Large language models (LLMs) have demonstrated remarkable progress in leveraging diverse knowledge sources. This study investigates how nine widely used LLMs allocate knowledge between local context and global parameters when answering open-ended questions in knowledge-consistent scenarios. We introduce a novel dataset, WikiAtomic, and systematically vary context sizes to analyze how LLMs prioritize and utilize the provided information and their parametric knowledge in knowledge-consistent scenarios. Additionally, we also study their tendency to hallucinate under varying context sizes. Our findings reveal consistent patterns across models, including a consistent reliance on both contextual (around 70%) and parametric (around 30%) knowledge, and a decrease in hallucinations with increasing context. These insights highlight the importance of more effective context organization and developing models that use input more deterministically for robust performance.

This study investigates the factors influencing the performance of multilingual large language models (MLLMs) across diverse languages. We study 6 MLLMs, including masked language models, autoregressive models, and instruction-tuned LLMs, on the SIB-200 dataset, a topic classification dataset encompassing 204 languages. Our analysis considers three scenarios: ALL languages, SEEN languages (present in the model’s pretraining data), and UNSEEN languages (not present or documented in the model’s pretraining data in any meaningful way). We examine the impact of factors such as pretraining data size, general resource availability, language family, and script type on model performance. Decision tree analysis reveals that pretraining data size is the most influential factor for SEEN languages. However, interestingly, script type and language family become more crucial for UNSEEN languages, highlighting the importance of cross-lingual transfer learning. Notably, model size and architecture do not significantly alter the most important features identified. Our findings provide valuable insights into the strengths and limitations of current MLLMs and hope to guide the development of more effective and equitable multilingual NLP systems.

Text summarization models have typically focused on optimizing aspects of quality such as fluency, relevance, and coherence, particularly in the context of news articles. However, summarization models are increasingly being used to summarize diverse sources of text, such as social media data, that encompass a wide demographic user base. It is thus crucial to assess not only the quality of the generated summaries, but also the extent to which they can fairly represent the opinions of diverse social groups. Position bias, a long-known issue in news summarization, has received limited attention in the context of social multi-document summarization. We deeply investigate this phenomenon by analyzing the effect of group ordering in input documents when summarizing tweets from three distinct linguistic communities: African-American English, Hispanic-aligned Language, and White-aligned Language. Our empirical analysis shows that although the textual quality of the summaries remains consistent regardless of the input document order, in terms of fairness, the results vary significantly depending on how the dialect groups are presented in the input data. Our results suggest that position bias manifests differently in social multi-document summarization, severely impacting the fairness of summarization models.

Conversation systems accommodate diverse users with unique personalities and distinct writing styles. Within the domain of multi-turn dialogue modeling, this work studies the impact of varied utterance lengths on the quality of subsequent responses generated by conversation models. Using GPT-3 as the base model, multiple dialogue datasets, and several metrics, we conduct a thorough exploration of this aspect of conversational models. Our analysis sheds light on the complex relationship between utterance lengths and the quality of follow-up responses generated by dialogue systems. Empirical findings suggests that, for certain types of conversations, utterance lengths can be reduced by up to 72% without any noticeable difference in the quality of follow-up responses.

Most existing fact-checking systems are unable to explain their decisions by providing relevant rationales (justifications) for their predictions. It highlights a lack of transparency that poses significant risks, such as the prevalence of unexpected biases, which may increase political polarization due to limitations in impartiality. To address this critical gap, we introduce SEntence-Level FActual Reasoning (SELFAR), aimed at improving explainable fact-checking. SELFAR relies on fact extraction and verification by predicting the news source reliability and factuality (veracity) of news articles or claims at the sentence level, generating post-hoc explanations using SHAP/LIME and zero-shot prompts. Our experiments show that unreliable news stories predominantly consist of subjective statements, in contrast to reliable ones. Consequently, predicting unreliable news articles at the sentence level by analyzing impartiality and subjectivity is a promising approach for fact extraction and improving explainable fact-checking. Furthermore, LIME outperforms SHAP in explaining predictions on reliability. Additionally, while zero-shot prompts provide highly readable explanations and achieve an accuracy of 0.71 in predicting factuality, their tendency to hallucinate remains a challenge. Lastly, this paper also presents the first study on explainable fact-checking in the Portuguese language.

Recent large language models (LLMs) demonstrate impressive capabilities in handling long contexts, some exhibiting near-perfect recall on synthetic retrieval tasks. However, these evaluations have mainly focused on English text and involved a single target sentence within lengthy contexts. Our work investigates how LLM performance generalizes to multilingual settings with multiple hidden target sentences. We create a new dataset – mLongRR – to comprehensively evaluate several multilingual long-context LLMs on retrieval and reasoning tasks across five languages: English, Vietnamese, Indonesian, Swahili, and Somali. These languages share the Latin script but belong to distinct language families and resource levels. Our analysis reveals a significant performance gap between languages. The best-performing models such as Gemini-1.5 and GPT-4o, achieve around 96% accuracy in English to around 36% in Somali with a single target sentence. However, this accuracy drops to 40% in English and 0% in Somali when dealing with three target sentences. Our findings highlight the challenges long-context LLMs face when processing longer contexts, an increase in the number of target sentences, or languages of lower resource levels.

