We introduce Bonito, an open-source model for conditional task generation that converts unannotated text into task-specific training datasets for instruction tuning. We aim to enable zero-shot task adaptation of large language models on users' specialized, private data. We train Bonito by fine-tuning a pretrained large language model on a new large-scale dataset with 1.65M examples created by remixing existing instruction tuning datasets into meta-templates. The meta-templates for a dataset produce training examples where the input is the unannotated text and the task attribute and the output consists of the instruction and the response. We use Bonito to generate synthetic tasks for seven datasets from specialized domains with unannotated text across three task types -- yes-no question answering, extractive question answering, and natural language inference -- and adapt language models. We show that Bonito significantly improves the average performance of pretrained and instruction tuned models over the de facto self supervised baseline. For example, adapting Mistral-Instruct-v2 and instruction tuned variants of Mistral and Llama2 with Bonito improves the strong zero-shot performance by 22.1 F1 points whereas the next word prediction objective undoes some of the benefits of instruction tuning and reduces the average performance by 0.8 F1 points. We conduct additional experiments with Bonito to understand the effects of the domain, the size of the training set, and the choice of alternative synthetic task generators. Overall, we show that learning with synthetic instruction tuning datasets is an effective way to adapt language models to new domains. The model, dataset, and code are available at https://github.com/BatsResearch/bonito.
The superior performance of supervised classification methods in the information extraction (IE) area heavily relies on a large amount of gold standard data. Recent zero-shot classification methods converted the task to other NLP tasks (e.g., textual entailment) and used off-the-shelf models of these NLP tasks to directly perform inference on the test data without using a large amount of IE annotation data. A potentially valuable by-product of these methods is the large-scale silver standard data, i.e., pseudo-labeled data by the off-the-shelf models of other NLP tasks. However, there is no further investigation into the use of these data. In this paper, we propose a new framework, Clean-LaVe, which aims to utilize silver standard data to enhance the zero-shot performance. Clean-LaVe includes four phases: (1) Obtaining silver data; (2) Identifying relatively clean data from silver data; (3) Finetuning the off-the-shelf model using clean data; (4) Inference on the test data. The experimental results show that Clean-LaVe can outperform the baseline by 5% and 6% on TACRED and Wiki80 dataset in the zero-shot relation classification task, and by 3%-7% on Smile (Korean and Polish) in the zero-shot cross-lingual relation classification task, and by 8% on ACE05-E+ in the zero-shot event argument classification task. The code is share in https://github.com/wjw136/Clean_LaVe.git.
We study existing approaches to leverage off-the-shelf Natural Language Inference (NLI) models for the evaluation of summary faithfulness and argue that these are sub-optimal due to the granularity level considered for premises and hypotheses. That is, the smaller content unit considered as hypothesis is a sentence and premises are made up of a fixed number of document sentences. We propose a novel approach, namely InFusE, that uses a variable premise size and simplifies summary sentences into shorter hypotheses. Departing from previous studies which focus on single short document summarisation, we analyse NLI based faithfulness evaluation for diverse summarisation tasks. We introduce DiverSumm, a new benchmark comprising long form summarisation (long documents and summaries) and diverse summarisation tasks (e.g., meeting and multi-document summarisation). In experiments, InFusE obtains superior performance across the different summarisation tasks. Our code and data are available at https://github.com/HJZnlp/infuse.
