Course Objectives

Harmonic Grammar: Models and methods

1.0 Overview

There now exist a number of variants of HG, along with a range of freely available computational tools for exploring these models of grammar. This course provides an introduction to both the models and the methods, assuming no background in either mathematics or computational modeling. Starting with a version of HG that closely resembles the well-known “classic” OT model of Prince and Smolensky (1993/2004), the course then proceeds to introduce more elaborate probabilistic models (Goldwater and Johnson’s 2003 Maximum Entropy Grammar, Boersma and Pater’s 2008 Noisy HG), as well as versions of HG that use serial derivations (as Harmonic Serialism - McCarthy 2007 et seq.). It also shows how HG learning algorithms can be used to model human language acquisition, and how they can be applied to simulations of language change through iterated “agent-based” learning. Step-by-step instructions for working with the associated software tools will be provided, along with example input files and scripts.

Description of course content

Background

Harmonic Grammar (HG) is a synthesis of generative and connectionist/statistical approaches to language. The models of linguistic knowledge combine representations and constraints (or rules) from generative linguistics with the numerical weights used in connectionist and statistical frameworks. The methods include the formal analysis of individual languages and of language typology familiar from generative work, along with mathematical simulations of language learning and other cognitive processes that draw on research in connectionist modeling and statistical learning. By building on these two usually distinct research traditions, HG combines the strengths of each: generative grammar’s ability to express the structural complexity of linguistic systems, with statistical models’ ability to capture probabilistic generalizations and their learning.

HG was first introduced twenty years ago (Legendre, Miyata and Smolensky 1990; see also Goldsmith 1990, 1991, 1993a and other papers collected in Goldsmith 1993b), but shortly thereafter it was mostly abandoned in favor of the closely related Optimality Theory (OT; Prince and Smolensky 1993/2004), which replaced HG’s constraint weighting with ranking. Recently, there has been a revival of interest in weighted constraint grammars, especially in probabilistic variants of HG and their associated learning algorithms (e.g. Goldwater and Johnson 2003, Soderstrom, Mathis and Smolensky 2006, Hayes and Wilson 2008; see Smolensky and Legendre 2006 and Pater 2009 for overviews of research in HG). Though twenty years old, research in HG is in many ways “a new line of inquiry”.

The revival of interest in HG has been accompanied by the development of computer programs that facilitate linguistic analysis and cognitive modeling. Because it can be difficult to construct “hand-crafted” analyses with weighted constraints, and to program the required algorithms from scratch, the earlier lack of these tools may well have contributed to the dearth of interest in HG compared with OT, whose ranked constraints are relatively easy to work with.

The course is aimed at graduate students and active scholars who are familiar with the basic goals and methods of generative linguistics, especially in phonology and OT. Research in HG is currently flourishing, but the primary literature is often not accessible to readers lacking a strong mathematical background. Furthermore, even for someone with some experience with the literature, the nature of the relationship amongst the various sub-species of HG may well remain unclear. This course highlights the similarities and differences amongst weighted constraint grammars and learning algorithms, and situates them in the context of better-known OT models. After taking this course, a participant will be able to identify a model that has the properties required to address a particular research question, and will know how to use the methodological tools needed to do so. The course also includes a summary of current research results, and topics for further exploration. As well as summarizing and explicating published work, this course will present a number of unpublished case studies, and some theoretical innovations, especially in the sections on serial HG and on agent-based learning.

2.0 Objectives

1. Present the state of the art of constraint grammars and weighted constraint grammars

2. To enable participants to approach a research problem by identifying a model which is best suited for the data in hand

3. Introduce a specific methodological tool, that is, Harmonic Grammar to participants interested in linguistic research