Projects

Active projects (2)


PIA 2026 1 publications

Promoció del valencià amb intel·ligència artificial

Generalitat Valenciana

L'objectiu del projecte és generar coneixement, recursos i aplicacions basats en IA que porten el valencià als nous entorns digitals. També vol facilitar-ne l’ús en l’administració, l’ensenyament i la comunicació amb la ciutadania, d’acord amb l’Estratègia d’Intel·ligència Artificial de la Comunitat Valenciana.

Generalitat Valenciana logo

«Activitat finançada per la Generalitat Valenciana» Línia nominativa S1678

BRAINS 2025–2028 13 publications

Building Real-world Applications with Intelligent, Neural, and Sustainable approaches

Agencia Estatal de Investigación; Ministerio de Ciencia, Innovación y Universidades

BRAINS is the ELiRF subproject of a coordinated national research project funded by MICIU/AEI and co-funded by the European Union. In BRAINS we apply neural language models to text and speech problems with a clear social impact. The work is organised in four lines: early detection of mental-health risk in Spanish, knowledge extraction from clinical and biomedical documents, automatic summarization and text simplification, and detection of sexism, misogyny and irony in multimodal content. We follow a Green AI approach. We use efficient models tailored to each domain rather than very large general-purpose systems. These models can run on the organisation’s own servers, so sensitive data such as clinical records are processed locally and never leave the institution. Smaller models also use less energy for both training and inference.

Agencia Estatal de Investigación; Ministerio de Ciencia, Innovación y Universidades logo

Finished projects (57)


Theses in progress (2)


ACIF BioMedIE 2023–2027

Aplicación de técnicas de procesamiento del lenguaje natural para la extracción de información en artículos e informes médicos

Pere Marco García · Supervision: Lluís Felip Hurtado Oliver, Antonio Molina Marco

Generalitat Valenciana

Predoctoral research contract of Pere Marco García, supervised by Lluís-F. Hurtado, within the ACIF programme of the Generalitat Valenciana, co-funded by the European Social Fund Plus. The thesis applies natural language processing to extract structured information from medical and scientific texts: identifying the sections of clinical documents, recognising genomic variants and the diseases they relate to, and classifying biomedical relations such as those between food and gut microbiota.

Generalitat Valenciana logo

Grant CIACIF/2022/234 funded by the Generalitat Valenciana and co-funded by the European Social Fund Plus (ESF+).

FPU EarlyMH 2025–2028

Aplicación de técnicas de procesamiento del lenguaje natural para la ayuda a la detección y el seguimiento de personas con riesgo de enfermedad mental

Andreu Casamayor Segarra · Supervision: Lluís Felip Hurtado Oliver, Antonio Molina Marco

Ministerio de Ciencia, Innovación y Universidades

Predoctoral contract of Andreu Casamayor Segarra supervised by Lluís-F. Hurtado and Antonio Molina under the FPU programme (Formación de Profesorado Universitario) of the Spanish Ministry of Science, Innovation and Universities. His research focuses on the early detection of mental-health risks in Spanish texts, such as signs of depression, anorexia or gambling addiction in social media, using domain-adapted long-context transformers, and on multimodal irony detection in spontaneous spoken Spanish.

Ministerio de Ciencia, Innovación y Universidades logo

Grant FPU24/01938 funded by the Spanish Ministry of Science, Innovation and Universities.

Defended theses (10)


2025 FPU

Delving into the abstractivity of summaries

Vicent Ahuir Esteve · Supervision: Lluís Felip Hurtado Oliver, Encarna Segarra Soriano