Program-of-Thought (PoT) replaces natural language-based Chain-of-Thought (CoT) as the most popular method in Large Language Models (LLMs) mathematical reasoning tasks by utilizing external tool calls to circumvent computational errors. However, our evaluation of the GPT-4 and Llama series reveals that using PoT introduces more reasoning errors, such as incorrect formulas or flawed logic, compared to CoT. To address this issue, we propose Human-Think Language (HTL), which leverages a suite of strategies that help integrate PoT and CoT, encompassing: (1) a new generation paradigm that uses full CoT reasoning to control code generation. (2) Focus Attention, that directs model attention to the CoT reasoning during PoT to generate more logical code. (3) reinforcement learning that utilizes the accuracy of both CoT and PoT responses as rewards to prevent repetitive reasoning steps in LLMs when solving difficult math problems. Our method achieves an average improvement of 6.5% on the Llama-Base model and 4.3% on the Mistral-Base model across 8 mathematical calculation datasets. It also shows significant effectiveness on five out-of-domain datasets by controlling the model's information flow, exhibiting strong transferability. Additionally, HTL shows the most significant improvement in non-mathematical natural language inference task, contributing to a unified reasoning task framework
Large language models (LLMs) exhibit powerful general intelligence across diverse scenarios, including their integration into chatbots. However, a vital challenge of LLM-based chatbots is that they may produce hallucinated content in responses, which significantly limits their applicability. Various efforts have been made to alleviate hallucination, such as retrieval augmented generation and reinforcement learning with human feedback, but most of them require additional training and data annotation. In this paper, we propose a novel post-hoc Citation-Enhanced Generation (CEG) approach combined with retrieval argumentation. Unlike previous studies that focus on preventing hallucinations during generation, our method addresses this issue in a post-hoc way. It incorporates a retrieval module to search for supporting documents relevant to the generated content, and employs a natural language inference-based citation generation module. Once the statements in the generated content lack of reference, our model can regenerate responses until all statements are supported by citations. Note that our method is a training-free plug-and-play plugin that is capable of various LLMs. Experiments on various hallucination-related datasets show our framework outperforms state-of-the-art methods in both hallucination detection and response regeneration on three benchmarks. Our codes and dataset will be publicly available.
Online hate detection suffers from biases incurred in data sampling, annotation, and model pre-training. Therefore, measuring the averaged performance over all examples in held-out test data is inadequate. Instead, we must identify specific model weaknesses and be informed when it is more likely to fail. A recent proposal in this direction is HateCheck, a suite for testing fine-grained model functionalities on synthesized data generated using templates of the kind "You are just a [slur] to me." However, despite enabling more detailed diagnostic insights, the HateCheck test cases are often generic and have simplistic sentence structures that do not match the real-world data. To address this limitation, we propose GPT-HateCheck, a framework to generate more diverse and realistic functional tests from scratch by instructing large language models (LLMs). We employ an additional natural language inference (NLI) model to verify the generations. Crowd-sourced annotation demonstrates that the generated test cases are of high quality. Using the new functional tests, we can uncover model weaknesses that would be overlooked using the original HateCheck dataset.
Decoder-based large language models (LLMs) have shown high performance on many tasks in natural language processing. This is also true for sentence embedding learning, where a decoder-based model, PromptEOL, has achieved the best performance on semantic textual similarity (STS) tasks. However, PromptEOL requires a manually annotated natural language inference (NLI) dataset for fine-tuning. We aim to improve sentence embeddings without using large manually annotated datasets by automatically generating an NLI dataset with an LLM and using it for fine-tuning of PromptEOL. To achieve this, we explore methods of data generation suitable for sentence embedding learning in this study. Specifically, we will focus on automatic dataset generation through few-shot learning and explore the appropriate methods to leverage few-shot examples. Experimental results on the STS tasks demonstrate that our approach outperforms existing models in settings without large manually annotated datasets.
Credal sets are sets of probability distributions that are considered as candidates for an imprecisely known ground-truth distribution. In machine learning, they have recently attracted attention as an appealing formalism for uncertainty representation, in particular due to their ability to represent both the aleatoric and epistemic uncertainty in a prediction. However, the design of methods for learning credal set predictors remains a challenging problem. In this paper, we make use of conformal prediction for this purpose. More specifically, we propose a method for predicting credal sets in the classification task, given training data labeled by probability distributions. Since our method inherits the coverage guarantees of conformal prediction, our conformal credal sets are guaranteed to be valid with high probability (without any assumptions on model or distribution). We demonstrate the applicability of our method to natural language inference, a highly ambiguous natural language task where it is common to obtain multiple annotations per example.
