In the SELEXINI project, jointly working with Université de Lorraine and CNRS, I build and test NLP models for processing multi-word expressions (e.g., kick the bucket, take off). The main goal is to improve the interpretability and robustness of NLP models while enhancing the diversity in linguistic data.

In this project, DISINFTRUST, at Cardiff University, I developed a natural language processing (NLP) model to assess the trustworthiness of news articles on online platforms. Through online surveys, I identified factors that influence people's trust levels in online news articles. Based on these insights, I created datasets to train NLP models and tested them. Consequently, I developed an NLP model that accurately identifies untrustworthy news articles.

For my PhD project at University College London (UCL), I investigated how various factors interact and collectively affect learning probabilistic morpho-phonological patterns. 

From artificial language learning experiments, I found that learners' bias against homophony caused learners to be poorer at learning a neutralizing alternation compared to a non-neutralizing alternation. 

However, when learners were frequently exposed to neutralizations in their native language, they were equally good at learning both alternations.

I implement these findings in a MaxEnt model by discounting the homophonous words in the data.