In this paper, we analyse to what extent machine translation (MT) systems and humans base their gender translations and associations on role names and on stereotypicality in the absence of (generic) grammatical gender cues in language. We compare an MT system’s choice of gender for a certain word when translating from a notional gender language, English, into a grammatical gender language, German, with thegender associations of humans. We outline a comparative case study of gender translation and annotation of words in isolation, out-of-context, and words in sentence contexts. The analysis reveals patterns of gender (bias) by MT and gender associations by humans for certain (1) out-of-context words and (2) words in-context. Our findings reveal the impact of context on gender choice and translation and show that word-level analyses fall short in such studies.

This paper presents a detailed description and results of the first shared task on explainability for cross-lingual emotion in tweets. Given a tweet in one of the five target languages (Dutch, Russian, Spanish, English, and French), systems should predict the correct emotion label (Task 1), as well as the words triggering the predicted emotion label (Task 2). The tweets were collected based on a list of stop words to prevent topical or emotional bias and were subsequently manually annotated. For both tasks, only a training corpus for English was provided, obliging participating systems to design cross-lingual approaches. Our shared task received submissions from 14 teams for the emotion detection task and from 6 teams for the trigger word detection task. The highest macro F1-scores obtained for both tasks are respectively 0.629 and 0.616, demonstrating that cross-lingual emotion detection is still a challenging task.

This paper provides an overview of the Shared Task for Cross-lingual Classification of CSR Themes and Topics. We framed the task as two separate sub-tasks: one cross-lingual multi-class CSR theme recognition task for English, French and simplified Chinese and one multi-label fine-grained classification task of CSR topics for Environment (ENV) and Labor and Human Rights (LAB) themes in English. The participants were provided with URLs and annotations for both tasks. Several teams downloaded the data, of which two teams submitted a system for both sub-tasks. In this overview paper, we discuss the set-up of the task and our main findings.

In this paper, we examine the recognition of irony by both humans and automatic systems. We achieve this by enhancing the annotations of an English benchmark data set for irony detection. This enhancement involves a layer of human-annotated irony likelihood using a 7-point Likert scale that combines binary annotation with a confidence measure. Additionally, the annotators indicated the trigger words that led them to perceive the text as ironic, which leveraged necessary theoretical insights into the definition of irony and its various forms. By comparing these trigger word spans across annotators, we determine the extent to which humans agree on the source of irony in a text. Finally, we compare the human-annotated spans with sub-token importance attributions for fine-tuned transformers using Layer Integrated Gradients, a state-of-the-art interpretability metric. Our results indicate that our model achieves better performance on tweets that were annotated with high confidence and high agreement. Although automatic systems can identify trigger words with relative success, they still attribute a significant amount of their importance to the wrong tokens.

Human evaluation in often considered to be the gold standard method of evaluating a Natural Language Generation system. However, whilst its importance is accepted by the community at large, the quality of its execution is often brought into question. In this position paper, we argue that the generation of more esoteric forms of language - humour, irony and sarcasm - constitutes a subdomain where the characteristics of selected evaluator panels are of utmost importance, and every effort should be made to report demographic characteristics wherever possible, in the interest of transparency and replicability. We support these claims with an overview of each language form and an analysis of examples in terms of how their interpretation is affected by different participant variables. We additionally perform a critical survey of recent works in NLG to assess how well evaluation procedures are reported in this subdomain, and note a severe lack of open reporting of evaluator demographic information, and a significant reliance on crowdsourcing platforms for recruitment.

In this paper we investigate potential bias in fine-tuned transformer models for irony detection. Bias is defined in this research as spurious associations between word n-grams and class labels, that can cause the system to rely too much on superficial cues and miss the essence of the irony. For this purpose, we looked for correlations between class labels and words that are prone to trigger irony, such as positive adjectives, intensifiers and topical nouns. Additionally, we investigate our irony model’s predictions before and after manipulating the data set through irony trigger replacements. We further support these insights with state-of-the-art explainability techniques (Layer Integrated Gradients, Discretized Integrated Gradients and Layer-wise Relevance Propagation). Both approaches confirm the hypothesis that transformer models generally encode correlations between positive sentiments and ironic texts, with even higher correlations between vividly expressed sentiment and irony. Based on these insights, we implemented a number of modification strategies to enhance the robustness of our irony classifier.

This paper investigates whether adding data of typologically closer languages improves the performance of transformer-based models for three different downstream tasks, namely Part-of-Speech tagging, Named Entity Recognition, and Sentiment Analysis, compared to a monolingual and plain multilingual language model. For the presented pilot study, we performed experiments for the use case of Slovene, a low(er)-resourced language belonging to the Slavic language family. The experiments were carried out in a controlled setting, where a monolingual model for Slovene was compared to combined language models containing Slovene, trained with the same amount of Slovene data. The experimental results show that adding typologically closer languages indeed improves the performance of the Slovene language model, and even succeeds in outperforming the large multilingual XLM-RoBERTa model for NER and PoS-tagging. We also reveal that, contrary to intuition, distantly or unrelated languages also combine admirably with Slovene, often out-performing XLM-R as well. All the bilingual models used in the experiments are publicly available at https://github.com/pranaydeeps/BLAIR

This paper describes the system we developed for the shared task ‘Hero, Villain and Victim: Dissecting harmful memes for Semantic role labelling of entities’ organised in the framework of the Second Workshop on Combating Online Hostile Posts in Regional Languages during Emergency Situation (Constraint 2022). We present an ensemble approach combining transformer-based models and linguistic information, such as the presence of irony and implicit sentiment associated to the target named entities. The ensemble system obtains promising classification scores, resulting in a third place finish in the competition.

This paper presents the results of a replication experiment for automatic irony detection in Dutch social media text, investigating both a feature-based SVM classifier, as was done by Van Hee et al. (2017) and and a transformer-based approach. In addition to building a baseline model, an important goal of this research is to explore the implementation of common-sense knowledge in the form of implicit sentiment, as we strongly believe that common-sense and connotative knowledge are essential to the identification of irony and implicit meaning in tweets. We show promising results and the presented approach can provide a solid baseline and serve as a staging ground to build on in future experiments for irony detection in Dutch.