Laurence Cullen

Built an entire NLP pipeline to translate raw transliterated cuneiform script from the global archive of cuneiform writings, focusing initially on the Sumerian language.

This involved creating a tokenizing model from the text using Google's SentencePiece tool and building word and word fragment embeddings from the entire corpus to use as a pre-trained embedding layer for the translation model with Facebook's FastText encoding tool.

Finally, I built a sequence-to-sequence model using an encoder-decoder LSTM neural network in Keras and trained it on the fraction of the Sumerian corpus for which translations exist. Some of the better experiments have achieved BLEU scores of 12.8, giving pretty sensible results for a good fraction of untranslated sentences.

I undertook experiments aligning the SNOMED and HPO medical ontologies using sentence embeddings.

I used the recently released Google sentence embedding model BERT to build embeddings of SNOMED and HPO terms and matched them based on cosine similarity. The results were surprisingly effective, and in most cases, the system could correctly translate a term from one ontology to another.

Won a £4,000 tech fellowship with open-source research organization Bellingcat by winning their hackathon.

I developed a tool to make the army tribunal documents published by the Indian government searchable with keyword search, semantic search, and reverse document search. This was achieved by combining web scraping of tribunal documents with the embedding of these documents using OpenAI API calls and a vector database to calculate distances between document embeddings. I deployed it into a web application with Google Cloud Run as a FastAPI server wrapped in a Docker container.