Prompt-based learning, with its capability to tackle zero-shot and few-shot NLP tasks, has gained much attention in community. The main idea is to bridge the gap between NLP downstream tasks and language modeling (LM), by mapping these tasks into natural language prompts, which are then filled by pre-trained language models (PLMs). However, for prompt learning, there are still two salient gaps between NLP tasks and pretraining. First, prompt information is not necessarily sufficiently present during LM pretraining. Second, task-specific data are not necessarily well represented during pretraining. We address these two issues by proposing AdaPrompt, adaptively retrieving external data for continual pretraining of PLMs by making use of both task and prompt characteristics. In addition, we make use of knowledge in Natural Language Inference models for deriving adaptive verbalizers. Experimental results on five NLP benchmarks show that AdaPrompt can improve over standard PLMs in few-shot settings. In addition, in zero-shot settings, our method outperforms standard prompt-based methods by up to 26.35\% relative error reduction.
While the general idea of self-supervised learning is identical across modalities, the actual algorithms and objectives differ widely because they were developed with a single modality in mind. To get us closer to general self-supervised learning, we present data2vec, a framework that uses the same learning method for either speech, NLP or computer vision. The core idea is to predict latent representations of the full input data based on a masked view of the input in a self-distillation setup using a standard Transformer architecture. Instead of predicting modality-specific targets such as words, visual tokens or units of human speech which are local in nature, data2vec predicts contextualized latent representations that contain information from the entire input. Experiments on the major benchmarks of speech recognition, image classification, and natural language understanding demonstrate a new state of the art or competitive performance to predominant approaches.
There have been many successful applications of sentence embedding methods. However, it has not been well understood what properties are captured in the resulting sentence embeddings depending on the supervision signals. In this paper, we focus on two types of sentence embedding methods with similar architectures and tasks: one fine-tunes pre-trained language models on the natural language inference task, and the other fine-tunes pre-trained language models on word prediction task from its definition sentence, and investigate their properties. Specifically, we compare their performances on semantic textual similarity (STS) tasks using STS datasets partitioned from two perspectives: 1) sentence source and 2) superficial similarity of the sentence pairs, and compare their performances on the downstream and probing tasks. Furthermore, we attempt to combine the two methods and demonstrate that combining the two methods yields substantially better performance than the respective methods on unsupervised STS tasks and downstream tasks.
Aspect-based sentiment analysis (ABSA) typically requires in-domain annotated data for supervised training/fine-tuning. It is a big challenge to scale ABSA to a large number of new domains. This paper aims to train a unified model that can perform zero-shot ABSA without using any annotated data for a new domain. We propose a method called contrastive post-training on review Natural Language Inference (CORN). Later ABSA tasks can be cast into NLI for zero-shot transfer. We evaluate CORN on ABSA tasks, ranging from aspect extraction (AE), aspect sentiment classification (ASC), to end-to-end aspect-based sentiment analysis (E2E ABSA), which show ABSA can be conducted without any human annotated ABSA data.
Recently, flat-minima optimizers, which seek to find parameters in low-loss neighborhoods, have been shown to improve a neural network's generalization performance over stochastic and adaptive gradient-based optimizers. Two methods have received significant attention due to their scalability: 1. Stochastic Weight Averaging (SWA), and 2. Sharpness-Aware Minimization (SAM). However, there has been limited investigation into their properties and no systematic benchmarking of them across different domains. We fill this gap here by comparing the loss surfaces of the models trained with each method and through broad benchmarking across computer vision, natural language processing, and graph representation learning tasks. We discover several surprising findings from these results, which we hope will help researchers further improve deep learning optimizers, and practitioners identify the right optimizer for their problem.
