This work presents a novel framework for training Arabic nested embedding models through Matryoshka Embedding Learning, leveraging multilingual, Arabic-specific, and English-based models, to highlight the power of nested embeddings models in various Arabic NLP downstream tasks. Our innovative contribution includes the translation of various sentence similarity datasets into Arabic, enabling a comprehensive evaluation framework to compare these models across different dimensions. We trained several nested embedding models on the Arabic Natural Language Inference triplet dataset and assessed their performance using multiple evaluation metrics, including Pearson and Spearman correlations for cosine similarity, Manhattan distance, Euclidean distance, and dot product similarity. The results demonstrate the superior performance of the Matryoshka embedding models, particularly in capturing semantic nuances unique to the Arabic language. Results demonstrated that Arabic Matryoshka embedding models have superior performance in capturing semantic nuances unique to the Arabic language, significantly outperforming traditional models by up to 20-25\% across various similarity metrics. These results underscore the effectiveness of language-specific training and highlight the potential of Matryoshka models in enhancing semantic textual similarity tasks for Arabic NLP.
Hallucination, a phenomenon where large language models (LLMs) produce output that is factually incorrect or unrelated to the input, is a major challenge for LLM applications that require accuracy and dependability. In this paper, we introduce a reliable and high-speed production system aimed at detecting and rectifying the hallucination issue within LLMs. Our system encompasses named entity recognition (NER), natural language inference (NLI), span-based detection (SBD), and an intricate decision tree-based process to reliably detect a wide range of hallucinations in LLM responses. Furthermore, we have crafted a rewriting mechanism that maintains an optimal mix of precision, response time, and cost-effectiveness. We detail the core elements of our framework and underscore the paramount challenges tied to response time, availability, and performance metrics, which are crucial for real-world deployment of these technologies. Our extensive evaluation, utilizing offline data and live production traffic, confirms the efficacy of our proposed framework and service.
In this paper, we introduce the VerifAI project, a pioneering open-source scientific question-answering system, designed to provide answers that are not only referenced but also automatically vetted and verifiable. The components of the system are (1) an Information Retrieval system combining semantic and lexical search techniques over scientific papers (PubMed), (2) a Retrieval-Augmented Generation (RAG) module using fine-tuned generative model (Mistral 7B) and retrieved articles to generate claims with references to the articles from which it was derived, and (3) a Verification engine, based on a fine-tuned DeBERTa and XLM-RoBERTa models on Natural Language Inference task using SciFACT dataset. The verification engine cross-checks the generated claim and the article from which the claim was derived, verifying whether there may have been any hallucinations in generating the claim. By leveraging the Information Retrieval and RAG modules, Verif.ai excels in generating factual information from a vast array of scientific sources. At the same time, the Verification engine rigorously double-checks this output, ensuring its accuracy and reliability. This dual-stage process plays a crucial role in acquiring and confirming factual information, significantly enhancing the information landscape. Our methodology could significantly enhance scientists' productivity, concurrently fostering trust in applying generative language models within scientific domains, where hallucinations and misinformation are unacceptable.
Our collective attention span is shortened by the flood of online information. With \textit{FarFetched}, we address the need for automated claim validation based on the aggregated evidence derived from multiple online news sources. We introduce an entity-centric reasoning framework in which latent connections between events, actions, or statements are revealed via entity mentions and represented in a graph database. Using entity linking and semantic similarity, we offer a way for collecting and combining information from diverse sources in order to generate evidence relevant to the user's claim. Then, we leverage textual entailment recognition to quantitatively determine whether this assertion is credible, based on the created evidence. Our approach tries to fill the gap in automated claim validation for less-resourced languages and is showcased on the Greek language, complemented by the training of relevant semantic textual similarity (STS) and natural language inference (NLI) models that are evaluated on translated versions of common benchmarks.
Methods to evaluate Large Language Model (LLM) responses and detect inconsistencies, also known as hallucinations, with respect to the provided knowledge, are becoming increasingly important for LLM applications. Current metrics fall short in their ability to provide explainable decisions, systematically check all pieces of information in the response, and are often too computationally expensive to be used in practice. We present GraphEval: a hallucination evaluation framework based on representing information in Knowledge Graph (KG) structures. Our method identifies the specific triples in the KG that are prone to hallucinations and hence provides more insight into where in the response a hallucination has occurred, if at all, than previous methods. Furthermore, using our approach in conjunction with state-of-the-art natural language inference (NLI) models leads to an improvement in balanced accuracy on various hallucination benchmarks, compared to using the raw NLI models. Lastly, we explore the use of GraphEval for hallucination correction by leveraging the structure of the KG, a method we name GraphCorrect, and demonstrate that the majority of hallucinations can indeed be rectified.