Pretrained language models are long known to be subpar in capturing sentence and document-level semantics. Though heavily investigated, transferring perturbation-based methods from unsupervised visual representation learning to NLP remains an unsolved problem. This is largely due to the discreteness of subword units brought by tokenization of language models, limiting small perturbations of inputs to form semantics-preserved positive pairs. In this work, we conceptualize the learning of sentence-level textual semantics as a visual representation learning process. Drawing from cognitive and linguistic sciences, we introduce an unsupervised visual sentence representation learning framework, employing visually-grounded text perturbation methods like typos and word order shuffling, resonating with human cognitive patterns, and enabling perturbation to texts to be perceived as continuous. Our approach is further bolstered by large-scale unsupervised topical alignment training and natural language inference supervision, achieving comparable performance in semantic textual similarity (STS) to existing state-of-the-art NLP methods. Additionally, we unveil our method's inherent zero-shot cross-lingual transferability and a unique leapfrogging pattern across languages during iterative training. To our knowledge, this is the first representation learning method devoid of traditional language models for understanding sentence and document semantics, marking a stride closer to human-like textual comprehension. Our code is available at https://github.com/gowitheflow-1998/Pixel-Linguist
The increasing use of complex and opaque black box models requires the adoption of interpretable measures, one such option is extractive rationalizing models, which serve as a more interpretable alternative. These models, also known as Explain-Then-Predict models, employ an explainer model to extract rationales and subsequently condition the predictor with the extracted information. Their primary objective is to provide precise and faithful explanations, represented by the extracted rationales. In this paper, we take a semi-supervised approach to optimize for the plausibility of extracted rationales. We adopt a pre-trained natural language inference (NLI) model and further fine-tune it on a small set of supervised rationales ($10\%$). The NLI predictor is leveraged as a source of supervisory signals to the explainer via entailment alignment. We show that, by enforcing the alignment agreement between the explanation and answer in a question-answering task, the performance can be improved without access to ground truth labels. We evaluate our approach on the ERASER dataset and show that our approach achieves comparable results with supervised extractive models and outperforms unsupervised approaches by $> 100\%$.
The self-rationalising capabilities of LLMs are appealing because the generated explanations can give insights into the plausibility of the predictions. However, how faithful the explanations are to the predictions is questionable, raising the need to explore the patterns behind them further. To this end, we propose a hypothesis-driven statistical framework. We use a Bayesian network to implement a hypothesis about how a task (in our example, natural language inference) is solved, and its internal states are translated into natural language with templates. Those explanations are then compared to LLM-generated free-text explanations using automatic and human evaluations. This allows us to judge how similar the LLM's and the Bayesian network's decision processes are. We demonstrate the usage of our framework with an example hypothesis and two realisations in Bayesian networks. The resulting models do not exhibit a strong similarity to GPT-3.5. We discuss the implications of this as well as the framework's potential to approximate LLM decisions better in future work.
Efficiently fine-tuning Large Language Models (LLMs) for specific tasks presents a considerable challenge in natural language processing. Traditional methods, like prompt or prefix tuning, typically rely on arbitrary tokens for training, leading to prolonged training times and generalized token use across various class labels. To address these issues, this paper introduces L-Tuning, an efficient fine-tuning approach designed for classification tasks within the Natural Language Inference (NLI) framework. Diverging from conventional methods, L-Tuning focuses on the fine-tuning of label tokens processed through a pre-trained LLM, thereby harnessing its pre-existing semantic knowledge. This technique not only improves the fine-tuning accuracy and efficiency but also facilitates the generation of distinct label embeddings for each class, enhancing the model's training nuance. Our experimental results indicate a significant improvement in training efficiency and classification accuracy with L-Tuning compared to traditional approaches, marking a promising advancement in fine-tuning LLMs for complex language tasks.
