keyphrase-extraction-kbir-inspec
| Property | Value |
|---|---|
| License | MIT |
| Paper | Research Paper |
| Dataset | Inspec |
| F1 Score | 0.588 |
What is keyphrase-extraction-kbir-inspec?
This is a specialized transformer-based model designed for extracting key phrases from scientific texts. Built on the KBIR (Keyphrase Boundary Infilling with Replacement) architecture and fine-tuned on the Inspec dataset, it excels at identifying crucial concepts in academic papers and technical documents. The model treats keyphrase extraction as a token classification problem, labeling words as either beginning (B-KEY), inside (I-KEY), or outside (O) of keyphrases.
Implementation Details
The model implements a sophisticated approach to keyphrase extraction through a multi-task learning setup that combines Masked Language Modeling (MLM), Keyphrase Boundary Infilling (KBI), and Keyphrase Replacement Classification (KRC). It was trained with a learning rate of 1e-4 over 50 epochs with early stopping patience of 3.
- Utilizes the KBIR architecture for enhanced semantic understanding
- Processes text through token classification pipeline
- Supports batch processing with maximum sequence length of 512 tokens
- Implements sophisticated post-processing for keyphrase aggregation
Core Capabilities
- Automatic extraction of relevant keyphrases from scientific documents
- High accuracy with F1@M score of 0.56 on the Inspec test set
- Effective handling of long-term semantic dependencies
- Support for English language documents
- Specialized performance on scientific and technical content
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its use of the KBIR architecture, which combines multiple learning objectives to better understand keyphrase boundaries and context. It achieves state-of-the-art performance on scientific text analysis, with particularly strong results on the Inspec dataset.
Q: What are the recommended use cases?
The model is specifically designed for processing scientific papers and technical documents, particularly in the domains of Computers, Control, and Information Technology. It's most effective when analyzing academic abstracts and technical content, though it may not perform as well on general-domain texts.
