Visually Analyzing Contextualized Embeddings

TL;DR

This paper presents a visual analysis method for understanding what linguistic information is captured by deep neural language models through unsupervised clustering and visualization techniques.

Contribution

It introduces an unsupervised, visualization-based approach to analyze contextualized embeddings, moving beyond traditional supervised probing methods.

Findings

User feedback shows the visualization aids in discovering linguistic structures.

Clustering reveals different types of linguistic information in embeddings.

The method provides insights into the functionality and relationships of embedding clusters.

Abstract

In this paper we introduce a method for visually analyzing contextualized embeddings produced by deep neural network-based language models. Our approach is inspired by linguistic probes for natural language processing, where tasks are designed to probe language models for linguistic structure, such as parts-of-speech and named entities. These approaches are largely confirmatory, however, only enabling a user to test for information known a priori. In this work, we eschew supervised probing tasks, and advocate for unsupervised probes, coupled with visual exploration techniques, to assess what is learned by language models. Specifically, we cluster contextualized embeddings produced from a large text corpus, and introduce a visualization design based on this clustering and textual structure - cluster co-occurrences, cluster spans, and cluster-word membership - to help elicit the functionality of, and relationship between, individual clusters. User feedback highlights the benefits of our design in discovering different types of linguistic structures.