The realm of Large Language Models (LLMs) is undergoing a continuous and dynamic transformation. These state-of-the-art LLMs showcase an impressive ability to craft narratives based on contextual cues, highlighting their skill in comprehending and producing text resembling human writing. However, there exists a potential risk: the potential inclination of LLMs to create gossips when prompted with specific contexts. These LLMs possess the capacity to generate stories rooted in the context provided by the prompts. Yet, this very capability carries a risk of generating gossips given the context as input. To mitigate this, we introduce a dataset named “French GossipPrompts” designed for identifying prompts that lead to the creation of gossipy content in the French language. This dataset employs binary classification, categorizing whether a given prompt generates gossip or not. The dataset comprises a total of 7253 individual prompts. We have developed classification models and achieved an accuracy of 89.95%.

This paper explores an empirical approach to learn more discriminantive sentence representations in an unsupervised fashion. Leveraging semantic graph smoothing, we enhance sentence embeddings obtained from pretrained models to improve results for the text clustering and classification tasks. Our method, validated on eight benchmarks, demonstrates consistent improvements, showcasing the potential of semantic graph smoothing in improving sentence embeddings for the supervised and unsupervised document categorization tasks.

This paper explores the task of language-agnostic speaker replication, a novel endeavor that seeks to replicate a speaker’s voice irrespective of the language they are speaking. Towards this end, we introduce a multi-level attention aggregation approach that systematically probes and amplifies various speaker-specific attributes in a hierarchical manner. Through rigorous evaluations across a wide range of scenarios including seen and unseen speakers conversing in seen and unseen lingua, we establish that our proposed model is able to achieve substantial speaker similarity, and is able to generalize to out-of-domain (OOD) cases.

Hallucinations and off-target translation remain unsolved problems in MT, especially for low-resource languages and massively multilingual models. In this paper, we introduce two related methods to mitigate these failure cases with a modified decoding objective, without either requiring retraining or external models. In source-contrastive decoding, we search for a translation that is probable given the correct input, but improbable given a random input segment. In language-contrastive decoding, we search for a translation that is probable, but improbable given the wrong language indicator token. Experiments on the massively multilingual models M2M-100 (418M) and SMaLL-100 show that these methods suppress hallucinations and off-target translations, reducing the number of translations with segment-level chrF2 below 10 by 67-83% on average across 57 tested translation directions. In a proof of concept on out-of-English translation, we also show that we can suppress off-target translations with large language models. We release code upon acceptance.

While static word embeddings are blind to context, for lexical semantics tasks context is rather too present in contextual word embeddings, vectors of same-meaning occurrences being too different (Ethayarajh, 2019). Fine-tuning pre-trained language models (PLMs) using contrastive learning was proposed, leveraging automatically self-augmented examples (Liu et al., 2021b). In this paper, we investigate how to inject a lexicon as an alternative source of supervision, using the English Wiktionary. We also test how dimensionality reduction impacts the resulting contextual word embeddings. We evaluate our approach on the Word-In-Context (WiC) task, in the unsupervised setting (not using the training set). We achieve new SoTA result on the original WiC test set. We also propose two new WiC test sets for which we show that our fine-tuning method achieves substantial improvements. We also observe improvements, although modest, for the semantic frame induction task. Although we experimented on English to allow comparison with related work, our method is adaptable to the many languages for which large Wiktionaries exist.

While information from the field of linguistic typology has the potential to improve performance on NLP tasks, reliable typological data is a prerequisite. Existing typological databases, including WALS and Grambank, suffer from inconsistencies primarily caused by their categorical format. Furthermore, typological categorisations by definition differ significantly from the continuous nature of phenomena, as found in natural language corpora. In this paper, we introduce a new seed dataset made up of continuous-valued data, rather than categorical data, that can better reflect the variability of language. While this initial dataset focuses on word-order typology, we also present the methodology used to create the dataset, which can be easily adapted to generate data for a broader set of features and languages.

