Mutual-Attention Net: A Deep Attentional Neural Network for Keyphrase Generation

Area of Science:

• Natural Language Processing
• Information Retrieval
• Machine Learning

Background:

• Neural keyphrase generation (NKG) automatically extracts keyphrases from documents.
• Unlike traditional methods, NKG can generate novel keyphrases not present in the source text.
• Existing NKG models are often hindered by noise in the source document, impacting performance.

Purpose of the Study:

• To introduce a new denoising architecture, the mutual-attention network (MA-net), for neural keyphrase generation.
• To address the limitations of existing NKG models that do not account for source document denoising.
• To improve the accuracy and quality of generated keyphrases.

Main Methods:

• Developed a mutual-attention network (MA-net) incorporating multihead attention to identify title-abstract relevance for denoising.
• Utilized multihead attention for content vector computation, enhancing keyphrase generation accuracy.
• Implemented a hybrid network to address the out-of-vocabulary (OOV) problem by enabling word generation and copying from the source document.

Main Results:

• The MA-net architecture effectively denoises the source document by leveraging title-abstract relevance.
• The use of multihead attention for content vectors leads to more accurate high-quality keyphrase generation.
• The hybrid network successfully mitigates the OOV challenge, improving overall model robustness.

Conclusions:

• The proposed MA-net significantly enhances neural keyphrase generation by incorporating effective denoising strategies.
• The model demonstrates superior performance compared to state-of-the-art methods across five benchmark datasets.
• This work advances the field of automatic keyphrase extraction by providing a robust and accurate denoising approach.