With the exorbitant amount of content continuously generated by today's technological society, to which one has access almost instantaneously, a person can spend excessive time searching and filtering content to perform a task. Therefore, being able to automatically identify, extract, and condense the most relevant information from any document is vital. In Natural Language Processing (NLP) this problem is addressed by the automatic summary generation or summarization field. Advances in the design of neural network architectures, most notably the Transformers architecture, have significantly increased the capabilities of language models. These improvements have enabled language models to generate more coherent texts. Therefore, in the automatic summarization field, the scientific community has begun to create summarization models that adopt an abstractive strategy. With the abstractive strategy, we obtain summaries where the summary is written by identifying the most relevant information of the document and reinterpreting the text of the original document to obtain a coherent, informative, and concise summary as possible. Therefore, the abstractive summarization strategy brings the models closer to how human beings summarize content. This work delves into different challenges within the abstractive strategy of summarization, focusing our efforts on the composition or writing process of summaries. The thesis aims to define in more detail what characterizes the perception of abstractivity in summaries and whether there are grades of variation in that perception, from complete extractive summaries to purely abstract summaries. We also explore ways to determine when two summaries are valid for the exact text despite having notably different wording and structure. Additionally, we study the emotions that arise from certain words in summaries, how the emotions vary from the document to the summary, and how automatic summarization models are influenced by this aspect.

Ministerio de Ciencia, Innovación y Universidades logo

Grant FPU21/05288 funded by the Spanish Ministry of Universities.

2023

User eXperience evaluation on university virtual learning through sentiment analysis

Rosario Sanchis Font · Supervision: María José Castro Bleda, María Begoña Jordá Albiñana

The use of new technologies and the number of users of university online learning systems have spread around the world in the last decades, showing a further increase with the spread of the Covid-19 pandemic since 2020. Additionally, ISO 9241-210:2019 sets international quality standards for designing human-computer interaction products, services, and systems that meet usability, accessibility, and User eXperience (UX) requirements. Therefore, the concept of UX has become very important as a quality requirement. For several authors, UX is a multidimensional concept that includes the motivations, feelings, and needs of end users. On the other hand, the United Nations' (UN) Sustainable Development Goal (SDG) 4 for 2030 aims to ensure inclusive, equitable, quality education for all globally. In this sense, in order to design interfaces and learning experiences in university environments that respect all quality specifications, it is necessary to evaluate the user experience of these environments automatically and accurately beforehand. Thus, the main objective of this thesis is to identify the most relevant specific characteristics in the user experience of university e-learning environments that allow specific and automatic analysis of the students' feelings in order to lay the foundations for the design of user-centered e-learning platforms. To this end, the study proposes to analyse the needs and feelings of online university students with digital, advanced, and efficient artificial intelligence methods. Therefore, this project investigates the application of machine learning models of sentiment analysis for the evaluation of user experience. These artificial intelligence techniques have been applied to the responses received from more than 2,000 university students surveyed from postgraduate online studies and massive open online courses (MOOCs). The results present the basis of a model that allows ontologically classifying categories or aspects of university online education and knowing the users' polarity of feeling about their e-learning experience in an automatic way. In this way, it has been possible to find out the students' opinions in an automated way with regard to key categories of digital teaching. In addition, student comments have been classified into several UX e-learning or UXEL dimensions. Also, it has been identified the polarity of sentiment for each dimension. To sum up, this work has generated four major contributions to the scientific community. Firstly, an adaptation of the validated questionnaire UEQ-S integrated and adapted to three e-learning platforms for specific postgraduate courses at the Universitat de València and at the Universidad Rey Juan Carlos; and for MOOCs at the Universitat Politècnica de València. Secondly, this thesis has generated an innovative application of sentiment analysis and machine learning methods through natural language processing for the evaluation of the user experience of university online students. Therefore, this method provides the analysis of learners opinions and classifies them according to their polarity in positive, negative or neutral. And thirdly, this scientific work brings out a proprietary ontology of aspects for the virtual learning experience associated with UX dimensions. This ontology used with sentiment analysis tools, allows classifying the polarity of student opinions (positive, neutral, negative) by key categories of e-learning (VLE, Teacher, Student, Sound, Image, Material, Exercise, Evaluation and Communication) and group the comments in three dimensions UX e-learning or UXEL (VLE, Social Connections, and Learning Resources and Tools). Finally, these contributions will help to evaluate in an automatic and accurate way several university e-learning environments in order to design user-centered virtual learning experiences more personalised and inclusive for all which suit quality standards and meet UN SDG 4 for 2030.