Recent work has shown that large language models (LLMs) are capable of generating summaries zero-shot—i.e., without explicit supervision—that, under human assessment, are often comparable or even preferred to manually composed reference summaries. However, this prior work has focussed almost exclusively on evaluating news article summarization. How do zero-shot summarizers perform in other (potentially more specialized) domains?In this work we evaluate zero-shot generated summaries across specialized domains including: biomedical articles, and legal bills (in addition to standard news benchmarks for reference). We focus especially on the factuality of outputs. We acquire annotations from domain experts to identify inconsistencies in summaries and systematically categorize these errors. We analyze whether the prevalence of a given domain in the pretraining corpus affects extractiveness and faithfulness of generated summaries of articles in this domain. We release all collected annotations to facilitate additional research toward measuring and realizing factually accurate summarization, beyond news articles (The dataset can be downloaded from https://anonymous.4open.science/r/zero_shot_faceval_domains-9B83)

We study improving social conversational agents by learning from natural dialogue between users and a deployed model, without extra annotations. To implicitly measure the quality of a machine-generated utterance, we leverage signals like user response length, sentiment and reaction of the future human utterances in the collected dialogue episodes. Our experiments use the publicly released deployment data from BlenderBot (Xu et al., 2023). Human evaluation indicates improvements in our new models over baseline responses; however, we find that some proxy signals can lead to more generations with undesirable properties as well. For example, optimizing for conversation length can lead to more controversial or unfriendly generations compared to the baseline, whereas optimizing for positive sentiment or reaction can decrease these behaviors.

There has recently been a growing interest in using Large Language Models (LLMs) to evaluate NLP tasks automatically. Considerable research effort has been put into improving such systems towards achieving high correlations with human judgement. However, it is still unclear what level of correlation is good enough for practical applications of LLM-based automatic evaluation systems. This paper characterizes these LLM evaluators’ confidence in ranking candidate NLP models and develops a configurable Monte Carlo simulation method. We show that even automatic metrics with low correlation with human judgement can reach high-confidence rankings of candidate models with reasonable evaluation set sizes (100s of examples). Further, we describe tradeoff curves between the LLM evaluator performance (i.e., correlation with humans) and evaluation set size; loss in correlation can be compensated with modest increases in the evaluation set size. We validate our results on RoSE, a text summarization dataset, and find our estimates of confidence align with empirical observations.Code available at https://github.com/rickardstureborg/llm-eval-confidence

Large language models (LLMs) have exhibited an array of reasoning capabilities but face challenges like error propagation and hallucination, particularly in specialised areas like finance, where data is heterogeneous, and precision is paramount. We explore the potential of language model augmentation with external tools to mitigate these limitations and offload certain reasoning steps to external tools that are more suited for the task, instead of solely depending on the LLM’s inherent abilities. More concretely, using financial domain question answering datasets, we apply supervised finetuning on a LLAMA-2 13B CHAT model to act both as a task router and task solver. The task router dynamically directs a question to either be answered internally by the LLM or externally via the right tool from the tool set. Our tool-equipped SFT model, RAVEN, demonstrates an improvement of 35.2% and 5.06% over the base model and SFT-only baselines, respectively, and is highly competitive with strong GPT-3.5 results. To the best of our knowledge, our work is the first that investigates tool augmentation of language models for the finance domain.

Memorizing and utilizing speakers’ personas is a common practice for response generation in long-term conversations. Yet, human-authored datasets often provide uninformative persona sentences that hinder response quality. This paper presents a novel framework that leverages commonsense-based persona expansion to address such issues in long-term conversation.While prior work focuses on not producing personas that contradict others, we focus on transforming contradictory personas into sentences that contain rich speaker information, by refining them based on their contextual backgrounds with designed strategies. As the pioneer of persona expansion in multi-session settings, our framework facilitates better response generation via human-like persona refinement. The supplementary video of our work is available at https://caffeine-15bbf.web.app/.

Cross-lingual transfer has recently been tackled through modular, parameter-efficient fine-tuning methods which allow arbitrary combinations of language and task modules for transfer of any task to any language. Concurrently, task arithmetic has emerged as a powerful and modular tool for editing pretrained models using multiple full fine-tunings. In this work, we connect the paradigms of task arithmetic and cross-lingual transfer, demonstrating that modularity for cross-lingual transfer can be achieved even with full model fine-tuning. Our approach displays strong performance on a range of multilingual benchmarks encompassing both high-resource and low-resource languages.

