Large Language Models (LLMs) exhibit remarkable text classification capabilities, excelling in zero- and few-shot learning (ZSL and FSL) scenarios. However, since they are trained on different datasets, performance varies widely across tasks between those models. Recent studies emphasize the importance of considering human label variation in data annotation. However, how this human label variation also applies to LLMs remains unexplored. Given this likely model specialization, we ask: Do aggregate LLM labels improve over individual models (as for human annotators)? We evaluate four recent instruction-tuned LLMs as “annotators” on five subjective tasks across four languages. We use ZSL and FSL setups and label aggregation from human annotation. Aggregations are indeed substantially better than any individual model, benefiting from specialization in diverse tasks or languages. Surprisingly, FSL does not surpass ZSL, as it depends on the quality of the selected examples. However, there seems to be no good information-theoretical strategy to select those. We find that no LLM method rivals even simple supervised models. We also discuss the tradeoffs in accuracy, cost, and moral/ethical considerations between LLM and human annotation.

With the rise of Large Language Models (LLMs), the NLP community is increasingly aware of the environmental consequences of model development due to the energy consumed for training and running these models. This study investigates the energy consumption and environmental impact of systems participating in the MentalRiskES shared task, at the Iberian Language Evaluation Forum (IberLEF) in the year 2023, which focuses on early risk identification of mental disorders in Spanish comments. Participants were asked to submit, for each prediction, a set of efficiency metrics, being carbon dioxide emissions among them. We conduct an empirical analysis of the data submitted considering model architecture, task complexity, and dataset characteristics, covering a spectrum from traditional Machine Learning (ML) models to advanced LLMs. Our findings contribute to understanding the ecological footprint of NLP systems and advocate for prioritizing environmental impact assessment in shared tasks to foster sustainability across diverse model types and approaches, being evaluation campaigns an adequate framework for this kind of analysis.

Emotions are a central aspect of communication. Consequently, emotion analysis (EA) is a rapidly growing field in natural language processing (NLP). However, there is no consensus on scope, direction, or methods. In this paper, we conduct a thorough review of 154 relevant NLP publications from the last decade. Based on this review, we address four different questions: (1) How are EA tasks defined in NLP? (2) What are the most prominent emotion frameworks and which emotions are modeled? (3) Is the subjectivity of emotions considered in terms of demographics and cultural factors? and (4) What are the primary NLP applications for EA? We take stock of trends in EA and tasks, emotion frameworks used, existing datasets, methods, and applications. We then discuss four lacunae: (1) the absence of demographic and cultural aspects does not account for the variation in how emotions are perceived, but instead assumes they are universally experienced in the same manner; (2) the poor fit of emotion categories from the two main emotion theories to the task; (3) the lack of standardized EA terminology hinders gap identification, comparison, and future goals; and (4) the absence of interdisciplinary research isolates EA from insights in other fields. Our work will enable more focused research into EA and a more holistic approach to modeling emotions in NLP.

With mental health issues on the rise on the Web, especially among young people, there is a growing need for effective identification and intervention. In this paper, we introduce a new open-sourced corpus for the early detection of mental disorders in Spanish, focusing on eating disorders, depression, and anxiety. It consists of user messages posted on groups within the Telegram message platform and contains over 1,300 subjects with more than 45,000 messages posted in different public Telegram groups. This corpus has been manually annotated via crowdsourcing and is prepared for its use in several Natural Language Processing tasks including text classification and regression tasks. The samples in the corpus include both text and time data. To provide a benchmark for future research, we conduct experiments on text classification and regression by using state-of-the-art transformer-based models.

This paper describes the participation of SINAI research team in the Explainable Detection of Online Sexism (EDOS) Shared Task at SemEval 2023. Specifically, we participate in subtask A (binary sexism detection), subtask B (category of sexism), and subtask C (fine-grained vector of sexism). For the three subtasks, we propose a system that integrates information related to emotions, sentiments, and irony in order to check whether these features help detect sexism content. Our team ranked 46th in subtask A, 37th in subtask B, and 29th in subtask C, achieving 0.8245, 0.6043, and 0.4376 of macro f1-score, respectively, among the participants.