Recent advances in interactive large language models like ChatGPT have revolutionized various domains; however, their behavior in natural and role-play conversation settings remains underexplored. In our study, we address this gap by deeply investigating how ChatGPT behaves during conversations in different settings by analyzing its interactions in both a normal way and a role-play setting. We introduce a novel dataset of broad range of human-AI conversations annotated with user motives and model naturalness to examine (i) how humans engage with the conversational AI model, and (ii) how natural are AI model responses. Our study highlights the diversity of user motives when interacting with ChatGPT and variable AI naturalness, showing not only the nuanced dynamics of natural conversations between humans and AI, but also providing new avenues for improving the effectiveness of human-AI communication.

As dialogue systems become more popular, evaluation of their response quality gains importance. Engagingness highly correlates with overall quality and creates a sense of connection that gives human participants a more fulfilling experience. Although qualities like coherence and fluency are readily measured with well-worn automatic metrics, evaluating engagingness often relies on human assessment, which is a costly and time-consuming process. Existing automatic engagingness metrics evaluate the response without the conversation history, are designed for one dataset, or have limited correlation with human annotations. Furthermore, they have been tested exclusively on English conversations. Given that dialogue systems are increasingly available in languages beyond English, multilingual evaluation capabilities are essential. We propose that large language models (LLMs) may be used for evaluation of engagingness in dialogue through prompting, and ask how prompt constructs and translated prompts compare in a multilingual setting. We provide a prompt-design taxonomy for engagingness and find that using selected prompt elements with LLMs, including our comprehensive definition of engagingness, outperforms state-of-the-art methods on evaluation of engagingness in dialogue across multiple languages.

Paraphrase generation, a.k.a. paraphrasing, is a common and important task in natural language processing. Emotional paraphrasing, which changes the emotion embodied in a piece of text while preserving its meaning, has many potential applications, including moderating online dialogues and preventing cyberbullying. We introduce a new task of fine-grained emotional paraphrasing along emotion gradients, that is, altering the emotional intensities of the paraphrases in fine-grained settings following smooth variations in affective dimensions while preserving the meaning of the original text. We reconstruct several widely used paraphrasing datasets by augmenting the input and target texts with their fine-grained emotion labels. Then, we propose a framework for emotion and sentiment guided paraphrasing by leveraging pre-trained language models for conditioned text generation. Extensive evaluation of the fine-tuned models suggests that including fine-grained emotion labels in the paraphrase task significantly improves the likelihood of obtaining high-quality paraphrases that reflect the desired emotions while achieving consistently better scores in paraphrase metrics such as BLEU, ROUGE, and METEOR.

Prior work typically describes out-of-domain (OOD) or out-of-distribution (OODist) samples as those that originate from dataset(s) or source(s) different from the training set but for the same task. When compared to in-domain (ID) samples, the models have been known to usually perform poorer on OOD samples, although this observation is not consistent. Another thread of research has focused on OOD detection, albeit mostly using supervised approaches. In this work, we first consolidate and present a systematic analysis of multiple definitions of OOD and OODist as discussed in prior literature. Then, we analyze the performance of a model under ID and OOD/OODist settings in a principled way. Finally, we seek to identify an unsupervised method for reliably identifying OOD/OODist samples without using a trained model. The results of our extensive evaluation using 12 datasets from 4 different tasks suggest the promising potential of unsupervised metrics in this task.

Online conversations are particularly susceptible to derailment, which can manifest itself in the form of toxic communication patterns like disrespectful comments or verbal abuse. Forecasting conversation derailment predicts signs of derailment in advance enabling proactive moderation of conversations. Current state-of-the-art approaches to address this problem rely on sequence models that treat dialogues as text streams. We propose a novel model based on a graph convolutional neural network that considers dialogue user dynamics and the influence of public perception on conversation utterances. Through empirical evaluation, we show that our model effectively captures conversation dynamics and outperforms the state-of-the-art models on the CGA and CMV benchmark datasets by 1.5\% and 1.7\%, respectively.

Recent advances in automatic text summarization have contemporaneously been accompanied by a great deal of new metrics of automatic evaluation. This in turn has inspired recent research to re-assess these evaluation metrics to see how well they correlate with each other as well as with human evaluation, mostly focusing on single-document summarization (SDS) tasks. Although many of these metrics are typically also used for evaluating multi-document summarization (MDS) tasks, so far, little attention has been paid to studying them under such a distinct scenario. To address this gap, we present a systematic analysis of the inter-metric correlations for MDS tasks, while comparing and contrasting the results with SDS models. Using datasets from a wide range of domains (news, peer reviews, tweets, dialogues), we thus study a unified set of metrics under both the task setups. Our empirical analysis suggests that while most reference-based metrics show fairly similar trends across both multi- and single-document summarization, there is a notable lack of correlation between reference-free metrics in multi-document summarization tasks.