Text summarization and simplification are among the most widely used applications of AI. However, such models are often prone to hallucination, which can result from training models on unaligned data. One efficient approach to address this issue is Loss Truncation (Kang and Hashimoto, 2020), an approach to modify the standard log loss to adaptively remove noisy examples during training. However, we find that LT alone yields a considerable number of hallucinated entities on various datasets. We study the behavior of the underlying losses between factual and non-factual examples, to understand and refine the performance of LT. We demonstrate that LT’s performance is limited when the underlying assumption that noisy targets have higher NLL loss is not satisfied, and find that word-level NLL among entities provides better signal for distinguishing factuality. We then leverage this to propose a fine-grained NLL loss and fine-grained data cleaning strategies, and observe improvements in hallucination reduction across some datasets. Our work is available at https://github.com/yale-nlp/Simplification-Projects.

Scientific abstracts provide a concise summary of research findings, making them a valuable resource for extracting scientific arguments. In this study, we assess various unsupervised approaches for extracting arguments as aligned premise-conclusion pairs: semantic similarity, text perplexity, and mutual information. We aggregate structured abstracts from PubMed Central Open Access papers published in 2022 and evaluate the argument aligners in terms of the performance of language models that we fine-tune to generate the conclusions from the extracted premise given as input prompts. We find that mutual information outperforms the other measures on this task, suggesting that the reasoning process in scientific abstracts hinges mostly on linguistic constructs beyond simple textual similarity.

Large language models (LLMs) can solve problems step-by-step.While this chain-of-thought (CoT) reasoning boosts LLMs’ performance, it is unclear if LLMs know when to use CoT and whether those CoT are always necessary to answer the question. This paper shows that LLMs tend to generate redundant calculations and reasoning on a manually constructed math QA dataset, GSM8K-Zero.GSM8K-Zero is constructed such that the questions can be answered without any calculations, but LLMs, including Llama-2 models and Claude-2, tend to generate lengthy and unnecessary calculations to answer the questions.We also conduct experiments to explain why LLMs generate redundant calculations and reasonings.

Recent advances in NLP suggest that some tasks, such as argument detection and relation classification, are better framed in a multimodal perspective. We propose multimodal argument mining for argumentative fallacy classification in political debates. To this end, we release the first corpus for multimodal fallacy classification. Our experiments show that the integration of the audio modality leads to superior classification performance. Our findings confirm that framing fallacy classification as a multimodal task is essential to capture paralinguistic aspects of fallacious arguments.

In the realm of Computational Social Science (CSS), practitioners often navigate complex, low-resource domains and face the costly and time-intensive challenges of acquiring and annotating data. We aim to establish a set of guidelines to address such challenges, comparing the use of human-labeled data with synthetically generated data from GPT-4 and Llama-2 in ten distinct CSS classification tasks of varying complexity. Additionally, we examine the impact of training data sizes on performance. Our findings reveal that models trained on human-labeled data consistently exhibit superior or comparable performance compared to their synthetically augmented counterparts. Nevertheless, synthetic augmentation proves beneficial, particularly in improving performance on rare classes within multi-class tasks. Furthermore, we leverage GPT-4 and Llama-2 for zero-shot classification and find that, while they generally display strong performance, they often fall short when compared to specialized classifiers trained on moderately sized training sets.

Controlled text generation (CTG) seeks to guide large language model (LLM) output, that statistical language generation would conform to desired criteria. The current study presents a novel CTG algorithm that enforces adherence toward specific rhetorical relations in an LLM sentence-completion context by a parser-driven decoding scheme that requires no model fine-tuning. The method is validated both with automatic and human evaluation.

Indigenous languages are historically under-served by Natural Language Processing (NLP) technologies, but this is changing for some languages with the recent scaling of large multilingual models and an increased focus by the NLP community on endangered languages. This position paper explores ethical considerations in building NLP technologies for Indigenous languages, based on the premise that such projects should primarily serve Indigenous communities. We report on interviews with 17 researchers working in or with Aboriginal and/or Torres Strait Islander communities on language technology projects in Australia. Drawing on insights from the interviews, we recommend practices for NLP researchers to increase attention to the process of engagements with Indigenous communities, rather than focusing only on decontextualised artefacts.

