Surprisal theory links human processing effort to the predictability of an upcoming linguistic unit, but empirical work often leaves the notion of a unit underspecified.In practice, experimental stimuli are segmented into linguistically motivated units (e.g., words), while pretrained language models assign probability mass to a fixed token alphabet that typically does not align with those units.As a result, surprisal-based predictors depend implicitly on ad hoc procedures that conflate two distinct modeling choices: the definition of the unit of analysis and the choice of regions of interest over which predictions are evaluated.In this paper, we disentangle these choices and give a unified framework for reasoning about surprisal over arbitrary unit inventories.We argue that surprisal-based analyses should make these choices explicit and treat tokenization as an implementation detail rather than a scientific primitive.https://github.com/samuki/ units-surprisal

When a vision–language model (VLM) is prompted to identify an entity depicted in an image, it may answer "I see a conifer," rather than the specific label NORWAY SPRUCE. This raises two issues for evaluation: Firstly, the unconstrained generated text needs to be mapped to the evaluation label space (i.e., CONIFER). Secondly, a useful classification measure should give partial credit to lessspecific, but not incorrect, answers (NORWAY SPRUCE being a type of CONIFER). To meet these requirements, we propose a framework for evaluating unconstrained text predictions such as those generated from a vision–language model against a taxonomy. Specifically, we propose the use of hierarchical precision and recall measures to assess the level of correctness and specificity of predictions with regard to a taxonomy. Experimentally, we first show that existing text similarity measures do not capture taxonomic similarity well. We then develop and compare different methods to map textual VLM predictions onto a taxonomy. This allows us to compute hierarchical similarity measures between the generated text and the ground truth labels. Finally, we analyze modern VLMs on fine-grained visual classification tasks based on our proposed taxonomic evaluation scheme. Data and code are made available at https://github.com/vesteinn/vlm-eval.

Understanding and manipulating the causal generation mechanisms in language models is essential for controlling their behavior. Previous work has primarily relied on techniques such as representation surgery---e.g., model ablations or manipulation of linear subspaces tied to specific concepts---to intervene on these models. To understand the impact of interventions precisely, it is useful to examine counterfactuals---e.g., how a given sentence would have appeared had it been generated by the model following a specific intervention. We highlight that counterfactual reasoning is conceptually distinct from interventions, as articulated in Pearl's causal hierarchy. Based on this observation, we propose a framework for generating true string counterfactuals by reformulating language models as a structural equation model using the Gumbel-max trick, which we called Gumbel counterfactual generation. This reformulation allows us to model the joint distribution over original strings and their counterfactuals resulting from the same instantiation of the sampling noise. We develop an algorithm based on hindsight Gumbel sampling that allows us to infer the latent noise variables and generate counterfactuals of observed strings. Our experiments demonstrate that the approach produces meaningful counterfactuals while at the same time showing that commonly used intervention techniques have considerable undesired side effects.

Kevin Du, Vésteinn Snæbjarnarson, Niklas Stoehr, Jennifer White, Aaron Schein, Ryan Cotterell. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Open-set recognition (OSR) methods aim to identify whether or not a test example belongs to a category observed during training. Depending on how visually similar a test example is to the training categories, the OSR task can be easy or extremely challenging. However, the vast majority of previous work has studied OSR in the presence of large, coarse-grained semantic shifts. In contrast, many real-world problems are inherently fine-grained, which means that test examples may be highly visually similar to the training categories. Motivated by this observation, we investigate three aspects of OSR: label granularity, similarity between the open- and closed-sets, and the role of hierarchical supervision during training. To study these dimensions, we curate new open-set splits of a large fine-grained visual categorization dataset. Our analysis results in several interesting findings, including: (i) the best OSR method to use is heavily dependent on the degree of semantic shift present, and (ii) hierarchical representation learning can improve coarse-grained OSR, but has little effect on fine-grained OSR performance. To further enhance fine-grained OSR performance, we propose a hierarchy-adversarial learning method to discourage hierarchical structure in the representation space, which results in a perhaps counter-intuitive behaviour, and a relative improvement in fine-grained OSR of up to 2% in AUROC and 7% in AUPR over standard training. Code and data are available: langnico.github. io/fine-grained-osr.

Despite advances using low-rank adapters and quantization, pretraining of large models on consumer hardware has not been possible without model sharding, offloading during training, or per-layer gradient updates. To address these limitations, we propose Low-Rank Adapters for Quantized Training (LoQT), a method for efficiently training quantized models. LoQT uses gradient-based tensor factorization to initialize low-rank trainable weight matrices that are periodically merged into quantized full-rank weight matrices. Our approach is suitable for both pretraining and fine-tuning models. We demonstrate this for language modeling and downstream task adaptation, finding that LoQT enables efficient training of models up to 7B parameters on a 24GB GPU. We also demonstrate the feasibility of training a 13B model using per-layer gradient updates on the same hardware.

