Large Language Models (LLMs) from the GPT family have become extremely popular, leading to a race towards reducing their inference costs to allow for efficient local computation. However, the vast majority of existing work focuses on weight-only quantization, which can reduce runtime costs in the memory-bound one-token-at-a-time generative setting, but does not address costs in compute-bound scenarios, such as batched inference or prompt processing.In this paper, we address the general quantization problem, where both weights and activations should be quantized, which leads to computational improvements in general. We show that the majority of inference computations for large generative models can be performed with both weights and activations being cast to 4 bits, while at the same time maintaining good accuracy. We achieve this via a hybrid quantization strategy called QUIK that compresses most of the weights and activations to 4-bit, while keeping a small fraction of “outlier” weights and activations in higher-precision. QUIK is that it is designed with computational efficiency in mind: we provide GPU kernels matching the QUIK format with highly-efficient layer-wise runtimes, which lead to practical end-to-end throughput improvements of up to 3.4x relative to FP16 execution. We provide detailed studies for models from the OPT, LLaMA-2 and Falcon families, as well as a first instance of accurate inference using quantization plus 2:4 sparsity.Anonymized code is available.

Toxic language remains an ongoing challenge on social media platforms, presenting significant issues for users and communities. This paper provides a cross-topic and cross-lingual analysis of toxicity in Reddit conversations. We collect 1.5 million comment threads from 481 communities in six languages. By aligning languages with topics, we thoroughly analyze how toxicity spikes within different communities. Our analysis targets six languages spanning different communities and topics such as Culture, Politics, and News. We observe consistent patterns across languages where toxicity increases within the same topics while also identifying significant differences where specific language communities exhibit notable variations in relation to certain topics.

We present the winning approach to the TRAC 2024 Shared Task on Offline Harm Potential Identification (HarmPot-ID). The task focused on low-resource Indian languages and consisted of two sub-tasks: 1a) predicting the offline harm potential and 1b) detecting the most likely target(s) of the offline harm. We explored low-source domain specific, cross-lingual, and monolingual transformer models and submitted the aggregate predictions from the MuRIL and BERT models. Our approach achieved 0.74 micro-averaged F1-score for sub-task 1a and 0.96 for sub-task 1b, securing the 1st rank for both sub-tasks in the competition.

Hate speech (HS) is a widely acknowledged societal problem with potentially grave effects on vulnerable individuals and minority groups. Developing counter-narratives (CNs) that confront biases and stereotypes driving hateful narratives is considered an impactful strategy. Current automatic methods focus on isolated utterances to detect and react to hateful content online, often omitting the conversational context where HS naturally occurs. In this work, we explore strategies for the incorporation of conversational history for CN generation, comparing text and graphical representations with varying degrees of context. Overall, automatic and human evaluations show that 1) contextualized representations are comparable to those of isolated utterances, and 2) models based on graph representations outperform text representations, thus opening new research directions for future work.

Cross-lingual transfer has become an effective way of transferring knowledge between languages. In this paper, we explore an often overlooked aspect in this domain: the influence of the source language of a language model on language transfer performance. We consider a case where the target language and its script are not part of the pre-trained model. We conduct a series of experiments on monolingual and multilingual models that are pre-trained on different tokenization methods to determine factors that affect cross-lingual transfer to a new language with a unique script. Our findings reveal the importance of the tokenizer as a stronger factor than the shared script, language similarity, and model size.

In the context of the proliferation of multimodal hate speech related to the Russia-Ukraine conflict, we introduce a unified multimodal fusion system for detecting hate speech and its targets in text-embedded images. Our approach leverages the Twitter-based RoBERTa and Swin Transformer V2 models to encode textual and visual modalities, and employs the Multilayer Perceptron (MLP) fusion mechanism for classification. Our system achieved macro F1 scores of 87.27% for hate speech detection and 80.05% for hate speech target detection in the Multimodal Hate Speech Event Detection Challenge 2024, securing the 1st rank in both subtasks. We open-source the trained models at https://huggingface.co/Yestin-Wang

We present the CLTL system designed for the ArAIEval Shared Task 2024 on multimodal propagandistic memes classification in Arabic. The challenge was divided into three subtasks: identifying propagandistic content from textual modality of memes (subtask 2A), from visual modality of memes (subtask 2B), and in a multimodal scenario when both modalities are combined (subtask 2C). We explored various unimodal transformer models for Arabic language processing (subtask 2A), visual models for image processing (subtask 2B), and concatenated text and image embeddings using the Multilayer Perceptron fusion module for multimodal propagandistic memes classification (subtask 2C). Our system achieved 77.96% for subtask 2A, 71.04% for subtask 2B, and 79.80% for subtask 2C, ranking 2nd, 1st, and 3rd on the leaderboard.

