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.

Group fairness is a central research topic in text classification, where reaching fair treatment between sensitive groups (e.g., women and men) remains an open challenge. We propose an approach that extends the use of the Wasserstein Dependency Measure for learning unbiased neural text classifiers. Given the challenge of distinguishing fair from unfair information in a text encoder, we draw inspiration from adversarial training by inducing independence between representations learned for the target label and those for a sensitive attribute. We further show that domain adaptation can be efficiently leveraged to remove the need for access to the sensitive attributes in the data set we cure. We provide both theoretical and empirical evidence that our approach is well-founded.

Group fairness is a central research topic in text classification, where reaching fair treatment between sensitive groups (e.g.women vs. men) remains an open challenge.This paper presents a novel method for mitigating biases in neural text classification, agnostic to the model architecture.Considering the difficulty to distinguish fair from unfair information in a text encoder, we take inspiration from adversarial training to induce Wasserstein independence between representations learned to predict our target label and the ones learned to predict some sensitive attribute.Our approach provides two significant advantages.Firstly, it does not require annotations of sensitive attributes in both testing and training data.This is more suitable for real-life scenarios compared to existing methods that require annotations of sensitive attributes at train time.Secondly, our approach exhibits a comparable or better fairness-accuracy trade-off compared to existing methods.Our implementation is available on Github 1 .

The progress introduced by pre-trained language models and their fine-tuning has resulted in significant improvements in most downstream NLP tasks. The unsupervised training of a language model combined with further target task finetuning has become the standard QA fine-tuning procedure. In this work, we demonstrate that this strategy is sub-optimal for fine-tuning QA models, especially under a low QA annotation budget, which is a usual setting in practice due to the extractive QA labeling cost. We draw our conclusions by conducting an exhaustive analysis of the performance of the alternatives of the sequential fine-tuning strategy on different QA datasets. Based on the experiments performed, we observed that the best strategy to fine-tune the QA model in low-budget settings is taking a pre-trained language model (PLM) and then fine-tuning PLM with a dataset composed of the target dataset and SQuAD dataset. With zero extra annotation effort, the best strategy outperforms the standard strategy by 2.28% to 6.48%. Our experiments provide one of the first investigations on how to best fine-tune a QA system under a low budget and are therefore of the utmost practical interest to the QA practitioners.

Wikidata has grown to a knowledge graph with an impressive size. To date, it contains more than 17 billion triples collecting information about people, places, films, stars, publications, proteins, and many more. On the other side, most of the information on the Web is not published in highly structured data repositories like Wikidata, but rather as unstructured and semi-structured content, more concretely in HTML pages containing text and tables. Finding, monitoring, and organizing this data in a knowledge graph is requiring considerable work from human editors. The volume and complexity of the data make this task difficult and time-consuming. In this work, we present a framework that is able to identify and extract new facts that are published under multiple Web domains so that they can be proposed for validation by Wikidata editors. The framework is relying on question-answering technologies. We take inspiration from ideas that are used to extract facts from textual collections and adapt them to extract facts from Web pages. For achieving this, we demonstrate that language models can be adapted to extract facts not only from textual collections but also from Web pages. By exploiting the information already contained in Wikidata the proposed framework can be trained without the need for any additional learning signals and can extract new facts for a wide range of properties and domains. Following this path, Wikidata can be used as a seed to extract facts on the Web. Our experiments show that we can achieve a mean performance of 84.07 at F1-score. Moreover, our estimations show that we can potentially extract millions of facts that can be proposed for human validation. The goal is to help editors in their daily tasks and contribute to the completion of the Wikidata knowledge graph.

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.