Neural abstractive summarization models make summaries in an end-to-end manner, and little is known about how the source information is actually converted into summaries. In this paper, we define input sentences that contain essential information in the generated summary as source sentences and study how abstractive summaries are made by analyzing the source sentences. To this end, we annotate source sentences for reference summaries and system summaries generated by PEGASUS on document-summary pairs sampled from the CNN/DailyMail and XSum datasets. We also formulate automatic source sentence detection and compare multiple methods to establish a strong baseline for the task. Experimental results show that the perplexity-based method performs well in highly abstractive settings, while similarity-based methods perform robustly in relatively extractive settings.

We study incremental constituent parsers to assess their capacity to output trees based on prefix representations alone. Guided by strictly left-to-right generative language models and tree-decoding modules, we build parsers that adhere to a strong definition of incrementality across languages. This builds upon work that asserted incrementality, but that mostly only enforced it on either the encoder or the decoder. Finally, we conduct an analysis against non-incremental and partially incremental models.

Language models (LMs) are statistical models trained to assign probability to human-generated text. As such, it is reasonable to question whether they approximate linguistic variability exhibited by humans well. This form of statistical assessment is difficult to perform at the passage level, for it requires acceptability judgments (i.e., human evaluation) or a robust automated proxy (which is non-trivial). At the word level, however, given some context, samples from an LM can be assessed via exact matching against a prerecorded dataset of alternative single-word continuations of the available context. We exploit this fact and evaluate the LM’s ability to reproduce variability that humans (in particular, a population of English speakers) exhibit in the ‘next word prediction’ task. This can be seen as assessing a form of calibration, which, in the context of text classification, Baan et al. (2022) termed calibration to human uncertainty. We assess GPT2, BLOOM and ChatGPT and find that they exhibit fairly low calibration to human uncertainty. We also verify the failure of expected calibration error (ECE) to reflect this, and as such, advise the community against relying on it in this setting.

Gender-neutral translation (GNT) that avoids biased and undue binary assumptions is a pivotal challenge for the creation of more inclusive translation technologies. Advancements for this task in Machine Translation (MT), however, are hindered by the lack of dedicated parallel data, which are necessary to adapt MT systems to satisfy neutral constraints. For such a scenario, large language models offer hitherto unforeseen possibilities, as they come with the distinct advantage of being versatile in various (sub)tasks when provided with explicit instructions. In this paper, we explore this potential to automate GNT by comparing MT with the popular GPT-4 model. Through extensive manual analyses, our study empirically reveals the inherent limitations of current MT systems in generating GNTs and provides valuable insights into the potential and challenges associated with prompting for neutrality.

With the rise of increasingly powerful and user-facing NLP systems, there is growing interest in assessing whether they have a good _representation of uncertainty_ by evaluating the quality of their predictive distribution over outcomes. We identify two main perspectives that drive starkly different evaluation protocols. The first treats predictive probability as an indication of model confidence; the second as an indication of human label variation. We discuss their merits and limitations, and take the position that both are crucial for trustworthy and fair NLP systems, but that exploiting a single predictive distribution is limiting. We recommend tools and highlight exciting directions towards models with disentangled representations of uncertainty about predictions and uncertainty about human labels.

As large language models are becoming more embedded in different user-facing services, it is important to be able to distinguish between human-written and machine-generated text to verify the authenticity of news articles, product reviews, etc. Thus, in this paper we set out to explore whether it is possible to use one language model to identify machine-generated text produced by another language model, in a zero-shot way, even if the two have different architectures and are trained on different data. We find that overall, smaller models are better universal machine-generated text detectors: they can more precisely detect text generated from both small and larger models, without the need for any additional training/data. Interestingly, we find that whether or not the detector and generator models were trained on the same data is not critically important to the detection success. For instance the OPT-125M model has an AUC of 0.90 in detecting GPT4 generations, whereas a larger model from the GPT family, GPTJ-6B, has AUC of 0.65.

