Performance impact of stop lists and morphological decomposition on word-word corpus-based semantic space models

Area of Science:

• Computational Linguistics
• Cognitive Science
• Psycholinguistics

Background:

• Corpus-based semantic space models use lexical co-occurrence statistics.
• Existing models are influenced by orthographic word frequency, potentially biasing results with high-frequency closed-class words.

Purpose of the Study:

• To investigate the impact of morphological decomposition and stop list usage on corpus-based semantic space models.
• To explore the role of sublexical information in lexical semantic processing.

Main Methods:

• Compiled co-occurrence statistics from corpora, manipulating morphological decomposition and the exclusion of closed-class words (stop lists).
• Evaluated model performance based on these manipulations.

Main Results:

• Morphological decomposition significantly improved performance in word-word co-occurrence semantic space models.
• Employing stop lists led to an overall decrease in model performance.
• Evidence suggests closed-class words may not solely provide syntactic information.

Conclusions:

• Sublexical information, specifically word morphology, appears to play a role in lexical semantic processing.
• The exclusion of closed-class words may not enhance, but rather hinder, semantic space model performance.
• The assumption that closed-class words are purely syntactic requires re-evaluation in the context of co-occurrence models.