Optimization of offensive content moderation models for different types of hateful messages is typically achieved through continued pre-training or fine-tuning on new hate speech benchmarks. However, existing benchmarks mainly address explicit hate toward protected groups and often overlook implicit or indirect hate, such as demeaning comparisons, calls for exclusion or violence, and subtle discriminatory language that still causes harm. While explicit hate can often be captured through surface features, implicit hate requires deeper, full-model semantic processing. In this work, we question the need for repeated fine-tuning and analyze the role of HatePrototypes, class-level vector representations derived from language models optimized for hate speech detection and safety moderation. We find that these prototypes, built from as few as 50 examples per class, enable cross-task transfer between explicit and implicit hate, with interchangeable prototypes across benchmarks. Moreover, we show that parameter-free early exiting with prototypes is effective for both hate types. We release the code, prototype resources, and evaluation scripts to support future research on efficient and transferable hate speech detection.

Thibaud Leteno, Irina Proskurina, Antoine Gourru, Julien Velcin, Charlotte Laclau, Guillaume Metzler, Christophe Gravier. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Explainable AI (XAI) is an important area but remains relatively understudied for clustering. We propose an explainable-by-design clustering approach that not only finds clusters but also exemplars to explain each cluster. The use of exemplars for understanding is supported by the exemplar-based school of concept definition in psychology. We show that finding a small set of exemplars to explain even a single cluster is computationally intractable; hence, the overall problem is challenging. We develop an approximation algorithm that provides provable performance guarantees with respect to clustering quality as well as the number of exemplars used. This basic algorithm explains all the instances in every cluster whilst another approximation algorithm uses a bounded number of exemplars to allow simpler explanations and provably covers a large fraction of all the instances. Experimental results show that our work is useful in domains involving difficult to understand deep embeddings of images and text.

Most real-world networks evolve over time. Existing literature proposes models for dynamic networks that are either unlabeled or assumed to have a single membership structure. On the other hand, a new family of Mixed Membership Stochastic Block Models (MMSBM) allows to model static labeled networks under the assumption of mixed-membership clustering. In this work, we propose to extend this later class of models to infer dynamic labeled networks under a mixed membership assumption. Our approach takes the form of a temporal prior on the model's parameters. It relies on the single assumption that dynamics are not abrupt. We show that our method significantly differs from existing approaches, and allows to model more complex systems --dynamic labeled networks. We demonstrate the robustness of our method with several experiments on both synthetic and real-world datasets. A key interest of our approach is that it needs very few training data to yield good results. The performance gain under challenging conditions broadens the variety of possible applications of automated learning tools --as in social sciences, which comprise many fields where small datasets are a major obstacle to the introduction of machine learning methods.

Last years have seen a regain of interest for the use of stochastic block modeling (SBM) in recommender systems. These models are seen as a flexible alternative to tensor decomposition techniques that are able to handle labeled data. Recent works proposed to tackle discrete recommendation problems via SBMs by considering larger contexts as input data and by adding second order interactions between contexts’ related elements. In this work, we show that these models are all special cases of a single global framework: the Serialized Interacting Mixed membership Stochastic Block Model (SIMSBM). It allows to model an arbitrarily large context as well as an arbitrarily high order of interactions. We demonstrate that SIMSBM generalizes several recent SBM-based baselines. Besides, we demonstrate that our formulation allows for an increased predictive power on five real-world datasets.1

Authors publish documents in a dynamic manner. Their topic of interest and writing style might shift over time. Tasks such as author classification, author identification or link prediction are difficult to solve in such complex data settings. We propose a new representation learning model, DGEA (for Dynamic Gaussian Embedding of Authors), that is more suited to solve these tasks by capturing this temporal evolution. We formulate a general embedding framework: author representation at time t is a Gaussian distribution that leverages pre-trained document vectors, and that depends on the publications observed until t. The representations should retain some form of multi-topic information and temporal smoothness. We propose two models that fit into this framework. The first one, K-DGEA, uses a first order Markov model optimized with an Expectation Maximization Algorithm with Kalman Equations. The second, R-DGEA, makes use of a Recurrent Neural Network to model the time dependence. We evaluate our method on several quantitative tasks: author identification, classification, and co-authorship prediction, on two datasets written in English. In addition, our model is language agnostic since it only requires pre-trained document embeddings. It outperforms existing baselines by up to 18% on an author classification task on a news articles dataset.