Transformers-based models, such as BERT, have dramatically improved the performance for various natural language processing tasks. The clinical knowledge enriched model, namely ClinicalBERT, also achieved state-of-the-art results when performed on clinical named entity recognition and natural language inference tasks. One of the core limitations of these transformers is the substantial memory consumption due to their full self-attention mechanism. To overcome this, long sequence transformer models, e.g. Longformer and BigBird, were proposed with the idea of sparse attention mechanism to reduce the memory usage from quadratic to the sequence length to a linear scale. These models extended the maximum input sequence length from 512 to 4096, which enhanced the ability of modeling long-term dependency and consequently achieved optimal results in a variety of tasks. Inspired by the success of these long sequence transformer models, we introduce two domain enriched language models, namely Clinical-Longformer and Clinical-BigBird, which are pre-trained from large-scale clinical corpora. We evaluate both pre-trained models using 10 baseline tasks including named entity recognition, question answering, and document classification tasks. The results demonstrate that Clinical-Longformer and Clinical-BigBird consistently and significantly outperform ClinicalBERT as well as other short-sequence transformers in all downstream tasks. We have made our source code available at [https://github.com/luoyuanlab/Clinical-Longformer] the pre-trained models available for public download at: [https://huggingface.co/yikuan8/Clinical-Longformer].
Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer the relationship between the sentence pair (premise and hypothesis). Many recent works have used contrastive learning by incorporating the relationship of the sentence pair from NLI datasets to learn sentence representation. However, these methods only focus on comparisons with sentence-level representations. In this paper, we propose a Pair-level Supervised Contrastive Learning approach (PairSCL). We adopt a cross attention module to learn the joint representations of the sentence pairs. A contrastive learning objective is designed to distinguish the varied classes of sentence pairs by pulling those in one class together and pushing apart the pairs in other classes. We evaluate PairSCL on two public datasets of NLI where the accuracy of PairSCL outperforms other methods by 2.1% on average. Furthermore, our method outperforms the previous state-of-the-art method on seven transfer tasks of text classification.
In this paper, we examine several methods of acquiring Czech data for automated fact-checking, which is a task commonly modeled as a classification of textual claim veracity w.r.t. a corpus of trusted ground truths. We attempt to collect sets of data in form of a factual claim, evidence within the ground truth corpus, and its veracity label (supported, refuted or not enough info). As a first attempt, we generate a Czech version of the large-scale FEVER dataset built on top of Wikipedia corpus. We take a hybrid approach of machine translation and document alignment; the approach and the tools we provide can be easily applied to other languages. We discuss its weaknesses and inaccuracies, propose a future approach for their cleaning and publish the 127k resulting translations, as well as a version of such dataset reliably applicable for the Natural Language Inference task - the CsFEVER-NLI. Furthermore, we collect a novel dataset of 3,097 claims, which is annotated using the corpus of 2.2M articles of Czech News Agency. We present its extended annotation methodology based on the FEVER approach, and, as the underlying corpus is kept a trade secret, we also publish a standalone version of the dataset for the task of Natural Language Inference we call CTKFactsNLI. We analyze both acquired datasets for spurious cues - annotation patterns leading to model overfitting. CTKFacts is further examined for inter-annotator agreement, thoroughly cleaned, and a typology of common annotator errors is extracted. Finally, we provide baseline models for all stages of the fact-checking pipeline and publish the NLI datasets, as well as our annotation platform and other experimental data.
Existing textual adversarial attacks usually utilize the gradient or prediction confidence to generate adversarial examples, making it hard to be deployed in real-world applications. To this end, we consider a rarely investigated but more rigorous setting, namely hard-label attack, in which the attacker can only access the prediction label. In particular, we find we can learn the importance of different words via the change on prediction label caused by word substitutions on the adversarial examples. Based on this observation, we propose a novel adversarial attack, termed Text Hard-label attacker (TextHacker). TextHacker randomly perturbs lots of words to craft an adversarial example. Then, TextHacker adopts a hybrid local search algorithm with the estimation of word importance from the attack history to minimize the adversarial perturbation. Extensive evaluations for text classification and textual entailment show that TextHacker significantly outperforms existing hard-label attacks regarding the attack performance as well as adversary quality.
In this article, we explore the shallow heuristics used by transformer-based pre-trained language models (PLMs) that are fine-tuned for natural language inference (NLI). To do so, we construct or own dataset based on syllogistic, and we evaluate a number of models' performance on our dataset. We find evidence that the models rely heavily on certain shallow heuristics, picking up on symmetries and asymmetries between premise and hypothesis. We suggest that the lack of generalization observable in our study, which is becoming a topic of lively debate in the field, means that the PLMs are currently not learning NLI, but rather spurious heuristics.
