Towards Broad-Coverage AMR Parsing with Dynamic Continuized CCG

Dynamic continuized CCG is a method of compositional parsing that allows indefinites to take scope out of scope islands in a predictable and formalized manner. For example, it allows us to achieve the correct reading of the sentence "If one of my relatives dies, I'll inherit a fortune" (wherein the speaker is identifying a specific relative). At the same time, it prevents the universal from taking wide scope in the sentence "If all of my relatives die, I'll inherit a fortune" - if it did, we would get the reading that as long as any of my relatives die, I will inherit the fortune.

Unfortunately, previous work in dynamic continuized CCG (DC-CCG for short) has not provided a way to allow it to learn from datasets. Rather, it can only be used with a strictly specified lexicon, making it not useful for practical purposes. However, if we were able to reverse the rules of DC-CCG, we would be able to implement lexicon generation, and thus be able to learn from large datasets. As such, this thesis will demonstrate how these rules can be reversed, and how this extended system can be implemented in an existing semantic parsing framework.