Thomas Dopierre, Christophe Gravier, Wilfried Logerais. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Modern classification models tend to struggle when the amount of annotated data is scarce.To overcome this issue, several neural fewshot classification models have emerged, yielding significant progress over time, both in Computer Vision and Natural Language Processing.In the latter, such models used to rely on fixed word embeddings before the advent of transformers.Additionally, some models used in Computer Vision are yet to be tested in NLP applications.In this paper, we compare all these models, first adapting those made in the field of image processing to NLP, and second providing them access to transformers.We then test these models equipped with the same transformer-based encoder on the intent detection task, known for having a large number of classes.Our results reveal that while methods perform almost equally on the ARSC dataset, this is not the case for the Intent Detection task, where the most recent and supposedly best competitors perform worse than older and simpler ones (while all are given access to transformers).We also show that a simple baseline is surprisingly strong.All the new developed models, as well as the evaluation framework, are made publicly available 1 .

In this paper, we introduce a state-of-the-art pseudo-labeling technique for few-shot intent detection.We devise a folding/unfolding hierarchical clustering algorithm which assigns weighted pseudo-labels to unlabeled user utterances.We show that our two-step method yields significant improvement over existing solutions.This performance is achieved on multiple intent detection datasets, even in more challenging situations where the number of classes is large or when the dataset is highly imbalanced.Moreover, we confirm these results on the more general text classification task.We also demonstrate that our approach nicely complements existing solutions, thereby providing an even stronger state-of-the-art ensemble method.We make our code publicly available 1 for future research.

International audience

Word embeddings are commonly used as a starting point in many NLP models to achieve state-of-the-art performances. However, with a large vocabulary and many dimensions, these floating-point representations are expensive both in terms of memory and calculations which makes them unsuitable for use on low-resource devices. The method proposed in this paper transforms real-valued embeddings into binary embeddings while preserving semantic information, requiring only 128 or 256 bits for each vector. This leads to a small memory footprint and fast vector operations. The model is based on an autoencoder architecture, which also allows to reconstruct original vectors from the binary ones. Experimental results on semantic similarity, text classification and sentiment analysis tasks show that the binarization of word embeddings only leads to a loss of ∼2% in accuracy while vector size is reduced by 97%. Furthermore, a top-k benchmark demonstrates that using these binary vectors is 30 times faster than using real-valued vectors.

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.

Hady Elsahar, Christophe Gravier, Frederique Laforest. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Most people need textual or visual interfaces in order to make sense of Semantic Web data. In this paper, we investigate the problem of generating natural language summaries for Semantic Web data using neural networks. Our end-to-end trainable architecture encodes the information from a set of triples into a vector of fixed dimensionality and generates a textual summary by conditioning the output on the encoded vector. We explore a set of different approaches that enable our models to verbalise entities from the input set of triples in the generated text. Our systems are trained and evaluated on two corpora of loosely aligned Wikipedia snippets with triples from DBpedia and Wikidata, with promising results.

Learning word embeddings on large unlabeled corpus has been shown to be successful in improving many natural language tasks.The most efficient and popular approaches learn or retrofit such representations using additional external data.Resulting embeddings are generally better than their corpus-only counterparts, although such resources cover a fraction of words in the vocabulary.In this paper, we propose a new approach, Dict2vec, based on one of the largest yet refined datasource for describing words -natural language dictionaries.Dict2vec builds new word pairs from dictionary entries so that semantically-related words are moved closer, and negative sampling filters out pairs whose words are unrelated in dictionaries.We evaluate the word representations obtained using Dict2vec on eleven datasets for the word similarity task and on four datasets for a text classification task.

Top-k queries over data streams is a well studied problem. There exists numerous systems allowing to process continuous queries over sliding windows. At the opposite, non-append only streams call for ad-hoc solutions, e.g. tailor-made solutions implemented in a mainstream programming language. In the meantime, the Stream API and lambda expressions have been added in Java 8, thus gaining powerful operations for data stream processing. However, the Java Collections Framework does not provide data structures to safely and conveniently support sorted collections of evolving data. In this paper, we demonstrate Upsortable, an annotation-based approach that allows to use existing sorted collections from the standard Java API for dynamic data management. Our approach relies on a combination of pre-compilation abstract syntax tree modifications and runtime analysis of bytecode. Upsortable offers the developer a safe and time-efficient solution for developing top-k queries on data streams while keeping a full compatibility with standard Java.

