130 research outputs found
Norms of Conversation in a Framework for Agent Communication Languages
No full textAbstract. In open and heterogeneous environments offered by the Internet, where agents are designed by different vendors, the development of standards for agent communication needs to keep abreast of new dynamic interaction modalities. The objective of this paper is to contribute to FIPA’s standardization effort by proposing a pragmatic approach to the design of agent communication languages (ACLs) in which the meaning of messages is the combination of its semantics and pragmatics. First, we present a reformulation of FIPA’s communicative acts (ACL semantics) using a grounded specification language which overcomes some of the usual problems attributed to FIPA’s ACL semantics. Then the ACL pragmatics aims to account for the contextual factors that enriches the semantics, such agents ’ roles, turn-taking, and the satisfiability of messages’ perlocutionary effects. We claim that the ACL pragmatics is best specified by means of norms related to agents ’ obligations, permissions and rights
VaxxStance@IberLEF 2021: Descripción de la tarea de detección de actitudes basada en el uso de información más allá del texto
No full textThis paper describes the VaxxStance task at IberLEF 2021. The task proposes to detect stance in Tweets referring to vaccines, a relevant and controversial topic in the current pandemia. The task is proposed in a multilingual setting, providing data for Basque and Spanish languages. The objective is to explore crosslingual approaches which also complement textual information with contextual features obtained from the social network. The results demonstrate that contextual information is crucial to obtain competitive results, especially across languages.En este artículo se describe la tarea VaxxStance celebrada en el marco de IberLEF 2021. La tarea propone detectar la actitud de un conjunto de tweets relativos a las vacunas, a un tema muy actual y polémico en estos tiempos de pandemia. La tarea se ha propuesto en un marco multilingüe, euskera y español. Además del texto de cada tweet, se ha proporcionado además información relacionada con la red social de los usuarios autores de los tweets. Los resultados de los participantes han corroborado que el uso de información de la red social permite mejorar el rendimiento en esta tarea, particularmente en un entorno crosslingüe.This work has been partially supported by the European Social Fund through the Youth Employment Initiative (YEI 2019) and the Spanish Ministry of Science, Innovation and Universities (DeepReading RTI2018-096846-B-C21, MCIU/AEI/FEDER, UE), and by the DeepText project (KK-2020/00088), funded by the Basque Government. Rodrigo Agerri is also funded by the RYC-2017-23647 fellowship
Conclusiones de la evaluación de Modelos del Lenguaje en Español
No full textGiven the impact of language models on the field of Natural Language Processing, a number of Spanish encoder-only masked language models (aka BERTs) have been trained and released. These models were developed either within large projects using very large private corpora or by means of smaller scale academic efforts leveraging freely available data. In this paper we present a comprehensive head-to-head comparison of language models for Spanish with the following results: (i) Previously ignored multilingual models from large companies fare better than monolingual models, substantially changing the evaluation landscape of language models in Spanish; (ii) Results across the monolingual models are not conclusive, with supposedly smaller and inferior models performing competitively. Based on these empirical results, we argue for the need of more research to understand the factors underlying them. In this sense, the effect of corpus size, quality and pre-training techniques need to be further investigated to be able to obtain Spanish monolingual models significantly better than the multilingual ones released by large private companies, specially in the face of rapid ongoing progress in the field. The recent activity in the development of language technology for Spanish is to be welcomed, but our results show that building language models remains an open, resource-heavy problem which requires to marry resources (monetary and/or computational) with the best research expertise and practice.Actualmente existen varios modelos del lenguaje en español (también conocidos como BERTs) los cuales han sido desarrollados tanto en el marco de grandes proyectos que utilizan corpus privados de gran tamaño, como mediante esfuerzos académicos de menor escala aprovechando datos de libre acceso. En este artículo presentamos una comparación exhaustiva de modelos de lenguaje en español con los siguientes resultados: (i) La inclusión de modelos multilingües previamente ignorados altera sustancialmente el panorama de la evaluación para el español, ya que resultan ser en general mejores que sus homólogos monolingües; (ii) Las diferencias en los resultados entre los modelos monolingües no son concluyentes, ya que aquellos supuestamente más pequeños e inferiores obtienen resultados más que competitivos. El resultado de nuestra evaluación demuestra que es necesario seguir investigando para comprender los factores que subyacen a estos resultados. En este sentido, es necesario seguir investigando el efecto del tamaño del corpus, su calidad y las técnicas de preentrenamiento para poder obtener modelos monolingües en español significativamente mejores que los multilingües ya existentes. Aunque esta actividad reciente demuestra un creciente interés en el desarrollo de la tecnología lingüística para el español, nuestros resultados ponen de manifiesto que el desarrollo de modelos de lenguaje sigue siendo un problema abierto que requiere conjugar recursos (monetarios y/o computacionales) con los mejores conocimientos y prácticas de investigación en PLN.This work has been partially supported by the HiTZ center and the Basque Government (Research group funding IT-1805-22). We also acknowledge the funding from the following projects: (i) DeepKnowledge (PID2021-127777OB-C21) MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe; (ii) Disargue (TED2021-130810B-C21), MCIN/AEI/10.13039/501100011033 and European Union NextGenerationEU/PRTR (iii) Antidote (PCI2020-120717-2), MCIN/AEI/10.13039/501100011033 and by European Union NextGenerationEU/PRTR; (iv) DeepR3 (TED2021-130295B-C31) by MCIN/AEI/10.13039/501100011033 and EU NextGeneration programme EU/PRTR. Rodrigo Agerri currently holds the RYC-2017-23647 fellowship (MCIN/AEI/10.13039/501100011033 and by ESF Investing in your future)
