Representing Documents via Latent Keyphrase Inference.

Area of Science:

• Natural Language Processing
• Information Retrieval
• Machine Learning

Background:

• Traditional text mining uses bag-of-words or n-grams, limiting deep understanding.
• Concept-based models rely on external knowledge bases, which lack coverage in specialized domains.

Purpose of the Study:

• To introduce a data-driven model, Latent Keyphrase Inference (LAKI), for document representation.
• To overcome limitations of existing methods in vertical domains by using domain-specific keyphrases.

Main Methods:

• LAKI learns topical content units for domain keyphrases from in-domain documents.
• Documents are represented as vectors of related domain keyphrases, moving beyond surface forms.
• The model infers latent keyphrases, enhancing computer comprehension of text.

Main Results:

• LAKI effectively bridges the gap between bag-of-words and concept-based models.
• The approach eliminates the need for external knowledge bases, offering interpretable representations.
• LAKI demonstrated superior performance against eight other methods on two text mining tasks.

Conclusions:

• Latent Keyphrase Inference (LAKI) offers a robust and interpretable method for document representation.
• This data-driven approach enhances text mining capabilities, particularly in specialized domains.