Large pretrained language models have been performing increasingly well in a variety of downstream tasks via prompting. However, it remains unclear from where the model learns the task-specific knowledge, especially in a zero-shot setup. In this work, we want to find evidence of the model's task-specific competence from pretraining and are specifically interested in locating a very small subset of pretraining data that directly supports the model in the task. We call such a subset supporting data evidence and propose a novel method ORCA to effectively identify it, by iteratively using gradient information related to the downstream task. This supporting data evidence offers interesting insights about the prompted language models: in the tasks of sentiment analysis and textual entailment, BERT shows a substantial reliance on BookCorpus, the smaller corpus of BERT's two pretraining corpora, as well as on pretraining examples that mask out synonyms to the task verbalizers.
Generics express generalizations about the world (e.g., birds can fly) that are not universally true (e.g., newborn birds and penguins cannot fly). Commonsense knowledge bases, used extensively in NLP, encode some generic knowledge but rarely enumerate such exceptions and knowing when a generic statement holds or does not hold true is crucial for developing a comprehensive understanding of generics. We present a novel framework informed by linguistic theory to generate exemplars -- specific cases when a generic holds true or false. We generate ~19k exemplars for ~650 generics and show that our framework outperforms a strong GPT-3 baseline by 12.8 precision points. Our analysis highlights the importance of linguistic theory-based controllability for generating exemplars, the insufficiency of knowledge bases as a source of exemplars, and the challenges exemplars pose for the task of natural language inference.
Large language models trained on a mixture of NLP tasks that are converted into a text-to-text format using prompts, can generalize into novel forms of language and handle novel tasks. A large body of work within prompt engineering attempts to understand the effects of input forms and prompts in achieving superior performance. We consider an alternative measure and inquire whether the way in which an input is encoded affects social biases promoted in outputs. In this paper, we study T0, a large-scale multi-task text-to-text language model trained using prompt-based learning. We consider two different forms of semantically equivalent inputs: question-answer format and premise-hypothesis format. We use an existing bias benchmark for the former BBQ and create the first bias benchmark in natural language inference BBNLI with hand-written hypotheses while also converting each benchmark into the other form. The results on two benchmarks suggest that given two different formulations of essentially the same input, T0 conspicuously acts more biased in question answering form, which is seen during training, compared to premise-hypothesis form which is unlike its training examples. Code and data are released under https://github.com/feyzaakyurek/bbnli.
Policy Compliance Detection (PCD) is a task we encounter when reasoning over texts, e.g. legal frameworks. Previous work to address PCD relies heavily on modeling the task as a special case of Recognizing Textual Entailment. Entailment is applicable to the problem of PCD, however viewing the policy as a single proposition, as opposed to multiple interlinked propositions, yields poor performance and lacks explainability. To address this challenge, more recent proposals for PCD have argued for decomposing policies into expression trees consisting of questions connected with logic operators. Question answering is used to obtain answers to these questions with respect to a scenario. Finally, the expression tree is evaluated in order to arrive at an overall solution. However, this work assumes expression trees are provided by experts, thus limiting its applicability to new policies. In this work, we learn how to infer expression trees automatically from policy texts. We ensure the validity of the inferred trees by introducing constrained decoding using a finite state automaton to ensure the generation of valid trees. We determine through automatic evaluation that 63% of the expression trees generated by our constrained generation model are logically equivalent to gold trees. Human evaluation shows that 88% of trees generated by our model are correct.
Combining the visual modality with pretrained language models has been surprisingly effective for simple descriptive tasks such as image captioning. More general text generation however remains elusive. We take a step back and ask: How do these models work for more complex generative tasks, i.e. conditioning on both text and images? Are multimodal models simply visually adapted language models, or do they combine they reason jointly over modalities? We investigate these questions in the context of self-rationalization (jointly generating task labels/answers and free-text explanations) of three tasks: (i) visual question answering in VQA-X, (ii) visual commonsense reasoning in VCR, and (iii) visual-textual entailment in e-SNLI-VE. We show that recent unimodal advances, CLIP image representations and scaling of language models, do not consistently improve self-rationalization in multimodal tasks. We find that no single model type works universally best across tasks, datasets, and finetuning data sizes. Our findings motivate the need for novel general backbones approach that move text generation from images and text beyond image captioning.