Slow emerging topic detection is a task between event detection, where we aggregate behaviors of different words on short period of time, and language evolution, where we monitor their long term evolution.In this work, we tackle the problem of early detection of slowly emerging new topics.To this end, we gather evidence of weak signals at the word level.We propose to monitor the behavior of words representation in an embedding space and use one of its geometrical properties to characterize the emergence of topics.As evaluation is typically hard for this kind of task, we present a framework for quantitative evaluation.We show positive results that outperform state-ofthe-art methods on two public datasets of press and scientific articles.

The textual content of a document and its publication date are intertwined. For example, the publication of a news article on a topic is influenced by previous publications on similar issues, according to underlying temporal dynamics. However, it can be challenging to retrieve meaningful information when textual information conveys little information or when temporal dynamics are hard to unveil. Furthermore, the textual content of a document is not always linked to its temporal dynamics. We develop a flexible method to create clusters of textual documents according to both their content and publication time, the Powered Dirichlet-Hawkes process (PDHP). We show PDHP yields significantly better results than state-of-the-art models when temporal information or textual content is weakly informative. The PDHP also alleviates the hypothesis that textual content and temporal dynamics are always perfectly correlated. PDHP allows retrieving textual clusters, temporal clusters, or a mixture of both with high accuracy when they are not. We demonstrate that PDHP generalizes previous work –such as the Dirichlet-Hawkes process (DHP) and Uniform process (UP). Finally, we illustrate the changes induced by PDHP over DHP and UP in a real-world application using Reddit data.

In most real-world applications, it is seldom the case that a result appears independently from an environment. In social networks, users’ behavior results from the people they interact with, news in their feed, or trending topics. In natural language, the meaning of phrases emerges from the combination of words. In general medicine, a diagnosis is established on the basis of the interaction of symptoms. Here, we propose the Interacting Mixed Membership Stochastic Block Model (IMMSBM), which investigates the role of interactions between entities (hashtags, words, memes, etc.) and quantifies their importance within the aforementioned corpora. We find that in inference tasks, taking them into account leads to average relative changes with respect to non-interacting models of up to 150% in the probability of an outcome and greatly improves the predictions performances. Furthermore, their role greatly improves the predictive power of the model. Our findings suggest that neglecting interactions when modeling real-world phenomena might lead to incorrect conclusions being drawn.

Gaussian Embedding of Linked Documents (GELD) is a new method that embeds linked documents (e.g., citation networks) onto a pretrained semantic space (e.g., a set of word embeddings). We formulate the problem in such a way that we model each document as a Gaussian distribution in the word vector space. We design a generative model that combines both words and links in a consistent way. Leveraging the variance of a document allows us to model the uncertainty related to word and link generation. In most cases, our method outperforms state-of-the-art methods when using our document vectors as features for usual downstream tasks. In particular, GELD achieves better accuracy in classification and link prediction on Cora and Dblp. In addition, we demonstrate qualitatively the convenience of several properties of our method. We provide the implementation of GELD and the evaluation datasets to the community (https://github.com/AntoineGourru/DNEmbedding).

Most network embedding algorithms consist in measuring co-occur-rences of nodes via random walks then learning the embeddings using Skip-Gram with Negative Sampling. While it has proven to be a relevant choice, there are alternatives, such as GloVe, which has not been investigated yet for network embedding. Even though SGNS better handles non co-occurrence than GloVe, it has a worse time-complexity. In this paper, we propose a matrix factorization approach for network embedding, inspired by GloVe, that better handles non co-occurrence with a competitive time-complexity. We also show how to extend this model to deal with networks where nodes are documents, by simultaneously learning word, node and document representations. Quantitative evaluations show that our model achieves state-of-the-art performance, while not being so sensitive to the choice of hyper-parameters. Qualitatively speaking, we show how our model helps exploring a network of documents by generating complementary network-oriented and content-oriented keywords.

Readitopics provides a new tool for browsing a textual corpus that showcases several recent work on topic labeling and topic coherence. We demonstrate the potential of these techniques to get a deeper understanding of the topics that structure different datasets. This tool is provided as a Web demo but it can be installed to experiment with your own dataset. It can be further extended to deal with more advanced topic modeling techniques.