Large language models (LLMs) are very proficient text generators. We leverage this capability of LLMs to generate task-specific data via zero-shot prompting and promote cross-lingual transfer for low-resource target languages. Given task-specific data in a source language and a teacher model trained on this data, we propose using this teacher to label LLM generations and employ a set of simple data selection strategies that use the teacher's label probabilities. Our data selection strategies help us identify a representative subset of diverse generations that help boost zero-shot accuracies while being efficient, in comparison to using all the LLM generations (without any subset selection). We also highlight other important design choices that affect cross-lingual performance such as the use of translations of source data and what labels are best to use for the LLM generations. We observe significant performance gains across sentiment analysis and natural language inference tasks (of up to a maximum of 7.13 absolute points and 1.5 absolute points on average) across a number of target languages (Hindi, Marathi, Urdu, Swahili) and domains.
Large language models (LLMs) exhibit hallucinations in long-form question-answering tasks across various domains and wide applications. Current hallucination detection and mitigation datasets are limited in domains and sizes, which struggle to scale due to prohibitive labor costs and insufficient reliability of existing hallucination annotators. To facilitate the scalable oversight of LLM hallucinations, this paper introduces an iterative self-training framework that simultaneously and progressively scales up the hallucination annotation dataset and improves the accuracy of the hallucination annotator. Based on the Expectation Maximization (EM) algorithm, in each iteration, the framework first applies a hallucination annotation pipeline to annotate a scaled dataset and then trains a more accurate hallucination annotator on the dataset. This new hallucination annotator is adopted in the hallucination annotation pipeline used for the next iteration. Extensive experimental results demonstrate that the finally obtained hallucination annotator with only 7B parameters surpasses the performance of GPT-4 and obtains new state-of-the-art hallucination detection results on HaluEval and HalluQA by zero-shot inference. Such an annotator can not only evaluate the hallucination levels of various LLMs on the large-scale dataset but also help to mitigate the hallucination of LLMs generations, with the Natural Language Inference (NLI) metric increasing from 25% to 37% on HaluEval.
Large Language Models (LLMs) are widely used for writing economic analysis reports or providing financial advice, but their ability to understand economic knowledge and reason about potential results of specific economic events lacks systematic evaluation. To address this gap, we propose a new dataset, natural language inference on economic events (EconNLI), to evaluate LLMs' knowledge and reasoning abilities in the economic domain. We evaluate LLMs on (1) their ability to correctly classify whether a premise event will cause a hypothesis event and (2) their ability to generate reasonable events resulting from a given premise. Our experiments reveal that LLMs are not sophisticated in economic reasoning and may generate wrong or hallucinated answers. Our study raises awareness of the limitations of using LLMs for critical decision-making involving economic reasoning and analysis. The dataset and codes are available at https://github.com/Irenehere/EconNLI.
Each new generation of English-oriented Large Language Models (LLMs) exhibits enhanced cross-lingual transfer capabilities and significantly outperforms older LLMs on low-resource languages. This prompts the question: Is there a need for LLMs dedicated to a particular low-resource language? We aim to explore this question for Bengali, a low-to-moderate resource Indo-Aryan language native to the Bengal region of South Asia. We compare the performance of open-weight and closed-source LLMs such as LLaMA-3 and GPT-4 against fine-tuned encoder-decoder models across a diverse set of Bengali downstream tasks, including translation, summarization, paraphrasing, question-answering, and natural language inference. Our findings reveal that while LLMs generally excel in reasoning tasks, their performance in tasks requiring Bengali script generation is inconsistent. Key challenges include inefficient tokenization of Bengali script by existing LLMs, leading to increased computational costs and potential performance degradation. Additionally, we highlight biases in machine-translated datasets commonly used for Bengali NLP tasks. We conclude that there is a significant need for a Bengali-oriented LLM, but the field currently lacks the high-quality pretraining and instruction-tuning datasets necessary to develop a highly effective model.
