We introduce a general-purpose agentic architecture with expert-in-the-loop guidance that iteratively learns to create tools for searching information and generating documents while minimizing time for task domain adaptation and human feedback. We illustrate it with preliminary experiments on scientific synthesis processes (e.g. state-of-the-art research papers, patents) and could be applied to various problems requiring long structured and nuanced answers.

The perception of the readability of biomedical texts varies depending on the reader's profile, a disparity further amplified by the intrinsic complexity of these documents and the unequal distribution of health literacy within the population. Although 72% of Internet users consult medical information online, a significant proportion have difficulty understanding it. To ensure that texts are accessible to a diverse audience, it is essential to assess readability. However, conventional readability formulas, designed for general texts, do not take this diversity into account, underlining the need to adapt evaluation tools to the specific needs of biomedical texts and the heterogeneity of readers. To address this gap, we propose a novel readability assessment method tailored to three distinct audiences: expert adults, non-expert adults, and children. Our approach is built upon a structured, bilingual biomedical corpus of 20,008 documents (8,854 in French, 11,154 in English), compiled from multiple sources to ensure diversity in both content and audience. Specifically, the French corpus combines texts from Cochrane and Wikipedia/Vikidia, both of which are subsets of the CLEAR corpus, while the English corpus merges documents from the Cochrane Library, Plaba, and Science Journal for Kids. For each original expert-level text, domain specialists produced simplified variants calibrated specifically to the comprehension abilities of non-expert adults or children. Every document is therefore explicitly labeled by its target audience. Leveraging this resource, we trained a diverse suite of classifiers, from classical approaches (e.g., XGBoost, SVM) to classifiers built upon language models (e.g., BERT, CamemBERT, BioBERT, DrBERT). We then designed a hybrid architecture "BioReadNet" that integrates transformer embeddings with expert-driven linguistic features, achieving a macro-averaged F1 score of 0.987.

We introduce in this paper a new approach to improve deep learning-based architectures for multi-label document classification. Dependencies between labels are an essential factor in the multi-label context. Our proposed strategy takes advantage of the knowledge extracted from label co-occurrences. The proposed method consists in adding a regularization term to the loss function used for training the model, in a way that incorporates the label similarities given by the label co-occurrences to encourage the model to jointly predict labels that are likely to co-occur, and and not consider labels that are rarely present with each other. This allows the neural model to better capture label dependencies. Our approach was evaluated on three datasets: the standard AAPD dataset, a corpus of scientific abstracts and Reuters-21578, a collection of news articles, and a newly proposed multi-label dataset called arXiv-ACM. Our method demonstrates improved performance, setting a new state-of-the-art on all three datasets.

Citation analysis is considered as major and one of the most popular branches of bibliometrics. Citation analysis is based on the assumption that all citations have similar values and weights each equally. Specific research fields like content-based citation analysis (CCA) seeks to explain the "how" and "why" of citation behavior. In this paper we tackle to explain the "how" from a centrality indicator based on factors which are built automatically according to the authors' citation behavior. This indicator allows to evaluate bibliographical references' importance for reading the paper with which user interacts. From objective quantitative measurements, factors are computed in order to characterize the level of granularity where citations are used. By the setting of the centrality indicator's factors we can highlight citations which tend towards a partial or a global construction of the authors' discourse. We carry out a pilot study in which we test our approach on some papers and discuss the challenges in carrying out the citation analysis in this context. Our results show interesting and consistent correlations between the level of granularity and the significance of citation influences.

In this paper, new schemas for feature categorization in different kinds of reviews, in the domain of books, are presented, so aspect extraction techniques could be later applied. We deal here with two types of reviews: formal reviews about scholarly books, written by experts, and informal ones about fiction books, written by readers which are not necessarily highly qualified. Our final goal is to extract the most relevant aspects or features to which any opinion is expressed in these reviews, along with the sentiment associated, for later integrating it to book recommender systems, improving the quality of the recommendations. Throughout this paper, the need for different annotation schemas is proved, by developing a new review classification system, as well as making an analysis at lexical and semantic levels on both kinds of reviews, for finally concluding with the presentation of the new categorization schemas.

Verbose query reduction and query term weighting are automatic techniques to deal with verbose queries. The objective is either to assign an appropriate weight to query terms according to their importance in the topic, or outright remove unsuitable terms from the query and keep only the suitable terms to the topic and user's need. These techniques improve performance and provide good results for ad hoc information retrieval. In this paper we propose a new approach to deal with long verbose queries in Social Information Retrieval (SIR) by taking Social Book Search as an example. In this approach, a new statistical measure was introduced to reduce and weight terms of verbose queries. Next, we expand the query by exploiting the similar books mentioned by users in their queries. We find that the proposed approach improves significantly the results.

Filtering a time-ordered corpus for documents that are highly relevant to an entity is a task receiving more and more attention over the years. One application is to reduce the delay between the moment an information about an entity is being first observed and the moment the entity entry in a knowledge base is being updated. Current state-of-the-art approaches are highly supervised and require training examples for each entity monitored. We propose an approach which does not require new training data when processing a new entity. To capture intrinsic characteristics of highly relevant documents our approach relies on three types of features: document centric features, entity profile related features and time features. Evaluated within the framework of the "Knowledge Base Acceleration" track at TREC 2012, it outperforms current state-of-the-art approaches.

Improving query understanding is crucial for providing the user with information that suits her needs. To this end, the retrieval system must be able to deal with several sources of knowledge from which it could infer a topical context. The use of external sources of information for improving document retrieval has been extensively studied. Improvements with either structured or large sets of data have been reported. However, in these studies resources are often used separately and rarely combined together. We experiment in this paper a method that discounts documents based on their weighted divergence from a set of external resources. We present an evaluation of the combination of four resources on two standard TREC test collections. Our proposed method significantly outperforms a state-of-the-art Mixture of Relevance Models on one test collection, while no significant differences are detected on the other one.

Automatic bibliographic reference annotation involves the tokenization and identification of reference fields. Recent methods use machine learning techniques such as Conditional Random Fields to tackle this problem. On the other hand, the state of the art methods always learn and evaluate their systems with a well structured data having simple format such as bibliography at the end of scientific articles. And that is a reason why the parsing of new reference different from a regular format does not work well. In our previous work, we have established a standard for the tokenization and feature selection with a less formulaic data such as notes. In this paper, we evaluate our system BILBO with other popular online reference parsing tools on a new data from totally different source. BILBO is constructed with our own corpora extracted and annotated from real world data, digital humanities articles of Revues.org site (90% in French) of OpenEdition. The robustness of BILBO system allows a language independent tagging result. We expect that this first attempt of evaluation will motivate the development of other efficient techniques for the scattered and less formulaic bibliographic references.

Identifying and extracting named entities from web pages has been the subject of many researches. In this paper, we propose and evaluate some new unsupervised language modeling approaches to determine the membership level of a candidate answer, a named entity, to a natural language question to a very fine-grained conceptual class of entity. We propose to address this issue by using the Web or DBPedia hierarchy as sources of evidence. Then, this level of membership can be used to improve the ranking of candidate answers in a question-answering task. Lastly, we present the results we obtained by participating in TREC 2010 Entity track.

International audience