Without proper safeguards, large language models will readily follow malicious instructions and generate toxic content. This risk motivates safety efforts such as red-teaming and large-scale feedback learning, which aim to make models both helpful and harmless. However, there is a tension between these two objectives, since harmlessness requires models to refuse to comply with unsafe prompts, and thus not be helpful. Recent anecdotal evidence suggests that some models may have struck a poor balance, so that even clearly safe prompts are refused if they use similar language to unsafe prompts or mention sensitive topics. In this paper, we introduce a new test suite called XSTest to identify such eXaggerated Safety behaviours in a systematic way. XSTest comprises 250 safe prompts across ten prompt types that well-calibrated models should not refuse to comply with, and 200 unsafe prompts as contrasts that models, for most applications, should refuse. We describe XSTest’s creation and composition, and then use the test suite to highlight systematic failure modes in state-of-the-art language models as well as more general challenges in building safer language models.

Online hate speech is a dangerous phenomenon that can (and should) be promptly counteracted properly. While Natural Language Processing supplies appropriate algorithms for trying to reach this objective, all research efforts are directed toward the English language. This strongly limits the classification power on non-English languages. In this paper, we test several learning frameworks for identifying hate speech in Italian text. We release HATE-ITA, a multi-language model trained on a large set of English data and available Italian datasets. HATE-ITA performs better than mono-lingual models and seems to adapt well also on language-specific slurs. We hope our findings will encourage the research in other mid-to-low resource communities and provide a valuable benchmarking tool for the Italian community.

Product discovery is a crucial component for online shopping. However, item-to-item recommendations today do not allow users to explore changes along selected dimensions: given a query item, can a model suggest something similar but in a different color? We consider item recommendations of the comparative nature (e.g. “something darker”) and show how CLIP-based models can support this use case in a zero-shot manner. Leveraging a large model built for fashion, we introduce GradREC and its industry potential, and offer a first rounded assessment of its strength and weaknesses.

Meaning is context-dependent, but many properties of language (should) remain the same even if we transform the context. For example, sentiment or speaker properties should be the same in a translation and original of a text. We introduce language invariant properties: i.e., properties that should not change when we transform text, and how they can be used to quantitatively evaluate the robustness of transformation algorithms. Language invariant properties can be used to define novel benchmarks to evaluate text transformation methods. In our work we use translation and paraphrasing as examples, but our findings apply more broadly to any transformation. Our results indicate that many NLP transformations change properties. We additionally release a tool as a proof of concept to evaluate the invariance of transformation applications.

Current language technology is ubiquitous and directly influences individuals’ lives worldwide. Given the recent trend in AI on training and constantly releasing new and powerful large language models (LLMs), there is a need to assess their biases and potential concrete consequences. While some studies have highlighted the shortcomings of these models, there is only little on the negative impact of LLMs on LGBTQIA+ individuals. In this paper, we investigated a state-of-the-art template-based approach for measuring the harmfulness of English LLMs sentence completion when the subjects belong to the LGBTQIA+ community. Our findings show that, on average, the most likely LLM-generated completion is an identity attack 13% of the time. Our results raise serious concerns about the applicability of these models in production environments.

Detecting emotion in text allows social and computational scientists to study how people behave and react to online events. However, developing these tools for different languages requires data that is not always available. This paper collects the available emotion detection datasets across 19 languages. We train a multilingual emotion prediction model for social media data, XLM-EMO. The model shows competitive performance in a zero-shot setting, suggesting it is helpful in the context of low-resource languages. We release our model to the community so that interested researchers can directly use it.

The maturity level of language models is now at a stage in which many companies rely on them to solve various tasks. However, while research has shown how biased and harmful these models are, systematic ways of integrating social bias tests into development pipelines are still lacking. This short paper suggests how to use these verification techniques in development pipelines. We take inspiration from software testing and suggest addressing social bias evaluation as software testing. We hope to open a discussion on the best methodologies to handle social bias testing in language models.

Hate speech is a global phenomenon, but most hate speech datasets so far focus on English-language content. This hinders the development of more effective hate speech detection models in hundreds of languages spoken by billions across the world. More data is needed, but annotating hateful content is expensive, time-consuming and potentially harmful to annotators. To mitigate these issues, we explore data-efficient strategies for expanding hate speech detection into under-resourced languages. In a series of experiments with mono- and multilingual models across five non-English languages, we find that 1) a small amount of target-language fine-tuning data is needed to achieve strong performance, 2) the benefits of using more such data decrease exponentially, and 3) initial fine-tuning on readily-available English data can partially substitute target-language data and improve model generalisability. Based on these findings, we formulate actionable recommendations for hate speech detection in low-resource language settings.