This paper explores a question-answer driven approach to reveal affirmative interpretations from verbal negations (i.e., when a negation cue grammatically modifies a verb). We create a new corpus consisting of 4,472 verbal negations and discover that 67.1% of them convey that an event actually occurred. Annotators generate and answer 7,277 questions for the 3,001 negations that convey an affirmative interpretation. We first cast the problem of revealing affirmative interpretations from negations as a natural language inference (NLI) classification task. Experimental results show that state-of-the-art transformers trained with existing NLI corpora are insufficient to reveal affirmative interpretations. We also observe, however, that fine-tuning brings small improvements. In addition to NLI classification, we also explore the more realistic task of generating affirmative interpretations directly from negations with the T5 transformer. We conclude that the generation task remains a challenge as T5 substantially underperforms humans.
Current Natural Language Inference (NLI) models achieve impressive results, sometimes outperforming humans when evaluating on in-distribution test sets. However, as these models are known to learn from annotation artefacts and dataset biases, it is unclear to what extent the models are learning the task of NLI instead of learning from shallow heuristics in their training data. We address this issue by introducing a logical reasoning framework for NLI, creating highly transparent model decisions that are based on logical rules. Unlike prior work, we show that improved interpretability can be achieved without decreasing the predictive accuracy. We almost fully retain performance on SNLI, while also identifying the exact hypothesis spans that are responsible for each model prediction. Using the e-SNLI human explanations, we verify that our model makes sensible decisions at a span level, despite not using any span labels during training. We can further improve model performance and span-level decisions by using the e-SNLI explanations during training. Finally, our model is more robust in a reduced data setting. When training with only 1,000 examples, out-of-distribution performance improves on the MNLI matched and mismatched validation sets by 13% and 16% relative to the baseline. Training with fewer observations yields further improvements, both in-distribution and out-of-distribution.
Natural Language Inference Generation task is to generate a text hypothesis given a text premise and a logical relation between the two. This task can be used in data augmentation and controllable text generation in practice. In this paper, we propose language models with prompt and dynamic demonstration (LM-PDD) to tackle this problem in few-shot settings. Our framework outperforms standard fine-tuned models with low resource, achieving an average 8% absolute improvement on SNLI and MNLI datasets, and the results on 13 natural language classification tasks also show that our dynamic demonstration method has good generalizability.
The ever-growing model size and scale of compute have attracted increasing interests in training deep learning models over multiple nodes. However, when it comes to training on cloud clusters, especially across remote clusters, huge challenges are faced. In this work, we introduce a general framework, Nebula-I, for collaboratively training deep learning models over remote heterogeneous clusters, the connections between which are low-bandwidth wide area networks (WANs). We took natural language processing (NLP) as an example to show how Nebula-I works in different training phases that include: a) pre-training a multilingual language model using two remote clusters; and b) fine-tuning a machine translation model using knowledge distilled from pre-trained models, which run through the most popular paradigm of recent deep learning. To balance the accuracy and communication efficiency, in Nebula-I, parameter-efficient training strategies, hybrid parallel computing methods and adaptive communication acceleration techniques are jointly applied. Meanwhile, security strategies are employed to guarantee the safety, reliability and privacy in intra-cluster computation and inter-cluster communication. Nebula-I is implemented with the PaddlePaddle deep learning framework, which can support collaborative training over heterogeneous hardware, e.g. GPU and NPU. Experiments demonstrate that the proposed framework could substantially maximize the training efficiency while preserving satisfactory NLP performance. By using Nebula-I, users can run large-scale training tasks over cloud clusters with minimum developments, and the utility of existed large pre-trained models could be further promoted. We also introduced new state-of-the-art results on cross-lingual natural language inference tasks, which are generated based upon a novel learning framework and Nebula-I.
Incorporating information from other languages can improve the results of tasks in low-resource languages. A powerful method of building functional natural language processing systems for low-resource languages is to combine multilingual pre-trained representations with cross-lingual transfer learning. In general, however, shared representations are learned separately, either across tasks or across languages. This paper proposes a meta-learning approach for inferring natural language in Persian. Alternately, meta-learning uses different task information (such as QA in Persian) or other language information (such as natural language inference in English). Also, we investigate the role of task augmentation strategy for forming additional high-quality tasks. We evaluate the proposed method using four languages and an auxiliary task. Compared to the baseline approach, the proposed model consistently outperforms it, improving accuracy by roughly six percent. We also examine the effect of finding appropriate initial parameters using zero-shot evaluation and CCA similarity.
Cross-lingual natural language processing relies on translation, either by humans or machines, at different levels, from translating training data to translating test sets. However, compared to original texts in the same language, translations possess distinct qualities referred to as translationese. Previous research has shown that these translation artifacts influence the performance of a variety of cross-lingual tasks. In this work, we propose a novel approach to reducing translationese by extending an established bias-removal technique. We use the Iterative Null-space Projection (INLP) algorithm, and show by measuring classification accuracy before and after debiasing, that translationese is reduced at both sentence and word level. We evaluate the utility of debiasing translationese on a natural language inference (NLI) task, and show that by reducing this bias, NLI accuracy improves. To the best of our knowledge, this is the first study to debias translationese as represented in latent embedding space.