Recent instruction fine-tuned models can solve multiple NLP tasks when prompted to do so, with machine translation (MT) being a prominent use case. However, current research often focuses on standard performance benchmarks, leaving compelling fairness and ethical considerations behind. In MT, this might lead to misgendered translations, resulting, among other harms, in the perpetuation of stereotypes and prejudices. In this work, we address this gap by investigating whether and to what extent such models exhibit gender bias in machine translation and how we can mitigate it. Concretely, we compute established gender bias metrics on the WinoMT corpus from English to German and Spanish. We discover that IFT models default to male-inflected translations, even disregarding female occupational stereotypes. Next, using interpretability methods, we unveil that models systematically overlook the pronoun indicating the gender of a target occupation in misgendered translations. Finally, based on this finding, we propose an easy-to-implement and effective bias mitigation solution based on few-shot learning that leads to significantly fairer translations.

Hate speech detection faces two significant challenges: 1) the limited availability of labeled data and 2) the high variability of hate speech across different contexts and languages. Prompting brings a ray of hope to these challenges. It allows injecting a model with task-specific knowledge without relying on labeled data. This paper explores zero-shot learning with prompting for hate speech detection. We investigate how well zero-shot learning can detect hate speech in 3 languages with limited labeled data. We experiment with various large language models and verbalizers on 8 benchmark datasets. Our findings highlight the impact of prompt selection on the results. They also suggest that prompting, specifically with recent large language models, can achieve performance comparable to and surpass fine-tuned models, making it a promising alternative for under-resourced languages. Our findings highlight the potential of prompting for hate speech detection and show how both the prompt and the model have a significant impact on achieving more accurate predictions in this task.

In this paper we present SHARE, a new lexical resource with 10,125 offensive terms and expressions collected from Spanish speakers. We retrieve this vocabulary using an existing chatbot developed to engage a conversation with users and collect insults via Telegram, named Fiero. This vocabulary has been manually labeled by five annotators obtaining a kappa coefficient agreement of 78.8%. In addition, we leverage the lexicon to release the first corpus in Spanish for offensive span identification research named OffendES_spans. Finally, we show the utility of our resource as an interpretability tool to explain why a comment may be considered offensive.

Within textual emotion classification, the set of relevant labels depends on the domain and application scenario and might not be known at the time of model development. This conflicts with the classical paradigm of supervised learning in which the labels need to be predefined. A solution to obtain a model with a flexible set of labels is to use the paradigm of zero-shot learning as a natural language inference task, which in addition adds the advantage of not needing any labeled training data. This raises the question how to prompt a natural language inference model for zero-shot learning emotion classification. Options for prompt formulations include the emotion name anger alone or the statement “This text expresses anger”. With this paper, we analyze how sensitive a natural language inference-based zero-shot-learning classifier is to such changes to the prompt under consideration of the corpus: How carefully does the prompt need to be selected? We perform experiments on an established set of emotion datasets presenting different language registers according to different sources (tweets, events, blogs) with three natural language inference models and show that indeed the choice of a particular prompt formulation needs to fit to the corpus. We show that this challenge can be tackled with combinations of multiple prompts. Such ensemble is more robust across corpora than individual prompts and shows nearly the same performance as the individual best prompt for a particular corpus.

This paper describes the participation of SINAI team at Task 5: Toxic Spans Detection which consists of identifying spans that make a text toxic. Although several resources and systems have been developed so far in the context of offensive language, both annotation and tasks have mainly focused on classifying whether a text is offensive or not. However, detecting toxic spans is crucial to identify why a text is toxic and can assist human moderators to locate this type of content on social media. In order to accomplish the task, we follow a deep learning-based approach using a Bidirectional variant of a Long Short Term Memory network along with a stacked Conditional Random Field decoding layer (BiLSTM-CRF). Specifically, we test the performance of the combination of different pre-trained word embeddings for recognizing toxic entities in text. The results show that the combination of word embeddings helps in detecting offensive content. Our team ranks 29th out of 91 participants.

Offensive language detection and analysis has become a major area of research in Natural Language Processing. The freedom of participation in social media has exposed online users to posts designed to denigrate, insult or hurt them according to gender, race, religion, ideology, or other personal characteristics. Focusing on young influencers from the well-known social platforms of Twitter, Instagram, and YouTube, we have collected a corpus composed of 47,128 Spanish comments manually labeled on offensive pre-defined categories. A subset of the corpus attaches a degree of confidence to each label, so both multi-class classification and multi-output regression studies are possible. In this paper, we introduce the corpus, discuss its building process, novelties, and some preliminary experiments with it to serve as a baseline for the research community.

