Extracting relations from texts using vector language models and a neural network classifier

Area of Science:

• Natural Language Processing
• Knowledge Representation
• Information Retrieval

Background:

• Simple Knowledge Organization System (SKOS) is crucial for organizing information.
• Identifying semantic relations between concepts is challenging.
• Vector representations in language models offer potential for relation extraction.

Purpose of the Study:

• To investigate the feasibility of identifying SKOS relations using term vector representations.
• To develop and evaluate an artificial neural network (ANN) classifier for SKOS relations.
• To compare the performance of different language model families (Word2Vec, GloVe) for this task.

Main Methods:

• Utilized vector representations from Word2Vec and GloVe language models.
• Developed an ANN-based classifier for SKOS relations.
• Trained and tested the classifier using DBPedia and EuroVoc thesauri datasets.

Main Results:

• The ANN classifier trained with GloVe models demonstrated high efficiency in identifying SKOS relations.
• Training the classifier with Word2Vec models was found to be infeasible within the current ANN architecture.
• The study highlights the critical role of global term context for SKOS relation identification.

Conclusions:

• GloVe models are suitable for building effective SKOS relation classifiers.
• Word2Vec models, in their current application, are not viable for this specific ANN-based SKOS relation identification task.
• Understanding the global context of term usage is essential for accurate SKOS relation extraction.