The robustness of image classifiers is essential to their deployment in the real world. The ability to assess this resilience to manipulations or deviations from the training data is thus crucial. These modifications have traditionally consisted of minimal changes that still manage to fool classifiers, and modern approaches are increasingly robust to them. Semantic manipulations that modify elements of an image in meaningful ways have thus gained traction for this purpose. However, they have primarily been limited to style, color, or attribute changes. While expressive, these manipulations do not make use of the full capabilities of a pretrained generative model. In this work, we aim to bridge this gap. We show how a pretrained image generator can be used to semantically manipulate images in a detailed, diverse, and photorealistic way while still preserving the class of the original image. Inspired by recent GAN-based image inversion methods, we propose a method called Adversarial Pivotal Tuning (APT). Given an image, APT first finds a pivot latent space input that reconstructs the image using a pretrained generator. It then adjusts the generator’s weights to create small yet semantic manipulations in order to fool a pretrained classifier. APT preserves the full expressive editing capabilities of the generative model. We demonstrate that APT is capable of a wide range of class-preserving semantic image manipulations that fool a variety of pretrained classifiers. Finally, we show that classifiers that are robust to other benchmarks are not robust to APT manipulations and suggest a method to improve them.

Svanhvít Lilja Ingólfsdóttir, Petur Ragnarsson, Haukur Jónsson, Haukur Simonarson, Vilhjalmur Thorsteinsson, Vésteinn Snæbjarnarson. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Recent advances in text-to-image diffusion models have enabled the generation of diverse and high-quality images. While impressive, the images often fall short of depicting subtle details and are susceptible to errors due to ambiguity in the input text. One way of alleviating these issues is to train diffusion models on class-labeled datasets. This approach has two disadvantages: (i) supervised datasets are generally small compared to large-scale scraped text-image datasets on which text-to-image models are trained, affecting the quality and diversity of the generated images, or (ii) the input is a hard-coded label, as opposed to free-form text, limiting the control over the generated images.In this work, we propose a non-invasive fine-tuning technique that capitalizes on the expressive potential of freeform text while achieving high accuracy through discriminative signals from a pretrained classifier. This is done by iteratively modifying the embedding of an added input token of a text-to-image diffusion model, by steering generated images toward a given target class according to a classifier. Our method is fast compared to prior fine-tuning methods and does not require a collection of in-class images or retraining of a noise-tolerant classifier. We evaluate our method extensively, showing that the generated images are: (i) more accurate and of higher quality than standard diffusion models, (ii) can be used to augment training data in a low-resource setting, and (iii) reveal information about the data used to train the guiding classifier. The code is available at https://github.com/idansc/discriminative_class_tokens.

We present the first extractive question answering (QA) dataset for Icelandic, Natural Questions in Icelandic (NQiI). Developing such datasets is important for the development and evaluation of Icelandic QA systems. It also aids in the development of QA methods that need to work for a wide range of morphologically and grammatically different languages in a multilingual setting. The dataset was created by asking contributors to come up with questions they would like to know the answer to. Later, they were tasked with finding answers to each others questions following a previously published methodology. The questions are Natural in the sense that they are real questions posed out of interest in knowing the answer. The complete dataset contains 18 thousand labeled entries of which 5,568 are directly suitable for training an extractive QA system for Icelandic. The dataset is a valuable resource for Icelandic which we demonstrate by creating and evaluating a system capable of extractive QA in Icelandic.

We train several language models for Icelandic, including IceBERT, that achieve state-of-the-art performance in a variety of downstream tasks, including part-of-speech tagging, named entity recognition, grammatical error detection and constituency parsing. To train the models we introduce a new corpus of Icelandic text, the Icelandic Common Crawl Corpus (IC3), a collection of high quality texts found online by targeting the Icelandic top-level-domain .is. Several other public data sources are also collected for a total of 16GB of Icelandic text. To enhance the evaluation of model performance and to raise the bar in baselines for Icelandic, we manually translate and adapt the WinoGrande commonsense reasoning dataset. Through these efforts we demonstrate that a properly cleaned crawled corpus is sufficient to achieve state-of-the-art results in NLP applications for low to medium resource languages, by comparison with models trained on a curated corpus. We further show that initializing models using existing multilingual models can lead to state-of-the-art results for some downstream tasks.