Hallucinations in large language models (LLMs), where they generate fluent but factually incorrect outputs, pose challenges for applications requiring strict truthfulness. This work proposes a multi-faceted approach to detect such hallucinations across various language tasks. We leverage automatic data annotation using a proprietary LLM, fine-tuning of the Mistral-7B-instruct-v0.2 model on annotated and benchmark data, role-based and rationale-based prompting strategies, and an ensemble method combining different model outputs through majority voting. This comprehensive framework aims to improve the robustness and reliability of hallucination detection for LLM generations.

State-of-the-art conversational AI systems raise concerns due to their potential risks of generating unsafe, toxic, unethical, or dangerous content. Previous works have developed datasets to teach conversational agents the appropriate social paradigms to respond effectively to specifically designed hazardous content. However, models trained on these adversarial datasets still struggle to recognize subtle unsafe situations that appear naturally in conversations or introduce an inappropriate response in a casual context. To understand the extent of this problem, we study prosociality in both adversarial and casual dialog contexts and audit the response quality of general-purpose language models in terms of propensity to produce unsafe content. We propose a dual-step fine-tuning process to address these issues using a socially aware n-pair contrastive loss. Subsequently, we train a base model that integrates prosocial behavior by leveraging datasets like Moral Integrity Corpus (MIC) and ProsocialDialog. Experimental results on several dialog datasets demonstrate the effectiveness of our approach in generating socially appropriate responses.

Hateful comments are prevalent on social media platforms. Although tools for automatically detecting, flagging, and blocking such false, offensive, and harmful content online have lately matured, such reactive and brute force methods alone provide short-term and superficial remedies while the perpetrators persist. With the public availability of large language models which can generate articulate synthetic and engaging content at scale, there are concerns about the rapid growth of dissemination of such malicious content on the web. There is now a need to focus on deeper, long-term solutions that involve engaging with the human perpetrator behind the source of the content to change their viewpoint or at least bring down the rhetoric using persuasive means. To do that, we propose defining and experimenting with controllable strategies for generating counterarguments to hateful comments in online conversations. We experiment with controlling response generation using features based on (i) argument structure and reasoning-based Walton argument schemes, (ii) counter-argument speech acts, and (iii) human characteristicsbased qualities such as Big-5 personality traits and human values. Using automatic and human evaluations, we determine the best combination of features that generate fluent, argumentative, and logically sound arguments for countering hate. We further share the developed computational models for automatically annotating text with such features, and a silver-standard annotated version of an existing hate speech dialog corpora.

With the increasing number of users on social media platforms, the detection and categorization of abusive comments have become crucial, necessitating effective strategies to mitigate their impact on online discussions. However, the intricate and diverse nature of lowresource Indic languages presents a challenge in developing reliable detection methodologies. This research focuses on the task of classifying YouTube comments written in Tamil language into various categories. To achieve this, our research conducted experiments utilizing various multi-lingual transformer-based models along with data augmentation approaches involving back translation approaches and other pre-processing techniques. Our work provides valuable insights into the effectiveness of various preprocessing methods for this classification task. Our experiments showed that the Multilingual Representations for Indian Languages (MURIL) transformer model, coupled with round-trip translation and lexical replacement, yielded the most promising results, showcasing a significant improvement of over 15 units in macro F1-score compared to existing baselines. This contribution adds to the ongoing research to mitigate the adverse impact of abusive content on online platforms, emphasizing the utilization of diverse preprocessing strategies and state-of-the-art language models.

The subtle human values we acquire through life experiences govern our thoughts and gets reflected in our speech. It plays an integral part in capturing the essence of our individuality and making it imperative to identify such values in computational systems that mimic human actions. Computational argumentation is a field that deals with the argumentation capabilities of humans and can benefit from identifying such values. Motivated by that, we present an ensemble approach for detecting human values from argument text. Our ensemble comprises three models: (i) An entailment-based model for determining the human values based on their descriptions, (ii) A Roberta-based classifier that predicts the set of human values from an argument. (iii) A Roberta-based classifier to predict a reduced set of human values from an argument. We experiment with different ways of combining the models and report our results. Furthermore, our best combination achieves an overall F1 score of 0.48 on the main test set.

