Transformers have emerged as dominant play- ers in various scientific fields, especially NLP. However, their inner workings, like many other neural networks, remain opaque. In spite of the widespread use of model-agnostic interpretability techniques, including gradient-based and occlusion-based, their shortcomings are becoming increasingly apparent for Transformer interpretation, making the field of interpretability more demanding today. In this tutorial, we will present Transformer-specific interpretability methods, a new trending approach, that make use of specific features of the Transformer architecture and are deemed more promising for understanding Transformer-based models. We start by discussing the potential pitfalls and misleading results model-agnostic approaches may produce when interpreting Transformers. Next, we discuss Transformer-specific methods, including those designed to quantify context- mixing interactions among all input pairs (as the fundamental property of the Transformer architecture) and those that combine causal methods with low-level Transformer analysis to identify particular subnetworks within a model that are responsible for specific tasks. By the end of the tutorial, we hope participants will understand the advantages (as well as current limitations) of Transformer-specific interpretability methods, along with how these can be applied to their own research.

Recent studies have shown that transformer models like BERT rely on number information encoded in their representations of sentences’ subjects and head verbs when performing subject-verb agreement. However, probing experiments suggest that subject number is also encoded in the representations of all words in such sentences. In this paper, we use causal interventions to show that BERT only uses the subject plurality information encoded in its representations of the subject and words that agree with it in number. We also demonstrate that current probing metrics are unable to determine which words’ representations contain functionally relevant information. This both provides a revised view of subject-verb agreement in language models, and suggests potential pitfalls for current probe usage and evaluation.

Language models (LMs) exhibit and amplify many types of undesirable biases learned from the training data, including gender bias. However, we lack tools for effectively and efficiently changing this behavior without hurting general language modeling performance. In this paper, we study three methods for identifying causal relations between LM components and particular output: causal mediation analysis, automated circuit discovery and our novel, efficient method called DiffMask+ based on differential masking. We apply the methods to GPT-2 small and the problem of gender bias, and use the discovered sets of components to perform parameter-efficient fine-tuning for bias mitigation. Our results show significant overlap in the identified components (despite huge differences in the computational requirements of the methods) as well as success in mitigating gender bias, with less damage to general language modeling compared to full model fine-tuning. However, our work also underscores the difficulty of defining and measuring bias, and the sensitivity of causal discovery procedures to dataset choice. We hope our work can contribute to more attention for dataset development, and lead to more effective mitigation strategies for other types of bias.

Animacy—whether an entity is alive and sentient—is fundamental to cognitive processing, impacting areas such as memory, vision, and language. However, animacy is not always expressed directly in language: in English it often manifests indirectly, in the form of selectional constraints on verbs and adjectives. This poses a potential issue for transformer language models (LMs): they often train only on text, and thus lack access to extralinguistic information from which humans learn about animacy. We ask: how does this impact LMs’ animacy processing—do they still behave as humans do? We answer this question using open-source LMs. Like previous studies, we find that LMs behave much like humans when presented with entities whose animacy is typical. However, we also show that even when presented with stories about atypically animate entities, such as a peanut in love, LMs adapt: they treat these entities as animate, though they do not adapt as well as humans. Even when the context indicating atypical animacy is very short, LMs pick up on subtle clues and change their behavior. We conclude that despite the limited signal through which LMs can learn about animacy, they are indeed sensitive to the relevant lexical semantic nuances available in English.

Artificial agents are nowadays challenged to perform embodied AI tasks. To succeed, agents must understand the meaning of verbs and how their corresponding actions transform the surrounding world. In this work, we propose ACT-Thor, a novel controlled benchmark for embodied action understanding. We use the AI2-THOR simulated environment to produce a controlled setup in which an agent, given a before-image and an associated action command, has to determine what the correct after-image is among a set of possible candidates. First, we assess the feasibility of the task via a human evaluation that resulted in 81.4% accuracy, and very high inter-annotator agreement (84.9%). Second, we design both unimodal and multimodal baselines, using state-of-the-art visual feature extractors. Our evaluation and error analysis suggest that only models that have a very structured representation of the actions together with powerful visual features can perform well on the task. However, they still fall behind human performance in a zero-shot scenario where the model is exposed to unseen (action, object) pairs. This paves the way for a systematic way of evaluating embodied AI agents that understand grounded actions.

The high performance of large pretrained language models (LLMs) such as BERT on NLP tasks has prompted questions about BERT’s linguistic capabilities, and how they differ from humans’. In this paper, we approach this question by examining BERT’s knowledge of lexical semantic relations. We focus on hypernymy, the “is-a” relation that relates a word to a superordinate category. We use a prompting methodology to simply ask BERT what the hypernym of a given word is. We find that, in a setting where all hypernyms are guessable via prompting, BERT knows hypernyms with up to 57% accuracy. Moreover, BERT with prompting outperforms other unsupervised models for hypernym discovery even in an unconstrained scenario. However, BERT’s predictions and performance on a dataset containing uncommon hyponyms and hypernyms indicate that its knowledge of hypernymy is still limited.

BERTScore, a recently proposed automatic metric for machine translation quality, uses BERT, a large pre-trained language model to evaluate candidate translations with respect to a gold translation. Taking advantage of BERT’s semantic and syntactic abilities, BERTScore seeks to avoid the flaws of earlier approaches like BLEU, instead scoring candidate translations based on their semantic similarity to the gold sentence. However, BERT is not infallible; while its performance on NLP tasks set a new state of the art in general, studies of specific syntactic and semantic phenomena have shown where BERT’s performance deviates from that of humans more generally. This naturally raises the questions we address in this paper: what are the strengths and weaknesses of BERTScore? Do they relate to known weaknesses on the part of BERT? We find that while BERTScore can detect when a candidate differs from a reference in important content words, it is less sensitive to smaller errors, especially if the candidate is lexically or stylistically similar to the reference.