ESR projects

A two-step approach is planned to extract fine-grained event locations from large-scale knowledge graphs, such as Wikidata, DBpedia, and YAGO. The method first integrates a geographic knowledge graph into an event knowledge graph to counteract the lack of detailed location information, and then employs a graph neural network on the combined event knowledge graph to extract precise event locations.

Fact validation is composed of the sub-tasks of parsing inputs, retrieving evidence, and inferring relations between the two. In related applications, approaches that compose a design – after breaking problems down into steps – are better suited to tasks with composite structure. Learning on multiple modalities introduces an additional level of complexity for monolithic architectures. This research aims to implement and evaluate systems optimised to process diverse data and adapt to the multiple sub-tasks comprising the fact validation pipeline.

• • We provided a semi-automated framework for generating multiple paraphrase responses for a question by utilizing methods such as back-translation. Furthermore, we release the first KGQA dataset with multiple paraphrased verbalized responses.
  • We presented the first multi-task-based answer verbalization framework that employs questions and logical forms as inputs and simultaneously trains four modules for generating natural language answers.
  • We developed two multi-task neural semantic parsing approaches for (complex) conversational question answering over knowledge graphs. We distinguish the works from the number of sub-tasks they perform and the different deep neural architectural modules they utilize.

For Conversational QA, we studied whether the availability of entire dialog history, domain information, and verbalized answers can act as context sources in determining the ranking of KG paths while retrieving correct answers. We proposed an approach that models conversational context and KG paths in a shared space by jointly learning the embeddings for homogeneous representation.

• • There are empirical evaluations that for AI systems, the explanations regarding the retrieved answers improve trustworthiness, especially in wrong prediction. Hence, how an answer verbalization can be explained remains an important open research direction.
  • For Conversational QA, path ranking is a relatively new method for approaching the task. While it is more difficult to comprehend the predictions compared to semantic parsing models, path ranking will be favored since it does not depend on any annotation process or data. We see an emergence of future ConvQA systems towards that direction; however, the explainability of ranking strategies remains a tough challenge and is eventually to be addressed.
  • For the developed approaches, we employ multi-task learning via hard parameter sharing. Multi-task learning can be crucial for approaching complicated tasks such as question answering. We expect the development of massive and flexible multi-task learners able to perform numerous tasks. Question answering will be achieved by combining parts of them, and for any new knowledge learned, the model will update a comparatively small number of parameters. Likewise, we believe that multi-task learning is essential in developing artificial intelligence with more human-like qualities.

• • A novel method for text-based claim detection was proposed, as well as a novel approach for hate speech detection in multimodal memes (a multi-task learning model).
  • A new multimodal claim detection dataset was created including a multi-topic and diverse set of claims with annotations for multiple tasks.
  • A theoretical framework was proposed for multimodal news analysis based on research in semiotics (image-text relations), journalism (news values) and computational science.

• • Data enrichment: Contextualizing news documents helps researchers enrich their datasets with valuable information such as the geographic source of an event.
  • Improved data analysis: Geolocation estimation and event type classification provide additional insights for analysing news photos. For instance researchers can explore the spatial and temporal patterns in news coverage.
  • Verification and fact-checking: Geolocation estimation and event type classification can aid researchers in verifying the accuracy of news documents. By integrating these meta information researchers can further explore misinformation detection, and identification of manipulated images. 
  • Multimedia representation: Geolocation-based image-text relations can help social scientists in analysing media content and representation. For instance they can investigate how different locations are depicted in images and the accompanying text. Furthermore they can explore cultural stereotypes, or media bias.

• • Study geolocation explainability between image and text pairs of news
  • Study the impact of geolocation estimation on fake news detection and news recommendation

• • Development of a framework to combine qualitative and quantitative methods for the study of media coverage of attitudes to Europe in UK and Portuguese media. The focus was on accessibility and reproducibility of the method by historians with limited digital research skills.
  • Development of a new (and easily reproducible) method for building a multilingual corpus derived from online newspapers, both open access and subscription content, based on thematic criteria
  • The study of media coverage of the European Union has contributed to understandings of how discussion in the digital public sphere influences the attitudes of European citizens to the EU, including pre- and post-Brexit in the UK. It has uncovered significant similarities but also significant differences between the topics, institutions and individuals who feature prominently in UK and Portuguese media, and particularly in how the two countries mobilise their histories and national traditions in relation to the EU.

• • Expanding debates in multilingual Digital Humanities, with special focus on democratising access to resources and data.
  • Introducing reproducible digital methods to historians, including through publication of results in mainstream history journals.
  • Encouraging reflection on the nature of the digital public sphere in the European Union.

