Modeling metaphor perception with distributional semantics vector space models

Kat R. Agres, Stephen McGregor, Karolina Rataj, Matthew Purver, Geraint A. WigginsC3gi@esslli 2016

Abstract

Traditional distributional semantic models represent words as fixed vectors, often failing to capture the dynamic, context-dependent nature of metaphors. We evaluate a Concept Discovery Model (CDM) that constructs dynamic subspaces to represent verb-noun metaphors. Our findings show that models representing the distinct properties of both the source and target words (Independent Subspaces) outperform static models and better align with human psycholinguistic judgments, effectively isolating metaphoricity from simple word familiarity.

Introduction

Metaphors like "the city sleeps" involve a conceptual transfer from a source (living beings) to a target (a geographic location). While standard Distributional Semantic Models (DSMs) excel at identifying synonyms through fixed high-dimensional vectors, they struggle with the fluid meanings inherent in metaphorical language. This study investigates whether dynamic vector space models can more accurately predict human ratings of metaphoricity, familiarity, and meaningfulness.

Methods

The researchers compared a baseline static model, Word2Vec, against a novel Concept Discovery Model (CDM). While Word2Vec uses fixed 300-dimensional vectors, the CDM starts with a massive 7.5-million-dimensional base space and dynamically constructs 200-dimensional subspaces for specific word pairs.

Three subspace construction strategies were tested within the CDM framework:

  1. Noun-Only: Dimensions are selected based solely on the target noun's salience.
  2. Joint: Dimensions are selected based on features common to both the verb and noun.
  3. Independent: Dimensions are selected by merging the top 100 salient features of the verb and the top 100 of the noun.
Comparison of Noun-Only, Joint, and Independent Subspace constructions in the CDM.

Datasets and Human Norming

Models were trained on a 1.1-billion-token corpus from English Wikipedia. Evaluation was performed on 228 verb-noun dyads categorized as novel metaphors (e.g., "harvest courage"), conventional metaphors (e.g., "gather courage"), or literal expressions (e.g., "experience courage"). To establish a ground truth, four normative studies were conducted with over 400 participants to gather ratings for cloze probability, meaningfulness, familiarity, and metaphoricity.

Summary of normative studies used to gather human ratings for model evaluation.

Key Results

The study found that metaphoricity is inversely correlated with mathematical similarity; highly metaphorical phrases exhibit greater "distance" in the vector space. The CDM Independent Subspace model emerged as the superior approach, accounting for 27.1% of the variance in human ratings ($R^2 = .271$), compared to 24.9% for Word2Vec.

Crucially, while Word2Vec's performance was confounded by word familiarity, the CDM Independent model's predictions were driven uniquely by metaphoricity ($p < .001$). Qualitatively, literal word pairs had the highest similarity scores, followed by conventional metaphors, with novel metaphors having the lowest similarity.

Discussion and Future Work

The success of the Independent Subspace method highlights the "power of imbalance" in metaphor processing. It suggests that human perception requires maintaining the distinct, salient properties of both words separately before bringing them together. While the model shows significant improvement over static baselines, the remaining unexplained variance suggests that geometric distance alone cannot fully capture the complex cognitive processes of metaphorical interpretation.

Future directions include focusing on specific property transfer from the verb to the noun and moving beyond simple verb-noun pairs to more complex metaphorical sentences.

Summary updated

BibTeX
@inproceedings{agres2016modeling,
  author = {Kat R. Agres and Stephen McGregor and Karolina Rataj and Matthew Purver and Geraint A. Wiggins},
  title = {{Modeling metaphor perception with distributional semantics vector space models}},
  booktitle = {C3gi@esslli},
  year = {2016},
}
Figure