One of the prevalent learning tasks involving images is content-based image classification. This is a difficult task especially because the low-level features used to digitally describe images usually capture little information about the semantics of the images. In this paper, we tackle this diffic ulty by enriching the semantic content of the image representation by using external knowledge. The underlying hypothesis of our work is that creating a more semantically rich representation for images would yield higher machine learning performances, without the need to modify the learning algorithms themselves. The external semantic information is presented under the form of non-positional image labels, therefore positioning our work in a weakly supervised context. Two approaches are proposed: the first one leverages the labels into the visual vocabulary construction algorithm, the result being dedicated visual vocabularies. The second approach adds a filtering phase as a pre-processing of the vocabulary construction. Known positive and known negative sets are constructed and features that are unlikely to be associated with the objects denoted by the labels are filtered. We apply our proposition to the task of content-based image classification and we show that semantically enriching the image representation yields higher classification performances than the baseline representation.

Most existing topic models focus either on extracting static topic-sentiment conjunctions or topic-wise evolution over time leaving out topic-sentiment dynamics and missing the opportunity to provide a more in-depth analysis of textual data. In this paper, we propose an LDA-based topic model for analyzing topic-sentiment evolution over time by modeling time jointly with topics and sentiments. We derive inference algorithm based on Gibbs Sampling process. Finally, we present results on reviews and news datasets showing interpretable trends and strong correlation with ground truth in particular for topic-sentiment evolution over time.

Online discussions became increasingly widespread with the Web 2.0: no matter the distance, whether you know the person or not, you can discuss and exchange ideas with people all over the world through forums, blogs, and newsgroups. The news websites have extensively used forums in order to encourage the reader being a real participant in the information media. This paper aims at automatically extracting the celebrities from such discussions. We propose certain meta-criteria and we provide an evaluation on a dataset of 35,175 posts written by 14,443 users. The results show that one of the proposed meta-criteria succeeds in extracting celebrities and allows for further improvements.

We propose an original approach for mapping the choreographic data into a new representation language adapted to data mining techniques. Our approach relies mainly on the notion of "dance tags" that we took from the NLP community by analogy with Part-of-Speech tagging. The process starts from scores described in Labanotation and produces in a fully automatic manner a high-level, comprehensive representation of the choreographic sequence. Our experiments show that we succeed in retrieving manually translated scores with an accuracy of 85% to 94%. Using this new representation of the choreographic data, one can then perform several useful tasks in an efficient manner. Among these are: music recommendation, automated detection of dance style or genre, and ultimately any task that requires a deeper understanding of the meaning of choreographic information than traditional processing can provide. In this paper, we demonstrate the usefulness of our approach with a simple example for discriminating between classical ballet, modern ballet, and folkloric dances.

In this paper, we propose a new time-aware dissimilarity measure that takes into account the temporal dimension. Observations that are close in the description space, but distant in time are considered as dissimilar. We also propose a method to enforce the segmentation contiguity, by introducing, in the objective function, a penalty term inspired from the Normal Distribution Function. We combine the two propositions into a novel time-driven constrained clustering algorithm, called TDCK-Means, which creates a partition of coherent clusters, both in the multidimensional space and in the temporal space. This algorithm uses soft semi-supervised constraints, to encourage adjacent observations belonging to the same entity to be assigned to the same cluster. We apply our algorithm to a Political Studies dataset in order to detect typical evolution phases. We adapt the Shannon entropy in order to measure the entity contiguity, and we show that our proposition consistently improves temporal cohesion of clusters, without any significant loss in the multidimensional variance.

The expansion of web user roles is, nowadays, a fact due to the ability of users to interact, discuss, exchange ideas and opinions, and form social networks through the web. The interaction level among users leads to the appearance of several social

Web forums are a huge data source. They allow people to interact with unknown individuals. Studying forums shows that the interaction is not obvious only through the structure but also through the content of the post. Taking into account this observation, we extract a social network with different kinds of relationships i.e. the structural relation, the name and the text quotations relation. We present here the promising results we obtain, and the difficulties we face while extracting the quotations in this kind of textual content. These results are obtained from real data (from two information websites) which make the validation difficult. So, we create a validation protocol composed of two steps and based on human raters. Finally, we will see the objective of this work which is understanding interactions in order to extract the social roles of individuals.

International audience

Online discussion systems in the form of forums have recently been analyzed by using graphs and social network techniques. Each forum is regarded as a social network and it is modeled by a graph whose vertices represent forum participants. In this paper, we focus on the structure and the opinion content of the forum posts and we are looking at the social network that is developed from a semantics point of view. We formally define an opinion-oriented model whose purpose is to provide complementary information to the knowledge extracted by the social network model. We define and present measures that can give important information regarding the opinion flow as well as the general attitude of users and towards users throughout the whole forum. Applying our model to a real forum found on the Web shows the additional information that can be extracted.