 94.14, 85.72, and 87.13, respectively. The results substantiate MSF-Net's superiority over baseline models in capturing semantic information and reinforcing paraphrase recognition tasks.
- Subjects
PARAPHRASE; FEATURE extraction; INFORMATION architecture
- Publication
IAENG International Journal of Computer Science, 2023, Vol 50, Issue 4, p1391
- ISSN
1819-656X
- Publication type
Article