Daoxin Li will be defending their dissertation, titled "Distributional learning of syntactic generalizations", on Monday June 17th at 10:00 AM EST.

The defense will take place in person in the Linguistics Library (open to all), and on Zoom.

The dissertation is attached, and the abstract is provided below.


Title: Distributional learning of syntactic generalizations

Supervisor: Charles Yang

Committee: Kathryn Schuler, Julie Anne Legate, Marlyse Baptista, John Trueswell



During language acquisition, children are tasked with the challenge of determining which words can appear in which syntactic constructions. This has been long recognized as a serious learnability problem. On one hand, there are productive generalizations that children must learn. On the other hand, language is known for its arbitrariness, so children also need to decide when not to generalize and just resort to memorization. Finally, the picture is further complicated by the lack of negative evidence of what sentences are ungrammatical in a language. In this dissertation, by applying a threshold-based generalization learning model, The Tolerance/Sufficiency Principle, I provide novel approaches to the acquisition of a range of syntactic generalizations across languages.

In Chapter 2, I start with the acquisition of verb argument structure, where systematic mappings between syntax and semantics have been well-documented. I argue that even in this case where systematic syntax-semantics mapping is present, knowledge of such syntax-semantics mapping should not and need not be innate. Instead, I propose a computational model that can learn these mappings distributionally from modest-sized input data. I also conduct model comparisons to illustrate that the proposed model yields learning outcomes that are more accurate and more consistent with human behavior than another model which relies on Bayesian inference.

Chapter 3 moves on to a case where the relation between syntax and semantics is far less systematic - the acquisition of recursive structures. The rules for recursion differ across languages. Some structures are freely recursive; some structures are regulated by semantic constraints which must be learned from language specific experience; and some structures in general cannot recurse. In addition, even for structures where recursion is productive, examples of recursive embedding are rare in children’s input, which poses another acquisition challenge. In this work, I propose a new conceptualization of recursion based on its formal properties in non-recursively embedded data, which leads to a new theory of how it can be acquired. Through corpus analyses of different recursive structures across languages, I demonstrate that the rules for recursive embedding can be established through purely formal analyses of one-level embedding data, and the core semantic properties such as alienable possession vs. inalienable possession can be identified subsequently.

In Chapter 4, to determine whether the proposed distributional learning mechanism can indeed be helpful during language acquisition, I conduct a series of artificial language learning experiments, which find that both adults and children can use purely distributional cues to acquire recursive structures: They will allow recursive embedding in an artificial grammar when there are sufficient cues in the exposure supporting the generalization, even though they never hear recursively embedded sentences in the exposure phase.

Ultimately, this dissertation aims to contribute quantitatively rigorous and psychologically real solutions to a well-known learning problem, offering new perspectives for the mechanisms of learning generalizations.