UCLA

With the rapid development of neural models in natural language processing (NLP), large and deep models achieve state-of-the-art across NLP tasks, and are deployed in real-world applications. Models become black-box to our human. Therefore, effective approaches controlling NLP moedls are demanding. Controlling helps model solve particular tasks.For example, when we ask the model to generate a recipe, we have a constraint about what ingredients we want the recipe to contain. In addition, as NLP researchers, we are responsible for preventing models from generating offensive or other unpredictable outputs, otherwise deploying them in real-world applications may cause society issues. To control the NLP models, my research focus on injecting constraints, a set of rules that the model must follow, to control the model behaviour via constrained inference and decoding. My research goal is to develop techniques leveraging different kinds of constraints in various scenarios for structure prediction models and large language models. Generally, constraints represent human knowledge and expectation to the model outputs, and constrained inference is the bridge between human beings and the neural models.