Skip to main content

I am interested in explaining how children learn the phonologies of their native languages. Such explanations are empirically accountable on multiple fronts: they should be consistent with what is known about the phonologies of the world’s languages, they should be consistent with plausible limits on the psychological resources available to children (the amount of language data they hear, the amount of computation their brains can perform in the allotted time), and they should be consistent with broad patterns of language development behavior exhibited by children. Working within cognitive science, I take language learning to be an algorithmic mental process, and my work focuses on the implications of modern linguistic theory for the language learning algorithm (and vice-versa). I identify specific formal issues, such as structural ambiguity, grammatical restrictiveness and the learning of underlying forms for morphemes, that must be dealt with by a language learner, typically across multiple chronological stages of development, and then construct and evaluate proposals for how a learner could deal with those issues.

Output-Driven Phonology

The book Output-Driven Phonology (Tesar 2014) as the culmination of much of my research over the preceding decade. The work introduces the formal concept of an output-driven map, and pursues the consequences of that concept in both phonological theory and phonological learning. The definition of output-driven map specifies conditions under which all phonological activity may be said to occur solely to satisfy restrictions on the output. The major issues in phonological theory addressed in this work include process opacity, correspondence, and contrast.

While work on output-driven maps addresses central issues in phonological theory, the idea has its origin in work on learning. The structure that output-driven maps impose on the space of possible underlying forms enables large reductions in the computational effort required for the learning both of underlying forms and of certain kinds of ranking information.

Work by Nyman and Tesar proposes a way for extending learning in output-driven maps to systems that include deletion and insertion.

The learning proposals presented in Output-Driven Phonology contain within them proposals addressing phonotactic restrictiveness and the learning of phonological underlying forms. Learning with structural ambiguity in the context of output-driven maps is not discussed in detail in the book, but has been investigated elsewhere. Crystal Akers’ doctoral dissertation combined the learning of underlying forms, enforcing restrictiveness, and contention with structural ambiguity into one simultaneous learning algorithm.

Learning Phonological Underlying Forms

The work that introduced the concept of a contrast pair, and the Faithful Contrastive Feature property:

The use of contrast pairs in learning was advanced substantially by work with Nazarré Merchant, work that formed a key component of Merchant’s doctoral dissertation:

Cycling between the use of morphemic contrast to set underlying features and the learning of ranking information from morphemic alternation:

  • Tesar, Bruce. (2006). Learning from Paradigmatic Information. In Eric Bakovic, Junko Ito, and John McCarthy (eds.), Wondering at the Natural Fecundity of Things: Essays in Honor of Alan Prince. Linguistics Research Center, University of California, Santa Cruz. 293-310. (also appeared in the proceedings of NELS36).

Structural Ambiguity in Learning

Most of my work of contending with structural ambiguity has focused on metrical foot assignment for prosodic word level stress. The more recent work on this has focused on applying inconsistency detection to foot interpretations of multiple words simultaneously, to eliminate foot assignments (interpretations) of an ambiguous word that are inconsistent with other words. This work introduced the MultiRecursive Constraint Demotion (MRCD) algorithm.

Another paper examined the interaction between structural ambiguity and grammatical restrictiveness, in the context of ambiguity in syllabification.

Earlier work approached structural ambiguity in a more iterative fashion, inspired by the Expectation-Maximization (EM) approach popular in statistical learning. That included the use of Robust Interpretive Parsing (RIP) as an Optimality Theory-style way of estimating the full structural description of an ambiguous form.

  • Tesar, Bruce, and Paul Smolensky. (2000). Learnability in Optimality Theory. Cambridge, MA: MIT Press.
  • Tesar, Bruce. (1998). An iterative strategy for language learning. Lingua 104. 131-145.
  • Tesar, Bruce. (1999). Robust interpretive parsing in metrical stress theory. Proceedings of the Seventeenth West Coast Conference on Formal Linguistics (1998). 625-639.
  • Tesar, Bruce. (1997). An iterative strategy for learning metrical stress in Optimality TheoryProceedings of the Twenty-First Annual Boston University Conference on Language Acquisition (1996). 615-626.

Restrictiveness in Learning

The Biased Constraint Demotion (BCD) algorithm combines effectively with MRCD to create an argument that learners should actively enforce restrictiveness throughout learning, rather than relying on a restrictive initial state and ‘conservative’ grammar modifications, and without any obligation to gather all relevant data before constraint reranking.

  • Prince, Alan, & Bruce Tesar. (2004). Learning phonotactic distributions. In Rene Kager, Joe Pater, and Wim Zonneveld (eds.), Fixing Priorities: Constraints in Phonological Acquisition. Cambridge University Press. 245-291. ROA-353.

More recent work has confronted the phenomenon of paradigmatic subsets. The phenomenon and one approach to contending with it are discussed in Output-Driven Phonology, chapter 8. More recent work by Moyer and Tesar explores a complementary approach to the problem.

Learning Constraint Rankings in Optimality Theory

The earliest work in learnability in Optimality Theory focused on the learning of constraint rankings. This is the source of the original Constraint Demotion algorithms, both Recursive Constraint Demotion (RCD) and “online” Constraint Demotion.

  • Tesar, Bruce, and Paul Smolensky. (2000). Learnability in Optimality Theory. Cambridge, MA: MIT Press.
  • Tesar, Bruce, & Paul Smolensky. (1998). Learnability in Optimality Theory. Linguistic Inquiry 29:2. 229-268.
  • Smolensky, Paul, Géraldine Legendre, & Bruce Tesar. (2006). Optimality Theory: The structure, use, and acquisition of grammatical knowledge. In Paul Smolensky & Géraldine Legendre, The Harmonic Mind, Chapter 12. MIT Press. 453-544.

Standard approaches to learning rankings use winner-loser pairs, which requires some method for loser selection; an observed datum is the basis for a winner, but a losing candidate must be constructed and hypothesized by the learner. The concept of using production-directed parsing for loser selection was introduced in the following:

  • Tesar, Bruce. (1998). Error-driven learning in Optimality Theory via the efficient computation of optimal forms. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, & David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax. MIT Press and MITWPL. 421-435.

For historical interest, this is the first paper written on constraint demotion (note the ROA number):

Computation of Optimal Forms in Optimality Theory

  • Tesar, Bruce. (1995). Computational Optimality Theory. Doctoral Dissertation, University of Colorado, Boulder. 127 pages.
  • Tesar, Bruce. (1995a). Computing Optimal Forms in Optimality Theory: Basic Syllabification. Technical Report CU-CS-763-95, University of Colorado, Boulder. 28 pages. ROA-52.
  • Tesar, Bruce. (1996). Computing optimal descriptions for Optimality Theory grammars with context-free position structures. Proceedings of the Thirty-Fourth Annual Meeting of the Association for Computational Linguistics (1996). 101-107.
  • Tesar, Bruce. (1998). Error-driven learning in Optimality Theory via the efficient computation of optimal forms. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, & David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax. MIT Press and MITWPL. 421-435.

Symbolic Computation in Connectionist Models

  • Smolensky, Paul, & Bruce Tesar. (2006). Symbolic computation with activation patterns. In Paul Smolensky & Géraldine Legendre, The Harmonic Mind, Chapter 7. MIT Press. 235-270.
  • Tesar, Bruce, & Paul Smolensky. (1994). Synchronous firing variable binding is a tensor product representation with temporal role vectors. Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society (1994). 870-75.