论文笔记[2] Re-Representing Metaphor: Modeling Metaphor Perception Using Dynamically Contextual ...

论文题目：Re-Representing Metaphor: Modeling Metaphor Perception Using Dynamically Contextual Distributional Semantics
论文地址： https://www.researchgate.net/publication/332431541
论文作者：Stephen McGregor, Kat Agres, Karolina Rataj, Matthew Purver, Geraint Wiggins

Introduction & Background

• Metaphor : words take on new semantic roles in a particular communicative context, one conceptual domain to another

• Introduce a dynamically contextual distributional semantic model

• metaphoricity of verb-object compositions, by statistical analysis ways

• Focus more on Geometry

Computational Methodology

• standard distributional semantic view of geometric semantic representation

• Data Cleaning & Matrix Building, $2\times 2$ window, 200,000 most frequent word types & 9 million columns (co-occurrence word types)

• 3 different techniques for selecting subspaces : $\Rightarrow$ k

  • MEAN : We take the co-occurrence terms with the highest arithmetic mean PMI value across input words
    
  • GEOM : highest geometric mean PMI
    
  • INDY : We take a concatenation of the co-occurrence terms with the highest PMI values for each word independently
    

• Word-vectors :

  • generic-vectors
  • mean-vector
  • maximum-vector
  • central-vector

• 48 geometric features(Distances, Means, Ratios, Fractions, Angles, Areas)
  
  

Human Metaphor Judgements

• novel metaphors (e.g., to harvest courage), conventional metaphors (e.g., to gather courage), and literal expressions (e.g., to experience courage)
• 228($76\times 3$) English verb-noun word dyads, human judge(1-7) in :
  • Cloze Probability
  • Meaningfulness
  • Familiarity
  • Metaphoricity
    
    

Experimental Methodology

• Learning Linear mappings between geometric features and human scores, Logistic Regressions designed to predict metaphoric class

• Wiki $\Rightarrow$ Co-occurrence matrix $\Rightarrow$ as Base space $\Rightarrow$ 200d Subspace $\Rightarrow$ Word vector projection $\Rightarrow$ Compute geometry features

• 48$d$ + 3 human judgements $\Rightarrow$ Least Squares Regression

• normalized matrix of geometry features $\Rightarrow$ LR $\Rightarrow$ metaphoric class(choose highest out of 3)

• Use variance inflation factor to eliminate collinearity

Rusults

Conclusion

• Metaphor involves shifting a concept to suit a situation, and new meaning is produced

• Model follows a regular progression from literal to conventional to novel compositions

  • wish happiness, raise happiness, and collect happiness
  • enjoy wonder, provoke wonder, and murder wonder
    $\Rightarrow$Metaphoricity ↑, meaningfulness and familiarity ↓

• Practical applications in neurolinguistic and clinical contexts