• • Development of a framework to combine quantitative (digital) and qualitative methods for the study of the media coverage of events of significant importance for society. The focus was especially on mechanisms of critical data and digital methods thinking by using different approaches to review the implications of their use in the social sciences. Results are published in the paper (https://doi.org/10.1007/s42803-022-00052-9)
  • development of a tutorial (forthcoming) on the Programming Historian on how to retrieve and analyse data publicly provided by the BL. Results were shared in the UK Web Archive conference 2022, organised internally by the institution for its collaborators. Documented in video (https://www.youtube.com/watch?v=9XbCcVqXVeo)
  • The study on the media coverage of the Olympic legacy has contributed to the understanding of the research community of the different dimensions in which the event is narrated and cities are impacted as a result of hosting mega events. The results were shared and discussed in the event ‘Documenting the Olympics and the Paralympics’, organised by me (ESR10) in partnership with the British Library, British Society of Sports History (BSSH), the International Centre for Sports History and Culture at De Montfort University (ICSHC) and the School of Advanced Study (SAS). Documented as podcast (https://open.spotify.com/episode/5zDSk4JsfOvO8Jq4Aq1cXL?go=1&sp_cid=0e0eeb5dba04330acafd5dd9774ffa5c&utm_source=embed_player_p&utm_medium=desktop&nd=1)

• • Expanding debates in multilingual Digital Humanities, with special focus on democratising access to resources and data.
  • Deepening research on digital methods, their potentialities and limitations. It would be important to investigate how explainable AI can help Digital Humanities scholars to improve our understanding of the impact of using such tools.
  • Provoking reflections and debates on interdisciplinary research and how to develop mechanisms to improve collaboration across disciplines in aspects such as joint publications, productive discussions and development of projects.

• • A new methodology to analyse news spreading barriers: a corpus has been collected and annotated containing news articles related to three different domains: natural disasters, sports, and climate change. The collected corpus has been extended with background information related to six barriers (linguistic, economic, time-zone, geographic, political, and cultural). For the mono-lingual news, network analysis is performed, whereas for the temporal analysis cascading analysis is performed.
  • A new approach to enhance the topic modelling technique and understand political and economic differences in news reporting: enhanced topic modelling technique that uses LDA with a combination of 1-6 grams and article pooling based on queries to improve the quality of topics without modifying the structure of LDA. The proposed approach has been applied to news articles related to COVID-19.

• • Extension of the proposed framework to deal with low-resource multilingual headlines under imbalanced sample distribution in a language-agnostic setting with comprehensive qualitative analysis. Our proposed language-agnostic solution is adaptable to real world scenarios where a system is expected to deal with low-resource situations.
  • Design and implementation of flexible data generation methods to generate custom-labelled datasets for related tasks. A major issue that frequently arises in real-world scenarios is a lack of annotated data. Our proposed data generation strategy, which is capable of working with low-resource languages with an imbalanced distribution, is intended to address this issue. In addition, it will facilitate future expansion and the creation of custom datasets for related tasks. The generated data would not only be ideal for automated systems like the “bias flipper” but would also be beneficial for social scientists and researchers interested in analysing news reporting bias.

• • We plan to analyze the speed of reporting, time-span, and importance given to the events by the outlets. In addition, we also are looking into how the outlets change their coverage style over time.
  • We intend to diversify our additional knowledge sources. In particular, we intend to investigate how knowledge sources such as Wiktionary and ConceptNet influence the task of polarity prediction. 
  • We are planning to extend this study beyond polarity prediction to its quantification and correction. We believe it would be interesting to experiment with auxiliary task involving news headlines in a multitask learning paradigm
  • We intend to have a distribution of articles with positive and negative sentiment for specific events and outlets. This would reveal not only the outlet’s political orientation but also the editorial’s overall attitude.

In my PhD, I research the creation and development of current event pages on Wikipedia and perform large-scale temporal analysis of events across different European Wikipedia language versions. I specifically research events that have a social or cultural significance for the European Union or European cultures such as the European Refugee Crisis, the Arab Spring, and Brexit.

I also work on cultural recommendation of page creations based on cultural and social relevance, preferences and importance, as well as the relationship between events, location and related entities.

• • Study of various large language models using probing mechanisms (negation) to improve the final sentiment analysis performance for low-resourced Slavic languages
  • Improving the overall scores using multitask learning and resources from the same and distant family languages in low- and high-resource settings
  • Study of various data augmentation techniques for the task of sentiment analysis and their performance in low-resourced language
  • Development of new resource(s) for SA in Slavic language

• • Experiment with datasets from multiple language families and the transfer of sentiment for low-resourced languages.
  • Introduction of multiple modalities in the form of audio and video annotation from other language families for SA.
  • More sophisticated data augmentation tools for low-resourced languages.
  • Focus on individual instances’ importance and domain sensitivity when using the source and target languages resources in cross-lingual experimentation.
  • Study of noisy labels in low-resourced settings using confident learning is another field to be explored.

• • Improving applicability of method by considering related evidence for example user credibility, Geo location, etc…and extending the methods to be applicable on other modalities which can be used for spreading fake news
  • providing a dataset for chart fact checking in order to validate a textual claim based on scientific charts (e.g bar charts, line charts, pie charts, etc)