Pre-trained language models (PLMs) have outperformed other NLP models on a wide range of tasks. Opting for a more thorough understanding of their capabilities and inner workings, researchers have established the extend to which they capture lower-level knowledge like grammaticality, and mid-level semantic knowledge like factual understanding. However, there is still little understanding of their knowledge of higher-level aspects of language. In particular, despite the importance of sociodemographic aspects in shaping our language, the questions of whether, where, and how PLMs encode these aspects, e.g., gender or age, is still unexplored. We address this research gap by probing the sociodemographic knowledge of different single-GPU PLMs on multiple English data sets via traditional classifier probing and information-theoretic minimum description length probing. Our results show that PLMs do encode these sociodemographics, and that this knowledge is sometimes spread across the layers of some of the tested PLMs. We further conduct a multilingual analysis and investigate the effect of supplementary training to further explore to what extent, where, and with what amount of pre-training data the knowledge is encoded. Our overall results indicate that sociodemographic knowledge is still a major challenge for NLP. PLMs require large amounts of pre-training data to acquire the knowledge and models that excel in general language understanding do not seem to own more knowledge about these aspects.

Well-annotated data is a prerequisite for good Natural Language Processing models. Too often, though, annotation decisions are governed by optimizing time or annotator agreement. We make a case for nuanced efforts in an interdisciplinary setting for annotating offensive online speech. Detecting offensive content is rapidly becoming one of the most important real-world NLP tasks. However, most datasets use a single binary label, e.g., for hate or incivility, even though each concept is multi-faceted. This modeling choice severely limits nuanced insights, but also performance.We show that a more fine-grained multi-label approach to predicting incivility and hateful or intolerant content addresses both conceptual and performance issues.We release a novel dataset of over 40,000 tweets about immigration from the US and UK, annotated with six labels for different aspects of incivility and intolerance.Our dataset not only allows for a more nuanced understanding of harmful speech online, models trained on it also outperform or match performance on benchmark datasets

Twitter data have become essential to Natural Language Processing (NLP) and social science research, driving various scientific discoveries in recent years. However, the textual data alone are often not enough to conduct studies: especially, social scientists need more variables to perform their analysis and control for various factors. How we augment this information, such as users’ location, age, or tweet sentiment, has ramifications for anonymity and reproducibility, and requires dedicated effort. This paper describes Twitter-Demographer, a simple, flow-based tool to enrich Twitter data with additional information about tweets and users. \tool is aimed at NLP practitioners, psycho-linguists, and (computational) social scientists who want to enrich their datasets with aggregated information, facilitating reproducibility, and providing algorithmic privacy-by-design measures for pseudo-anonymity. We discuss our design choices, inspired by the flow-based programming paradigm, to use black-box components that can easily be chained together and extended. We also analyze the ethical issues related to the use of this tool, and the built-in measures to facilitate pseudo-anonymity.

In this paper, we describe the system proposed by the MilaNLP team for the Multimedia Automatic Misogyny Identification (MAMI) challenge. We use Perceiver IO as a multimodal late fusion over unimodal streams to address both sub-tasks A and B. We build unimodal embeddings using Vision Transformer (image) and RoBERTa (text transcript). We enrich the input representation using face and demographic recognition, image captioning, and detection of adult content and web entities. To the best of our knowledge, this work is the first to use Perceiver IO combining text and image modalities. The proposed approach outperforms unimodal and multimodal baselines.

Language models have revolutionized the field of NLP. However, language models capture and proliferate hurtful stereotypes, especially in text generation. Our results show that 4.3% of the time, language models complete a sentence with a hurtful word. These cases are not random, but follow language and gender-specific patterns. We propose a score to measure hurtful sentence completions in language models (HONEST). It uses a systematic template- and lexicon-based bias evaluation methodology for six languages. Our findings suggest that these models replicate and amplify deep-seated societal stereotypes about gender roles. Sentence completions refer to sexual promiscuity when the target is female in 9% of the time, and in 4% to homosexuality when the target is male. The results raise questions about the use of these models in production settings.

We investigate grounded language learning through real-world data, by modelling a teacher-learner dynamics through the natural interactions occurring between users and search engines; in particular, we explore the emergence of semantic generalization from unsupervised dense representations outside of synthetic environments. A grounding domain, a denotation function and a composition function are learned from user data only. We show how the resulting semantics for noun phrases exhibits compositional properties while being fully learnable without any explicit labelling. We benchmark our grounded semantics on compositionality and zero-shot inference tasks, and we show that it provides better results and better generalizations than SOTA non-grounded models, such as word2vec and BERT.

We present Query2Prod2Vec, a model that grounds lexical representations for product search in product embeddings: in our model, meaning is a mapping between words and a latent space of products in a digital shop. We leverage shopping sessions to learn the underlying space and use merchandising annotations to build lexical analogies for evaluation: our experiments show that our model is more accurate than known techniques from the NLP and IR literature. Finally, we stress the importance of data efficiency for product search outside of retail giants, and highlight how Query2Prod2Vec fits with practical constraints faced by most practitioners.