As more personal and interactive applications are moving to the cloud, modeling the end-user expectations and satisfaction is becoming necessary for any SaaS provider to survive and thrive in today's competitive market. However, most of existing works addressing cloud elasticity management adopt a centralized approach where user preferences are mostly overlooked. Based on evidence from the fields of customer expectation management and psychophysics, in this article we propose a personal user model to represent end-user satisfaction and her expectations. To integrate the end-user into the decision loop we develop multi-agent negotiation architecture in which the end-user model is embodied by a personal agent who negotiates on her behalf. The results of the evaluation process show that automated negotiation provides a useful platform to empower the user choices, fulfill her expectations, and maximize her satisfaction hereby outperforming centralized approaches where the provider acts in a unilateral manner.

The advent of semantic data on the Web requires efficient reasoning systems to infer RDF and OWL data. The linked nature and the huge volume of data entail efficiency and scalability challenges when designing productive inference systems. This paper presents Inferray, an implementation of RDFS, ρdf, and RDFS-Plus inference with improved performance over existing solutions. The main features of Inferray are 1) a storage layout based on vertical partitioning that guarantees sequential access and efficient sort-merge join inference; 2) efficient sorting of pairs of 64-bit integers using ad-hoc optimizations on MSD radix and a custom counting sort; 3) a dedicated temporary storage to perform efficient graph closure computation. Our measurements on synthetic and real-world datasets show improvements over competitors on RDFS-Plus, and up to several orders of magnitude for transitivity closure.

Real-time recommendation of Twitter users based on the content of their profiles is a very challenging task. Traditional IR methods such as TF-IDF fail to handle efficiently large datasets. In this paper we present a scalable approach that allows real time recommendation of users based on their tweets. Our model builds a graph of terms, driven by the fact that users sharing similar interests will share similar terms. We show how this model can be encoded as a compact binary footprint, that allows very fast comparison and ranking, taking full advantage of modern CPU architectures. We validate our approach through an empirical evaluation against the Apache Lucene’s implementation of TF-IDF. We show that our approach is in average two hundred times faster than standard optimized implementation of TF-IDF with a precision of 58%.

The Semantic Web has gained substantial momentum over the last decade. It contributes to the manifestation of knowledge from data, and leverages implicit knowledge through reasoning algorithms. The main drawbacks of current reasoning methods over ontologies are two-fold: first they struggle to provide scalability for large datasets, and second, the batch processing reasoners who provide the best scalability so far are unable to infer knowledge from evolving data. We contribute to solving these problems by introducing Slider, an efficient incremental reasoner. Slider goes a significant step beyond existing system, including i) performance, by more than a 70% improvement in average compared to the fastest reasoner available to the best of our knowledge, and ii) inferences on streams of semantic data, by using intrinsic features that are themselves streams-oriented. Slider is fragment agnostic and conceived to handle expanding data with a growing background knowledge base. It natively supports pdf and RDFS, and its architecture allows to extend it to more complex fragments with a minimal effort. In this demo a web-based interface allows the users to visualize the internal behaviour of Slider during the inference, to better understand its design and principles.

Microblogging websites such as Twitter produce tremendous amount of data each second. Consequently, real-time recommendation systems require very efficient algorithm to quickly proceed this massive amount of data. In this paper we propose a scalable and extensible way of building content-based user profiles. Scalability refers to the relative complexity of algorithms involved in building the users profiles with respect to state-of-the-art solutions. Extensibility consider avoiding to recompute the model for newcomers. We present a tractable algorithm to build user profiles out of their tweets. Our model is a graph of terms cooccurency, driven by the fact that user sharing similar interests will share similar terms. We then present how this model can be encoded as a binary footprint, hence boosting comparison of users. We provide an empirical study to measure how the distance between users in the hash space differs from distance between users using standard Information Retrieval techniques. This experiment is based on a Twitter dataset we crawled, and represents 25K users and 1 million tweets. Our approach is driven by real-time analysis requirements and is thus oriented on a trade-off between expressivity and efficiency. Experimental results shows that our approach outperforms vector space model by three orders of magnitude, with a precision of 58%.

