Natural language processing postgraduate personal statement example

Personal statement example

My final-year project began with a spreadsheet rather than a research question. A community centre near my university had collected several hundred free-text comments from residents and had no practical way to read them all. I built a classifier to sort the comments into broad themes, starting with TF-IDF and logistic regression as a baseline and then fine-tuning a small pre-trained transformer. The accuracy improvement was modest, but what stayed with me was how much of the work was linguistic rather than computational: deciding whether "the lift is out again" belonged with maintenance or with accessibility, and noticing that my labels smuggled in assumptions about what residents were complaining about. I want to study natural language processing at postgraduate level because that project showed me the limits of what I could do with borrowed architectures and no real understanding of how they represent meaning.

My degree gave me solid foundations in algorithms, probability and machine learning, and I took optional modules in information retrieval and human-computer interaction. Since graduating I have kept reading and building. I worked through Jurafsky and Martin's chapters on sequence labelling and semantics alongside the Hugging Face course, and I have been reimplementing components from scratch to check that I actually understand them: a bigram language model with add-k smoothing, then a small attention layer in NumPy before trusting a library version. I have also read Bender and Koller's argument that form alone does not give a system access to meaning, which I found genuinely unsettling given how confidently I had described my classifier as "understanding" feedback. I would like to study evaluation and interpretability properly rather than relying on a single accuracy figure.

During my final year I spent a supervised week shadowing the two-person data team at a housing charity. My role was firmly that of an observer and occasional helper: I sat in on their weekly planning call, was given a copy of an anonymised dataset to explore, and wrote a short script that flagged duplicate records for a colleague to check. Most of what I learned came from watching them work. They spent far longer agreeing definitions with caseworkers than they did writing code, and they were careful about what they would not automate, particularly anything touching vulnerable tenants. When I suggested a keyword search for urgent cases, the analyst pointed out that the people in most difficulty often wrote the least. That correction has shaped how I think about text data.

I now work as a library assistant at a further education college, which is ordinary work that has turned out to be relevant. Students arrive with a vague phrase and need it turned into something a catalogue can match, and I have become good at asking the questions that narrow it down. I also tutor GCSE maths on Saturdays, where explaining probability to fifteen-year-olds has made me clearer in my own writing. Growing up speaking Twi at home makes me conscious that most NLP resources assume English, and I am interested in low-resource approaches, particularly the practical question of what can be achieved with small, carefully built corpora.

At postgraduate level I hope to work on evaluation of language models beyond benchmark scores, and to take a dissertation topic involving text from a real organisation, with the messiness that entails. In the longer term I would like to work as a research engineer somewhere that treats language data as something produced by people rather than scraped from them.