The pluralization palette: unveiling semantic clusters in English nominal pluralization through distributional...

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Semantics

Background:

• English nominal pluralization, while lacking extensive morphological marking, exhibits semantic variations based on word class (e.g., fruits vs. animals).
• Existing computational methods for predicting plural semantics often fail to capture these intricate class-specific differences.

Purpose of the Study:

• To investigate semantic clustering in English nominal pluralization.
• To introduce and evaluate a novel semantically informed method (CosClassAvg) for generating plural word vectors.
• To compare CosClassAvg against existing methods in terms of faithfulness to corpus data and computational model performance.

Main Methods:

• Distributional semantics and vector space models were employed.
• A new method, CosClassAvg, was developed to predict plural vectors based on semantic class.
• CosClassAvg was compared to a fixed shift method and a linear mapping method (FRACSS) using corpus-extracted plural vectors.

Main Results:

• CosClassAvg predicted plural vectors with greater similarity in vector length compared to FRACSS, though orientation similarity was slightly lower.
• Both CosClassAvg and FRACSS significantly outperformed the fixed shift method in capturing English plural semantics.
• Computational modeling demonstrated that the semantic nuances captured by CosClassAvg improved a listener model's ability to understand novel plural forms.

Conclusions:

• English plural semantics are class-dependent, even without explicit morphological markers.
• CosClassAvg offers a balanced approach, effectively modeling plural semantics while maintaining the ability to generalize to new forms.
• The findings highlight the importance of incorporating semantic class information into computational models of language.