Multilingual pre-trained models are known to suffer from the curse of multilinguality, which causes per-language performance to drop as they cover more languages. We address this issue by introducing language-specific modules, which allows us to grow the total capacity of the model, while keeping the total number of trainable parameters per language constant. In contrast with prior work that learns language-specific components post-hoc, we pre-train the modules of our Cross-lingual Modular (X-Mod) models from the start. Our experiments on natural language inference, named entity recognition and question answering show that our approach not only mitigates the negative interference between languages, but also enables positive transfer, resulting in improved monolingual and cross-lingual performance. Furthermore, our approach enables adding languages post-hoc with no measurable drop in performance, no longer limiting the model usage to the set of pre-trained languages.
Neural abstractive summarization models are prone to generate summaries which are factually inconsistent with their source documents. Previous work has introduced the task of recognizing such factual inconsistency as a downstream application of natural language inference (NLI). However, state-of-the-art NLI models perform poorly in this context due to their inability to generalize to the target task. In this work, we show that NLI models can be effective for this task when the training data is augmented with high-quality task-oriented examples. We introduce Falsesum, a data generation pipeline leveraging a controllable text generation model to perturb human-annotated summaries, introducing varying types of factual inconsistencies. Unlike previously introduced document-level NLI datasets, our generated dataset contains examples that are diverse and inconsistent yet plausible. We show that models trained on a Falsesum-augmented NLI dataset improve the state-of-the-art performance across four benchmarks for detecting factual inconsistency in summarization. The code to obtain the dataset is available online at https://github.com/joshbambrick/Falsesum
Prior research in the area of Natural Language Processing (NLP) has shown that including the syntactic structure of a sentence using a dependency parse tree while training a representation learning model improves the performance on downstream tasks. However, most of these modeling approaches make use of the dependency parse tree of sentences for learning task-specific word representations rather than considering that for learning generic representations. In this paper, we propose a new model named DIBERT which stands for Dependency Injected Bidirectional Encoder Representations from Transformers. DIBERT is a variation of the BERT, that apart from Masked Language Modeling (MLM) and Next Sentence Prediction (NSP) also incorporates an additional third objective called Parent Prediction (PP). PP injects the syntactic structure of a dependency tree while pre-training the DIBERT, which generates syntax-aware generic representations. We use the WikiText-103 benchmark dataset to pre-train both the original BERT (BERT-Base) and the proposed DIBERT models. After fine-tuning, we observe that DIBERT performs better than BERT-Base on various NLP downstream tasks including Semantic Similarity, Natural Language Inference and Sentiment Analysis hinting at the fact that incorporating dependency information when learning textual representations can improve the quality of the learned representations.
Existing pre-trained models are generally geared towards a particular class of problems. To date, there seems to be still no consensus on what the right architecture and pre-training setup should be. This paper presents a unified framework for pre-training models that are universally effective across datasets and setups. We begin by disentangling architectural archetypes with pre-training objectives -- two concepts that are commonly conflated. Next, we present a generalized & unified perspective for self-supervision in NLP and show how different pre-training objectives can be cast as one another and how interpolating between different objectives can be effective. We then propose Mixture-of-Denoisers (MoD), a pre-training objective that combines diverse pre-training paradigms together. We furthermore introduce a notion of mode switching, wherein downstream fine-tuning is associated with specific pre-training schemes. We conduct extensive ablative experiments to compare multiple pre-training objectives and find that our method pushes the Pareto-frontier by outperforming T5 & GPT-like models across multiple diverse setups. By scaling our model up to 20B parameters, we achieve SOTA performance on 50 well-established supervised finetuning based NLP tasks. Our model also achieve strong results at in-context learning, outperforming 175B GPT-3 on zero-shot SuperGLUE and tripling the performance of T5-XXL on one-shot summarization. On 0-shot MMLU, UL2 20B outperforms T0 and T5 models. UL2 20B also works well with chain-of-thought prompting and reasoning, making it an appealing choice for research into reasoning at a small to medium scale of 20B parameters. Finally, we apply FLAN instruction tuning to the UL2 20B model, achieving MMLU and Big-Bench scores competitive to FLAN-PaLM 62B. We release Flax-based T5X checkpoints for the UL2 20B & Flan-UL2 20B.