Natural language reasoning plays an increasingly important role in improving language models’ ability to solve complex language understanding tasks. An interesting use case for reasoning is the resolution of context-dependent ambiguity. But no resources exist to evaluate how well Large Language Models can use explicit reasoning to resolve ambiguity in language. We propose to use ambiguous definite descriptions for this purpose and create and publish the first benchmark dataset consisting of such phrases. Our method includes all information required to resolve the ambiguity in the prompt, which means a model does not require anything but reasoning to do well. We find this to be a challenging task for recent LLMs. Code and data available at: https://github.com/sfschouten/exploiting-ambiguity

The spread of hate speech (HS) in the digital age poses significant challenges, with online platforms becoming breeding grounds for harmful content. While many natural language processing (NLP) studies have focused on identifying hate speech, few have explored the generation of counter narratives (CNs) as means to combat it. Previous studies have shown that computational models often generate CNs that are dull and generic, and therefore do not resonate with hate speech authors. In this paper, we explore the personalization capabilities of computational models for generating more targeted and engaging CNs. This paper investigates various strategies for incorporating author profiling information into GPT-2 and GPT-3.5 models to enhance the personalization of CNs to combat online hate speech. We investigate the effectiveness of incorporating author profiling aspects, more specifically the age and gender information of HS authors, in tailoring CNs specifically targeted at HS spreaders. We discuss the challenges, opportunities, and future directions for incorporating user profiling information into CN interventions.

Conventional techniques for detecting online hate speech rely on the availability of a sufficient number of annotated instances, which can be costly and time consuming. For this reason, zero-shot or few-shot detection can offer an attractive alternative. In this paper, we explore a zero-shot detection approach based on natural language inference (NLI) models. Since the performance of the models in this approach depends heavily on the choice of a hypothesis, our goal is to determine which factors affect the quality of detection. We conducted a set of experiments with three NLI models and four hate speech datasets. We demonstrate that a zero-shot NLI-based approach is competitive with approaches that require supervised learning, yet they are highly sensitive to the choice of hypothesis. In addition, our experiments indicate that the results for a set of hypotheses on different model-data pairs are positively correlated, and that the correlation is higher for different datasets when using the same model than it is for different models when using the same dataset. These results suggest that if we find a hypothesis that works well for a specific model and domain or for a specific type of hate speech, we can use that hypothesis with the same model also within a different domain. While, another model might require different suitable hypotheses in order to demonstrate high performance.

Online hate speech detection is an inherently challenging task that has recently received much attention from the natural language processing community. Despite a substantial increase in performance, considerable challenges remain and include encoding contextual information into automated hate speech detection systems. In this paper, we focus on detecting the target of hate speech in Dutch social media: whether a hateful Facebook comment is directed against migrants or not (i.e., against someone else). We manually annotate the relevant conversational context and investigate the effect of different aspects of context on performance when adding it to a Dutch transformer-based pre-trained language model, BERTje. We show that performance of the model can be significantly improved by integrating relevant contextual information.

We study the usefulness of hateful metaphorsas features for the identification of the type and target of hate speech in Dutch Facebook comments. For this purpose, all hateful metaphors in the Dutch LiLaH corpus were annotated and interpreted in line with Conceptual Metaphor Theory and Critical Metaphor Analysis. We provide SVM and BERT/RoBERTa results, and investigate the effect of different metaphor information encoding methods on hate speech type and target detection accuracy. The results of the conducted experiments show that hateful metaphor features improve model performance for the both tasks. To our knowledge, it is the first time that the effectiveness of hateful metaphors as an information source for hatespeech classification is investigated.

Hate speech detection is an actively growing field of research with a variety of recently proposed approaches that allowed to push the state-of-the-art results. One of the challenges of such automated approaches – namely recent deep learning models – is a risk of false positives (i.e., false accusations), which may lead to over-blocking or removal of harmless social media content in applications with little moderator intervention. We evaluate deep learning models both under in-domain and cross-domain hate speech detection conditions, and introduce an SVM approach that allows to significantly improve the state-of-the-art results when combined with the deep learning models through a simple majority-voting ensemble. The improvement is mainly due to a reduction of the false positive rate.

In this paper, we describe experiments designed to evaluate the impact of stylometric and emotion-based features on hate speech detection: the task of classifying textual content into hate or non-hate speech classes. Our experiments are conducted for three languages – English, Slovene, and Dutch – both in in-domain and cross-domain setups, and aim to investigate hate speech using features that model two linguistic phenomena: the writing style of hateful social media content operationalized as function word usage on the one hand, and emotion expression in hateful messages on the other hand. The results of experiments with features that model different combinations of these phenomena support our hypothesis that stylometric and emotion-based features are robust indicators of hate speech. Their contribution remains persistent with respect to domain and language variation. We show that the combination of features that model the targeted phenomena outperforms words and character n-gram features under cross-domain conditions, and provides a significant boost to deep learning models, which currently obtain the best results, when combined with them in an ensemble.

Native language identification (NLI) – identifying the native language (L1) of a person based on his/her writing in the second language (L2) – is useful for a variety of purposes, including marketing, security, and educational applications. From a traditional machine learning perspective,NLI is usually framed as a multi-class classification task, where numerous designed features are combined in order to achieve the state-of-the-art results. We introduce a deep generative language modelling (LM) approach to NLI, which consists in fine-tuning a GPT-2 model separately on texts written by the authors with the same L1, and assigning a label to an unseen text based on the minimum LM loss with respect to one of these fine-tuned GPT-2 models. Our method outperforms traditional machine learning approaches and currently achieves the best results on the benchmark NLI datasets.

