Adrian-Gabriel Chifu

Adaptation des systèmes de recherche d'information aux contextes : le cas des requêtes difficiles: Le domaine de la recherche d'information (RI) étudie la façon de trouver des informations pertinentes dans un ou plusieurs corpus, pour répondre à un besoin d'information. Dans un Système de Recherche d'Information (SRI) les informations cherchées sont des " documents " et un besoin d'information prend la forme d'une " requête " formulée par l'utilisateur. La performance d'un SRI est dépendante de la requête. Les requêtes pour lesquelles les SRI échouent (pas ou peu de documents pertinents retrouvés) sont appelées dans la littérature des " requêtes difficiles ". Cette difficulté peut être causée par l'ambiguïté des termes, la formulation peu claire de la requête, le manque de contexte du besoin d'information, la nature et la structure de la collection de documents, etc. Cette thèse vise à adapter les systèmes de recherche d'information à des contextes, en particulier dans le cadre de requêtes difficiles. Le manuscrit est structuré en cinq chapitres principaux, outre les remerciements, l'introduction générale et les conclusions et perspectives. Le premier chapitre représente une introduction à la RI. Nous développons le concept de pertinence, les modèles de recherche de la littérature, l'expansion de requêtes et le cadre d'évaluation utilisé dans les expérimentations qui ont servi à valider nos propositions. Chacun des chapitres suivants présente une de nos contributions. Les chapitres posent les problèmes, indiquent l'état de l'art, nos propositions théoriques et leur validation sur des collections de référence. Dans le chapitre deux, nous présentons nos recherche sur la prise en compte du caractère ambigu des requêtes. L'ambiguïté des termes des requêtes peut en effet conduire à une mauvaise sélection de documents par les moteurs. Dans l'état de l'art, les méthodes de désambiguïsation qui donnent des bonnes performances sont supervisées, mais ce type de méthodes n'est pas applicable dans un contexte réel de RI, car elles nécessitent de l'information normalement indisponible. De plus, dans la littérature, la désambiguïsation de termes pour la RI est déclarée comme sous optimale.

Query performance prediction (QPP) aims to forecast the effectiveness of a search engine across a range of queries and documents. While state-of-the-art predictors offer a certain level of precision, their accuracy is not flawless. Prior research has recognized the challenges inherent in QPP but often lacks a thorough qualitative analysis. In this paper, we delve into QPP by examining the factors that influence the predictability of query performance accuracy. We propose the working hypothesis that while some queries are readily predictable, others present significant challenges. By focusing on outliers, we aim to identify the queries that are particularly challenging to predict. To this end, we employ multivariate outlier detection method. Our results demonstrate the effectiveness of this approach in identifying queries on which QPP do not perform well, yielding less reliable predictions. Moreover, we provide evidence that excluding these hard-to-predict queries from the analysis significantly enhances the overall accuracy of QPP.

One of the challenges of natural language understanding is to deal with the subjectivity of sentences, which may express opinions and emotions that add layers of complexity and nuance. Sentiment analysis is a field that aims to extract and analyze these subjective elements from text, and it can be applied at different levels of granularity, such as document, paragraph, sentence, or aspect. Aspect-based sentiment analysis is a well-studied topic with many available data sets and models. However, there is no clear definition of what makes a sentence difficult for aspect-based sentiment analysis. In this paper, we explore this question by conducting an experiment with three data sets: "Laptops", "Restaurants", and "MTSC" (Multi-Target-dependent Sentiment Classification), and a merged version of these three datasets. We study the impact of domain diversity and syntactic diversity on difficulty. We use a combination of classifiers to identify the most difficult sentences and analyze their characteristics. We employ two ways of defining sentence difficulty. The first one is binary and labels a sentence as difficult if the classifiers fail to correctly predict the sentiment polarity. The second one is a six-level scale based on how many of the top five best-performing classifiers can correctly predict the sentiment polarity. We also define 9 linguistic features that, combined, aim at estimating the difficulty at sentence level.

