Abstract

1) Distributional Semantic Models Suited To Morphologically Rich Languages

Paula Czarnowska

During my MPhil I worked on a modified version of the Skip-gram model in which the target-word’s contexts were defined as words connected to it through a dependency relation. The model associated meaning representations with all three: target-words, context-words and dependency labels which allowed it to learn important interactions between those during training. In the next three years I intend to continue working in the area of distributional semantics. More specifically, I plan to work on cross-linguistically applicable semantic models, with a focus on morphologically rich languages.

——————————————————————————————————————

2) Multimodality and Language Acquisition

Chris Davis

My MPhil focused on differentiating visual input in the context of semantic similarity and predicting brain activity. My PhD will investigate multimodal models for language acquisition.

——————————————————————————————————————

3) Predicting the limits of the Zone of Proximal Development: a vector based approach

Russell Moore

According to Vygotsky, the `Zone of Proximal Development’ (ZPD) is the set of learning tasks that can be successfully completed by a student, with assistance from a human or from non-human scaffolding. In this study we use a vector-based encoding of student achievement as the input to a machine-learning classification task which aims to predict the limits of a student’s ZPD at a given point in their learning career. We evaluate this encoding with data from 1,000 eligible users taken from a large-scale online learning environment. The model in this context is able to predict whether a question falls inside a student’s ZPD , with a combined F1 score of 0.80.