Chidananda AV

Collected, cleaned and built a huge corpus of legal data. Using this trained In-house word2vec, fasttext embeddings and fine tuned ELMO, BERT. Using these instead of original embeddings resulted in huge improvements in accuracy across all models. Extensively worked on contextual sentence similarity. Experimented with SIF, Siamese networks, USE, Sent2Vec, SentBERT. Built and deployed >10 NER models to extract multiple entities(parties, values, timelines) from various parts of documents. Built and deployed a classifier to identify and link master agreements and their child agreements(addendum, amendment). Experimented with ELMO, BERT, RoBERTa, DistillBERT and achieved a robust solution with >90% accuracy. Built an asynchronous service to automatically train an NER model. The Intention was to allow anyone to build an NER model instantly. Only annotated data was to be provided by the user in an excel and the service would train, tune and save the model.

Used Attention mechanism to improve >300 class paragraph classifier. The approach resulted in >20% increase in accuracy while making the model more robust to OCR errors. Built and deployed a model to identify Intent, liabilities and risks present in the contract. Used unsupervised approach to solve the problem with recall value of ~90% with almost no annotated data available. Extensively worked on document clustering building many models for multiple clients using K-means, KNN, GMM etc., further converted this into a service with sufficient flexibility to decrease redundancy.

Built a POC RNN based language model from scratch to predict the next probable phrase based on a user prompt to help draft contracts easily, the language model was trained using > 100k contracts consisting of ~500M tokens. Built and deployed many multiclass document classifiers for Ad Hoc customer requests for easy document processing. Used algorithms such as SGD, SVM, Naive bayes, LR, multinomial depending on effectiveness.