A recurring challenge of crowdsourcing NLP datasets at scale is that human writers often rely on repetitive patterns when crafting examples, leading to a lack of linguistic diversity. We introduce a novel approach for dataset creation based on worker and AI collaboration, which brings together the generative strength of language models and the evaluative strength of humans. Starting with an existing dataset, MultiNLI for natural language inference (NLI), our approach uses dataset cartography to automatically identify examples that demonstrate challenging reasoning patterns, and instructs GPT-3 to compose new examples with similar patterns. Machine generated examples are then automatically filtered, and finally revised and labeled by human crowdworkers. The resulting dataset, WANLI, consists of 107,885 NLI examples and presents unique empirical strengths over existing NLI datasets. Remarkably, training a model on WANLI improves performance on eight out-of-domain test sets we consider, including by 11% on HANS and 9% on Adversarial NLI, compared to training on the 4x larger MultiNLI. Moreover, it continues to be more effective than MultiNLI augmented with other NLI datasets. Our results demonstrate the promise of leveraging natural language generation techniques and re-imagining the role of humans in the dataset creation process.
The paper describes the open Russian medical language understanding benchmark covering several task types (classification, question answering, natural language inference, named entity recognition) on a number of novel text sets. Given the sensitive nature of the data in healthcare, such a benchmark partially closes the problem of Russian medical dataset absence. We prepare the unified format labeling, data split, and evaluation metrics for new tasks. The remaining tasks are from existing datasets with a few modifications. A single-number metric expresses a model's ability to cope with the benchmark. Moreover, we implement several baseline models, from simple ones to neural networks with transformer architecture, and release the code. Expectedly, the more advanced models yield better performance, but even a simple model is enough for a decent result in some tasks. Furthermore, for all tasks, we provide a human evaluation. Interestingly the models outperform humans in the large-scale classification tasks. However, the advantage of natural intelligence remains in the tasks requiring more knowledge and reasoning.
Recursive neural networks (Tree-RNNs) based on dependency trees are ubiquitous in modeling sentence meanings as they effectively capture semantic relationships between non-neighborhood words. However, recognizing semantically dissimilar sentences with the same words and syntax is still a challenge to Tree-RNNs. This work proposes an improvement to Dependency Tree-RNN (DT-RNN) using the grammatical relationship type identified in the dependency parse. Our experiments on semantic relatedness scoring (SRS) and recognizing textual entailment (RTE) in sentence pairs using SICK (Sentence Involving Compositional Knowledge) dataset show encouraging results. The model achieved a 2% improvement in classification accuracy for the RTE task over the DT-RNN model. The results show that Pearson's and Spearman's correlation measures between the model's predicted similarity scores and human ratings are higher than those of standard DT-RNNs.
Predictive uncertainty estimation of pre-trained language models is an important measure of how likely people can trust their predictions. However, little is known about what makes a model prediction uncertain. Explaining predictive uncertainty is an important complement to explaining prediction labels in helping users understand model decision making and gaining their trust on model predictions, while has been largely ignored in prior works. In this work, we propose to explain the predictive uncertainty of pre-trained language models by extracting uncertain words from existing model explanations. We find the uncertain words are those identified as making negative contributions to prediction labels, while actually explaining the predictive uncertainty. Experiments show that uncertainty explanations are indispensable to explaining models and helping humans understand model prediction behavior.
Despite recent breakthroughs in Machine Learning for Natural Language Processing, the Natural Language Inference (NLI) problems still constitute a challenge. To this purpose we contribute a new dataset that focuses exclusively on the factivity phenomenon; however, our task remains the same as other NLI tasks, i.e. prediction of entailment, contradiction or neutral (ECN). The dataset contains entirely natural language utterances in Polish and gathers 2,432 verb-complement pairs and 309 unique verbs. The dataset is based on the National Corpus of Polish (NKJP) and is a representative sample in regards to frequency of main verbs and other linguistic features (e.g. occurrence of internal negation). We found that transformer BERT-based models working on sentences obtained relatively good results ($\approx89\%$ F1 score). Even though better results were achieved using linguistic features ($\approx91\%$ F1 score), this model requires more human labour (humans in the loop) because features were prepared manually by expert linguists. BERT-based models consuming only the input sentences show that they capture most of the complexity of NLI/factivity. Complex cases in the phenomenon - e.g. cases with entitlement (E) and non-factive verbs - remain an open issue for further research.
