Semantic Similarity Computation for Abstract and Concrete Nouns Using Network-based Distributional Semantic Models

Elias Iosif, Alexandros Potamianos, Maria Giannoudaki, Kalliopi ZervanouIWCS 2013 · OpenReview

Abstract

Distributional Semantic Models (DSMs) typically treat all lexical items uniformly, ignoring qualitative differences in how concepts are organized. However, cognitive research suggests a "concreteness effect," where abstract and concrete concepts are processed differently by the human brain. This paper explores whether computational models should reflect this distinction. By employing network-based DSMs, we demonstrate that concrete nouns are best modeled using a "maximum sense" approach, whereas abstract nouns are more effectively captured via "attributional similarity" within their semantic neighborhoods. Our findings, consistent across English and Greek datasets, suggest that the optimal neighborhood size and similarity metric depend significantly on a noun's concreteness.

Introduction

In Natural Language Processing, word similarity is often calculated based on distributional properties—the contexts in which words appear. While standard DSMs have proven effective, they often ignore the qualitative differences between abstract concepts (e.g., "justice", "belief") and concrete ones (e.g., "car", "rooster"). Humans process concrete concepts faster and store them differently, a phenomenon known as the concreteness effect. This research investigates whether tailoring computational similarity metrics to the type of noun can improve model performance and better reflect human cognitive processes.

Methodology

The study utilizes network-based DSMs where words are represented as nodes and relationships as edges. Two distinct hypotheses for calculating similarity were tested:

  1. Maximum Similarity of Neighborhoods ($M_n$): Based on the "maximum sense" hypothesis, this metric assumes similarity is defined by the single most similar sense or neighbor between two words.
  2. Attributional Neighborhood Similarity ($R_n$): This metric assumes a word's meaning is an aggregate of its attributes, calculating the correlation between the entire distribution of neighbors for both words.

These metrics were evaluated using two base network construction methods: simple sentence-level co-occurrence and context-based cosine similarity of feature vectors. Testing was performed on the English WordSim-353 (WS353) dataset and a newly developed Greek (GIP) dataset, with specific focus on isolated abstract and concrete noun pairs.

Results and Discussion

The experimental results confirm that network-based approaches outperform standard baseline DSMs. Crucially, the "concreteness effect" was observed in the computational models: performance was generally higher for concrete noun pairs than for abstract ones.

Performance of similarity metrics across English (WS353) and Greek (GIP) datasets for varying neighborhood sizes. Subplots (a) and (c) show trends for the maximum sense approach, while (b) and (d) illustrate the attributional approach.

As shown in the figure above, the optimal strategy differs by noun category:

  • Concrete Nouns: These nouns typically have "tight" neighborhoods of synonyms or class members. They performed best with the $M_n$ metric and required larger neighborhoods (30-40 neighbors) to reach peak accuracy.
  • Abstract Nouns: These nouns possess "broad" neighborhoods that function as attributes. They reached peak performance with very small neighborhoods (around 10 neighbors) and were best captured by the $R_n$ metric.

The consistency of these results across both English and Greek suggests that the organizational differences between abstract and concrete concepts may be a universal linguistic feature.

Conclusion

This research demonstrates that the semantic representation of nouns is not monolithic. Concrete and abstract nouns exhibit distinct structural properties in semantic networks, necessitating different computational treatments. Future work should focus on developing hybrid similarity metrics that dynamically adjust their mathematical approach based on the estimated concreteness of the target words, moving AI closer to the nuanced linguistic processing of the human brain.

Summary updated

BibTeX
@inproceedings{iosif2013semantic,
  author = {Elias Iosif and Alexandros Potamianos and Maria Giannoudaki and Kalliopi Zervanou},
  title = {{Semantic Similarity Computation for Abstract and Concrete Nouns Using Network-based Distributional Semantic Models}},
  booktitle = {IWCS},
  year = {2013},
  url = {https://openreview.net/forum?id=td0iglW3PF},
}
Figure