In this paper, we present emotion lexicons of Croatian, Dutch and Slovene, based on manually corrected automatic translations of the English NRC Emotion lexicon. We evaluate the impact of the translation changes by measuring the change in supervised classification results of socially unacceptable utterances when lexicon information is used for feature construction. We further showcase the usage of the lexicons by calculating the difference in emotion distributions in texts containing and not containing socially unacceptable discourse, comparing them across four languages (English, Croatian, Dutch, Slovene) and two topics (migrants and LGBT). We show significant and consistent improvements in automatic classification across all languages and topics, as well as consistent (and expected) emotion distributions across all languages and topics, proving for the manually corrected lexicons to be a useful addition to the severely lacking area of emotion lexicons, the crucial resource for emotive analysis of text.

We present an ensemble approach for the detection of sarcasm in Reddit and Twitter responses in the context of The Second Workshop on Figurative Language Processing held in conjunction with ACL 2020. The ensemble is trained on the predicted sarcasm probabilities of four component models and on additional features, such as the sentiment of the comment, its length, and source (Reddit or Twitter) in order to learn which of the component models is the most reliable for which input. The component models consist of an LSTM with hashtag and emoji representations; a CNN-LSTM with casing, stop word, punctuation, and sentiment representations; an MLP based on Infersent embeddings; and an SVM trained on stylometric and emotion-based features. All component models use the two conversational turns preceding the response as context, except for the SVM, which only uses features extracted from the response. The ensemble itself consists of an adaboost classifier with the decision tree algorithm as base estimator and yields F1-scores of 67% and 74% on the Reddit and Twitter test data, respectively.

We present the INRIA approach to the suggestion mining task at SemEval 2019. The task consists of two subtasks: suggestion mining under single-domain (Subtask A) and cross-domain (Subtask B) settings. We used the Support Vector Machines algorithm trained on handcrafted features, function words, sentiment features, digits, and verbs for Subtask A, and handcrafted features for Subtask B. Our best run archived a F1-score of 51.18% on Subtask A, and ranked in the top ten of the submissions for Subtask B with 73.30% F1-score.

In this paper, we present experiments that estimate the impact of specific lexical choices of people writing in a second language (L2). In particular, we look at misspelled words that indicate lexical uncertainty on the part of the author, and separate them into three categories: misspelled cognates, “L2-ed” (in our case, anglicized) words, and all other spelling errors. We test the assumption that such errors contain clues about the native language of an essay’s author through the task of native language identification. The results of the experiments show that the information brought by each of these categories is complementary. We also note that while the distribution of such features changes with the proficiency level of the writer, their contribution towards native language identification remains significant at all levels.

In this paper, we describe experiments designed to explore and evaluate the impact of punctuation marks on the task of native language identification. Punctuation is specific to each language, and is part of the indicators that overtly represent the manner in which each language organizes and conveys information. Our experiments are organized in various set-ups: the usual multi-class classification for individual languages, also considering classification by language groups, across different proficiency levels, topics and even cross-corpus. The results support our hypothesis that punctuation marks are persistent and robust indicators of the native language of the author, which do not diminish in influence even when a high proficiency level in a non-native language is achieved.

We explore the hypothesis that emotion is one of the dimensions of language that surfaces from the native language into a second language. To check the role of emotions in native language identification (NLI), we model emotion information through polarity and emotion load features, and use document representations using these features to classify the native language of the author. The results indicate that emotion is relevant for NLI, even for high proficiency levels and across topics.

This paper presents the cic_ualg’s system that took part in the Discriminating between Similar Languages (DSL) shared task, held at the VarDial 2017 Workshop. This year’s task aims at identifying 14 languages across 6 language groups using a corpus of excerpts of journalistic texts. Two classification approaches were compared: a single-step (all languages) approach and a two-step (language group and then languages within the group) approach. Features exploited include lexical features (unigrams of words) and character n-grams. Besides traditional (untyped) character n-grams, we introduce typed character n-grams in the DSL task. Experiments were carried out with different feature representation methods (binary and raw term frequency), frequency threshold values, and machine-learning algorithms – Support Vector Machines (SVM) and Multinomial Naive Bayes (MNB). Our best run in the DSL task achieved 91.46% accuracy.

We present the CIC-FBK system, which took part in the Native Language Identification (NLI) Shared Task 2017. Our approach combines features commonly used in previous NLI research, i.e., word n-grams, lemma n-grams, part-of-speech n-grams, and function words, with recently introduced character n-grams from misspelled words, and features that are novel in this task, such as typed character n-grams, and syntactic n-grams of words and of syntactic relation tags. We use log-entropy weighting scheme and perform classification using the Support Vector Machines (SVM) algorithm. Our system achieved 0.8808 macro-averaged F1-score and shared the 1st rank in the NLI Shared Task 2017 scoring.