The GPT-3.5 models have demonstrated impressive performance in various Natural Language Processing (NLP) tasks, showcasing their strong understanding and reasoning capabilities. However, their robustness and abilities to handle various complexities of the open world have yet to be explored, which is especially crucial in assessing the stability of models and is a key aspect of trustworthy AI. In this study, we perform a comprehensive experimental analysis of GPT-3.5, exploring its robustness using 21 datasets (about 116K test samples) with 66 text transformations from TextFlint that cover 9 popular Natural Language Understanding (NLU) tasks. Our findings indicate that while GPT-3.5 outperforms existing fine-tuned models on some tasks, it still encounters significant robustness degradation, such as its average performance dropping by up to 35.74\% and 43.59\% in natural language inference and sentiment analysis tasks, respectively. We also show that GPT-3.5 faces some specific robustness challenges, including robustness instability, prompt sensitivity, and number sensitivity. These insights are valuable for understanding its limitations and guiding future research in addressing these challenges to enhance GPT-3.5's overall performance and generalization abilities.
Recent studies reveal that various biases exist in different NLP tasks, and over-reliance on biases results in models' poor generalization ability and low adversarial robustness. To mitigate datasets biases, previous works propose lots of debiasing techniques to tackle specific biases, which perform well on respective adversarial sets but fail to mitigate other biases. In this paper, we propose a new debiasing method Sparse Mixture-of-Adapters (SMoA), which can mitigate multiple dataset biases effectively and efficiently. Experiments on Natural Language Inference and Paraphrase Identification tasks demonstrate that SMoA outperforms full-finetuning, adapter tuning baselines, and prior strong debiasing methods. Further analysis indicates the interpretability of SMoA that sub-adapter can capture specific pattern from the training data and specialize to handle specific bias.
In order to reveal the rationale behind model predictions, many works have exploited providing explanations in various forms. Recently, to further guarantee readability, more and more works turn to generate sentence-level human language explanations. However, current works pursuing sentence-level explanations rely heavily on annotated training data, which limits the development of interpretability to only a few tasks. As far as we know, this paper is the first to explore this problem smoothly from weak-supervised learning to unsupervised learning. Besides, we also notice the high latency of autoregressive sentence-level explanation generation, which leads to asynchronous interpretability after prediction. Therefore, we propose a non-autoregressive interpretable model to facilitate parallel explanation generation and simultaneous prediction. Through extensive experiments on Natural Language Inference task and Spouse Prediction task, we find that users are able to train classifiers with comparable performance $10-15\times$ faster with parallel explanation generation using only a few or no annotated training data.
OpenAI has released the Chat Generative Pre-trained Transformer (ChatGPT) and revolutionized the approach in artificial intelligence to human-model interaction. Several publications on ChatGPT evaluation test its effectiveness on well-known natural language processing (NLP) tasks. However, the existing studies are mostly non-automated and tested on a very limited scale. In this work, we examined ChatGPT's capabilities on 25 diverse analytical NLP tasks, most of them subjective even to humans, such as sentiment analysis, emotion recognition, offensiveness, and stance detection. In contrast, the other tasks require more objective reasoning like word sense disambiguation, linguistic acceptability, and question answering. We also evaluated GPT-4 model on five selected subsets of NLP tasks. We automated ChatGPT and GPT-4 prompting process and analyzed more than 49k responses. Our comparison of its results with available State-of-the-Art (SOTA) solutions showed that the average loss in quality of the ChatGPT model was about 25% for zero-shot and few-shot evaluation. For GPT-4 model, a loss for semantic tasks is significantly lower than for ChatGPT. We showed that the more difficult the task (lower SOTA performance), the higher the ChatGPT loss. It especially refers to pragmatic NLP problems like emotion recognition. We also tested the ability to personalize ChatGPT responses for selected subjective tasks via Random Contextual Few-Shot Personalization, and we obtained significantly better user-based predictions. Additional qualitative analysis revealed a ChatGPT bias, most likely due to the rules imposed on human trainers by OpenAI. Our results provide the basis for a fundamental discussion of whether the high quality of recent predictive NLP models can indicate a tool's usefulness to society and how the learning and validation procedures for such systems should be established.
