[Paper Review] Title-Guided Encoding for Keyphrase Generation

Wang Chen, Yifan Gao+3 more|arXiv (Cornell University)|Aug 26, 2018

Advanced Text Analysis Techniques55 references3 citations

TL;DR

This paper proposes Title-Guided Network (TG-Net), a novel encoder-decoder model for keyphrase generation that treats the document title as a query-like input and uses attention mechanisms to guide context encoding with title information. TG-Net significantly outperforms state-of-the-art models, achieving up to 9.4% higher F1@10 and 19.1% higher Recall@50 on absent keyphrase prediction, especially on documents with very short or very long titles.

ABSTRACT

Keyphrase generation (KG) aims to generate a set of keyphrases given a document, which is a fundamental task in natural language processing (NLP). Most previous methods solve this problem in an extractive manner, while recently, several attempts are made under the generative setting using deep neural networks. However, the state-of-the-art generative methods simply treat the document title and the document main body equally, ignoring the leading role of the title to the overall document. To solve this problem, we introduce a new model called Title-Guided Network (TG-Net) for automatic keyphrase generation task based on the encoder-decoder architecture with two new features: (i) the title is additionally employed as a query-like input, and (ii) a title-guided encoder gathers the relevant information from the title to each word in the document. Experiments on a range of KG datasets demonstrate that our model outperforms the state-of-the-art models with a large margin, especially for documents with either very low or very high title length ratios.

Motivation & Objective

• To address the limitation of existing generative keyphrase models that treat titles and main bodies equally, ignoring the title's summarizing role.
• To improve keyphrase generation by explicitly leveraging the title as a query to guide encoding of the main body.
• To enhance performance on both present and absent keyphrase prediction, particularly for documents with very short or very long titles.
• To investigate whether title-guided encoding can better capture semantic relevance between titles and keyphrases.

Proposed method

• The model uses a dual-GRU encoder: one for the main body and one for the title, both producing contextual representations.
• An attention-based matching layer computes relevance between each context word and the title, enabling title information to guide encoding of each word in the context.
• The title-guided representation is formed by concatenating original context representations with title-attended representations through an additional bi-directional GRU layer.
• The decoder uses both an attention mechanism and a copy mechanism to generate keyphrases from a predefined vocabulary and copy phrases directly from the source text.
• The title is treated as a query-like input, allowing the model to focus on title-relevant content in the main body regardless of distance.
• The model is trained end-to-end on five benchmark datasets using cross-entropy loss for sequence generation.

Experimental results

Research questions

• RQ1 Can treating the title as a query-like input improve keyphrase generation performance compared to standard encoder-decoder models?
• RQ2 How does title-guided encoding affect keyphrase generation on documents with extreme title length ratios?
• RQ3 To what extent does title-guided attention improve the generation of absent keyphrases not explicitly present in the text?
• RQ4 Does the inclusion of a copy mechanism in the decoder further enhance performance when combined with title-guided encoding?

Key findings

• TG-Net achieves a 9.4% improvement in F1@10 for present keyphrase prediction and a 19.1% improvement in Recall@50 for absent keyphrase prediction on the largest dataset compared to the best baseline.
• The model shows the largest performance gains on documents with title length ratios below 3% and above 12%, where title information is most critical due to long context or short titles.
• Ablation studies confirm that removing the title-guided encoding component reduces performance significantly, proving its critical role in both present and absent keyphrase generation.
• The copy mechanism contributes substantially to performance, as its removal leads to a sharp drop in F1 and recall scores across all tasks.
• Case studies show TG-Net successfully generates absent keyphrases like 'time-delay systems' by leveraging title cues such as 'stochastic delay systems', while CopyRNN fails.
• The model consistently outperforms strong baselines across all five benchmark datasets, demonstrating robustness and generalization across diverse document types.