NLP benchmarks have largely focused on short texts, such as sentences and paragraphs, even though long texts comprise a considerable amount of natural language in the wild. We introduce SCROLLS, a suite of tasks that require reasoning over long texts. We examine existing long-text datasets, and handpick ones where the text is naturally long, while prioritizing tasks that involve synthesizing information across the input. SCROLLS contains summarization, question answering, and natural language inference tasks, covering multiple domains, including literature, science, business, and entertainment. Initial baselines, including Longformer Encoder-Decoder, indicate that there is ample room for improvement on SCROLLS. We make all datasets available in a unified text-to-text format and host a live leaderboard to facilitate research on model architecture and pretraining methods.
Long-form question answering (LFQA) tasks require retrieving the documents pertinent to a query, using them to form a paragraph-length answer. Despite considerable progress in LFQA modeling, fundamental issues impede its progress: i) train/validation/test dataset overlap, ii) absence of automatic metrics and iii) generated answers not being "grounded" in retrieved documents. This work addresses every one these critical bottlenecks, contributing natural language inference/generation (NLI/NLG) methods and metrics that make significant strides to their alleviation.
The fine-tuning of pre-trained language models has a great success in many NLP fields. Yet, it is strikingly vulnerable to adversarial examples, e.g., word substitution attacks using only synonyms can easily fool a BERT-based sentiment analysis model. In this paper, we demonstrate that adversarial training, the prevalent defense technique, does not directly fit a conventional fine-tuning scenario, because it suffers severely from catastrophic forgetting: failing to retain the generic and robust linguistic features that have already been captured by the pre-trained model. In this light, we propose Robust Informative Fine-Tuning (RIFT), a novel adversarial fine-tuning method from an information-theoretical perspective. In particular, RIFT encourages an objective model to retain the features learned from the pre-trained model throughout the entire fine-tuning process, whereas a conventional one only uses the pre-trained weights for initialization. Experimental results show that RIFT consistently outperforms the state-of-the-arts on two popular NLP tasks: sentiment analysis and natural language inference, under different attacks across various pre-trained language models.
Building natural language inference (NLI) benchmarks that are both challenging for modern techniques, and free from shortcut biases is difficult. Chief among these biases is "single sentence label leakage," where annotator-introduced spurious correlations yield datasets where the logical relation between (premise, hypothesis) pairs can be accurately predicted from only a single sentence, something that should in principle be impossible. We demonstrate that despite efforts to reduce this leakage, it persists in modern datasets that have been introduced since its 2018 discovery. To enable future amelioration efforts, introduce a novel model-driven technique, the progressive evaluation of cluster outliers (PECO) which enables both the objective measurement of leakage, and the automated detection of subpopulations in the data which maximally exhibit it.
Large-scale generative language models such as GPT-3 are competitive few-shot learners. While these models are known to be able to jointly represent many different languages, their training data is dominated by English, potentially limiting their cross-lingual generalization. In this work, we train multilingual generative language models on a corpus covering a diverse set of languages, and study their few- and zero-shot learning capabilities in a wide range of tasks. Our largest model with 7.5 billion parameters sets new state of the art in few-shot learning in more than 20 representative languages, outperforming GPT-3 of comparable size in multilingual commonsense reasoning (with +7.4% absolute accuracy improvement in 0-shot settings and +9.4% in 4-shot settings) and natural language inference (+5.4% in each of 0-shot and 4-shot settings). On the FLORES-101 machine translation benchmark, our model outperforms GPT-3 on 171 out of 182 directions with 32 training examples, while surpassing the official supervised baseline in 45 directions. We conduct an in-depth analysis of different multilingual prompting approaches, showing in particular that strong few-shot learning performance across languages can be achieved via cross-lingual transfer through both templates and demonstration examples. Finally, we evaluate our models in social value tasks such as hate speech detection in five languages and find it has limitations similar to comparable sized GPT-3 models.