Effective argumentation is essential towards a purposeful conversation with a satisfactory outcome. For example, persuading someone to reconsider smoking might involve empathetic, well founded arguments based on facts and expert opinions about its ill-effects and the consequences on one’s family. However, the automatic generation of high-quality factual arguments can be challenging. Addressing existing controllability issues can make the recent advances in computational models for argument generation a potential solution. In this paper, we introduce ArgU: a neural argument generator capable of producing factual arguments from input facts and real-world concepts that can be explicitly controlled for stance and argument structure using Walton’s argument scheme-based control codes. Unfortunately, computational argument generation is a relatively new field and lacks datasets conducive to training. Hence, we have compiled and released an annotated corpora of 69,428 arguments spanning six topics and six argument schemes, making it the largest publicly available corpus for identifying argument schemes; the paper details our annotation and dataset creation framework. We further experiment with an argument generation strategy that establishes an inference strategy by generating an “argument template” before actual argument generation. Our results demonstrate that it is possible to automatically generate diverse arguments exhibiting different inference patterns for the same set of facts by using control codes based on argument schemes and stance.

In this paper we detail the implementation of Proto-Gen, an end-to-end neural response generator capable of selecting appropriate persona and fact sentences from available options, and generating persona and fact grounded responses. Incorporating a novel interaction layer in an encoder-decoder architecture, Proto-Gen facilitates learning dependencies between facts, persona and the context, and outperforms existing baselines on the FoCus dataset for both the sub-tasks of persona and fact selection, and response generation. We further fine tune Proto-Gen’s hyperparameters, and share our results and findings.

Personality traits influence human actions and thoughts, which is manifested in day to day conversations. Although glimpses of personality traits are observable in existing open domain conversation corpora, leveraging generic language modelling for response generation overlooks the interlocutor idiosyncrasies, resulting in non-customizable personality agnostic responses. With the motivation of enabling stylistically configurable response generators, in this paper we experiment with end-to-end mechanisms to ground neural response generators based on both (i) interlocutor Big-5 personality traits, and (ii) discourse intent as stylistic control codes. Since most of the existing large scale open domain chat corpora do not include Big-5 personality traits and discourse intent, we employ automatic annotation schemes to enrich the corpora with noisy estimates of personality and intent annotations, and further assess the impact of using such features as control codes for response generation using automatic evaluation metrics, ablation studies and human judgement. Our experiments illustrate the effectiveness of this strategy resulting in improvements to existing benchmarks. Additionally, we yield two silver standard annotated corpora with intents and personality traits annotated, which can be of use to the research community.

Generative neural conversational systems are typically trained by minimizing the entropy loss between the training “hard” targets and the predicted logits. Performance gains and improved generalization are often achieved by employing regularization techniques like label smoothing, which converts the training “hard” targets to soft targets. However, label smoothing enforces a data independent uniform distribution on the incorrect training targets, leading to a false assumption of equiprobability. In this paper, we propose and experiment with incorporating data-dependent word similarity-based weighing methods to transform the uniform distribution of the incorrect target probabilities in label smoothing to a more realistic distribution based on semantics. We introduce hyperparameters to control the incorrect target distribution and report significant performance gains over networks trained using standard label smoothing-based loss on two standard open-domain dialogue corpora.

Here we discuss our implementation of two tasks in the Social Media Mining for Health Applications (SMM4H) 2022 shared tasks – classification, detection, and normalization of Adverse Events (AE) mentioned in English tweets (Task 1) and classification of English tweets self-reporting exact age (Task 4). We have explored different methods and models for binary classification, multi-class classification and named entity recognition (NER) for these tasks. We have also processed the provided dataset for noise, imbalance, and creative language expression from data. Using diverse NLP methods we classified tweets for mentions of adverse drug effects (ADEs) and self-reporting the exact age in the tweets. Further, extracted reactions from the tweets and normalized these adverse effects to a standard concept ID in the MedDRA vocabulary.

This paper analyzes data from the 2021 Amazon Alexa Prize Socialbot Grand Challenge 4, in order to better understand the differences between human-computer interactions (HCI) in a socialbot setting and conventional human-to-human interactions. We find that because socialbots are a new genre of HCI, we are still negotiating norms to guide interactions in this setting. We present several notable patterns in user behavior toward socialbots, which have important implications for guiding future work in the development of conversational agents.

Knowledge Graph(KG) grounded conversations often use large pre-trained models and usually suffer from fact hallucination. Frequently entities with no references in knowledge sources and conversation history are introduced into responses, thus hindering the flow of the conversation—existing work attempt to overcome this issue by tweaking the training procedure or using a multi-step refining method. However, minimal effort is put into constructing an entity-level hallucination detection system, which would provide fine-grained signals that control fallacious content while generating responses. As a first step to address this issue, we dive deep to identify various modes of hallucination in KG-grounded chatbots through human feedback analysis. Secondly, we propose a series of perturbation strategies to create a synthetic dataset named FADE (FActual Dialogue Hallucination DEtection Dataset). Finally, we conduct comprehensive data analyses and create multiple baseline models for hallucination detection to compare against human-verified data and already established benchmarks.