Counterfactually-Augmented Data (CAD) has the potential to improve language models' Out-Of-Distribution (OOD) generalization capability, as CAD induces language models to exploit causal features and exclude spurious correlations. However, the empirical results of OOD generalization on CAD are not as efficient as expected. In this paper, we attribute the inefficiency to Myopia Phenomenon caused by CAD: language models only focus on causal features that are edited in the augmentation and exclude other non-edited causal features. As a result, the potential of CAD is not fully exploited. Based on the structural properties of CAD, we design two additional constraints to help language models extract more complete causal features contained in CAD, thus improving the OOD generalization capability. We evaluate our method on two tasks: Sentiment Analysis and Natural Language Inference, and the experimental results demonstrate that our method could unlock CAD's potential and improve language models' OOD generalization capability.
This technical report briefly describes our JDExplore d-team's submission Vega v1 on the General Language Understanding Evaluation (GLUE) leaderboard, where GLUE is a collection of nine natural language understanding tasks, including question answering, linguistic acceptability, sentiment analysis, text similarity, paraphrase detection, and natural language inference. [Method] We investigate several effective strategies and choose their best combination setting as the training recipes. As for model structure, we employ the vanilla Transformer with disentangled attention as the basic block encoder. For self-supervised training, we employ the representative denoising objective (i.e., replaced token detection) in phase 1 and combine the contrastive objective (i.e., sentence embedding contrastive learning) with it in phase 2. During fine-tuning, several advanced techniques such as transductive fine-tuning, self-calibrated fine-tuning, and adversarial fine-tuning are adopted. [Results] According to our submission record (Jan. 2022), with our optimized pretraining and fine-tuning strategies, our 1.3 billion model sets new state-of-the-art on 4/9 tasks, achieving the best average score of 91.3. Encouragingly, our Vega v1 is the first to exceed powerful human performance on the two challenging tasks, i.e., SST-2 and WNLI. We believe our empirically successful recipe with a bag of tricks could shed new light on developing efficient discriminative large language models.
We explore incorporating natural language inference (NLI) into the text generative pipeline by using a pre-trained NLI model to assess whether a generated sentence entails, contradicts, or is neutral to the prompt and preceding text. First, we show that the NLI task is predictive of generation errors made by GPT-3. We use these results to develop an NLI-informed generation procedure for GPT-J. Then, we evaluate these generations by obtaining human annotations on error types and overall quality. We find that an NLI strategy of maximizing entailment improves text generation when the nucleus sampling randomness parameter value is high, while one which maximizes contradiction is in fact productive when the parameter value is low. Overall, though, we demonstrate that an NLI strategy of maximizing the neutral class provides the highest quality of generated text (significantly better than the vanilla generations), regardless of parameter value.
In recent years, active learning has been successfully applied to an array of NLP tasks. However, prior work often assumes that training and test data are drawn from the same distribution. This is problematic, as in real-life settings data may stem from several sources of varying relevance and quality. We show that four popular active learning schemes fail to outperform random selection when applied to unlabelled pools comprised of multiple data sources on the task of natural language inference. We reveal that uncertainty-based strategies perform poorly due to the acquisition of collective outliers, i.e., hard-to-learn instances that hamper learning and generalization. When outliers are removed, strategies are found to recover and outperform random baselines. In further analysis, we find that collective outliers vary in form between sources, and show that hard-to-learn data is not always categorically harmful. Lastly, we leverage dataset cartography to introduce difficulty-stratified testing and find that different strategies are affected differently by example learnability and difficulty.
Investigating whether pre-trained language models (LMs) can function as knowledge bases (KBs) has raised wide research interests recently. However, existing works focus on simple, triple-based, relational KBs, but omit more sophisticated, logic-based, conceptualised KBs such as OWL ontologies. To investigate an LM's knowledge of ontologies, we propose OntoLAMA, a set of inference-based probing tasks and datasets from ontology subsumption axioms involving both atomic and complex concepts. We conduct extensive experiments on ontologies of different domains and scales, and our results demonstrate that LMs encode relatively less background knowledge of Subsumption Inference (SI) than traditional Natural Language Inference (NLI) but can improve on SI significantly when a small number of samples are given. We will open-source our code and datasets.
Automatic headline generation enables users to comprehend ongoing news events promptly and has recently become an important task in web mining and natural language processing. With the growing need for news headline generation, we argue that the hallucination issue, namely the generated headlines being not supported by the original news stories, is a critical challenge for the deployment of this feature in web-scale systems Meanwhile, due to the infrequency of hallucination cases and the requirement of careful reading for raters to reach the correct consensus, it is difficult to acquire a large dataset for training a model to detect such hallucinations through human curation. In this work, we present a new framework named ExHalder to address this challenge for headline hallucination detection. ExHalder adapts the knowledge from public natural language inference datasets into the news domain and learns to generate natural language sentences to explain the hallucination detection results. To evaluate the model performance, we carefully collect a dataset with more than six thousand labeled <article, headline> pairs. Extensive experiments on this dataset and another six public ones demonstrate that ExHalder can identify hallucinated headlines accurately and justifies its predictions with human-readable natural language explanations.
