PhD Studentship: Who does what to whom - Syntactic parsing in man and machine

Project supervisors: Professor Uta Noppeney - School of Psychology

University of registration: University of Birmingham

Non-Academic partner: Bernd Bohnet and Ryan McDonald - Google

Project title: Who does what to whom - Syntactic parsing in man and machine

Project outline:

Language processing is one of the most complex tasks facing the human brain in everyday life. In order to understand ‘who does what to whom’, the brain needs to assign a syntactic structure to each sentence – a process coined ‘syntactic parsing’. Human neuroimaging research to date has investigated syntactic processing predominantly by comparing neural activity elicited by different classes of sentences (e.g. wh-clauses, garden path sentences). While these studies were able to delineate the neural systems involved in sentence processing, they provided only limited insights into the underlying computational operations and neural mechanisms. By contrast, computational linguistics has developed a range of machine learning algorithms (e.g. dependency parsing) that are trained on large annotated linguistic corpora to assign grammatical classes and syntactic structures automatically to novel natural sentences.Yet, while these algorithms are relatively efficient, they do not obtain the accuracy of human readers.

In this research project we exploit synergies between these two research fields. Moving beyond traditional structure-function mapping approaches, we will employ advanced machine learning algorithms to characterize the computational operations and neural mechanisms underlying syntactic processing in the human brain. Conversely, the insights obtained from human neuroimaging and eye tracking will provide critical constraints on the parameters and algorithms used in machine learning. The synergistic potential of machine learning and human cognitive neuroscience has recently been illustrated by the elegant work of Huth et al. revealing distributed organization of semantic representations in the human brain (Huth et al., 2016, Nature).

During the psychophysics, EEG and fMRI studies, participants will be presented with sentences from syntactically annotated corpora i. in rapid serial visual presentation mode under central fixation and ii. under natural reading conditions with informative eye movements. Comparing human and machine learning performance will provide insights into similarities and differences between parsing strategies. Importantly, we will also train parsing algorithms jointly on annotated corpora, human eye movement and neuroimaging data. This will enable us to directly associate neural processes with computational operations within the parsing algorithms. For instance, we will compare dependency parsing models with different working memory load or beam size (i.e. the number of words that the algorithm can look ahead). Moreover, from the machine learning perspective we will determine whether the accuracy of parsing algorithms can be increased by i. mimicking parameters employed by human parsing or ii. training on both annotated corpora an human neuroimaging data.

Closing date for applications: 8th January 2017

Check eligibility and apply here: https://www2.warwick.ac.uk/fac/cross_fac/mibtp/pgstudy/phd_opportunities/application/

Please note:

iCASE students must fulfil the MIBTP entry requirements and will join the MIBTP cohort for the taught modules and masterclasses during the first term. iCASE students can then start their PhD project in Jan 2018 but must complete a 3-month miniproject (at a non-home institution) before the end of their first year. They will remain as an integral part of the MIBTP cohort and take part in the core networking activities and transferable skills training. MIBTP iCase.