Knowledge Graphs (KGs) are directed labeled graphs representing entities and the relationships between them. Most prior work focuses on supervised or semi-supervised approaches which require large amounts of annotated data. While unsupervised approaches do not need labeled training data, most existing methods either generate too many redundant relations or require manual mapping of the extracted relations to a known schema. To address these limitations, we propose an unsupervised method for KG generation that requires neither labeled data nor manual mapping to the predefined relation schema. Instead, our method leverages sentence-level semantic similarity for automatically generating relations between pairs of entities. Our proposed method outperforms two baseline systems when evaluated over four datasets.

Social media posts provide a compelling, yet challenging source of data of diverse perspectives from many socially salient groups. Automatic text summarization algorithms make this data accessible at scale by compressing large collections of documents into short summaries that preserve salient information from the source text. In this work, we take a complementary approach to analyzing and improving the quality of summaries generated from social media data in terms of their ability to represent salient as well as diverse perspectives. We introduce a novel dataset, DivSumm, of dialect diverse tweets and human-written extractive and abstractive summaries. Then, we study the extent of dialect diversity reflected in human-written reference summaries as well as system-generated summaries. The results of our extensive experiments suggest that humans annotate fairly well-balanced dialect diverse summaries, and that cluster-based pre-processing approaches seem beneficial in improving the overall quality of the system-generated summaries without loss in diversity.

Language models pretrained on vast corpora of unstructured text using self-supervised learning framework are used in numerous natural language understanding and generation tasks. Many studies show that language acquisition in humans follows a rather structured simple-to-complex pattern and guided by this intuition, curriculum learning, which enables training of computational models in a meaningful order, such as processing easy samples before hard ones, has been shown to potentially reduce training time. The question remains whether curriculum learning can benefit pretraining of language models. In this work, we perform comprehensive experiments involving multiple curricula strategies varying the criteria for complexity and the training schedules. Empirical results of training transformer language models on English corpus and evaluating it intrinsically as well as after fine-tuning across eight tasks from the GLUE benchmark, show consistent improvement gains over conventional vanilla training. Interestingly, in our experiments, when evaluated on one epoch, the best model following a document-level hard-to-easy curriculum, outperforms the vanilla model by 1.7 points (average GLUE score) and it takes the vanilla model twice as many training steps to reach comparable performance.

Affective tasks such as sentiment analysis, emotion classification, and sarcasm detection have been popular in recent years due to an abundance of user-generated data, accurate computational linguistic models, and a broad range of relevant applications in various domains. At the same time, many studies have highlighted the importance of text preprocessing, as an integral step to any natural language processing prediction model and downstream task. While preprocessing in affective systems is well-studied, preprocessing in word vector-based models applied to affective systems, is not. To address this limitation, we conduct a comprehensive analysis of the role of preprocessing techniques in affective analysis based on word vector models. Our analysis is the first of its kind and provides useful insights of the importance of each preprocessing technique when applied at the training phase, commonly ignored in pretrained word vector models, and/or at the downstream task phase.

Most word representation learning methods are based on the distributional hypothesis in linguistics, according to which words that are used and occur in the same contexts tend to possess similar meanings. As a consequence, emotionally dissimilar words, such as “happy” and “sad” occurring in similar contexts would purport more similar meaning than emotionally similar words, such as “happy” and “joy”. This complication leads to rather undesirable outcome in predictive tasks that relate to affect (emotional state), such as emotion classification and emotion similarity. In order to address this limitation, we propose a novel method of obtaining emotion-enriched word representations, which projects emotionally similar words into neighboring spaces and emotionally dissimilar ones far apart. The proposed approach leverages distant supervision to automatically obtain a large training dataset of text documents and two recurrent neural network architectures for learning the emotion-enriched representations. Through extensive evaluation on two tasks, including emotion classification and emotion similarity, we demonstrate that the proposed representations outperform several competitive general-purpose and affective word representations.

Emotion classification from text typically requires some degree of word-emotion association, either gathered from pre-existing emotion lexicons or calculated using some measure of semantic relatedness. Most emotion lexicons contain a fixed number of emotion categories and provide a rather limited coverage. Current measures of computing semantic relatedness, on the other hand, do not adapt well to the specific task of word-emotion association and therefore, yield average results. In this work, we propose an unsupervised method of learning word-emotion association from large text corpora, called Selective Co-occurrences (SECO), by leveraging the property of mutual exclusivity generally exhibited by emotions. Extensive evaluation, using just one seed word per emotion category, indicates the effectiveness of the proposed approach over three emotion lexicons and two state-of-the-art models of word embeddings on three datasets from different domains.