Abstract: Currently, Named Entity Recognition (NER) tasks use shallow machine learning models like Maximum Entropy Markov Models or Multi-class logistic regression to get decent accuracies on Natural Language Processing tasks. More recently, deep neural models, especially recurrent neural networks like Long short-term memory (LSTM) have given state-of-the-art accuracies on NER data. We present a hybrid model that uses gazetteer data with a bi-directional LSTM model to beat current state-of-the-art accuracies, enabling our systems to power IOT type applications which have a high transcription error rate.
Bio: Vijay Ramakrishnan is a machine learning researcher at Cisco. He is a core member of the Mindmeld team within Cisco, developing Artificial Intelligence (AI) and Natural Language Processing (NLP) applications for Cisco’s flagship products. He is an expert practitioner in developing NLP models and building deep domain conversational AI products. He has built voice and chat assistants for fortune 100 companies at Mindmeld Inc. before it was acquired by Cisco in July 2017. His work combines basic research and software to build state-of-the art AI models for conversational products.