Probabilistic Keyphrase Generation From Copy and Generating Spaces.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Artificial Intelligence

Background:

• Existing keyphrase generation models often optimize negative log-likelihood loss without directly addressing copy and generation spaces, potentially limiting decoder generability.
• Current models struggle with dynamically determining the number of keyphrases, often producing them implicitly.
• Fundamental challenges in NLP include accurate and controllable keyphrase extraction.

Purpose of the Study:

• To propose a probabilistic keyphrase generation model that explicitly manipulates copy and generating spaces.
• To enable dynamic control over the number of generated keyphrases.
• To improve the generability and accuracy of keyphrase prediction.

Main Methods:

• Developed a model based on the variational encoder-decoder (VED) framework with two separate latent variables for copy and generating spaces.
• Utilized a von Mises-Fisher (vMF) distribution for the generating space to modify vocabulary probability distributions.
• Employed a clustering module for Gaussian Mixture learning in the copy space and used filtered components to determine keyphrase quantity, trained via latent variable modeling, neural variational inference, and self-supervised learning.

Main Results:

• The proposed model outperforms state-of-the-art baselines on social media and scientific article datasets.
• Achieved accurate keyphrase predictions.
• Demonstrated controllable generation of keyphrase numbers.

Conclusions:

• The novel probabilistic approach effectively models separate copy and generating spaces for improved keyphrase generation.
• The method provides explicit control over the number of keyphrases, addressing a limitation in existing models.
• This work advances NLP by offering a more flexible and accurate keyphrase generation system.