PhD in Computer Science and Informatics: Learning structured vector space embeddings for intelligent web search
|Application deadline||1 May 2018|
|Start date||1 October 2018|
|Level of study||Postgraduate research|
|Award type||PhD studentship|
|Number of studentships||1|
Many approaches in the field of artificial intelligence (AI) rely on the availability of knowledge bases that encode facts and rules about the world.
For example, when interpreting a query about 'UK prime ministers', search engines such as Google can infer that pages about 'David Cameron' are relevant, because their knowledge base contains the fact that David Cameron has been the UK prime minister. Similarly, this knowledge base may contain the rule that prime ministers are politicians.
While useful, such knowledge bases have a number of important limitations. First, any knowledge base will necessarily be highly incomplete, as there are simply too many things to know about the world that may be relevant to a search engine. Second, existing knowledge bases are currently limited to objective, factual knowledge, whereas some of the most useful knowledge for answering queries is actually subjective or vague; e.g. consider what is needed for answering a query that asks for a list of 'affordable hotels in the centre of Cardiff'. Finally, existing knowledge bases mostly encode facts, and typically only encode a small number of rules, or even none at all. This is related to the fact that most of the rules that we would intuitively want to express are not “hard rules”, but rather encode typical scenarios (e.g. the fact that most people live in the same city where they work), which can be difficult to formalise precisely.
In the last few years, a different paradigm for representing knowledge has emerged, called vector space embeddings. The main idea is that objects, relations, and concepts are represented using geometric constructs in high-dimensional vector spaces. For example, a standard approach is to represent objects as points and relations as translation vectors, such that e.g. we have plondon + rcaptial = puk, where plondon and puk are the points representing London and UK, and rcaptial is the translation vector representing the 'capital of' relationship. An important advantage of these embeddings is that they can be learned from data that is available on the web (e.g. Wikipedia articles). Furthermore, they address several of the shortcomings of knowledge bases, e.g. they can naturally capture subjective knowledge as well as knowledge about the similarity between different objects.
However, compared to knowledge bases, current vector space embedding models are still very restrictive in the kind of knowledge they can encode. For example, they cannot represent rules, and are very limited in how they represent events and causal relationships. The aim of this PhD will be to study ways of learning vector space embeddings that encode such types of knowledge, and to explore ways in which they can be used for developing more intelligent search engines.
|Tuition fee support||Full UK/EU tuition fees|
|Maintenance stipend||Doctoral stipend matching UK Research Council National Minimum|
|Additional funding offered||Additional funding will cover research consumables, training, and travel to conferences. Students earn additional income by supporting the School’s teaching.|
|Residency||UK Research Council eligibility conditions apply|
You should have a 1st or Upper 2nd Class degree or Master’s degree in a relevant discipline.
Maintenance stipend at UK Research Council rate of £14,553 per annum (2017/18). Tuition fees paid at the UK/EU rate: £4,121 per annum (2017/18). Overseas students must fund the difference between the tuition fees at the UK/EU rate (see above) and the overseas rate (£18,980 per annum 2017/18).
Consideration is automatic upon application for admission to the PhD in Computer Science and Informatics with an October 2018 start date.
In the funding section of your application, please select 'I will be applying for a scholarship/grant' and specify that you are applying for the advertised funding. In the research proposal section, include the project description contained here.
We reserve the right to close applications early should sufficient applications be received.