Leveraging word embeddings to enhance co-occurrence networks: A statistical analysis

Area of Science:

• Computational Linguistics
• Network Science
• Natural Language Processing

Background:

• Recent studies explore virtual edges in word co-occurrence networks using word embeddings to enhance graph representations, especially for short texts.
• The impact of semantic edges on traditional co-occurrence networks is not fully understood.
• Investigating statistical properties of text-based network models is crucial for understanding text meaning.

Purpose of the Study:

• To assess if network metrics can differentiate between meaningful and meaningless texts.
• To determine if network metrics are more sensitive to syntactic or semantic text aspects.
• To provide guidelines for selecting appropriate network metrics based on text characteristics and task requirements.

Main Methods:

• Enriching word co-occurrence networks with virtual edges derived from FastText word embeddings.
• Analyzing statistical properties of traditional and enriched text-based network models.
• Evaluating the informativeness of network metrics like average shortest path, closeness centrality, and clustering coefficient.

Main Results:

• Incorporating virtual edges has mixed effects on network metrics; some improve, others decrease in informativeness.
• Average shortest path and closeness centrality become more informative for short texts with virtual edges.
• Clustering coefficient's informativeness decreases with the addition of virtual edges.
• The inclusion of stopwords influences the statistical properties of enriched networks.

Conclusions:

• Virtual edges can enhance text network analysis but require careful selection of network metrics.
• Network metrics' sensitivity to semantic versus syntactic information varies.
• Findings offer guidance for optimizing text network analysis based on text length and specific application needs.