Does prompting a large language model (LLM) like GPT-3 with explanations improve in-context learning? We study this question on two NLP tasks that involve reasoning over text, namely question answering and natural language inference. We test the performance of four LLMs on three textual reasoning datasets using prompts that include explanations in multiple different styles. For these tasks, we find that including explanations in the prompts for OPT, GPT-3 (davinci), and InstructGPT (text-davinci-001) only yields small to moderate accuracy improvements over standard few-show learning. However, text-davinci-002 is able to benefit more substantially. We further show that explanations generated by the LLMs may not entail the models' predictions nor be factually grounded in the input, even on simple tasks with extractive explanations. However, these flawed explanations can still be useful as a way to verify LLMs' predictions post-hoc. Through analysis in our three settings, we show that explanations judged by humans to be good--logically consistent with the input and the prediction--more likely cooccur with accurate predictions. Following these observations, we train calibrators using automatically extracted scores that assess the reliability of explanations, allowing us to improve performance post-hoc across all of our datasets.
We present a comprehensive work on automated veracity assessment from dataset creation to developing novel methods based on Natural Language Inference (NLI), focusing on misinformation related to the COVID-19 pandemic. We first describe the construction of the novel PANACEA dataset consisting of heterogeneous claims on COVID-19 and their respective information sources. The dataset construction includes work on retrieval techniques and similarity measurements to ensure a unique set of claims. We then propose novel techniques for automated veracity assessment based on Natural Language Inference including graph convolutional networks and attention based approaches. We have carried out experiments on evidence retrieval and veracity assessment on the dataset using the proposed techniques and found them competitive with SOTA methods, and provided a detailed discussion.
Generating diverse, interesting responses to chitchat conversations is a problem for neural conversational agents. This paper makes two substantial contributions to improving diversity in dialogue generation. First, we propose a novel metric which uses Natural Language Inference (NLI) to measure the semantic diversity of a set of model responses for a conversation. We evaluate this metric using an established framework (Tevet and Berant, 2021) and find strong evidence indicating NLI Diversity is correlated with semantic diversity. Specifically, we show that the contradiction relation is more useful than the neutral relation for measuring this diversity and that incorporating the NLI model's confidence achieves state-of-the-art results. Second, we demonstrate how to iteratively improve the semantic diversity of a sampled set of responses via a new generation procedure called Diversity Threshold Generation, which results in an average 137% increase in NLI Diversity compared to standard generation procedures.
Recent work has shown that NLP tasks such as Relation Extraction (RE) can be recasted as Textual Entailment tasks using verbalizations, with strong performance in zero-shot and few-shot settings thanks to pre-trained entailment models. The fact that relations in current RE datasets are easily verbalized casts doubts on whether entailment would be effective in more complex tasks. In this work we show that entailment is also effective in Event Argument Extraction (EAE), reducing the need of manual annotation to 50% and 20% in ACE and WikiEvents respectively, while achieving the same performance as with full training. More importantly, we show that recasting EAE as entailment alleviates the dependency on schemas, which has been a road-block for transferring annotations between domains. Thanks to the entailment, the multi-source transfer between ACE and WikiEvents further reduces annotation down to 10% and 5% (respectively) of the full training without transfer. Our analysis shows that the key to good results is the use of several entailment datasets to pre-train the entailment model. Similar to previous approaches, our method requires a small amount of effort for manual verbalization: only less than 15 minutes per event argument type is needed, and comparable results can be achieved with users with different level of expertise.
Several popular Transformer based language models have been found to be successful for text-driven brain encoding. However, existing literature leverages only pretrained text Transformer models and has not explored the efficacy of task-specific learned Transformer representations. In this work, we explore transfer learning from representations learned for ten popular natural language processing tasks (two syntactic and eight semantic) for predicting brain responses from two diverse datasets: Pereira (subjects reading sentences from paragraphs) and Narratives (subjects listening to the spoken stories). Encoding models based on task features are used to predict activity in different regions across the whole brain. Features from coreference resolution, NER, and shallow syntax parsing explain greater variance for the reading activity. On the other hand, for the listening activity, tasks such as paraphrase generation, summarization, and natural language inference show better encoding performance. Experiments across all 10 task representations provide the following cognitive insights: (i) language left hemisphere has higher predictive brain activity versus language right hemisphere, (ii) posterior medial cortex, temporo-parieto-occipital junction, dorsal frontal lobe have higher correlation versus early auditory and auditory association cortex, (iii) syntactic and semantic tasks display a good predictive performance across brain regions for reading and listening stimuli resp.