Disentangling Linguistic Features with Dimension-Wise Analysis of Vector Embeddings

2025-04-20

AI Summary

Neural embedding models like BERT suffer from limited interpretability and poorly understood mechanisms for encoding linguistic attributes. Method: We propose a dimension-level interpretability analysis framework, introducing (i) LDSP-10—the first fine-grained, linguistically annotated dataset for disentangling language features; (ii) Embedding Dimension Importance (EDI), a novel metric quantifying dimension-specific contributions; and (iii) a multimodal attribution analysis combining Wilcoxon signed-rank tests, mutual information estimation, and recursive feature elimination. Results: We find that certain linguistic properties—e.g., negation and polarity—are robustly encoded in a small, consistent subset of dimensions, whereas synonymy exhibits distributed representation across many dimensions. Moreover, several dimensions demonstrate cross-task selectivity for specific linguistic attributes. These findings establish a new paradigm for interpretable embedding modeling and provide an actionable foundation for bias diagnosis and controllable embedding editing.

Technology Category

• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Machine Learning: Dimensionality Reduction/Feature Selection
• Computer Vision: Language and Vision

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Understanding the inner workings of neural embeddings, particularly in models such as BERT, remains a challenge because of their high-dimensional and opaque nature. This paper proposes a framework for uncovering the specific dimensions of vector embeddings that encode distinct linguistic properties (LPs). We introduce the Linguistically Distinct Sentence Pairs (LDSP-10) dataset, which isolates ten key linguistic features such as synonymy, negation, tense, and quantity. Using this dataset, we analyze BERT embeddings with various methods, including the Wilcoxon signed-rank test, mutual information, and recursive feature elimination, to identify the most influential dimensions for each LP. We introduce a new metric, the Embedding Dimension Impact (EDI) score, which quantifies the relevance of each embedding dimension to a LP. Our findings show that certain properties, such as negation and polarity, are robustly encoded in specific dimensions, while others, like synonymy, exhibit more complex patterns. This study provides insights into the interpretability of embeddings, which can guide the development of more transparent and optimized language models, with implications for model bias mitigation and the responsible deployment of AI systems.

Problem

• Identify embedding dimensions encoding distinct linguistic properties
• Analyze BERT embeddings to uncover influential linguistic feature dimensions
• Quantify dimension relevance to linguistic features using EDI score

Innovation

• Dimension-wise analysis of BERT embeddings
• Linguistically Distinct Sentence Pairs dataset
• Embedding Dimension Impact score metric