ICLR 2023 Papers

As large language models (LLMs) grow larger and more sophisticated, assessing their "reasoning" capabilities in natural language grows more challenging. Recent question answering (QA) benchmarks that attempt to assess reasoning are often limited by a narrow scope of covered situations and subject matters. We introduce WikiWhy, a QA dataset built around a novel auxiliary task: explaining why an answer is true in natural language. WikiWhy contains over 9,000 "why" question-answer-rationale triples, grounded on Wikipedia facts across a diverse set of topics. Each rationale is a set of supporting statements connecting the question to the answer. WikiWhy serves as a benchmark for the reasoning capabilities of LLMs because it demands rigorous explicit rationales for each answer to demonstrate the acquisition of implicit commonsense knowledge, which is unlikely to be easily memorized. GPT-3 baselines achieve only 38.7% human-evaluated correctness in the end-to-end answer & explain condition, leaving significant room for future improvements.

In the mode connectivity literature, it is widely accepted that there are common circumstances in which two neural networks, trained similarly on the same data, will maintain loss when interpolated in the weight space. In particular, transfer learning is presumed to ensure the necessary conditions for linear mode connectivity across training runs. In contrast to existing results from image classification, we find that among text classifiers (trained on MNLI, QQP, and CoLA), some pairs of finetuned models have large barriers of increasing loss on the linear paths between them. On each task, we find distinct clusters of models which are linearly connected on the test loss surface, but are disconnected from models outside the cluster---models that occupy separate basins on the surface. By measuring performance on specially-crafted diagnostic datasets, we find that these clusters correspond to different generalization strategies. For example, on MNLI, one cluster behaves like a bag of words model under domain shift, while another cluster uses syntactic heuristics. Our work demonstrates how the geometry of the loss surface can guide models towards different heuristic functions in standard finetuning settings.

Extractive summarization produces summaries by identifying and concatenating the most important sentences in a document. Since most summarization datasets do not come with gold labels indicating whether document sentences are summary-worthy, different labeling algorithms have been proposed to extrapolate oracle extracts for model training. In this work, we identify two flaws with the widely used greedy labeling approach: it delivers suboptimal and deterministic oracles. To alleviate both issues, we propose a simple yet effective labeling algorithm that creates soft, expectation-based sentence labels. We define a new learning objective for extractive summarization which incorporates learning signals from multiple oracle summaries and prove it is equivalent to estimating the oracle expectation for each document sentence. Without any architectural modifications, the proposed labeling scheme achieves superior performance on a variety of summarization benchmarks across domains and languages, in both supervised and zero-shot settings.

Language models are defined over a finite set of inputs, which creates a vocabulary bottleneck when we attempt to scale the number of supported languages. Tackling this bottleneck results in a trade-off between what can be represented in the embedding matrix and computational issues in the output layer. This paper introduces PIXEL, the Pixel-based Encoder of Language, which suffers from neither of these issues. PIXEL is a pretrained language model that renders text as images, making it possible to transfer representations across languages based on orthographic similarity or the co-activation of pixels. PIXEL is trained to reconstruct the pixels of masked patches instead of predicting a distribution over tokens. We pretrain the 86M parameter PIXEL model on the same English data as BERT and evaluate on syntactic and semantic tasks in typologically diverse languages, including various non-Latin scripts. We find that PIXEL substantially outperforms BERT on syntactic and semantic processing tasks on scripts that are not found in the pretraining data, but PIXEL is slightly weaker than BERT when working with Latin scripts. Furthermore, we find that PIXEL is more robust than BERT to orthographic attacks and linguistic code-switching, further confirming the benefits of modelling language with pixels.