Overview of TESTLINK at IberLEF 2023: Linking Results to Clinical Laboratory Tests and Measurements
No full textThe TESTLINK task conducted at IberLEF2023 focuses on relation extraction from clinical cases in Spanish and Basque. The task consists in identifying clinical results and measures and linking them to the tests and measurements from which they were obtained. Three teams took part in the task and various (supervised) deep learning models were evaluated; interestingly, none of the teams explored the use of few-shot learning. The evaluation shows that in-domain fine-tuning and larger training datasets improve the results. In fact, the fact that supervised models significantly outperformed the baseline based on few-shot learning shows the crucial role still played by the availability of annotated training data
Doris Martin at SemEval-2019 Task 4: Hyperpartisan News Detection with Generic Semi-supervised Features
No full textGeneric Framework for the Multidimensional Processing and Analysis of Social Media Content: A Proxemic Approach.
No full text286 p.In recent decades, significant growth and diversification in sources of User-Generated Content(UGC) have been observed. Social media emerges as one of the primary sources of UGC, offeringnumerous advantages over traditional data sources, such as affordability, vastness, and diversityacross various domains of application (for example, tourism, health, public policies). However, thehighly unstructured nature of social media posts introduces several challenges. The languagediversity and specificity of social media posts, characterized by features such as brevity, frequentgrammatical errors, and the use of special characters, combined with the substantial volume andnoisy nature of the data, make analyzing social media data a complex endeavour.This thesis introduces a novel multilingual framework, the APs Framework, designed tostreamline the processing and analysis of social media data. This framework is generic in two aspects:it can be applied across various social media platforms and is adaptable to different applicationdomains. The genericity of the application domain is supported by semantic representations ofdomain knowledge (for example, through thesaurus or ontologies). The APs Framework aimsto provide domain-independent insights from social media to non-computer scientists, such asstakeholders in various domains (for example, tourism offices in the tourism domain), therebyenhancing their analytical capabilities. The APs Framework is structured into four phases: Collect,Transform, Analyze, and Valorize.In the Collect phase, a generic and iterative methodology for constructing thematic datasetsfrom social media is proposed. This approach seeks to mitigate the challenges of creating accurateand representative datasets amidst the voluminous and noisy nature of social media. The objectiveis to shift from ad hoc extraction techniques, prevalent in existing studies, to a more systematic,semi-automatic process. This methodology incorporates human feedback at various stages andutilizes both content-based and metadata-based filtering techniques, alongside semantic domaindescriptions, to offer a standardized and reusable method for thematic dataset building fromsocial media. The methodology was evaluated both qualitatively and quantitatively through thedevelopment of an X/Twitter dataset focused on tourism in the Basque Country region.The Transform phase tackles the challenge of converting multilingual, unstructured text datainto structured knowledge within a given application domain. It concentrates on three pivotalknowledge extraction tasks: (1) Sentiment Analysis, (2) Named Entity Recognition (NER) forLocations, and (3) Fine-grained Thematic Concept Extraction. Given the scarcity of multilingualtraining resources in the tourism domain, the process of manually generating a novel annotatedtraining dataset for this domain is detailed. Subsequently, the thesis explores optimal strategiesfor the multilingual analysis of social media content in tourism, comparing rule-based and deepiiilearning-based approaches (including fine-tuning and prompting-based few-shot learning withvarious language models). This exploration aims to identify the minimal number of annotatedexamples necessary for achieving competitive results across these tasks, leveraging various trainingtechniques and language models. This phase addresses the challenge of minimizing manualannotation efforts without compromising the results¿ quality, considering the time-consuming andexpensive nature of manual data annotation.In the Analyze phase, we hypothesize that adapting the theory of proxemics, traditionallyapplied in physical contexts, to social media could offer a novel approach to crafting meaningful,domain-adaptable indicators for various end-users. The theory is formally redefined, leadingto the development of a modular and extensible proxemic data model. This model is capableof representing social media entities and their