Subcategorial Considerations in Statistical Categorial Parsing

Research in categorial grammars (CGs) is fortunate to have had a storied, decades-long history, with contributions from scholars of diverse research disciplines such as linguistics, mathematical logic, and computer science, among others. In natural language processing, categorial grammars have played critical roles in the development of efficient, wide-coverage statistical parsers, and have further provided an essential compositional framework for various semantic parsers and corpora. One aspect that has been missed in the vast majority of computational work in statistical CG parsing, however, is a treatment of CG lexical categories that reflects their structured nature. In this dissertation, I argue for a decomposed, subcategorial consideration of CG lexical categories for statistical CG parsing. In particular, my work focuses on combinatory categorial grammar (CCG) and Lambek categorial grammar (LCG), two important members of the categorial family. I demonstrate how subcategorial awareness is uniquely beneficial to three important aspects of statistical CG parsing. First, I introduce an LSTM-based CCG supertagger that can predict supertags one primitive at a time. This enables more effective incorporation of prediction history than is otherwise possible, resulting in a supertagger with improved accuracy, a parser with improved coverage and F1, and even allows for the generation of novel supertags, an entirely new capability for supervised CCG parsers. Second, for LCG parsing, I show how to express LCG proof net validity conditions as neural network loss functions, all of which are critically enabled by structured decomposition of the lexical categories. I apply these loss functions to the training of a Transformer-based LCG parser, thereby presenting the first statistical parser for LCG. I further show that the loss functions enable training the parser without ground-truth derivations. Finally, I investigate CCG parser evaluation, and show that the standard metric is prone to overamplifying minor errors. I introduce a new, decomposed version of the metric that relies on subcategorial labelling and alignment. Expert judges unanimously agree that the decomposed method better isolates parser errors. In examining their judgements, I find that expert judges show difficulty agreeing with each other when comparing parses for different sentences, raising important questions for statistical parser evaluations more generally.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.