Generalization refers to the ability of machine learning models to perform well on dataset distributions different from the one it was trained on. While several pre-existing works have characterized the generalizability of NLP models across different dimensions, such as domain shift, adversarial perturbations, or compositional variations, most studies were carried out in a stand-alone setting, emphasizing a single dimension of interest. We bridge this gap by systematically investigating the generalizability of pre-trained language models across different architectures, sizes, and training strategies, over multiple dimensions for the task of natural language inference and question answering. Our results indicate that model instances typically exhibit consistent generalization trends, i.e., they generalize equally well (or poorly) across most scenarios, and this ability is correlated with model architecture, base dataset performance, size, and training mechanism. We hope this research motivates further work in a) developing a multi-dimensional generalization benchmark for systematic evaluation and b) examining the reasons behind models’ generalization abilities. The code and data are available at https://github.com/sagnik/md-gen-nlp, and the trained models are released at https://huggingface.co/varun-v-rao.

This paper reports findings from the First Shared Task on Automated Prediction of Difficulty and Response Time for Multiple-Choice Questions. The task was organized as part of the 19th Workshop on Innovative Use of NLP for Building Educational Applications (BEA’24), held in conjunction with NAACL 2024, and called upon the research community to contribute solutions to the problem of modeling difficulty and response time for clinical multiple-choice questions (MCQs). A set of 667 previously used and now retired MCQs from the United States Medical Licensing Examination (USMLE®) and their corresponding difficulties and mean response times were made available for experimentation. A total of 17 teams submitted solutions and 12 teams submitted system report papers describing their approaches. This paper summarizes the findings from the shared task and analyzes the main approaches proposed by the participants.

Edge probing tests are classification tasks that test for grammatical knowledge encoded in token representations coming from contextual encoders such as large language models (LLMs). Many LLM encoders have shown high performance in EP tests, leading to conjectures about their ability to encode linguistic knowledge. However, a large body of research claims that the tests necessarily do not measure the LLM’s capacity to encode knowledge, but rather reflect the classifiers’ ability to learn the problem. Much of this criticism stems from the fact that often the classifiers have very similar accuracy when an LLM vs a random encoder is used. Consequently, several modifications to the tests have been suggested, including information theoretic probes. We show that commonly used edge probing test datasets have various biases including memorization. When these biases are removed, the LLM encoders do show a significant difference from the random ones, even with the simple non-information theoretic probes.

Reasoning over spans of tokens from different parts of the input is essential for natural language understanding (NLU) tasks such as fact-checking (FC), machine reading comprehension (MRC) or natural language inference (NLI). However, existing highlight-based explanations primarily focus on identifying individual important features or interactions only between adjacent tokens or tuples of tokens. Most notably, there is a lack of annotations capturing the human decision-making process with respect to the necessary interactions for informed decision-making in such tasks. To bridge this gap, we introduce SpanEx, a multi-annotator dataset of human span interaction explanations for two NLU tasks: NLI and FC. We then investigate the decision-making processes of multiple fine-tuned large language models in terms of the employed connections between spans in separate parts of the input and compare them to the human reasoning processes. Finally, we present a novel community detection based unsupervised method to extract such interaction explanations. We make the code and the dataset available on [Github](https://github.com/copenlu/spanex). The dataset is also available on [Huggingface datasets](https://huggingface.co/datasets/copenlu/spanex).

Two of the most fundamental issues in Natural Language Understanding (NLU) at present are: (a) how it can established whether deep learning-based models score highly on NLU benchmarks for the ”right” reasons; and (b) what those reasons would even be. We investigate the behavior of reading comprehension models with respect to two linguistic ”skills”: coreference resolution and comparison. We propose a definition for the reasoning steps expected from a system that would be ”reading slowly”, and compare that with the behavior of five models of the BERT family of various sizes, observed through saliency scores and counterfactual explanations. We find that for comparison (but not coreference) the systems based on larger encoders are more likely to rely on the ”right” information, but even they struggle with generalization, suggesting that they still learn specific lexical patterns rather than the general principles of comparison.

There have been many efforts to try to understand what grammatical knowledge (e.g., ability to understand the part of speech of a token) is encoded in large pre-trained language models (LM). This is done through ‘Edge Probing’ (EP) tests: supervised classification tasks to predict the grammatical properties of a span (whether it has a particular part of speech) using only the token representations coming from the LM encoder. However, most NLP applications fine-tune these LM encoders for specific tasks. Here, we ask: if an LM is fine-tuned, does the encoding of linguistic information in it change, as measured by EP tests? Specifically, we focus on the task of Question Answering (QA) and conduct experiments on multiple datasets. We find that EP test results do not change significantly when the fine-tuned model performs well or in adversarial situations where the model is forced to learn wrong correlations. From a similar finding, some recent papers conclude that fine-tuning does not change linguistic knowledge in encoders but they do not provide an explanation. We find that EP models are susceptible to exploiting spurious correlations in the EP datasets. When this dataset bias is corrected, we do see an improvement in the EP test results as expected.

Complex natural language understanding modules in dialog systems have a richer understanding of user utterances, and thus are critical in providing a better user experience. However, these models are often created from scratch, for specific clients and use cases and require the annotation of large datasets. This encourages the sharing of annotated data across multiple clients. To facilitate this we introduce the idea of intent features: domain and topic agnostic properties of intents that can be learnt from the syntactic cues only, and hence can be shared. We introduce a new neural network architecture, the Global-Local model, that shows significant improvement over strong baselines for identifying these features in a deployed, multi-intent natural language understanding module, and more generally in a classification setting where a part of an utterance has to be classified utilizing the whole context.

Current state-of-the-art models for named entity recognition (NER) are neural models with a conditional random field (CRF) as the final layer. Entities are represented as per-token labels with a special structure in order to decode them into spans. Current work eschews prior knowledge of how the span encoding scheme works and relies on the CRF learning which transitions are illegal and which are not to facilitate global coherence. We find that by constraining the output to suppress illegal transitions we can train a tagger with a cross-entropy loss twice as fast as a CRF with differences in F1 that are statistically insignificant, effectively eliminating the need for a CRF. We analyze the dynamics of tag co-occurrence to explain when these constraints are most effective and provide open source implementations of our tagger in both PyTorch and TensorFlow.

We introduce Baseline: a library for reproducible deep learning research and fast model development for NLP. The library provides easily extensible abstractions and implementations for data loading, model development, training and export of deep learning architectures. It also provides implementations for simple, high-performance, deep learning models for various NLP tasks, against which newly developed models can be compared. Deep learning experiments are hard to reproduce, Baseline provides functionalities to track them. The goal is to allow a researcher to focus on model development, delegating the repetitive tasks to the library.