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.

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.

Self-attention weights and their transformed variants have been the main source of information for analyzing token-to-token interactions in Transformer-based models. But despite their ease of interpretation, these weights are not faithful to the models’ decisions as they are only one part of an encoder, and other components in the encoder layer can have considerable impact on information mixing in the output representations. In this work, by expanding the scope of analysis to the whole encoder block, we propose Value Zeroing, a novel context mixing score customized for Transformers that provides us with a deeper understanding of how information is mixed at each encoder layer. We demonstrate the superiority of our context mixing score over other analysis methods through a series of complementary evaluations with different viewpoints based on linguistically informed rationales, probing, and faithfulness analysis.

Transformers have become a key architecture in speech processing, but our understanding of how they build up representations of acoustic and linguistic structure is limited. In this study, we address this gap by investigating how measures of ‘context-mixing’ developed for text models can be adapted and applied to models of spoken language. We identify a linguistic phenomenon that is ideal for such a case study: homophony in French (e.g. livre vs livres), where a speech recognition model has to attend to syntactic cues such as determiners and pronouns in order to disambiguate spoken words with identical pronunciations and transcribe them while respecting grammatical agreement. We perform a series of controlled experiments and probing analyses on Transformer-based speech models. Our findings reveal that representations in encoder-only models effectively incorporate these cues to identify the correct transcription, whereas encoders in encoder-decoder models mainly relegate the task of capturing contextual dependencies to decoder modules.

We study feature interactions in the context of feature attribution methods for post-hoc interpretability. In interpretability research, getting to grips with feature interactions is increasingly recognised as an important challenge, because interacting features are key to the success of neural networks. Feature interactions allow a model to build up hierarchical representations for its input, and might provide an ideal starting point for the investigation into linguistic structure in language models. However, uncovering the exact role that these interactions play is also difficult, and a diverse range of interaction attribution methods has been proposed. In this paper, we focus on the question which of these methods most faithfully reflects the inner workings of the target models. We work out a grey box methodology, in which we train models to perfection on a formal language classification task, using PCFGs. We show that under specific configurations, some methods are indeed able to uncover the grammatical rules acquired by a model. Based on these findings we extend our evaluation to a case study on language models, providing novel insights into the linguistic structure that these models have acquired.

We present a setup for training, evaluating and interpreting neural language models, that uses artificial, language-like data. The data is generated using a massive probabilistic grammar (based on state-split PCFGs), that is itself derived from a large natural language corpus, but also provides us complete control over the generative process. We describe and release both grammar and corpus, and test for the naturalness of our generated data. This approach allows us define closed-form expressions to efficiently compute exact lower bounds on obtainable perplexity using both causal and masked language modelling. Our results show striking differences between neural language modelling architectures and training objectives in how closely they allow approximating the lower bound on perplexity. Our approach also allows us to directly compare learned representations to symbolic rules in the underlying source. We experiment with various techniques for interpreting model behaviour and learning dynamics. With access to the underlying true source, our results show striking differences and outcomes in learning dynamics between different classes of words.

Detecting and mitigating harmful biases in modern language models are widely recognized as crucial, open problems. In this paper, we take a step back and investigate how language models come to be biased in the first place. We use a relatively small language model, using the LSTM architecture trained on an English Wikipedia corpus. With full access to the data and to the model parameters as they change during every step while training, we can map in detail how the representation of gender develops, what patterns in the dataset drive this, and how the model’s internal state relates to the bias in a downstream task (semantic textual similarity).We find that the representation of gender is dynamic and identify different phases during training. Furthermore, we show that gender information is represented increasingly locally in the input embeddings of the model and that, as a consequence, debiasing these can be effective in reducing the downstream bias. Monitoring the training dynamics, allows us to detect an asymmetry in how the female and male gender are represented in the input embeddings. This is important, as it may cause naive mitigation strategies to introduce new undesirable biases. We discuss the relevance of the findings for mitigation strategies more generally and the prospects of generalizing our methods to larger language models, the Transformer architecture, other languages and other undesirable biases.