Word embeddings (e.g., word2vec) have been applied successfully to eCommerce products through prod2vec. Inspired by the recent performance improvements on several NLP tasks brought by contextualized embeddings, we propose to transfer BERT-like architectures to eCommerce: our model - Prod2BERT - is trained to generate representations of products through masked session modeling. Through extensive experiments over multiple shops, different tasks, and a range of design choices, we systematically compare the accuracy of Prod2BERT and prod2vec embeddings: while Prod2BERT is found to be superior in several scenarios, we highlight the importance of resources and hyperparameters in the best performing models. Finally, we provide guidelines to practitioners for training embeddings under a variety of computational and data constraints.

Many data sets (e.g., reviews, forums, news, etc.) exist parallelly in multiple languages. They all cover the same content, but the linguistic differences make it impossible to use traditional, bag-of-word-based topic models. Models have to be either single-language or suffer from a huge, but extremely sparse vocabulary. Both issues can be addressed by transfer learning. In this paper, we introduce a zero-shot cross-lingual topic model. Our model learns topics on one language (here, English), and predicts them for unseen documents in different languages (here, Italian, French, German, and Portuguese). We evaluate the quality of the topic predictions for the same document in different languages. Our results show that the transferred topics are coherent and stable across languages, which suggests exciting future research directions.

Spotting a lie is challenging but has an enormous potential impact on security as well as private and public safety. Several NLP methods have been proposed to classify texts as truthful or deceptive. In most cases, however, the target texts’ preceding context is not considered. This is a severe limitation, as any communication takes place in context, not in a vacuum, and context can help to detect deception. We study a corpus of Italian dialogues containing deceptive statements and implement deep neural models that incorporate various linguistic contexts. We establish a new state-of-the-art identifying deception and find that not all context is equally useful to the task. Only the texts closest to the target, if from the same speaker (rather than questions by an interlocutor), boost performance. We also find that the semantic information in language models such as BERT contributes to the performance. However, BERT alone does not capture the implicit knowledge of deception cues: its contribution is conditional on the concurrent use of attention to learn cues from BERT’s representations.

Understanding differences of viewpoints across corpora is a fundamental task for computational social sciences. In this paper, we propose the Sliced Word Embedding Association Test (SWEAT), a novel statistical measure to compute the relative polarization of a topical wordset across two distributional representations. To this end, SWEAT uses two additional wordsets, deemed to have opposite valence, to represent two different poles. We validate our approach and illustrate a case study to show the usefulness of the introduced measure.

While emotions are universal aspects of human psychology, they are expressed differently across different languages and cultures. We introduce a new data set of over 530k anonymized public Facebook posts across 18 languages, labeled with five different emotions. Using multilingual BERT embeddings, we show that emotions can be reliably inferred both within and across languages. Zero-shot learning produces promising results for low-resource languages. Following established theories of basic emotions, we provide a detailed analysis of the possibilities and limits of cross-lingual emotion classification. We find that structural and typological similarity between languages facilitates cross-lingual learning, as well as linguistic diversity of training data. Our results suggest that there are commonalities underlying the expression of emotion in different languages. We publicly release the anonymized data for future research.

While sentiment analysis is a popular task to understand people’s reactions online, we often need more nuanced information: is the post negative because the user is angry or sad? An abundance of approaches have been introduced for tackling these tasks, also for Italian, but they all treat only one of the tasks. We introduce FEEL-IT, a novel benchmark corpus of Italian Twitter posts annotated with four basic emotions: anger, fear, joy, sadness. By collapsing them, we can also do sentiment analysis. We evaluate our corpus on benchmark datasets for both emotion and sentiment classification, obtaining competitive results. We release an open-source Python library, so researchers can use a model trained on FEEL-IT for inferring both sentiments and emotions from Italian text.

The paper describes the MilaNLP team’s submission (Bocconi University, Milan) in the WASSA 2021 Shared Task on Empathy Detection and Emotion Classification. We focus on Track 2 - Emotion Classification - which consists of predicting the emotion of reactions to English news stories at the essay-level. We test different models based on multi-task and multi-input frameworks. The goal was to better exploit all the correlated information given in the data set. We find, though, that empathy as an auxiliary task in multi-task learning and demographic attributes as additional input provide worse performance with respect to single-task learning. While the result is competitive in terms of the competition, our results suggest that emotion and empathy are not related tasks - at least for the purpose of prediction.

Topic models extract groups of words from documents, whose interpretation as a topic hopefully allows for a better understanding of the data. However, the resulting word groups are often not coherent, making them harder to interpret. Recently, neural topic models have shown improvements in overall coherence. Concurrently, contextual embeddings have advanced the state of the art of neural models in general. In this paper, we combine contextualized representations with neural topic models. We find that our approach produces more meaningful and coherent topics than traditional bag-of-words topic models and recent neural models. Our results indicate that future improvements in language models will translate into better topic models.

The main goal of machine translation has been to convey the correct content. Stylistic considerations have been at best secondary. We show that as a consequence, the output of three commercial machine translation systems (Bing, DeepL, Google) make demographically diverse samples from five languages “sound” older and more male than the original. Our findings suggest that translation models reflect demographic bias in the training data. This opens up interesting new research avenues in machine translation to take stylistic considerations into account.