We address the task of machine translation (MT) from extremely low-resource language (ELRL) to English by leveraging cross-lingual transfer from *closely-related* high-resource language (HRL). The development of an MT system for ELRL is challenging because these languages typically lack parallel corpora and monolingual corpora, and their representations are absent from large multilingual language models. Many ELRLs share lexical similarities with some HRLs, which presents a novel modeling opportunity. However, existing subword-based neural MT models do not explicitly harness this lexical similarity, as they only implicitly align HRL and ELRL latent embedding space. To overcome this limitation, we propose a novel, CharSpan, approach based on character-span noise augmentation into the training data of HRL. This serves as a regularization technique, making the model more robust to lexical divergences between the HRL and ELRL, thus facilitating effective cross-lingual transfer. Our method significantly outperformed strong baselines in zero-shot settings on closely related HRL and ELRL pairs from three diverse language families, emerging as the state-of-the-art model for ELRLs.

Neural machine translation (NMT) systems are vulnerable when trained on limited data. This is a common scenario in low-resource tasks in the real world. To increase robustness, a solution is to intently add realistic noise in the training phase. Noise simulation using text perturbation has been proven to be efficient in writing systems that use Latin letters. In this study, we further explore perturbation techniques on more complex abugida writing systems, for which the visual similarity of complex glyphs is considered to capture the essential nature of these writing systems. Besides the generated noise, we propose a training strategy to improve robustness. We conducted experiments on six languages: Bengali, Hindi, Myanmar, Khmer, Lao, and Thai. By overcoming the introduced noise, we obtained non-degenerate NMT systems with improved robustness for low-resource tasks for abugida glyphs.

Popular benchmarks (e.g., XNLI) used to evaluate cross-lingual language understanding consist of parallel versions of English evaluation sets in multiple target languages created with the help of professional translators. When creating such parallel data, it is critical to ensure high-quality translations for all target languages for an accurate characterization of cross-lingual transfer. In this work, we find that translation inconsistencies do exist and interestingly they disproportionally impact low-resource languages in XNLI. To identify such inconsistencies, we propose measuring the gap in performance between zero-shot evaluations on the human-translated and machine-translated target text across multiple target languages; relatively large gaps are indicative of translation errors. We also corroborate that translation errors exist for two target languages, namely Hindi and Urdu, by doing a manual reannotation of human-translated test instances in these two languages and finding poor agreement with the original English labels these instances were supposed to inherit.

Large Language Models (LLMs) have shown promising performance in summary evaluation tasks, yet they face challenges such as high computational costs and the Lost-in-the-Middle problem where important information in the middle of long documents is often overlooked. To address these issues, this paper introduces a novel approach, Extract-then-Evaluate, which involves extracting key sentences from a long source document and then evaluating the summary by prompting LLMs. The results reveal that the proposed method not only significantly reduces evaluation costs but also exhibits a higher correlation with human evaluations. Furthermore, we provide practical recommendations for optimal document length and sentence extraction methods, contributing to the development of cost-effective yet more accurate methods for LLM-based text generation evaluation.

With the advent of large language models (LLMs), there has been growing interest in exploring their potential for medical applications. This research aims to investigate the ability of LLMs, specifically ChatGPT, in the context of pharmacovigilance event extraction, of which the main goal is to identify and extract adverse events or potential therapeutic events from textual medical sources. We conduct extensive experiments to assess the performance of ChatGPT in the pharmacovigilance event extraction task, employing various prompts and demonstration selection strategies. The findings demonstrate that while ChatGPT demonstrates reasonable performance with appropriate demonstration selection strategies, it still falls short compared to fully fine-tuned small models. Additionally, we explore the potential of leveraging ChatGPT for data augmentation. However, our investigation reveals that the inclusion of synthesized data into fine-tuning may lead to a decrease in performance, possibly attributed to noise in the ChatGPT-generated labels. To mitigate this, we explore different filtering strategies and find that, with the proper approach, more stable performance can be achieved, although constant improvement remains elusive.

