Bruce Hayes

On the 17th, the talk will start at 6:00PM as usual and then we will have a reception in the seminar room. (The place will be Humanities 1006)

On the 18th, we will have lunch with the speaker at 12:30PM and the talk will be at SAC 303 at 3:30PM.
Around 6:00PM, we will have dinner at Lori's house. This will be like a colloquium reception.

If you are interested in having a meeting with the speaker, let us know.
And let the students in your classes know these talks and invite them, please.

We hope you will enjoy this semester's linguistics lecture series.

Hijo and Chih-hsiang.

---------------
The textsetting problem: the intersection of phonology, music cognition, and computation

Bruce Hayes, UCLA

Talk at Stony Brook University, Sept. 17, 2009

To understand the textsetting problem, imagine a song in multiple verses. Each syllable in a line is temporally aligned to an abstract rhythmic pattern, describable in a grid notation:

Often, people know only the first verse of a song, but when they read new verses, they intuitively sense what is the right way to arrange the new syllables in time—even when a different number of syllables is present. This is what is meant by “textsetting”. For example, a line from a later verse of the song just quoted would most likely be set like this:

Native speakers generally agree with one another on what settings should be preferred (Hayes and Kaun 1995). Thus, there is a clear analytical problem at hand: find and state the principles that tacitly guide people when they set text in their language.

The problem is not easy to solve, and I will here report some of the tentative progress made by myself and other researchers. In the specific approach I will take, the critical elements are:

· A phonological representation for each line, with stress, syllable weight, word boundaries, and phrasing

· On the musical side, a representation of rhythm and grouping, here borrowed from Lerdahl and Jackendoff’s (1983) Generative Theory of Tonal Music

· A grammar of constraints, regulating alignment of syllables to grid

· A computational system of constraint weighting, which tunes the grammar so that its predicted outputs best match what is seen in a data corpus. Ideally, the system permits not just prediction of single best outcomes, but also covers the variation and ambivalence seen with real speakers. My current grammars are based on maxent weighting.

I will demonstrate the approach with an application to textsetting data taken from Hayes and Kaun’s (1995) study of English folk song.

---------------
Maxent grammars for the metrics of Shakespeare and Milton

Bruce Hayes, UCLA[1]

Talk at Stony Brook University, Sept. 18, 2009

Metrics is a branch of phonology that studies how poets use stress, syllables, and phrasing in manifesting conventionalized rhythmic patterns, for example the iambic pentameter. Since Halle and Keyser (1969), generativist work in metrics has sought to establish explicit formal grammars that characterize how poets perform this task. Early research sought to locate inviolable constraints, for which the exceptions in the data corpus would be so rare as to be attributable to performance error. These efforts were unsuccessful: as critics were quick to point out, metrical constraints having enough substance to be theoretically interesting generally suffer from a fair number of counterexamples.

Later studies (e.g. Youmans 1989, Hayes 1989) suggested that inviolability was not an necessary criterion for metrical constraints, and that it would be reasonable to set up constraints whose violation results in gradiently reduced well-formedness, reflected empirically in the relatively small number of lines that violate them. While intuitively sensible, this proposal had a very undesirable consequence: it deprived the field of the ability to assess metrical grammars rigorously, resulting in great theoretical uncertainty. A particularly acute problem arises from the possibility of constraints being redundant without being detected as such: a proposed constraint C might have very few exceptions in the data corpus, but this would turn out to be solely the result of other constraints present in the system, which when combined make the same predictions as C.

Advances in the formal theory of gradient grammar offer a solution to this impasse. I discuss the application to metrics of maxent grammars, as used in the analysis of phonotactics by Hayes and Wilson (2008, Linguistic Inquiry). In a maxent system, we can develop complete metrical grammars that make explicit and testable gradient well-formedness predictions. The maxent approach also provides a rigorous method for detecting which constraints are redundant. These techniques are applied to two 2000-line verse corpora from Shakespeare and Milton, leading to a new assessment of some longstanding conflicting proposals about English metrics.

--------------------------------------------------------------------------------

[1] The talk describes joint work with Anne Shisko of UCLA and Colin Wilson of Johns Hopkins University