Disentangling Linguistic Features with Dimension-Wise Analysis of Vector Embeddings (2504.14766v1)

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 LLMs, with implications for model bias mitigation and the responsible deployment of AI systems.

Disentangling Linguistic Features in BERT Embeddings through Dimension-Wise Analysis

The paper "Disentangling Linguistic Features with Dimension-Wise Analysis of Vector Embeddings" offers a comprehensive framework to uncover which dimensions of vector embeddings encode distinct linguistic properties (LPs) using models such as BERT, GPT-2, and MPNet. The authors introduce a novel dataset, the Linguistically Distinct Sentence Pairs (LDSP-10), to isolate ten key LPs like synonymy, negation, tense, and quantity. Through this dataset, the paper employs a spectrum of statistical methods, including the Wilcoxon signed-rank test, mutual information, and recursive feature elimination, to evaluate the dimensional significance of various linguistic features.

Key to the findings is the introduction of the Embedding Dimension Importance (EDI) score, a metric designed to quantify the relevance of embedding dimensions to specific LPs. Utilizing the LDSP-10 dataset, the paper reveals that certain properties such as negation and polarity demonstrate robust encoding in distinct dimensions, whereas others like synonymy present more diffuse patterns.

The research significantly contributes to the interpretability of embeddings. For instance, the negation property yielded one of the highest EDI scores among tested features, affirming its strong and consistent representation across embedding dimensions. Conversely, the synonym category did not isolate any dimensions with significant encoding capacity, highlighting the inherent complexity in representing semantic similarity through discrete embeddings.

The approach is not without challenges, notably in quality assurance of the LDSP-10 database, which requires substantial input from LLM APIs for generation of high-quality outputs. Moreover, while the paper robustly handles smaller models, expanding analyses to more extensive, state-of-the-art networks could present further insights.

Practically, this framework allows for bias mitigation and customization in LLMs by offering a clearer view into the linguistic intricacies represented in embeddings. Theoretically, it paves the way for more transparent and refined LLM architectures, fostering further exploration into bias reduction and ethical AI deployment.

Future investigation could leverage these findings to refine embedding techniques in burgeoning AI models, extending the interpretative lenses offered in this paper to assist in the modular understanding of language representations. Additionally, analysis across various layers could provide a more granular understanding of how LPs propagate through model architectures, bolstering interpretability and enabling tailored interventions that preserve desired linguistic encodings while mitigating bias.