Black-box hard-label adversarial attack on text is a practical and challenging task, as the text data space is inherently discrete and non-differentiable, and only the predicted label is accessible. Research on this problem is still in the embryonic stage and only a few methods are available. Nevertheless, existing methods rely on the complex heuristic algorithm or unreliable gradient estimation strategy, which probably fall into the local optimum and inevitably consume numerous queries, thus are difficult to craft satisfactory adversarial examples with high semantic similarity and low perturbation rate in a limited query budget. To alleviate above issues, we propose a simple yet effective framework to generate high quality textual adversarial examples under the black-box hard-label attack scenarios, named HQA-Attack. Specifically, after initializing an adversarial example randomly, HQA-attack first constantly substitutes original words back as many as possible, thus shrinking the perturbation rate. Then it leverages the synonym set of the remaining changed words to further optimize the adversarial example with the direction which can improve the semantic similarity and satisfy the adversarial condition simultaneously. In addition, during the optimizing procedure, it searches a transition synonym word for each changed word, thus avoiding traversing the whole synonym set and reducing the query number to some extent. Extensive experimental results on five text classification datasets, three natural language inference datasets and two real-world APIs have shown that the proposed HQA-Attack method outperforms other strong baselines significantly.
An increasing amount of research in Natural Language Inference (NLI) focuses on the application and evaluation of Large Language Models (LLMs) and their reasoning capabilities. Despite their success, however, LLMs are still prone to factual errors and inconsistencies in their explanations, offering limited control and interpretability for inference in complex domains. In this paper, we focus on ethical NLI, investigating how hybrid neuro-symbolic techniques can enhance the logical validity and alignment of ethical explanations produced by LLMs. Specifically, we present an abductive-deductive framework named Logic-Explainer, which integrates LLMs with an external backward-chaining solver to refine step-wise natural language explanations and jointly verify their correctness, reduce incompleteness and minimise redundancy. An extensive empirical analysis demonstrates that Logic-Explainer can improve explanations generated via in-context learning methods and Chain-of-Thought (CoT) on challenging ethical NLI tasks, while, at the same time, producing formal proofs describing and supporting models' reasoning. As ethical NLI requires commonsense reasoning to identify underlying moral violations, our results suggest the effectiveness of neuro-symbolic methods for multi-step NLI more broadly, opening new opportunities to enhance the logical consistency, reliability, and alignment of LLMs.
Traditionally, Machine Translation (MT) Evaluation has been treated as a regression problem -- producing an absolute translation-quality score. This approach has two limitations: i) the scores lack interpretability, and human annotators struggle with giving consistent scores; ii) most scoring methods are based on (reference, translation) pairs, limiting their applicability in real-world scenarios where references are absent. In practice, we often care about whether a new MT system is better or worse than some competitors. In addition, reference-free MT evaluation is increasingly practical and necessary. Unfortunately, these two practical considerations have yet to be jointly explored. In this work, we formulate the reference-free MT evaluation into a pairwise ranking problem. Given the source sentence and a pair of translations, our system predicts which translation is better. In addition to proposing this new formulation, we further show that this new paradigm can demonstrate superior correlation with human judgments by merely using indirect supervision from natural language inference and weak supervision from our synthetic data. In the context of reference-free evaluation, MT-Ranker, trained without any human annotations, achieves state-of-the-art results on the WMT Shared Metrics Task benchmarks DARR20, MQM20, and MQM21. On a more challenging benchmark, ACES, which contains fine-grained evaluation criteria such as addition, omission, and mistranslation errors, MT-Ranker marks state-of-the-art against reference-free as well as reference-based baselines.
Text simplification aims to make technical texts more accessible to laypeople but often results in deletion of information and vagueness. This work proposes InfoLossQA, a framework to characterize and recover simplification-induced information loss in form of question-and-answer (QA) pairs. Building on the theory of Question Under Discussion, the QA pairs are designed to help readers deepen their knowledge of a text. We conduct a range of experiments with this framework. First, we collect a dataset of 1,000 linguist-curated QA pairs derived from 104 LLM simplifications of scientific abstracts of medical studies. Our analyses of this data reveal that information loss occurs frequently, and that the QA pairs give a high-level overview of what information was lost. Second, we devise two methods for this task: end-to-end prompting of open-source and commercial language models, and a natural language inference pipeline. With a novel evaluation framework considering the correctness of QA pairs and their linguistic suitability, our expert evaluation reveals that models struggle to reliably identify information loss and applying similar standards as humans at what constitutes information loss.
