Weight-based model editing methods update the parametric knowledge of language models post-training. However, these methods can unintentionally alter unrelated parametric knowledge representations, potentially increasing the risk of harm. In this work, we investigate how weight editing methods unexpectedly amplify model biases after edits. We introduce a novel benchmark dataset, Seesaw-CF, for measuring bias amplification of model editing methods for demographic traits such as race, geographic origin, and gender. We use Seesaw-CF to examine the impact of model editing on bias in five large language models. Our results demonstrate that edited models exhibit, to various degrees, more biased behavior for certain demographic groups than before they were edited, specifically becoming less confident in properties for Asian and African subjects. Additionally, editing facts about place of birth, country of citizenship, or gender has particularly negative effects on the model’s knowledge about unrelated properties, such as field of work, a pattern observed across multiple models.

Skill Extraction involves identifying skills and qualifications mentioned in documents such as job postings and resumes. The task is commonly tackled by training supervised models using a sequence labeling approach with BIO tags. However, the reliance on manually annotated data limits the generalizability of such approaches. Moreover, the common BIO setting limits the ability of the models to capture complex skill patterns and handle ambiguous mentions. In this paper, we explore the use of in-context learning to overcome these challenges, on a benchmark of 6 uniformized skill extraction datasets. Our approach leverages the few-shot learning capabilities of large language models (LLMs) to identify and extract skills from sentences. We show that LLMs, despite not being on par with traditional supervised models in terms of performance, can better handle syntactically complex skill mentions in skill extraction tasks.

Recent approaches in skill matching, employing synthetic training data for classification or similarity model training, have shown promising results, reducing the need for time-consuming and expensive annotations. However, previous synthetic datasets have limitations, such as featuring only one skill per sentence and generally comprising short sentences. In this paper, we introduce JobSkape, a framework to generate synthetic data that tackles these limitations, specifically designed to enhance skill-to-taxonomy matching. Within this framework, we create SkillSkape, a comprehensive open-source synthetic dataset of job postings tailored for skill-matching tasks. We introduce several offline metrics that show that our dataset resembles real-world data. Additionally, we present a multi-step pipeline for skill extraction and matching tasks using large language models (LLMs), benchmarking against known supervised methodologies. We outline that the downstream evaluation results on real-world data can beat baselines, underscoring its efficacy and adaptability.

Recent efforts in natural language processing (NLP) commonsense reasoning research have yielded a considerable number of new datasets and benchmarks. However, most of these datasets formulate commonsense reasoning challenges in artificial scenarios that are not reflective of the tasks which real-world NLP systems are designed to solve. In this work, we present CRoW, a manually-curated, multi-task benchmark that evaluates the ability of models to apply commonsense reasoning in the context of six real-world NLP tasks. CRoW is constructed using a multi-stage data collection pipeline that rewrites examples from existing datasets using commonsense-violating perturbations. We use CRoW to study how NLP systems perform across different dimensions of commonsense knowledge, such as physical, temporal, and social reasoning. We find a significant performance gap when NLP systems are evaluated on CRoW compared to humans, showcasing that commonsense reasoning is far from being solved in real-world task settings. We make our dataset and leaderboard available to the research community.

Understanding narratives requires reasoning about the cause-and-effect relationships between events mentioned in the text. While existing foundation models yield impressive results in many NLP tasks requiring reasoning, it is unclear whether they understand the complexity of the underlying network of causal relationships of events in narratives. In this work, we present CRAB, a new Causal Reasoning Assessment Benchmark designed to evaluate causal understanding of events in real-world narratives. CRAB contains fine-grained, contextual causality annotations for ~2.7K pairs of real-world events that describe various newsworthy event timelines (e.g., the acquisition of Twitter by Elon Musk). Using CRAB, we measure the performance of several large language models, demonstrating that most systems achieve poor performance on the task. Motivated by classical causal principles, we also analyze the causal structures of groups of events in CRAB, and find that models perform worse on causal reasoning when events are derived from complex causal structures compared to simple linear causal chains. We make our dataset and code available to the research community.

