One of the fundamental challenges in Natural Language Processing is the variety of expression permitted in English or any natural language. For example, there are vast numbers of words in any vocabulary and it is naïve to treat these as having completely distinct meanings. Distributed word representations, which represent each word in terms of it’s co-occurrences with other words, have long promised (Lee (1999),Weeds(2005)) to be useful in tackling this problem since they capture the similarity or relationships between the meanings of words in a continuous space.
The rise of neural network methods for learning low dimensional distributed word representations, since Mikolov’s seminal work on Word2Vec (Mikolov(2013)), has further increased the popularity of work in this area with applications being made in all areas of Natural Language Processing, including speech recognition, question-answering and machine translation. However, a number of questions remain unanswered; two of which will be considered in this project. First, distributed word representations tend to conflate a large number of different semantic relations including synonymy, antonymy and hyponymy; which may be appropriate in some but not all applications. Therefore, how can these different relations be distinguished? Second, given distributed representations of words, what is the most appropriate way to combine them to represent larger units of meaning such as phrases and sentences? These two questions are in fact related since the meaning of a phrase has a semantic relation with the meaning of each of its components. For example, a ‘sweet fruit’ is a ‘fruit’ and can be seen as a hyponym of ‘fruit’. This project will therefore explore the interaction between composition and lexical entailment in distributed word representations.
University of Sussex
* You will have good programming skills in at least one of Python/Java/C/C++ (essential).
* You will have experience in Natural Language Processing / Computational Linguistics (essential).
* You will have experience in Machine Learning (desirable).
Applicants will have an excellent academic record and should have received or be expected to receive a relevant first or upper-second class honours degree.
The full award is available to UK and to EU students who have been ordinarily resident in the UK for the previous 3 years. EU candidates who do not meet this criteria will be eligible for a fee waiver only and Overseas (non EU) students are not eligible to apply.
How to apply:
Please mention NLP People as a source when applying