Recommended reading for all projects:
Jesse Dodge, Suchin Gururangan, Dallas Card, Roy Schwartz & Noah A. Smith (2019). Show Your Work: Improved Reporting of Experimental Results. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)
Kyle Gorman & Steven Bedrick (2019). We Need to Talk about Standard Splits. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics
Integrated teaching and learning platforms are becoming increasingly sophisticated. Recent work has employed neural models to create vectors representing a learner’s skill-set and the learning tasks available to them (much like word embeddings). When these embeddings occupy the same vector space, they can be used to recommend tasks appropriate to the learner. Previous work by Moore et al (2019) has modelled latent user proficiencies and tasks as skills embeddings in the STEM domain. Building on work by Chen & Meurers (2019), the aim of this project is to apply a similar modelling approach to the domain of language learning. In particular this project would aim to model both the learner’s written language proficiency and their reading proficiency. It is anticipated that the vectors representing the proficiencies will not share the same vector space: therefore part of this project will involve mapping between the different vector spaces and being able to predict reading competence from writing competence.
Conversational question-answering involves a stimulus text, questions and answers as conversation turns, and character spans indicating which part of the text provides the evidence for the correct answer to the question. The normal objective for this dataset is to learn to generate answers correctly, a task which has been ‘solved’ by BERT-based models (Ju et al 2019). But what if we turn the task around – learning to generate questions based on the answers, the character spans and the stimulus text – do models perform equally well in the reverse direction? This project has the potential to involve natural language generation, evaluation methodology and error analysis on different question types.
In spoken language, words are not explicitly delimited: the listener is faced with the challenge of segmenting the speech stream, with limited cues available. At the same time they must recover from over- or under-segmentation errors, and accumulate a set of word senses which serves them well in their own use of language. This project involves multi-task machine learning to jointly learn word boundaries and word senses, making use of various data sources.
Supervisors: Andrew Caines, Ahmed Zaidi, Øistein Andersen, Helen Yannakoudakis, Marek Rei and Paula Buttery
Essay scoring systems (e.g. Taghipour & Ng, 2016) automatically assign a mark to some essay. However, it would be desirable to produce an explanation for the assigned score, both (i) as feedback to the learner and (ii) to provide users a “right to an explanation”, as required by the GDPR. There has been a lot of work in explainability/interpretability, with one strand being rationale generation (Yu et al, 2019). This project aims to apply recent ideas in rationale generation to obtain explanations for neural essay marking systems. Other potential extensions include exploring other explainability methods (Ribeiro et al, 2016; Shrikumar et al, 2017) and applying rationale generation to speech.
Kaveh Taghipour and Hwee Tou Ng. 2016. A Neural Approach to Automated Essay Scoring. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing
Mo Yu, Shiyu Chang, Yang Zhang, and Tommi Jaakkola. 2019. Rethinking Cooperative Rationalization: Introspective Extraction and Complement Control. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. In Proceedings of the 22Nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning Important Features Through Propagating Activation Differences. In Proceedings of the 34th International Conference on Machine Learning