Explaining Contextualization in Language Models using Visual Analytics

Abstract

Despite the success of contextualized language models on various NLP tasks, it is still unclear what these models really learn. In this paper, we contribute to the current efforts of explaining such models by exploring the continuum between function and content words with respect to contextualization in BERT, based on linguistically-informed insights. In particular, we utilize scoring and visual analytics techniques: we use an existing similarity-based score to measure contextualization and integrate it into a novel visual analytics technique, presenting the model's layers simultaneously and highlighting intra-layer properties and inter-layer differences. We show that contextualization is neither driven by polysemy nor by pure context variation. We also provide insights on why BERT fails to model words in the middle of the functionality continuum. * Contribution to the visualization part. † Equal contribution to the computational linguistics part.

Interactive optimization of embedding-based text similarity calculations

“…43 More specifically, this means that many words will need to have several different embeddings so that a context-dependent choice can be made for each situation. 44 To achieve this, the use of deep learning models is a popular choice, and approaches have, for example, been developed for recursive neural networks, 45 convolutional neural networks, 46 and recurrent neural networks. 47 Arguably, the current state-of-the-art technology for text embedding is the Universal Sentence Encoder (USE), 48 but the previously mentioned BERT algorithm also works for text of sentence length.…”

“…43 More specifically, this means that many words will need to have several different embeddings so that a context-dependent choice can be made for each situation. 44 To achieve this, the Figure 1. Using the EEVO tool to visualize the performance of embedding-based ensembles conducting text similarity calculations on a large set of scientific publications (see further Section Visualization).…”

LMFingerprints: Visual Explanations of Language Model Embedding Spaces through Layerwise Contextualization Scores

“…Most common explainability techniques either use supervised probing methods, i.e., linear classification models predicting specific linguistic properties (e.g., [Eth19]), or apply adversarial testing to conclude about models' capability of learning specific context properties (e.g., [MPL19]). However, the findings of these two strands of research are often contradictory [SKB∗21]. At the same time, visual analytics approaches are used for the explainability of embedding contextualization.…”

“…A further, main contribution of this paper is an interactive explanation workspace that visualizes the computed score values. The visual representation of the scores is crucial due to the huge amount of data that is generated and has to be investigated, and because the embedding contextualization differs depending on the token's role (e.g., meaning or function) in its context [SKB∗21]. Visualizations are effective means for generating insights into such (complex) data patterns [KAF∗08].…”

Untitled

“…Contextualized LMs are usually pre-trained on a language modeling task (e.g., next word prediction) and are used as transfer-learning methods in other NLP tasks [64]. Adaptation to tasks is typically carried out through fine-tuning of the model, or part of it, on domain-specific data.…”