Generating natural language text from graph-structured data is essential for conversational information seeking. Semantic triples derived from knowledge graphs can serve as a valuable source for grounding responses from conversational agents by providing a factual basis for the information they communicate. This is especially relevant in the context of large language models, which offer great potential for conversational interaction but are prone to hallucinating, omitting, or producing conflicting information. In this study, we conduct an empirical analysis of conversational large language models in generating natural language text from semantic triples. We compare four large language models of varying sizes with different prompting techniques. Through a series of benchmark experiments on the WebNLG dataset, we analyze the models’ performance and identify the most common issues in the generated predictions. Our findings show that the capabilities of large language models in triple verbalization can be significantly improved through few-shot prompting, post-processing, and efficient fine-tuning techniques, particularly for smaller models that exhibit lower zero-shot performance.

Although fine-tuning a pre-trained model with a conventional approach has shown to be effective in various downstream tasks, previous work has used only backpropagation to fine-tune the model, which causes a massive amount of computational resources and time. We propose Extreme Fine-Tuning (EFT), a novel approach for fine-tuning a pre-trained model effectively and efficiently. EFT uses backpropagation for a brief fine-tuning and an iterative extreme learning machine for training a classifier. We applied EFT to four text classification datasets, MELD, IEMOCAP, IMDb, and AG News, and compared its performance with state-of-the-art (SOTA) approaches. The results indicate that EFT noticeably outperformed the other approaches in training-time measurement with comparable model performance. We will release our code at https://github.com/up-33/extreme-fine-tuning.

Diffusion models are a promising tool for high-quality text generation. However, current models face multiple drawbacks including slow sampling, noise schedule sensitivity, and misalignment between the training and sampling stages. In this paper, we introduce FlowSeq, which bypasses all current drawbacks by leveraging flow matching for conditional text generation. FlowSeq can generate text in a few steps by training with a novel anchor loss, alleviating the need for expensive hyperparameter optimization of the noise schedule prevalent in diffusion models. We extensively evaluate our proposed method and show competitive performance in tasks such as question generation, open-domain dialogue, and paraphrasing tasks.

Recent studies demonstrate that query expansions generated by large language models (LLMs) can considerably enhance information retrieval systems by generating hypothetical documents that answer the queries as expansions. However, challenges arise from misalignments between the expansions and the retrieval corpus, resulting in issues like hallucinations and outdated information due to the limited intrinsic knowledge of LLMs. Inspired by Pseudo Relevance Feedback (PRF), we introduce Corpus-Steered Query Expansion (CSQE) to promote the incorporation of knowledge embedded within the corpus. CSQE utilizes the relevance assessing capability of LLMs to systematically identify pivotal sentences in the initially-retrieved documents. These corpus-originated texts are subsequently used to expand the query together with LLM-knowledge empowered expansions, improving the relevance prediction between the query and the target documents. Extensive experiments reveal that CSQE exhibits strong performance without necessitating any training, especially with queries for which LLMs lack knowledge.

Recent work in open-domain question answering (ODQA) has shown that adversarial poisoning of the search collection can cause large drops in accuracy for production systems. However, little to no work has proposed methods to defend against these attacks. To do so, we rely on the intuition that redundant information often exists in large corpora. To find it, we introduce a method that uses query augmentation to search for a diverse set of passages that could answer the original question but are less likely to have been poisoned. We integrate these new passages into the model through the design of a novel confidence method, comparing the predicted answer to its appearance in the retrieved contexts (what we call Confidence from Answer Redundancy, i.e. CAR). Together these methods allow for a simple but effective way to defend against poisoning attacks that provides gains of nearly 20% exact match across varying levels of data poisoning/knowledge conflicts.

Recent progress in sentence embedding, which represents a sentence’s meaning as a point in a vector space, has achieved high performance on several tasks such as the semantic textual similarity (STS) task.However, a sentence representation cannot adequately express the diverse information that sentences contain: for example, such representations cannot naturally handle asymmetric relationships between sentences.This paper proposes GaussCSE, a Gaussian-distribution-based contrastive learning framework for sentence embedding that can handle asymmetric inter-sentential relations, as well as a similarity measure for identifying entailment relations.Our experiments show that GaussCSE achieves performance comparable to that of previous methods on natural language inference (NLI) tasks, and that it can estimate the direction of entailment relations, which is difficult with point representations.

