Classifiers trained on auxiliary probing tasks are a popular tool to analyze the representations learned by neural sentence encoders such as BERT and ELMo. While many authors are aware of the difficulty to distinguish between “extracting the linguistic structure encoded in the representations” and “learning the probing task,” the validity of probing methods calls for further research. Using a neighboring word identity prediction task, we show that the token embeddings learned by neural sentence encoders contain a significant amount of information about the exact linear context of the token, and hypothesize that, with such information, learning standard probing tasks may be feasible even without additional linguistic structure. We develop this hypothesis into a framework in which analysis efforts can be scrutinized and argue that, with current models and baselines, conclusions that representations contain linguistic structure are not well-founded. Current probing methodology, such as restricting the classifier’s expressiveness or using strong baselines, can help to better estimate the complexity of learning, but not build a foundation for speculations about the nature of the linguistic structure encoded in the learned representations.

Previous work on probing word representations for linguistic knowledge has focused on interpolation tasks. In this paper, we instead analyse probes in an extrapolation setting, where the inputs at test time are deliberately chosen to be ‘harder’ than the training examples. We argue that such an analysis can shed further light on the open question whether probes actually decode linguistic knowledge, or merely learn the diagnostic task from shallow features. To quantify the hardness of an example, we consider scoring functions based on linguistic, statistical, and learning-related criteria, all of which are applicable to a broad range of NLP tasks. We discuss the relative merits of these criteria in the context of two syntactic probing tasks, part-of-speech tagging and syntactic dependency labelling. From our theoretical and experimental analysis, we conclude that distance-based and hard statistical criteria show the clearest differences between interpolation and extrapolation settings, while at the same time being transparent, intuitive, and easy to control.

Probing studies have extensively explored where in neural language models linguistic information is located. The standard approach to interpreting the results of a probing classifier is to focus on the layers whose representations give the highest performance on the probing task. We propose an alternative method that asks where the task-relevant information emerges in the model. Our framework consists of a family of metrics that explicitly model local information gain relative to the previous layer and each layer’s contribution to the model’s overall performance. We apply the new metrics to two pairs of syntactic probing tasks with different degrees of complexity and find that the metrics confirm the expected ordering only for one of the pairs. Our local metrics show a massive dominance of the first layers, indicating that the features that contribute the most to our probing tasks are not as high-level as global metrics suggest.

Models able to generate free-text rationales that explain their output have been proposed as an important step towards interpretable NLP for “reasoning” tasks such as natural language inference and commonsense question answering. However, the relative merits of different architectures and types of rationales are not well understood and hard to measure. In this paper, we contribute two insights to this line of research: First, we find that models trained on gold explanations learn to rely on these but, in the case of the more challenging question answering data set we use, fail when given generated explanations at test time. However, additional fine-tuning on generated explanations teaches the model to distinguish between reliable and unreliable information in explanations. Second, we compare explanations by a generation-only model to those generated by a self-rationalizing model and find that, while the former score higher in terms of validity, factual correctness, and similarity to gold explanations, they are not more useful for downstream classification. We observe that the self-rationalizing model is prone to hallucination, which is punished by most metrics but may add useful context for the classification step.