We investigate the extent to which modern neural language models are susceptible to structural priming, the phenomenon whereby the structure of a sentence makes the same structure more probable in a follow-up sentence. We explore how priming can be used to study the potential of these models to learn abstract structural information, which is a prerequisite for good performance on tasks that require natural language understanding skills. We introduce a novel metric and release Prime-LM, a large corpus where we control for various linguistic factors that interact with priming strength. We find that Transformer models indeed show evidence of structural priming, but also that the generalizations they learned are to some extent modulated by semantic information. Our experiments also show that the representations acquired by the models may not only encode abstract sequential structure but involve certain level of hierarchical syntactic information. More generally, our study shows that the priming paradigm is a useful, additional tool for gaining insights into the capacities of language models and opens the door to future priming-based investigations that probe the model’s internal states.1

In the Transformer model, “self-attention” combines information from attended embeddings into the representation of the focal embedding in the next layer. Thus, across layers of the Transformer, information originating from different tokens gets increasingly mixed. This makes attention weights unreliable as explanations probes. In this paper, we consider the problem of quantifying this flow of information through self-attention. We propose two methods for approximating the attention to input tokens given attention weights, attention rollout and attention flow, as post hoc methods when we use attention weights as the relative relevance of the input tokens. We show that these methods give complementary views on the flow of information, and compared to raw attention, both yield higher correlations with importance scores of input tokens obtained using an ablation method and input gradients.

Interpreting the inner workings of neural models is a key step in ensuring the robustness and trustworthiness of the models, but work on neural network interpretability typically faces a trade-off: either the models are too constrained to be very useful, or the solutions found by the models are too complex to interpret. We propose a novel strategy for achieving interpretability that – in our experiments – avoids this trade-off. Our approach builds on the success of using probability as the central quantity, such as for instance within the attention mechanism. In our architecture, DoLFIn (Distributions over Latent Features for Interpretability), we do no determine beforehand what each feature represents, and features go altogether into an unordered set. Each feature has an associated probability ranging from 0 to 1, weighing its importance for further processing. We show that, unlike attention and saliency map approaches, this set-up makes it straight-forward to compute the probability with which an input component supports the decision the neural model makes. To demonstrate the usefulness of the approach, we apply DoLFIn to text classification, and show that DoLFIn not only provides interpretable solutions, but even slightly outperforms the classical CNN and BiLSTM text classifiers on the SST2 and AG-news datasets.

In this paper, we define and apply representational stability analysis (ReStA), an intuitive way of analyzing neural language models. ReStA is a variant of the popular representational similarity analysis (RSA) in cognitive neuroscience. While RSA can be used to compare representations in models, model components, and human brains, ReStA compares instances of the same model, while systematically varying single model parameter. Using ReStA, we study four recent and successful neural language models, and evaluate how sensitive their internal representations are to the amount of prior context. Using RSA, we perform a systematic study of how similar the representational spaces in the first and second (or higher) layers of these models are to each other and to patterns of activation in the human brain. Our results reveal surprisingly strong differences between language models, and give insights into where the deep linguistic processing, that integrates information over multiple sentences, is happening in these models. The combination of ReStA and RSA on models and brains allows us to start addressing the important question of what kind of linguistic processes we can hope to observe in fMRI brain imaging data. In particular, our results suggest that the data on story reading from Wehbe et al./ (2014) contains a signal of shallow linguistic processing, but show no evidence on the more interesting deep linguistic processing.

Extensive research has recently shown that recurrent neural language models are able to process a wide range of grammatical phenomena. How these models are able to perform these remarkable feats so well, however, is still an open question. To gain more insight into what information LSTMs base their decisions on, we propose a generalisation of Contextual Decomposition (GCD). In particular, this setup enables us to accurately distil which part of a prediction stems from semantic heuristics, which part truly emanates from syntactic cues and which part arise from the model biases themselves instead. We investigate this technique on tasks pertaining to syntactic agreement and co-reference resolution and discover that the model strongly relies on a default reasoning effect to perform these tasks.

How do neural language models keep track of number agreement between subject and verb? We show that ‘diagnostic classifiers’, trained to predict number from the internal states of a language model, provide a detailed understanding of how, when, and where this information is represented. Moreover, they give us insight into when and where number information is corrupted in cases where the language model ends up making agreement errors. To demonstrate the causal role played by the representations we find, we then use agreement information to influence the course of the LSTM during the processing of difficult sentences. Results from such an intervention reveal a large increase in the language model’s accuracy. Together, these results show that diagnostic classifiers give us an unrivalled detailed look into the representation of linguistic information in neural models, and demonstrate that this knowledge can be used to improve their performance.

In this paper we describe FragmentSeeker, a tool which is capable to identify all those tree constructions which are recurring multiple times in a large Phrase Structure treebank. The tool is based on an efficient kernel-based dynamic algorithm, which compares every pair of trees of a given treebank and computes the list of fragments which they both share. We describe two different notions of fragments we will use, i.e. standard and partial fragments, and provide the implementation details on how to extract them from a syntactically annotated corpus. We have tested our system on the Penn Wall Street Journal treebank for which we present quantitative and qualitative analysis on the obtained recurring structures, as well as provide empirical time performance. Finally we propose possible ways our tool could contribute to different research fields related to corpus analysis and processing, such as parsing, corpus statistics, annotation guidance, and automatic detection of argument structure.