Deconstructing syntactic generalizations with minimalist grammars

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 435–444 November 10–11, 2021. ©2021 Association for Computational Linguistics 435 Deconstructing syntactic generalizations with minimalist grammars˚ Marina Ermolaeva University of Chicago / Chicago, IL mail@mermolaeva.com Abstract Within the currently dominant Minimalist framework for syntax (Chomsky, 1995, 2000), it is not uncommon to encounter multiple pro- posals for the same natural language pattern in the literature. We investigate the possibility of evaluating and comparing analyses of syntax phenomena, implemented as minimalist gram- mars (Stabler, 1997), from a quantitative point of view. This paper introduces a principled way of making linguistic generalizations by de- tecting and eliminating syntactic and phono- logical redundancies in the data. As proof of concept, we first provide a small step-by-step example transforming a naive grammar over unsegmented words into a linguistically moti- vated grammar over morphemes, and then dis- cuss a description of the English auxiliary sys- tem, passives, and raising verbs produced by a prototype implementation of a procedure for automated grammar optimization. 1 Introduction Marr’s (Marr, 1982) approach to understanding complex cognitive systems, including natural lan- guage, distinguishes between three levels of de- scription: computational (abstract specification of what the system computes), algorithmic (structures representing the data and algorithms that manip- ulate them), and implementational (concrete real- ization of the algorithms in the hardware or wet- ware). Linguistic theories are considered to be at the computational level (Johnson, 2017) or between the computational and algorithmic level (Peacocke, 1986). In particular, syntactic literature often omits lower-level details such as full descriptions of units and operations involved. This allows for a bird’s- eye view of the problem but leads to a lack of explicitness. While the body of work in this field holds tremendous empirical and theoretical value, there is no straightforward way to extract from it a ˚This paper is based on a chapter of the author’s disserta- tion (Ermolaeva, 2021). consensus on what constitutes the best description (grammar) for any given set of language data. How can one choose between various structures that could underlie the same linguistic construction based on some robust quantitative metric and in a way compatible with the insights of theoretical syntax? First, this task requires translating the data into an explicit formal representation. For this pur- pose, minimalist grammars (Stabler, 1997) are a natural choice. Designed as a formalization of the Minimalist Program (Chomsky, 1995, 2000), they are compatible with analyses of syntax phenomena found in theoretical literature and provide a way to implement them as grammar fragments, explicitly defining syntactic units and operations. There exists a substantial amount of work on learning formal grammars from unstructured strings; see e.g. (Clark and Eyraud, 2007; Yoshi- naka, 2011; Clark, 2017). These techniques are based on the notion of distributional similarity: in short, words found in the same contexts are as- signed the same syntactic category. Resulting algo- rithms can make plausible generalizations based on observable data; however, they are of limited appli- cability if the goal is to produce …

Content truncated to 500 words