Neural approaches to end-to-end argument mining (AM) are often formulated as dependency parsing (DP), which relies on token-level sequence labeling and intricate post-processing for extracting argumentative structures from text. Although such methods yield reasonable results, operating solely with tokens increases the possibility of discontinuous and overly segmented structures due to minor inconsistencies in token level predictions. In this paper, we propose EDU-AP, an end-to-end argument parser, that alleviates such problems in dependency-based methods by exploiting the intrinsic relationship between elementary discourse units (EDUs) and argumentative discourse units (ADUs) and operates at both token and EDU level granularity. Further, appropriately using contextual information, along with optimizing a novel objective function during training, EDU-AP achieves significant improvements across all four tasks of AM compared to existing dependency-based methods.

Personalized response selection systems are generally grounded on persona. However, a correlation exists between persona and empathy, which these systems do not explore well. Also, when a contradictory or off-topic response is selected, faithfulness to the conversation context plunges. This paper attempts to address these issues by proposing a suite of fusion strategies that capture the interaction between persona, emotion, and entailment information of the utterances. Ablation studies on the Persona-Chat dataset show that incorporating emotion and entailment improves the accuracy of response selection. We combine our fusion strategies and concept-flow encoding to train a BERT-based model which outperforms the previous methods by margins larger than 2.3% on original personas and 1.9% on revised personas in terms of hits@1 (top-1 accuracy), achieving a new state-of-the-art performance on the Persona-Chat dataset

While neural approaches to argument mining (AM) have advanced considerably, most of the recent work has been limited to parsing monologues. With an urgent interest in the use of conversational agents for broader societal applications, there is a need to advance the state-of-the-art in argument parsers for dialogues. This enables progress towards more purposeful conversations involving persuasion, debate and deliberation. This paper discusses Dialo-AP, an end-to-end argument parser that constructs argument graphs from dialogues. We formulate AM as dependency parsing of elementary and argumentative discourse units; the system is trained using extensive pre-training and curriculum learning comprising nine diverse corpora. Dialo-AP is capable of generating argument graphs from dialogues by performing all sub-tasks of AM. Compared to existing state-of-the-art baselines, Dialo-AP achieves significant improvements across all tasks, which is further validated through rigorous human evaluation.

We propose a novel, attention-based self-supervised approach to identify “claim-worthy” sentences in a fake news article, an important first step in automated fact-checking. We leverage aboutness of headline and content using attention mechanism for this task. The identified claims can be used for downstream task of claim verification for which we are releasing a benchmark dataset of manually selected compelling articles with veracity labels and associated evidence. This work goes beyond stylistic analysis to identifying content that influences reader belief. Experiments with three datasets show the strength of our model.

This paper details a system designed for Social Media Mining for Health Applications (SMM4H) Shared Task 2020. We specifically describe the systems designed to solve task 2: Automatic classification of multilingual tweets that report adverse effects, and task 3: Automatic extraction and normalization of adverse effects in English tweets. Fine tuning RoBERTa large for classifying English tweets enables us to achieve a F1 score of 56%, which is an increase of +10% compared to the average F1 score for all the submissions. Using BERT based NER and question answering, we are able to achieve a F1 score of 57.6% for extracting adverse reaction mentions from tweets, which is an increase of +1.2% compared to the average F1 score for all the submissions.

Popular fake news articles spread faster than mainstream articles on the same topic which renders manual fact checking inefficient. At the same time, creating tools for automatic detection is as challenging due to lack of dataset containing articles which present fake or manipulated stories as compelling facts. In this paper, we introduce manually verified corpus of compelling fake and questionable news articles on the USA politics, containing around 700 articles from Aug-Nov, 2016. We present various analyses on this corpus and finally implement classification model based on linguistic features. This work is still in progress as we plan to extend the dataset in the future and use it for our approach towards automated fake news detection.

Social media platforms play a crucial role in piecing together global news stories via their corresponding online discussions. Thus, in this work, we introduce the problem of automatically summarizing massively multilingual microblog text streams. We discuss the challenges involved in both generating summaries as well as evaluating them. We introduce a simple word graph based approach that utilizes node neighborhoods to identify keyphrases and thus in turn, pick summary candidates. We also demonstrate the effectiveness of our method in generating precise summaries as compared to other popular techniques.

Multiword Expressions (MWEs) are crucial lexico-semantic units in any language. However, most work on MWEs has been focused on standard monolingual corpora. In this work, we examine MWE usage on Twitter - an inherently multilingual medium with an extremely short average text length that is often replete with grammatical errors. In this work we present a new graph based, language agnostic method for automatically extracting MWEs from tweets. We show how our method outperforms standard Association Measures. We also present a novel unsupervised evaluation technique to ascertain the accuracy of MWE extraction.