Subjectivity is a key aspect of natural language understanding, especially in the context of user-generated text and conversational systems based on large language models. Natural language sentences often contain subjective elements, such as opinions and emotions, that make them more nuanced and complex. The level of detail at which the study of the text is performed determines the possible applications of sentiment analysis. The analysis can be done at the document or paragraph level, or, even more granularly, at the aspect level. Many researchers have studied this topic extensively. The field of aspect-based sentiment analysis has numerous data sets and models. In this work, we initiate the discussion around the definition of sentence difficulty in this context of aspect-based sentiment analysis. To assess and quantify the difficulty of the aspect-based sentiment analysis, we conduct an experiment using three data sets: “Laptops”, “Restaurants”, and “MTSC” (Multi-Target-dependent Sentiment Classification), along with 21 learning models from scikit-learn. We also use two textual representations, TF-IDF (Terms frequency-inverse document frequency) and BERT (Bidirectional Encoder Representations from Transformers), to analyze the difficulty faced by these models in performing aspect-based sentiment analysis. Additionally, we compare the models with a fine-tuned version of BERT on the three data sets. We identify the most challenging sentences using a combination of classifiers in order to better understand them. We propose two strategies for defining sentence difficulty. The first strategy is binary and considers sentences as difficult when the classifiers are unable to correctly assign the sentiment polarity. The second strategy uses a six-level difficulty scale based on how many of the top five best-performing classifiers can correctly identify sentiment polarity. These sentences with assigned difficulty classes are then used to create predictive models for early difficulty detection. The purpose of estimating the difficulty of aspect-based sentiment analysis is to enhance performance while minimizing resource usage.

The analysis of user-generated content on the Internet has become increasingly popular for a wide variety of applications. One particular type of content is represented by the user reviews for programs, multimedia, products, and so on. Investigating the opinion contained by reviews may help in following the evolution of the reviewed items and thus in improving their quality. Detecting contradictory opinions in reviews is crucial when evaluating the quality of the respective resource. This article aims to estimate the contradiction intensity (strength) in the context of online courses (MOOC). This estimation was based on review ratings and on sentiment polarity in the comments, with respect to specific aspects, such as “lecturer”, “presentation”, etc. Between course sessions, users stop reviewing, and also, the course contents may evolve. Thus, the reviews are time dependent, and this is why they should be considered grouped by the course sessions. Having this in mind, the contribution of this paper is threefold: (a) defining the notion of subjective contradiction around specific aspects and then estimating its intensity based on sentiment polarity, review ratings, and temporality; (b) developing a dataset to evaluate the contradiction intensity measure, which was annotated based on a user study; (c) comparing our unsupervised method with supervised methods with automatic feature selection, over the dataset. The dataset collected from coursera.org is in English. It includes 2244 courses and 73,873 user-generated reviews of those courses.The results proved that the standard deviation of the ratings, the standard deviation of the polarities, and the number of reviews are suitable features for predicting the contradiction intensity classes. Among the supervised methods, the J48 decision trees algorithm yielded the best performance, compared to the naive Bayes model and the SVM model.

In this paper, we introduce FreSaDa, a French Satire Data Set, which is composed of 11,570 articles from the news domain. In order to avoid reporting unreasonably high accuracy rates due to the learning of characteristics specific to publication sources, we divided our samples into training, validation and test, such that the training publication sources are distinct from the validation and test publication sources. This gives rise to a cross-domain (cross-source) satire detection task. We employ two classification methods as baselines for our new data set, one based on low-level features (character n-grams) and one based on high-level features (average of CamemBERT word embeddings). As an additional contribution, we present an unsupervised domain adaptation method based on regarding the pairwise similarities (given by the dot product) between the training samples and the validation samples as features. By including these domain-specific features, we attain significant improvements for both character n-grams and CamemBERT embeddings.

Whether or not word sense disambiguation (WSD) can improve information retrieval (IR) results represents a topic that has been intensely debated over the years, with many inconclusive or contradictory conclusions. The most rarely used type of WSD for this task is the unsupervised one, although it has been proven to be beneficial at a large scale. Our study builds on existing research and tries to improve the most recent unsupervised method which is based on spectral clustering. It investigates the possible benefits of “helping” spectral clustering through feature selection when it performs sense discrimination for IR. Results obtained so far, involving large data collections, encourage us to point out the importance of feature selection even in the case of this advanced, state of the art clustering technique that is known for performing its own feature weighting. By suggesting an improvement of what we consider the most promising approach to usage of WSD in IR, and by commenting on its possible extensions, we state that WSD still holds a promise for IR and hope to stimulate continuation of this line of research, perhaps at an even more successful level.