We present a new Hindi text-to-speech (TTS) dataset and demonstrate its utility for the expressive synthesis of children’s audio stories. The dataset comprises narration by a single female speaker who modifies her voice to produce different story characters. Annotation for dialogue identification, character labelling, and character attribution are provided, all of which are expected to facilitate the learning of character voice and speaking styles. Experiments are conducted using different versions of the annotated dataset that enable training a multi-speaker TTS model on the single-speaker data. Subjective tests show that the multi-speaker model improves expressiveness and character voice consistency compared to the baseline single-speaker TTS. With the multi-speaker model, objective evaluations show comparable word error rates, better speaker voice consistency, and higher correlations with ground-truth emotion attributes. We release a new 16.8 hours storytelling speech dataset in Hindi and propose effective solutions for expressive TTS with narrator voice modulation and character voice consistency.

We explore loss functions for fact verification in the FEVER shared task. While the cross-entropy loss is a standard objective for training verdict predictors, it fails to capture the heterogeneity among the FEVER verdict classes. In this paper, we develop two task-specific objectives tailored to FEVER. Experimental results confirm that the proposed objective functions outperform the standard cross-entropy. Performance is further improved when these objectives are combined with simple class weighting, which effectively overcomes the imbalance in the training data. The source code is available (https://github.com/yuta-mukobara/RLF-KGAT).

There is a large volume of late antique and medieval Hebrew texts. They represent a crucial linguistic and cultural bridge between Biblical and modern Hebrew. Poetry is prominent in these texts and one of its main characteristics is the frequent use of metaphor. Distinguishing figurative and literal language use is a major task for scholars of the Humanities, especially in the fields of literature, linguistics, and hermeneutics. This paper presents a new, challenging dataset of late antique and medieval Hebrew poetry with expert annotations of metaphor, as well as some baseline results, which we hope will facilitate further research in this area.

Social science NLP tasks, such as emotion or humor detection, are required to capture the semantics along with the implicit pragmatics from text, often with limited amounts of training data. Instruction tuning has been shown to improve the many capabilities of large language models (LLMs) such as commonsense reasoning, reading comprehension, and computer programming. However, little is known about the effectiveness of instruction tuning on the social domain where implicit pragmatic cues are often needed to be captured. We explore the use of instruction tuning for social science NLP tasks and introduce Socialite-Llama — an open-source, instruction-tuned Llama. On a suite of 20 social science tasks, Socialite-Llama improves upon the performance of Llama as well as matches or improves upon the performance of a state-of-the-art, multi-task finetuned model on a majority of them. Further, Socialite-Llama also leads to improvement on 5 out of 6 related social tasks as compared to Llama, suggesting instruction tuning can lead to generalized social understanding. All resources including our code, model and dataset can be found through [bit.ly/socialitellama](https://bit.ly/socialitellama/).

One interesting approach to Question Answering (QA) is to search for semantically similar questions, which have been answered before. This task is different from answer retrieval as it focuses on questions rather than only on the answers, therefore it requires different model training on different data.In this work, we introduce a novel unsupervised pre-training method specialized for retrieving and ranking questions. This leverages (i) knowledge distillation from a basic question retrieval model, and (ii) new pre-training task and objective for learning to rank questions in terms of their relevance with the query. Our experiments show that (i) the proposed technique achieves state-of-the-art performance on QRC and Quora-match datasets, and (ii) the benefit of combining re-ranking and retrieval models.

Most works on transformers trained with the Masked Language Modeling (MLM) objective use the original BERT model’s fixed masking rate of 15%. We propose to instead dynamically schedule the masking rate throughout training. We find that linearly decreasing the masking rate over the course of pretraining improves average GLUE accuracy by up to 0.46% and 0.25% in BERT-base and BERT-large, respectively, compared to fixed rate baselines. These gains come from exposure to both high and low masking rate regimes, providing benefits from both settings. Our results demonstrate that masking rate scheduling is a simple way to improve the quality of masked language models, achieving up to a 1.89x speedup in pretraining for BERT-base as well as a Pareto improvement for BERT-large.