Natural Language Process

Our goal is to solve the ambiguity of natural language processing (NLP) and to develop a deep-learning model for NLP at a level that can be used in the actual industry. Especially, we are interested in inventing natural language understanding (NLU) models to decipher the meaning of the text and applying those models on various types of NLU tasks such as multi-hop question answering or commonsense reasoning. Furthermore, we also fascinated by solving natural language generation (NLG) problems such as machine translation or dialogue systems using neural network models. We also study methods for structuring linguistic multi-modal representation with vision or knowledge graph in a way of neuro-symbolic approach.

Machine reading
comprehension and QA
Natural language understanding (NLU) is the main field of research in natural language processing.
In NLU, the goal is to make machines to read and understand text. DAVIAN LAB has been working on a neural network that can be reasoned for a given complex question such as multi-hop reasoning and answerability. Recently, we attempted to incorporate knowledge graphs into the neural network for neural symbolic AI.
Neural machine translation
Neural Machine Translation (NMT) is one approach for machine translation such that the model can predict a sequence of words. However, despite the recent success, the performance of NMT drops substantially against traditional statistical machine translation (SMT) when the model encounters the problems, e.g., domain mismatch, data scarcity, or long sentences. To overcome these issues, we are particularly interested in handling domain mismatch and data scarcity problems by using several learning methods, such as unsupervised learning, transfer learning, and meta-learning. Also, besides the problems of domain mismatch and data scarcity, we attempt to fulfill users' intent by leveraging post-editing data since each individual can have different translation styles.
Auto completion and
debugging in programming
DAVIAN's research is not limited to simple NLP tasks but also deals with unusual text processing.
We utilizes natural language models or reinforcement learning to handle the text requiring strict grammar or rule.
Auto-completion and debugging in programming languages are such areas covered by DAVIAN.
We are conducting research to predict the next line from the written codes and debug those codes automatically.