Keyphrase Generation with Correlation Constraints

Abstract

In this paper, we study automatic keyphrase generation.Although conventional approaches to this task show promising results, they neglect correlation among keyphrases, resulting in duplication and coverage issues. To solve these problems, we propose a new sequence-to-sequence architecture for keyphrase generation named CorrRNN, which captures correlation among multiple keyphrases in two ways. First, we employ a coverage vector to indicate whether the word in the source document has been summarized by previous phrases to improve the coverage for keyphrases. Second, preceding phrases are taken into account to eliminate duplicate phrases and improve result coherence. Experiment results show that our model significantly outperforms the state-of-the-art method on benchmark datasets in terms of both accuracy and diversity.

Neural Keyphrase Generation via Reinforcement Learning with Adaptive Rewards

“…These models include catSeq, catSeqD (Yuan et al, 2018), catSeqCorr (Chen et al, 2018a), and catSeqTG (Chen et al, 2018b). For all baselines, we use the method in Yuan et al (2018) Table 2: Results of present keyphrase prediction on five datasets.…”

“…With this setup, all baselines can determine the number of keyphrases to generate. The catSeqCorr and catSeqTG models are the CorrRNN (Chen et al, 2018a) and TG-Net (Chen et al, 2018b) models trained under this setup, respectively. For the reinforced models, we follow the method in Section 3.2 to concatenate keyphrases.…”

Keyphrase Generation: A Multi-Aspect Survey

“…A reduced vocabulary size is important to have decent AKG resutls within a reasonable computation time. For this reason, authors of many recent AKG studies like [12], [71], [68], [73], [75] and [79] replace all digit tokens with the symbol digit . Stemming is also commonly used in studies like [12], [75], [74], [24], [71] and [73] to have the predicted and golden keywords properly compared during evaluation.…”

A review of keyphrase extraction

“…Alternative types of context meaning can be captured by positional information incorporating the relative position of the phrase in a given text , that is, the position of the first occurrence of a phrase normalized by the length of the target publication (Caragea et al, ) (CeKE), (Hulth, ), (Zhang et al, ) (MIKE), and (Wang & Li, ) (PCU‐ICL). Recently, McIlraith and Weinberger () (SurfKE) as well as Meng et al (), Chen et al (), Wang et al (), Alzaidy et al () and Ye and Wang () go one step beyond the context‐based features by learning features/embeddings using graph structures and neural networks, respectively.…”