This paper describes the participation of SINAI team at Task 12: OffensEval 2: Multilingual Offensive Language Identification in Social Media. In particular, the participation in Sub-task A in English which consists of identifying tweets as offensive or not offensive. We preprocess the dataset according to the language characteristics used on social media. Then, we select a small set from the training set provided by the organizers and fine-tune different Transformerbased models in order to test their effectiveness. Our team ranks 20th out of 85 participants in Subtask-A using the XLNet model.

In recent years emotion detection in text has become more popular due to its potential applications in fields such as psychology, marketing, political science, and artificial intelligence, among others. While opinion mining is a well-established task with many standard data sets and well-defined methodologies, emotion mining has received less attention due to its complexity. In particular, the annotated gold standard resources available are not enough. In order to address this shortage, we present a multilingual emotion data set based on different events that took place in April 2019. We collected tweets from the Twitter platform. Then one of seven emotions, six Ekman’s basic emotions plus the “neutral or other emotions”, was labeled on each tweet by 3 Amazon MTurkers. A total of 8,409 in Spanish and 7,303 in English were labeled. In addition, each tweet was also labeled as offensive or no offensive. We report some linguistic statistics about the data set in order to observe the difference between English and Spanish speakers when they express emotions related to the same events. Moreover, in order to validate the effectiveness of the data set, we also propose a machine learning approach for automatically detecting emotions in tweets for both languages, English and Spanish.

Detecting emotions in textual conversation is a challenging problem in absence of nonverbal cues typically associated with emotion, like fa- cial expression or voice modulations. How- ever, more and more users are using message platforms such as WhatsApp or Telegram. For this reason, it is important to develop systems capable of understanding human emotions in textual conversations. In this paper, we carried out different systems to analyze the emotions of textual dialogue from SemEval-2019 Task 3: EmoContext for English language. Our main contribution is the integration of emotional and sentimental features in the classification using the SVM algorithm.

Misogyny and xenophobia are some of the most important social problems. With the in- crease in the use of social media, this feeling ofhatred towards women and immigrants can be more easily expressed, therefore it can cause harmful effects on social media users. For this reason, it is important to develop systems ca- pable of detecting hateful comments automatically. In this paper, we describe our system to analyze the hate speech in English and Spanish tweets against Immigrants and Women as part of our participation in SemEval-2019 Task 5: hatEval. Our main contribution is the integration of three individual algorithms of predic- tion in a model based on Vote ensemble classifier.

Offensive language has an impact across society. The use of social media has aggravated this issue among online users, causing suicides in the worst cases. For this reason, it is important to develop systems capable of identifying and detecting offensive language in text automatically. In this paper, we developed a system to classify offensive tweets as part of our participation in SemEval-2019 Task 6: OffensEval. Our main contribution is the integration of lexical features in the classification using the SVM algorithm.

Mental health is one of the main concerns of today’s society. Early detection of symptoms can greatly help people with mental disorders. People are using social networks more and more to express emotions, sentiments and mental states. Thus, the treatment of this information using NLP technologies can be applied to the automatic detection of mental problems such as eating disorders. However, the first step to solving the problem should be to provide a corpus in order to evaluate our systems. In this paper, we specifically focus on detecting anorexia messages on Twitter. Firstly, we have generated a new corpus of tweets extracted from different accounts including anorexia and non-anorexia messages in Spanish. The corpus is called SAD: Spanish Anorexia Detection corpus. In order to validate the effectiveness of the SAD corpus, we also propose several machine learning approaches for automatically detecting anorexia symptoms in the corpus. The good results obtained show that the application of textual classification methods is a promising option for developing this kind of system demonstrating that these tools could be used by professionals to help in the early detection of mental problems.

In this paper, we describe our participation in WASSA 2018 Implicit Emotion Shared Task (IEST 2018). We claim that the use of emotional external knowledge may enhance the performance and the capacity of generalization of an emotion classification system based on neural networks. Accordingly, we submitted four deep learning systems grounded in a sequence encoding layer. They mainly differ in the feature vector space and the recurrent neural network used in the sequence encoding layer. The official results show that the systems that used emotional external knowledge have a higher capacity of generalization, hence our claim holds.

Emotion classification is a new task that combines several disciplines including Artificial Intelligence and Psychology, although Natural Language Processing is perhaps the most challenging area. In this paper, we describe our participation in SemEval-2018 Task1: Affect in Tweets. In particular, we have participated in EI-oc, EI-reg and E-c subtasks for English and Spanish languages.