In this paper we present a Friend Recommender System for micro-blogging. Traditional batch processing of massive amounts of data makes it difficult to provide a near-real time friend recommender system or even a system that can properly scale to millions of users. In order to overcome these issues, we have designed a solution that represents user-generated micro posts as a set of pseudo-cliques. These graphs are assigned a hash value using an original Concept-Sensitive Hash function, a new sub-kind of Locally-Sensitive Hash functions. Finally, since the user profiles are represented as a binary footprint, the pairwise comparison of footprints using the Hamming distance provides scalability to the recommender system. The paper goes with an online application relying on a large Twitter dataset, so that the reader can freely experiment the system.

Over the last decade, two new facets of the Web emerged. On one hand, the Semantic Web provides data structures in order to break meanings barriers between distributed and heterogeneous data sources presents on the Web. On the other hand, the Social Web helps in clustering Web users into communities, and let them be prosumers of their Web experience. Both visions encompass many online applications, including Education. On the question of what is really the next stage of Web developments for Education, authors usually focus on a single of these two facets of the Web. As few studies have tried to combine them for an educational purpose this paper aims at giving an insight on the pioneers' works and the opportunities raised by mixing the Social and the Semantic Web for education.

This paper addresses ways users can remotely control industrial hardware devices over the Internet. We employ a J2EE based Remote Lab platform to describe the interface of the remote devices and to relay commands and results between users and devices. Collaboration is accomplished by embed ding the remote lab platform in an Open Wonderland virtual world.

Online Engineering allows users to perform tele-operation over the Internet. It is used in collaboratories, distance learning curriculum, and remote maintenance processes. Tele-operation over the Internet is however restrained by the development time and cost of ad hoc solutions. These solutions a) are hardly reusable, b) present a low fidelity of the human computer interface, and c) barely enable pervasive accesses (standalone clients or Web forms). We present in this paper a novel approach for building tele operation Human Computer Interfaces based on interactive multimedia. The end-user commands the remote device using interactive elements in the multimedia interface. The feedback information is seen as a combination of natural video content, produced by an IP camera capturing the instrument and synthetic graphics corresponding to elements of the device input interface (control panel). The proposed architecture fosters the reuses and gives a high level of interoperability between command terminals, since the interface can be displayed in any terminal able to lay out multimedia content.

Pervasive computing had brought the “anytime / anywhere” paradigm for information consumption. Moreover, given the tremendous amount of information and services available on the Web among the evolution of the social Web, more and more display units are being brought to the market (latest being the tablets family: iPad, kindle, etc.). Nevertheless, the Web contents were not designed to be properly displayed on interactive IPTVs (high resolution but poor interactive element: no mouse/keyboard, but a remote controller). A solution has been brought with dynamic adaptation of web content. The main drawback of researches in this domain is the default assumption that the SetTop Box capabilities are already known prior to the adaptation. In this paper, we propose an architecture not only to retrieve the display unit characteristics, but also a mean to propagate them in the browsing session. Consequently the Web content can be adapted by parts according to the device constraints, and each adaptation may depend on the previous adaptation (handling multi-part adaptation of Web pages for the interactive IPTV).

The users must face two main concerns in their digital lives, interactions and privacy, especially in identity federation architectures. Their environment and requestor client largely influence these concerns. A rich user environment issued from a user-centric vision on identity management architectures may be helpful for usability and privacy. In this paper we expose facts about the identity federation and the user digital life concepts. We then shape a user environment dedicated to identity management, and we establish some assumptions about future user environments and identity credentials implementations.