2021 Premio SEPLN 2022 a la mejor tesis doctoral en Procesamiento del Lenguaje Natural

Attention-based Approaches for Text Analytics in Social Media and Automatic Summarization

José Ángel González Barba · Supervision: Lluís Felip Hurtado Oliver, Emilio Sanchis Arnal

Nowadays, society has access, and the possibility to contribute, to large amounts of the content present on the internet, such as social networks, online newspapers, forums, blogs, or multimedia content platforms. These platforms have had, during the last years, an overwhelming impact on the daily life of individuals and organizations, becoming the predominant ways for sharing, discussing, and analyzing online content. Therefore, it is very interesting to work with these platforms, from different points of view, under the umbrella of Natural Language Processing. In this thesis, we focus on two broad areas inside this field, applied to analyze online content: text analytics in social media and automatic summarization. Neural networks are also a central topic in this thesis, where all the experimentation has been performed by using deep learning approaches, mainly based on attention mechanisms. Besides, we mostly work with the Spanish language, due to it is an interesting and underexplored language with a great interest in the research projects we participated in. On the one hand, for text analytics in social media, we focused on affective analysis tasks, including sentiment analysis and emotion detection, along with the analysis of the irony. In this regard, an approach based on Transformer Encoders, based on contextualizing pretrained Spanish word embeddings from Twitter, to address sentiment analysis and irony detection tasks, is presented. We also propose the use of evaluation metrics as loss functions, in order to train neural networks for reducing the impact of the class imbalance in multi-class and multi-label emotion detection tasks. Additionally, a specialization of BERT both for the Spanish language and the Twitter domain, that takes into account inter-sentence coherence in Twitter conversation flows, is presented. The performance of all these approaches has been tested with different corpora, from several reference evaluation benchmarks, showing very competitive results in all the tasks addressed. On the other hand, we focused on extractive summarization of news articles and TV talk shows. Regarding the summarization of news articles, a theoretical framework for extractive summarization, based on siamese hierarchical networks with attention mechanisms, is presented. Also, we present two instantiations of this framework: Siamese Hierarchical Attention Networks and Siamese Hierarchical Transformer Encoders. These systems were evaluated on the CNN/DailyMail and the NewsRoom corpora, obtaining competitive results in comparison to other contemporary extractive approaches. Concerning the TV talk shows, we proposed a text summarization task, for summarizing the transcribed interventions of the speakers, about a given topic, in the Spanish TV talk shows of the ``La Noche en 24 Horas" program. In addition, a corpus of news articles, collected from several Spanish online newspapers, is proposed, in order to study the domain transferability of siamese hierarchical approaches, between news articles and interventions of debate participants. This approach shows better results than other extractive techniques, along with a very promising domain transferability.

2016

Contributions to the joint segmentation and classification of sequences (My two cents on decoding and handwriting recognition)