interactions in a domain-independent manner.Leveraging this model, ProxMetrics, a toolkit and formula for generating adaptable indicators fromsocial media is introduced. These indicators, conceptualized as proxemic similarity measures, spanmultidimensional social media entities, including users, groups, places, themes, and temporalperiods. They are highly customizable, allowing for the adjustment of the five proxemic dimensions(Distance, Identity, Location, Movement and Orientation) to address various domain requirements.The toolkit and models underwent qualitative evaluations in collaboration with a local tourismoffice to model and address various local touristic requirements.Finally, the Valorize phase addresses the challenge of presenting social media indicators andanalyses to non-computer scientist users, such as domain stakeholders, in an accessible anddomain-independent manner. To this end, TextBI, a multimodal generic dashboard, is proposed.This tool is designed to display multidimensional annotations and indicators over volumes ofmultilingual social media data, focusing on four core dimensions: spatial, temporal, thematic,and personal, while also accommodating additional enrichment data, such as sentiment andengagement. The dashboard offers various visualization modes, including frequency, movement,association and, proxemics, combining features from Business Intelligence (interactivity, combinedfiltering, synchronization of visuals), Geographical Information Systems (spatial view at multiplegranularities), and Linguistic Information Visualization tools (text-based analyses). Unlike mostexisting dashboards, it is generic to operate across different domains, provided the data adheres tothe specified data model. The effectiveness of this dashboard was validated in the tourism domainthrough evaluations conducted by tourism offices, assessing its applicability and relevance.The framework¿s twofold genericity (application domain and data source) is demonstratedthrough the application of each phase in another domain of application: local public policies,leveraging data from municipality review platforms
Motivational attitudes and norms in a unified agent communication language for open multi-agent systems : a pragmatic approach
No full textEThOS - Electronic Theses Online ServiceGBUnited Kingdo
Cross-lingual Transfer for Low-Resource Natural Language Processing
No full text232 p.Natural Language Processing (NLP) has seen remarkable advances in recent years, particularly with the emergence of Large Language Models that have achieved unprecedented performance across many tasks. However, these developments have mainly benefited a small number of high-resource languages such as English. The majority of languages still face significant challenges due to the scarcity of training data and computational resources. To address this issue, this thesis focuses on cross-lingual transfer learning, a research area aimed at leveraging data and models from high-resource languages to improve NLP performance for low-resource languages. Specifically, we focus on Sequence Labeling tasks such as Named Entity Recognition, Opinion Target Extraction, and Argument Mining. The research is structured around three main objectives: (1) advancing data-based cross-lingual transfer learning methods through improved translation and annotation projection techniques, (2) developing enhanced model-based transfer learning approaches utilizing state-of-the-art multilingual models, and (3) applying these methods to real-world problems while creating open-source resources that facilitate future research in low-resource NLP. More specifically, this thesis presents a new method to improve data-based transfer with T-Projection, a state-of-the-art annotation projection method that leverages text-to-text multilingual models and machine translation systems. T-Projection significantly outperforms previous annotation projection methods by a wide margin. For model-based transfer, we introduce a constrained decoding algorithm that enhances cross-lingual Sequence Labeling in zero-shot settings using text-to-text models. Finally, we develop Medical mT5, the first multilingual text-to-text medical model, demonstrating the practical impact of our research on real-world applications
Automatic stance detection on political discourse in Twitter
No full textThe majority of opinion mining tasks in natural language processing (NLP) have been focused on sentiment analysis of texts about products and services while there is comparatively less research on automatic detection of political opinion. Almost all previous research work has been done for English, while this thesis is focused on the automatic detection of stance (whether he or she is favorable or not towards important political topic) from Twitter posts in Catalan, Spanish and English. The main objective of this work is to build and compare automatic stance detection systems using supervised both classic machine and deep learning techniques. We also study the influence of text normalization and perform experiments with differentt methods for word representations such as TF-IDF measures for unigrams, word embeddings, tweet embeddings, and contextual character-based embeddings. We obtain state-of-the-art results in the stance detection task on the IberEval 2018 dataset. Our research shows that text normalization and feature selection is important for the systems with unigram features, and does not affect the performance when working with word vector representations. Classic methods such as unigrams and SVM classifier still outperform deep learning techniques, but seem to be prone to overfitting. The classifiers trained using word vector representations and the neural network models encoded with contextual character-based vectors show greater robustness
- …