Advancing representation learning in specialized fields like medicine remains challenging due to the scarcity of expert annotations for text and images. To tackle this issue, we present a novel two-stage framework designed to extract high-quality factual statements from free-text radiology reports in order to improve the representations of text encoders and, consequently, their performance on various downstream tasks. In the first stage, we propose a \textit{Fact Extractor} that leverages large language models (LLMs) to identify factual statements from well-curated domain-specific datasets. In the second stage, we introduce a \textit{Fact Encoder} (CXRFE) based on a BERT model fine-tuned with objective functions designed to improve its representations using the extracted factual data. Our framework also includes a new embedding-based metric (CXRFEScore) for evaluating chest X-ray text generation systems, leveraging both stages of our approach. Extensive evaluations show that our fact extractor and encoder outperform current state-of-the-art methods in tasks such as sentence ranking, natural language inference, and label extraction from radiology reports. Additionally, our metric proves to be more robust and effective than existing metrics commonly used in the radiology report generation literature. The code of this project is available at \url{https://github.com/PabloMessina/CXR-Fact-Encoder}.
Trustworthy prediction in Deep Neural Networks (DNNs), including Pre-trained Language Models (PLMs) is important for safety-critical applications in the real world. However, DNNs often suffer from uncertainty estimation, such as miscalibration. In particular, approaches that require multiple stochastic inference can mitigate this problem, but the expensive cost of inference makes them impractical. In this study, we propose $k$-Nearest Neighbor Uncertainty Estimation ($k$NN-UE), which is an uncertainty estimation method that uses the distances from the neighbors and label-existence ratio of neighbors. Experiments on sentiment analysis, natural language inference, and named entity recognition show that our proposed method outperforms the baselines or recent density-based methods in confidence calibration, selective prediction, and out-of-distribution detection. Moreover, our analyses indicate that introducing dimension reduction or approximate nearest neighbor search inspired by recent $k$NN-LM studies reduces the inference overhead without significantly degrading estimation performance when combined them appropriately.
Reinforcement Learning (RL) has been used to finetune Large Language Models (LLMs) using a reward model trained from preference data, to better align with human judgment. The recently introduced direct alignment methods, which are often simpler, more stable, and computationally lighter, can more directly achieve this. However, these approaches cannot optimize arbitrary rewards, and the preference-based ones are not the only rewards of interest for LLMs (eg., unit tests for code generation or textual entailment for summarization, among others). RL-finetuning is usually done with a variation of policy gradient, which calls for on-policy or near-on-policy samples, requiring costly generations. We introduce Contrastive Policy Gradient, or CoPG, a simple and mathematically principled new RL algorithm that can estimate the optimal policy even from off-policy data. It can be seen as an off-policy policy gradient approach that does not rely on important sampling techniques and highlights the importance of using (the right) state baseline. We show this approach to generalize the direct alignment method IPO (identity preference optimization) and classic policy gradient. We experiment with the proposed CoPG on a toy bandit problem to illustrate its properties, as well as for finetuning LLMs on a summarization task, using a learned reward function considered as ground truth for the purpose of the experiments.
The development of Natural Language Processing (NLI) datasets and models has been inspired by innovations in annotation design. With the rapid development of machine learning models today, the performance of existing machine learning models has quickly reached state-of-the-art results on a variety of tasks related to natural language processing, including natural language inference tasks. By using a pre-trained model during the annotation process, it is possible to challenge current NLI models by having humans produce premise-hypothesis combinations that the machine model cannot correctly predict. To remain attractive and challenging in the research of natural language inference for Vietnamese, in this paper, we introduce the adversarial NLI dataset to the NLP research community with the name ViANLI. This data set contains more than 10K premise-hypothesis pairs and is built by a continuously adjusting process to obtain the most out of the patterns generated by the annotators. ViANLI dataset has brought many difficulties to many current SOTA models when the accuracy of the most powerful model on the test set only reached 48.4%. Additionally, the experimental results show that the models trained on our dataset have significantly improved the results on other Vietnamese NLI datasets.
Human label variation (HLV) is a valuable source of information that arises when multiple human annotators provide different labels for valid reasons. In Natural Language Inference (NLI) earlier approaches to capturing HLV involve either collecting annotations from many crowd workers to represent human judgment distribution (HJD) or use expert linguists to provide detailed explanations for their chosen labels. While the former method provides denser HJD information, obtaining it is resource-intensive. In contrast, the latter offers richer textual information but it is challenging to scale up to many human judges. Besides, large language models (LLMs) are increasingly used as evaluators ("LLM judges") but with mixed results, and few works aim to study HJDs. This study proposes to exploit LLMs to approximate HJDs using a small number of expert labels and explanations. Our experiments show that a few explanations significantly improve LLMs' ability to approximate HJDs with and without explicit labels, thereby providing a solution to scale up annotations for HJD. However, fine-tuning smaller soft-label aware models with the LLM-generated model judgment distributions (MJDs) presents partially inconsistent results: while similar in distance, their resulting fine-tuned models and visualized distributions differ substantially. We show the importance of complementing instance-level distance measures with a global-level shape metric and visualization to more effectively evaluate MJDs against human judgment distributions.