Knowledge base population seeks to expand knowledge graphs with facts that are typically extracted from a text corpus. Recently, language models pretrained on large corpora have been shown to contain factual knowledge that can be retrieved using cloze-style strategies. Such approach enables zero-shot recall of facts, showing competitive results in object prediction compared to supervised baselines. However, prompt-based fact retrieval can be brittle and heavily depend on the prompts and context used, which may produce results that are unintended or hallucinatory.We propose to use textual entailment to validate facts extracted from language models through cloze statements. Our results show that triple validation based on textual entailment improves language model predictions in different training regimes. Furthermore, we show that entailment-based triple validation is also effective to validate candidate facts extracted from other sources including existing knowledge graphs and text passages where named entities are recognized.
Recent studies of the emergent capabilities of transformer-based Natural Language Understanding (NLU) models have indicated that they have an understanding of lexical and compositional semantics. We provide evidence that suggests these claims should be taken with a grain of salt: we find that state-of-the-art Natural Language Inference (NLI) models are sensitive towards minor semantics preserving surface-form variations, which lead to sizable inconsistent model decisions during inference. Notably, this behaviour differs from valid and in-depth comprehension of compositional semantics, however does neither emerge when evaluating model accuracy on standard benchmarks nor when probing for syntactic, monotonic, and logically robust reasoning. We propose a novel framework to measure the extent of semantic sensitivity. To this end, we evaluate NLI models on adversarially generated examples containing minor semantics-preserving surface-form input noise. This is achieved using conditional text generation, with the explicit condition that the NLI model predicts the relationship between the original and adversarial inputs as a symmetric equivalence entailment. We systematically study the effects of the phenomenon across NLI models for $\textbf{in-}$ and $\textbf{out-of-}$ domain settings. Our experiments show that semantic sensitivity causes performance degradations of $12.92\%$ and $23.71\%$ average over $\textbf{in-}$ and $\textbf{out-of-}$ domain settings, respectively. We further perform ablation studies, analysing this phenomenon across models, datasets, and variations in inference and show that semantic sensitivity can lead to major inconsistency within model predictions.
Accurately typing entity mentions from text segments is a fundamental task for various natural language processing applications. Many previous approaches rely on massive human-annotated data to perform entity typing. Nevertheless, collecting such data in highly specialized science and engineering domains (e.g., software engineering and security) can be time-consuming and costly, without mentioning the domain gaps between training and inference data if the model needs to be applied to confidential datasets. In this paper, we study the task of seed-guided fine-grained entity typing in science and engineering domains, which takes the name and a few seed entities for each entity type as the only supervision and aims to classify new entity mentions into both seen and unseen types (i.e., those without seed entities). To solve this problem, we propose SEType which first enriches the weak supervision by finding more entities for each seen type from an unlabeled corpus using the contextualized representations of pre-trained language models. It then matches the enriched entities to unlabeled text to get pseudo-labeled samples and trains a textual entailment model that can make inferences for both seen and unseen types. Extensive experiments on two datasets covering four domains demonstrate the effectiveness of SEType in comparison with various baselines.
Instruction-tuned Large Language Models (LLMs) excel at many tasks and will even explain their reasoning, so-called self-explanations. However, convincing and wrong self-explanations can lead to unsupported confidence in LLMs, thus increasing risk. Therefore, it's important to measure if self-explanations truly reflect the model's behavior. Such a measure is called interpretability-faithfulness and is challenging to perform since the ground truth is inaccessible, and many LLMs only have an inference API. To address this, we propose employing self-consistency checks to measure faithfulness. For example, if an LLM says a set of words is important for making a prediction, then it should not be able to make its prediction without these words. While self-consistency checks are a common approach to faithfulness, they have not previously been successfully applied to LLM self-explanations for counterfactual, feature attribution, and redaction explanations. Our results demonstrate that faithfulness is explanation, model, and task-dependent, showing self-explanations should not be trusted in general. For example, with sentiment classification, counterfactuals are more faithful for Llama2, feature attribution for Mistral, and redaction for Falcon 40B.