We propose our solution to the multimodal semantic role labeling task from the CONSTRAINT’22 workshop. The task aims at classifying entities in memes into classes such as “hero” and “villain”. We use several pre-trained multi-modal models to jointly encode the text and image of the memes, and implement three systems to classify the role of the entities. We propose dynamic sampling strategies to tackle the issue of class imbalance. Finally, we perform qualitative analysis on the representations of the entities.

Cognates and borrowings carry different aspects of etymological evolution. In this work, we study semantic change of such items using multilingual word embeddings, both static and contextualised. We underline caveats identified while building and evaluating these embeddings. We release both said embeddings and a newly-built historical words lexicon, containing typed relations between words of varied Romance languages.

Zero-shot cross-lingual transfer learning has been shown to be highly challenging for tasks involving a lot of linguistic specificities or when a cultural gap is present between lan- guages, such as in hate speech detection. In this paper, we highlight this limitation for hate speech detection in several domains and languages using strict experimental settings. Then, we propose to train on multilingual auxiliary tasks – sentiment analysis, named entity recognition, and tasks relying on syntactic information – to improve zero-shot transfer of hate speech detection models across languages. We show how hate speech detection models benefit from a cross-lingual knowledge proxy brought by auxiliary tasks fine-tuning and highlight these tasks’ positive impact on bridging the hate speech linguistic and cultural gap between languages.

There is a global trend for responsible investing and the need for developing automated methods for analyzing and Environmental, Social and Governance (ESG) related elements in financial texts is raising. In this work we propose a solution to the FinSim4-ESG task, consisting of binary classification of sentences into sustainable or unsustainable. We propose a novel knowledge-based latent heterogeneous representation that is based on knowledge from taxonomies and knowledge graphs and multiple contemporary document representations. We hypothesize that an approach based on a combination of knowledge and document representations can introduce significant improvement over conventional document representation approaches. We consider ensembles on classifier as well on representation level late-fusion and early fusion. The proposed approaches achieve competitive accuracy of 89 and are 5.85 behind the best achieved score.

We tackle the problem of neural headline generation in a low-resource setting, where only limited amount of data is available to train a model. We compare the ideal high-resource scenario on English with results obtained on a smaller subset of the same data and also run experiments on two small news corpora covering low-resource languages, Croatian and Estonian. Two options for headline generation in a multilingual low-resource scenario are investigated: a pretrained multilingual encoder-decoder model and a combination of two pretrained language models, one used as an encoder and the other as a decoder, connected with a cross-attention layer that needs to be trained from scratch. The results show that the first approach outperforms the second one by a large margin. We explore several data augmentation and pretraining strategies in order to improve the performance of both models and show that while we can drastically improve the second approach using these strategies, they have little to no effect on the performance of the pretrained encoder-decoder model. Finally, we propose two new measures for evaluating the performance of the models besides the classic ROUGE scores.

La tâche de détection de contenus haineux est ardue, car elle nécessite des connaissances culturelles et contextuelles approfondies ; les connaissances nécessaires varient, entre autres, selon la langue du locateur ou la cible du contenu. Or, des données annotées pour des domaines et des langues spécifiques sont souvent absentes ou limitées. C’est là que les données dans d’autres langues peuvent être exploitées ; mais du fait de ces variations, le transfert cross-lingue est souvent difficile. Dans cet article, nous mettons en évidence cette limitation pour plusieurs domaines et langues et montrons l’impact positif de l’apprentissage de tâches auxiliaires multilingues - analyse de sentiments, reconnaissance des entités nommées et tâches reposant sur des informations morpho-syntaxiques - sur le transfert cross-lingue zéro-shot des modèles de détection de discours haineux, afin de combler ce fossé culturel.

Languages are dynamic systems: word usage may change over time, reflecting various societal factors. However, all languages do not evolve identically: the impact of an event, the influence of a trend or thinking, can differ between communities. In this paper, we propose to track these divergences by comparing the evolution of a word and its translation across two languages. We investigate several methods of building time-varying and bilingual word embeddings, using contextualised and non-contextualised embeddings. We propose a set of scenarios to characterize semantic divergence across two languages, along with a setup to differentiate them in a bilingual corpus. We evaluate the different methods by generating a corpus of synthetic semantic change across two languages, English and French, before applying them to newspaper corpora to detect bilingual semantic divergence and provide qualitative insight for the task. We conclude that BERT embeddings coupled with a clustering step lead to the best performance on synthetic corpora; however, the performance of CBOW embeddings is very competitive and more adapted to an exploratory analysis on a large corpus.