Large pretrained language models (LMs) have shown impressive In-Context Learning (ICL) ability, where the model learns to do an unseen task via a prompt consisting of input-output examples as the demonstration, without any parameter updates. The performance of ICL is highly dominated by the quality of the selected in-context examples. However, previous selection methods are mostly based on simple heuristics, leading to sub-optimal performance. In this work, we formulate in-context example selection as a subset selection problem. We propose CEIL (Compositional Exemplars for In-context Learning), which is instantiated by Determinantal Point Processes (DPPs) to model the interaction between the given input and in-context examples, and optimized through a carefully-designed contrastive learning objective to obtain preference from LMs. We validate CEIL on 12 classification and generation datasets from 7 distinct NLP tasks, including sentiment analysis, paraphrase detection, natural language inference, commonsense reasoning, open-domain question answering, code generation, and semantic parsing. Extensive experiments demonstrate not only the state-of-the-art performance but also the transferability and compositionality of CEIL, shedding new light on effective and efficient in-context learning. Our code is released at https://github.com/HKUNLP/icl-ceil.
Discrete prompts have been used for fine-tuning Pre-trained Language Models for diverse NLP tasks. In particular, automatic methods that generate discrete prompts from a small set of training instances have reported superior performance. However, a closer look at the learnt prompts reveals that they contain noisy and counter-intuitive lexical constructs that would not be encountered in manually-written prompts. This raises an important yet understudied question regarding the robustness of automatically learnt discrete prompts when used in downstream tasks. To address this question, we conduct a systematic study of the robustness of discrete prompts by applying carefully designed perturbations into an application using AutoPrompt and then measure their performance in two Natural Language Inference (NLI) datasets. Our experimental results show that although the discrete prompt-based method remains relatively robust against perturbations to NLI inputs, they are highly sensitive to other types of perturbations such as shuffling and deletion of prompt tokens. Moreover, they generalize poorly across different NLI datasets. We hope our findings will inspire future work on robust discrete prompt learning.
Recent work has shown how to prompt large language models with explanations to obtain strong performance on textual reasoning tasks, i.e., the chain-of-thought paradigm. However, subtly different explanations can yield widely varying downstream task accuracy. Explanations that have not been "tuned" for a task, such as off-the-shelf explanations written by nonexperts, may lead to mediocre performance. This paper tackles the problem of how to optimize explanation-infused prompts in a blackbox fashion. We first generate sets of candidate explanations for each example in the prompt using a leave-one-out scheme, then find an effective combination of these explanations with a two-stage framework. We first evaluate explanations for each in-context example in isolation according to two proxy metrics, log likelihood and accuracy on new examples. Then, we search over combinations of explanations to find one that yields high performance against a silver-labeled development set. Across four textual reasoning tasks spanning question answering, mathematical reasoning, and natural language inference, results show that our proxy metrics correlate with ground truth accuracy and our overall method can effectively improve prompts over crowdworker annotations and naive search strategies
In this paper, we explore Annotation Artifacts - the phenomena wherein large pre-trained NLP models achieve high performance on benchmark datasets but do not actually "solve" the underlying task and instead rely on some dataset artifacts (same across train, validation, and test sets) to figure out the right answer. We explore this phenomenon on the well-known Natural Language Inference task by first using contrast and adversarial examples to understand limitations to the model's performance and show one of the biases arising from annotation artifacts (the way training data was constructed by the annotators). We then propose a data augmentation technique to fix this bias and measure its effectiveness.
Specialised pre-trained language models are becoming more frequent in NLP since they can potentially outperform models trained on generic texts. BioBERT and BioClinicalBERT are two examples of such models that have shown promise in medical NLP tasks. Many of these models are overparametrised and resource-intensive, but thanks to techniques like Knowledge Distillation (KD), it is possible to create smaller versions that perform almost as well as their larger counterparts. In this work, we specifically focus on development of compact language models for processing clinical texts (i.e. progress notes, discharge summaries etc). We developed a number of efficient lightweight clinical transformers using knowledge distillation and continual learning, with the number of parameters ranging from 15 million to 65 million. These models performed comparably to larger models such as BioBERT and ClinicalBioBERT and significantly outperformed other compact models trained on general or biomedical data. Our extensive evaluation was done across several standard datasets and covered a wide range of clinical text-mining tasks, including Natural Language Inference, Relation Extraction, Named Entity Recognition, and Sequence Classification. To our knowledge, this is the first comprehensive study specifically focused on creating efficient and compact transformers for clinical NLP tasks. The models and code used in this study can be found on our Huggingface profile at https://huggingface.co/nlpie and Github page at https://github.com/nlpie-research/Lightweight-Clinical-Transformers, respectively, promoting reproducibility of our results.