Hyperbolic spaces have proven to be suitable for modeling data of hierarchical nature. As such we use the Poincare ball to embed sentences with the goal of proving how hyperbolic spaces can be used for solving Textual Entailment. To this end, apart from the standard datasets used for evaluating textual entailment, we developed two additional datasets. We evaluate against baselines of various backgrounds, including LSTMs, Order Embeddings and Euclidean Averaging, which comes as a natural counterpart to representing sentences into the Euclidean space. We consistently outperform the baselines on the SICK dataset and are second only to Order Embeddings on the SNLI dataset, for the binary classification version of the entailment task.
We explore the relationship between factuality and Natural Language Inference (NLI) by introducing FactRel -- a novel annotation scheme that models \textit{factual} rather than \textit{textual} entailment, and use it to annotate a dataset of naturally occurring sentences from news articles. Our analysis shows that 84\% of factually supporting pairs and 63\% of factually undermining pairs do not amount to NLI entailment or contradiction, respectively, suggesting that factual relationships are more apt for analyzing media discourse. We experiment with models for pairwise classification on the new dataset, and find that in some cases, generating synthetic data with GPT-4 on the basis of the annotated dataset can improve performance. Surprisingly, few-shot learning with GPT-4 yields strong results on par with medium LMs (DeBERTa) trained on the labelled dataset. We hypothesize that these results indicate the fundamental dependence of this task on both world knowledge and advanced reasoning abilities.
Scientific Natural Language Inference (NLI) is the task of predicting the semantic relation between a pair of sentences extracted from research articles. The automatic annotation method based on distant supervision for the training set of SciNLI (Sadat and Caragea, 2022b), the first and most popular dataset for this task, results in label noise which inevitably degenerates the performance of classifiers. In this paper, we propose a novel co-training method that assigns weights based on the training dynamics of the classifiers to the distantly supervised labels, reflective of the manner they are used in the subsequent training epochs. That is, unlike the existing semi-supervised learning (SSL) approaches, we consider the historical behavior of the classifiers to evaluate the quality of the automatically annotated labels. Furthermore, by assigning importance weights instead of filtering out examples based on an arbitrary threshold on the predicted confidence, we maximize the usage of automatically labeled data, while ensuring that the noisy labels have a minimal impact on model training. The proposed method obtains an improvement of 1.5% in Macro F1 over the distant supervision baseline, and substantial improvements over several other strong SSL baselines. We make our code and data available on Github.
This paper describes the inference system of FZI-WIM at the SemEval-2024 Task 2: Safe Biomedical Natural Language Inference for Clinical Trials. Our system utilizes the chain of thought (CoT) paradigm to tackle this complex reasoning problem and further improves the CoT performance with self-consistency. Instead of greedy decoding, we sample multiple reasoning chains with the same prompt and make the final verification with majority voting. The self-consistent CoT system achieves a baseline F1 score of 0.80 (1st), faithfulness score of 0.90 (3rd), and consistency score of 0.73 (12th). We release the code and data publicly https://github.com/jens5588/FZI-WIM-NLI4CT.
Negation is a common linguistic phenomenon. Yet language models face challenges with negation in many natural language understanding tasks such as question answering and natural language inference. In this paper, we experiment with seamless strategies that incorporate affirmative interpretations (i.e., paraphrases without negation) to make models more robust against negation. Crucially, our affirmative interpretations are obtained automatically. We show improvements with CondaQA, a large corpus requiring reasoning with negation, and five natural language understanding tasks.
Attributing answer text to its source document for information-seeking questions is crucial for building trustworthy, reliable, and accountable systems. We formulate a new task of post-hoc answer attribution for long document comprehension (LDC). Owing to the lack of long-form abstractive and information-seeking LDC datasets, we refactor existing datasets to assess the strengths and weaknesses of existing retrieval-based and proposed answer decomposition and textual entailment-based optimal selection attribution systems for this task. We throw light on the limitations of existing datasets and the need for datasets to assess the actual performance of systems on this task.