QUESTION ANSWERING, GROUNDING, AND GENERATION FOR VISION AND LANGUAGE

One ultimate goal of AI is to develop an artificial intelligent (AI) system that can communicate with people in a natural way. Such communication includes but is not limited to asking we humans questions, answering our questions, conducting dialogue with human beings, and performing some actions to better serve people. Imagine in the future where the service robot is everywhere, and we could ask our home robot to “grab me the red cup on the table.” To perform this command, the AI system needs to understand this spoken English sentence, perceive the visual world, navigate to the right place “table”, recognize the right object “the red cup”, then grab it and finally return it back to the commander. Just for this single command, it already involves many techniques, such as speech recognition, language understanding, scene understanding, embodied navigation, object recognition, pose estimation, robot manipulation, etc. Each of these techniques are not well solved yet, but we are on a rapid way toward the success. This thesis is in advancing our knowledge to explore various connections between vision, language and even beyond to push forward this ultimate goal. We study 3 popular vision and language tasks, including visual question answering, language grounding, and image-to-text language generation. Inside each, we will introduce our proposed novel task, accompanied with high-quality dataset and well-performing data-driven approaches. Specifically, we first introduce Visual Madlibs for image-based and region-based question answering. Then we introduce referring expressions, where we study both referring expression comprehension and generation, covering both language grounding and generation. Next, we study album summarization, which not only selects the key photos inside an album but also generates a natural language story describing the whole album. Last but not least, we describe multi-target embodied question answering, a task that is even closer to our ultimate goal that requires both language understanding and navigation ability from the AI system.Doctor of Philosoph