Language Models are the core for almost any Natural Language Processing system nowadays. One of their particularities is their contextualized representations, a game changer feature when a disambiguation between word senses is necessary. In this paper we aim to explore to what extent language models are capable of discerning among senses at inference time. We performed this analysis by prompting commonly used Languages Models such as BERT or RoBERTa to perform the task of Word Sense Disambiguation (WSD). We leverage the relation between word senses and domains, and cast WSD as a textual entailment problem, where the different hypothesis refer to the domains of the word senses. Our results show that this approach is indeed effective, close to supervised systems.
Recently, Language Models (LMs) instruction-tuned on multiple tasks, also known as multitask-prompted fine-tuning (MT), have shown the capability to generalize to unseen tasks. Previous work has shown that scaling the number of training tasks is the key component in making stronger MT LMs. In this work, we report an unexpected finding that an expert LM fine-tuned on just a single task can outperform an MT LM trained with 300+ different tasks on 11 different unseen datasets and on 13 datasets of the BIG-bench benchmark by a mean accuracy of 3.20% and 1.29%, respectively. This finding casts doubt on the previously held belief that simply scaling the number of tasks makes stronger MT LMs. Leveraging this finding, we further show that this distributed approach of training a separate expert LM per training task instead of a single MT LM for zero-shot inference possesses many benefits including (1) avoiding negative task transfer that often occurs during instruction tuning, (2) being able to continually learn new tasks without having to re-train on previous tasks to avoid catastrophic forgetting, and (3) showing compositional capabilities when merging individual experts together. The code is available at https://github.com/joeljang/ELM.
We propose two methods to make unsupervised domain adaptation (UDA) more parameter efficient using adapters, small bottleneck layers interspersed with every layer of the large-scale pre-trained language model (PLM). The first method deconstructs UDA into a two-step process: first by adding a domain adapter to learn domain-invariant information and then by adding a task adapter that uses domain-invariant information to learn task representations in the source domain. The second method jointly learns a supervised classifier while reducing the divergence measure. Compared to strong baselines, our simple methods perform well in natural language inference (MNLI) and the cross-domain sentiment classification task. We even outperform unsupervised domain adaptation methods such as DANN and DSN in sentiment classification, and we are within 0.85% F1 for natural language inference task, by fine-tuning only a fraction of the full model parameters. We release our code at https://github.com/declare-lab/domadapter
Several proposals have been put forward in recent years for improving out-of-distribution (OOD) performance through mitigating dataset biases. A popular workaround is to train a robust model by re-weighting training examples based on a secondary biased model. Here, the underlying assumption is that the biased model resorts to shortcut features. Hence, those training examples that are correctly predicted by the biased model are flagged as being biased and are down-weighted during the training of the main model. However, assessing the importance of an instance merely based on the predictions of the biased model may be too naive. It is possible that the prediction of the main model can be derived from another decision-making process that is distinct from the behavior of the biased model. To circumvent this, we introduce a fine-tuning strategy that incorporates the similarity between the main and biased model attribution scores in a Product of Experts (PoE) loss function to further improve OOD performance. With experiments conducted on natural language inference and fact verification benchmarks, we show that our method improves OOD results while maintaining in-distribution (ID) performance.
Ensuring that generated utterances are faithful to dialogue actions is crucial for Task-Oriented Dialogue Response Generation. Slot Error Rate (SER) only partially measures generation quality in that it solely assesses utterances generated from non-categorical slots whose values are expected to be reproduced exactly. Utterances generated from categorical slots, which are more variable, are not assessed by SER. We propose Schema-Guided Semantic Accuracy (SGSAcc) to evaluate utterances generated from both categorical and non-categorical slots by recognizing textual entailment. We show that SGSAcc can be applied to evaluate utterances generated from a wide range of dialogue actions in the Schema Guided Dialogue (SGD) dataset with good agreement with human judgment. We also identify a previously overlooked weakness in generating faithful utterances from categorical slots in unseen domains. We show that prefix tuning applied to T5 generation can address this problem. We further build an ensemble of prefix-tuning and fine-tuning models that achieves the lowest SER reported and high SGSAcc on the SGD dataset.