The extent to which text-only language models (LMs) learn to represent the physical, non-linguistic world is an open question. Prior work has shown that pretrained LMs can be taught to ``understand'' visual inputs when the models' parameters are updated on image captioning tasks. We test a stronger hypothesis: that the conceptual representations learned by text-only models are functionally equivalent (up to a linear transformation) to those learned by models trained on vision tasks. Specifically, we show that the image representations from vision models can be transferred as continuous prompts to frozen LMs by training only a single linear projection. Using these to prompt the LM achieves competitive performance on captioning and visual question answering tasks compared to models that tune both the image encoder and text decoder (such as the MAGMA model). We compare three image encoders with increasing amounts of linguistic supervision seen during pretraining: BEIT (no linguistic information), NF-ResNET (lexical category information), and CLIP (full natural language descriptions). We find that all three encoders perform equally well at transferring visual property information to the language model (e.g., whether an animal is large or small), but that image encoders pretrained with linguistic supervision more saliently encode category information (e.g., distinguishing hippo vs.\ elephant) and thus perform significantly better on benchmark language-and-vision tasks. Our results indicate that LMs encode conceptual information structurally similarly to vision-based models, even those that are solely trained on images.

Robustness evaluation against adversarial examples has become increasingly important to unveil the trustworthiness of the prevailing deep models in natural language processing (NLP). However, in contrast to the computer vision domain where the first-order projected gradient descent (PGD) is used as the benchmark approach to generate adversarial examples for robustness evaluation, there lacks a principled first-order gradient-based robustness evaluation framework in NLP. The emerging optimization challenges lie in 1) the discrete nature of textual inputs together with the strong coupling between the perturbation location and the actual content, and 2) the additional constraint that the perturbed text should be fluent and achieve a low perplexity under a language model. These challenges make the development of PGD-like NLP attacks difficult. To bridge the gap, we propose TextGrad, a new attack generator using gradient-driven optimization, supporting high-accuracy and high-quality assessment of adversarial robustness in NLP. Specifically, we address the aforementioned challenges in a unified optimization framework. And we develop an effective convex relaxation method to co-optimize the continuously-relaxed site selection and perturbation variables and leverage an effective sampling method to establish an accurate mapping from the continuous optimization variables to the discrete textual perturbations. Moreover, as a first-order attack generation method, TextGrad can be baked into adversarial training to further improve the robustness of NLP models. Extensive experiments are provided to demonstrate the effectiveness of TextGrad not only in attack generation for robustness evaluation but also in adversarial defense. From the attack perspective, we show that TextGrad achieves remarkable improvements in both the attack success rate and the perplexity score over five state-of-the-art baselines. From the defense perspective, TextGrad-enabled adversarial training yields the most robust NLP model against a wide spectrum of NLP attacks.

This paper focuses on data augmentation for low-resource NLP tasks where the training set is limited. The existing solutions either leverage task-independent heuristic rules (e.g., Synonym Replacement) or fine-tune general-purpose pre-trained language models (e.g., GPT2) using the limited training instances to produce new synthetic data. Consequently, they have trivial task-specific knowledge and are limited to yielding low-quality synthetic data. To combat this issue, we propose Knowledge Mixture Data Augmentation Model (KnowDA), a Seq2Seq language model pretrained on a mixture of diverse NLP tasks under a novel framework of Knowledge Mixture Training (KoMT). The goal of KoMT is to condense diverse NLP task-specific knowledge into the single KnowDA model (i.e., all-in-one). The resulting KnowDA could utilize these knowledge to quickly grasp the inherent synthesis law of the target task through limited training instances. Specifically, KoMT reformulates input examples from various heterogeneous NLP tasks into a unified text-to-text format and employs denoising training objectives in different granularity to learn to reconstruct partial or complete samples. To the best of our knowledge, we are the first to attempt to apply 100+ NLP multi-task training for data augmentation. Extensive experiments show that i) the synthetic data produced by KnowDA successfully improves the performance of the strong pre-trained language models (i.e., Bert, ALBert and Deberta) by a large margin on the low-resource NLP benchmark FewGLUE, CoNLL’03 and WikiAnn; ii) KnowDA successful transfer the task knowledge to NLP tasks whose types are seen and unseen in KoMT.