Salvador España Boquera · Supervision: María José Castro Bleda

This work is focused on problems (like automatic speech recognition (ASR) and handwritten text recognition (HTR)) that: 1) can be represented (at least approximately) in terms of one-dimensional sequences, and 2) solving these problems entails breaking the observed sequence down into segments which are associated to units taken from a finite repertoire. The required segmentation and classification tasks are so intrinsically interrelated ("Sayre's Paradox") that they have to be performed jointly. We have been inspired by what some works call the "successful trilogy", which refers to the synergistic improvements obtained when considering: - a good formalization framework and powerful algorithms; - a clever design and implementation taking the best profit of hardware; - an adequate preprocessing and a careful tuning of all heuristics. We describe and study "two stage generative models" (TSGMs) comprising two stacked probabilistic generative stages without reordering. This model not only includes Hidden Markov Models (HMMs, but also "segmental models" (SMs). "Two stage decoders" may be deduced by simply running a TSGM in reversed way, introducing non determinism when required: 1) A directed acyclic graph (DAG) is generated and 2) it is used together with a language model (LM). One-pass decoders constitute a particular case. A formalization of parsing and decoding in terms of semiring values and language equations proposes the use of recurrent transition networks (RTNs) as a normal form for Context Free Grammars (CFGs), using them in a parsing-as-composition paradigm, so that parsing CFGs result in a slight extension of regular ones. Novel transducer composition algorithms have been proposed that can work with RTNs and can deal with null transitions without resorting to filter-composition even in the presence of null transitions and non-idempotent semirings. A review of LMs is described and some contributions mainly focused on LM interfaces, LM representation and on the evaluation of Neural Network LMs (NNLMs) are provided. A review of SMs includes the combination of generative and discriminative segmental models and general scheme of frame emission and another one of SMs. Some fast cache-friendly specialized Viterbi lexicon decoders taking profit of particular HMM topologies are proposed. They are able to manage sets of active states without requiring dictionary look-ups (e.g. hashing). A dataflow architecture allowing the design of flexible and diverse recognition systems from a little repertoire of components has been proposed, including a novel DAG serialization protocol. DAG generators can take over-segmentation constraints into account, make use SMs other than HMMs, take profit of the specialized decoders proposed in this work and use a transducer model to control its behavior making it possible, for instance, to use context dependent units. Relating DAG decoders, they take profit of a general LM interface that can be extended to deal with RTNs. Some improvements for one pass decoders are proposed by combining the specialized lexicon decoders and the "bunch" extension of the LM interface, including an adequate parallelization. The experimental part is mainly focused on HTR tasks on different input modalities (offline, bimodal). We have proposed some novel preprocessing techniques for offline HTR which replace classical geometrical heuristics and make use of automatic learning techniques (neural networks). Experiments conducted on the IAM database using this new preprocessing and HMM hybridized with Multilayer Perceptrons (MLPs) have obtained some of the best results reported for this reference database. Among other HTR experiments described in this work, we have used over-segmentation information, tried lexicon free approaches, performed bimodal experiments and experimented with the combination of hybrid HMMs with holistic classifiers.

2016 FPU

A Strategy for Multilingual Spoken Language Understanding Based on Graphs of Linguistic Units

Marcos Calvo Lance · Supervision: Fernando García Granada, Emilio Sanchis Arnal

In this thesis, the problem of multilingual spoken language understanding is addressed using graphs to model and combine the different knowledge sources that take part in the understanding process. As a result of this work, a full multilingual spoken language understanding system has been developed, in which statistical models and graphs of linguistic units are used. One key feature of this system is its ability to combine and process multiple inputs provided by one or more sources such as speech recognizers or machine translators. A graph-based monolingual spoken language understanding system was developed as a starting point. The input to this system is a set of sentences that is provided by one or more speech recognition systems. First, these sentences are combined by means of a grammatical inference algorithm in order to build a graph of words. Next, the graph of words is processed to construct a graph of concepts by using a dynamic programming algorithm that identifies the lexical structures that represent the different concepts of the task. Finally, the graph of concepts is used to build the best sequence of concepts. The multilingual case happens when the user speaks a language different to the one natively supported by the system. In this thesis, a test-on-source approach was followed. This means that the input sentences are translated into the system's language, and then they are processed by the monolingual system. For this purpose, two speech translation systems were developed. The output of these speech translation systems are graphs of words that are then processed by the monolingual graph-based spoken language understanding system. Both in the monolingual case and in the multilingual case, the experimental results show that a combination of several inputs allows to improve the results obtained with a single input. In fact, this approach outperforms the current state of the art in many cases when several inputs are combined.

Other finished grants (1)


FPI 2011–2013

Statistical models for spoken dialogue management

Ximo Planells Lerma · Supervision: Lluís Felip Hurtado Oliver, Emilio Sanchis Arnal

Ministerio de Educación

Predoctoral contract of Joaquín Planells Lerma under the FPI programme of the Spanish Ministry of Education, on statistical models for spoken dialogue systems.