Several cluster-based methods for semantic change detection with contextual embeddings emerged recently. They allow a fine-grained analysis of word use change by aggregating embeddings into clusters that reflect the different usages of the word. However, these methods are unscalable in terms of memory consumption and computation time. Therefore, they require a limited set of target words to be picked in advance. This drastically limits the usability of these methods in open exploratory tasks, where each word from the vocabulary can be considered as a potential target. We propose a novel scalable method for word usage-change detection that offers large gains in processing time and significant memory savings while offering the same interpretability and better performance than unscalable methods. We demonstrate the applicability of the proposed method by analysing a large corpus of news articles about COVID-19.

Plusieurs méthodes de détection des changements sémantiques utilisant des plongements lexicaux contextualisés sont apparues récemment. Elles permettent une analyse fine du changement d’usage des mots, en agrégeant les plongements contextualisés en clusters qui reflètent les différents usages d’un mot. Nous proposons une nouvelle méthode basée sur le transport optimal. Nous l’évaluons sur plusieurs corpus annotés, montrant un gain de précision par rapport aux autres méthodes utilisant des plongements contextualisés, et l’illustrons sur un corpus d’articles de journaux.

This paper describes the approaches used by the Discovery Team to solve SemEval-2020 Task 1 - Unsupervised Lexical Semantic Change Detection. The proposed method is based on clustering of BERT contextual embeddings, followed by a comparison of cluster distributions across time. The best results were obtained by an ensemble of this method and static Word2Vec embeddings. According to the official results, our approach proved the best for Latin in Subtask 2.

Au sein d’une langue, l’usage des mots varie selon deux axes : diachronique (dimension temporelle) et synchronique (variation selon l’auteur, la communauté, la zone géographique... ). Dans ces travaux, nous proposons une méthode de détection et d’interprétation des variations d’usages des mots à travers ces différentes dimensions. Pour cela, nous exploitons les capacités d’une nouvelle ligne de plongements lexicaux contextualisés, en particulier le modèle BERT. Nous expérimentons sur un corpus de rapports financiers d’entreprises françaises, pour appréhender les enjeux et préoccupations propres à certaines périodes, acteurs et secteurs d’activités.

L’usage, le sens et la connotation des mots peuvent changer au cours du temps. Les plongements lexicaux diachroniques permettent de modéliser ces changements de manière non supervisée. Dans cet article nous étudions l’impact de plusieurs fonctions de coût sur l’apprentissage de plongements dynamiques, en comparant les comportements de variantes du modèle Dynamic Bernoulli Embeddings. Les plongements dynamiques sont estimés sur deux corpus couvrant les mêmes deux décennies, le New York Times Annotated Corpus en anglais et une sélection d’articles du journal Le Monde en français, ce qui nous permet de mettre en place un processus d’analyse bilingue de l’évolution de l’usage des mots.

This study is a preliminary exploration of the concept of informativeness –how much information a sentence gives about a word it contains– and its potential benefits to building quality word representations from scarce data. We propose several sentence-level classifiers to predict informativeness, and we perform a manual annotation on a set of sentences. We conclude that these two measures correspond to different notions of informativeness. However, our experiments show that using the classifiers’ predictions to train word embeddings has an impact on embedding quality.

Word meaning change can be inferred from drifts of time-varying word embeddings. However, temporal data may be too sparse to build robust word embeddings and to discriminate significant drifts from noise. In this paper, we compare three models to learn diachronic word embeddings on scarce data: incremental updating of a Skip-Gram from Kim et al. (2014), dynamic filtering from Bamler & Mandt (2017), and dynamic Bernoulli embeddings from Rudolph & Blei (2018). In particular, we study the performance of different initialisation schemes and emphasise what characteristics of each model are more suitable to data scarcity, relying on the distribution of detected drifts. Finally, we regularise the loss of these models to better adapt to scarce data.