A Rule-based/BPSO Approach to Produce Low-dimensional Semantic Basis Vectors Set

TL;DR

This paper introduces a novel rule-based and BPSO-driven method to generate low-dimensional, interpretable semantic vectors by selecting context words based on specific features, improving correlation with human judgments.

Contribution

It proposes a new approach combining decision trees and binary particle swarm optimization to create low-dimensional explicit semantic vectors with enhanced interpretability and performance.

Findings

Improved Spearman correlation on MEN, RG-65, and SimLex-999 datasets.

Effective selection of context words enhances semantic vector quality.

Demonstrated superiority over baseline methods using fixed window co-occurrence.

Abstract

We intend to generate low-dimensional explicit distributional semantic vectors. In explicit semantic vectors, each dimension corresponds to a word, so word vectors are interpretable. In this research, we propose a new approach to obtain low-dimensional explicit semantic vectors. First, the proposed approach considers the three criteria Word Similarity, Number of Zero, and Word Frequency as features for the words in a corpus. Then, we extract some rules for obtaining the initial basis words using a decision tree that is drawn based on the three features. Second, we propose a binary weighting method based on the Binary Particle Swarm Optimization algorithm that obtains N_B = 1000 context words. We also use a word selection method that provides N_S = 1000 context words. Third, we extract the golden words of the corpus based on the binary weighting method. Then, we add the extracted golden words to the context words that are selected by the word selection method as the golden context words. We use the ukWaC corpus for constructing the word vectors. We use MEN, RG-65, and SimLex-999 test sets to evaluate the word vectors. We report the results compared to a baseline that uses 5k most frequent words in the corpus as context words. The baseline method uses a fixed window to count the co-occurrences. We obtain the word vectors using the 1000 selected context words together with the golden context words. Our approach compared to the Baseline method increases the Spearman correlation coefficient for the MEN, RG-65, and SimLex-999 test sets by 4.66%, 14.73%, and 1.08%, respectively.