How syntactic distributions affect noun processing (dissertation)
My dissertation presents several novel methods for measuring the syntactic information carried by individual words. These measures, based on the core equations of Information Theory, aim at clarifying several points of contention in modern (psycho)linguistic models of the relationship between words and abstract structures. Current linguistic theories are divided with respect to (a) what syntactic information must be specified for words, (b) whether this information is constraint-based (i.e., categorical) or probabilistic, and (c) whether syntactic information plays any role in lexical processing (i.e., in non-syntactic contexts). Broadly put, the research seeks to nail down more precisely the nature of the syntax-lexis interface.
As a first step, I focus on nouns. I define syntactic distributions using a formalism based on Dependency Grammar. As such, these measures reflect relatively low-level syntactic relationships (i.e., those that bind pairs of words). I correlate these measures with performance on several classic psycholinguistic tasks from production (e.g., picture naming) and comprehension (e.g., visual lexical decision). Based on the behavioral data, I predict patterns of use for a well-established alternation phenomenon: conjunct ordering (e.g., the cat and the dog vs. the dog and the cat). Results shed new light on how syntactic resources are related to nouns, as well as how these resources are recruited during naturalistic production.
The interested reader will find preliminary results in Lester and Moscoso del Prado Martin (2015, 2016) and Lester, Feldman, and Moscoso del Prado Martin (2017), available under the Publications tab.
Null-alternations in second language speech and writing
In on-going collaboration with Stefanie Wulff and Stefan Th. Gries, we focus on morpho-syntactic alternation phenomena, in particular, "null" alternations. As an example, consider that the English complementizer that may be omitted without affecting grammaticality in finite complement clauses: I heard (that) the English complementizer may be omitted. The choice to include or omit that depends on a complex set of factors: semantic-pragmatic, information-theoretic, psycholinguistic, structural, and so on. Our work demonstrates that adult learners of English approximate the native behavior quite well, though with some predictable differences. We view these findings as support for the integrated or unified models of bilingual language representation and processing.
We are currently extending these methods to explore how the distributional profiles of cognate alternations (English that and German dass) co-determine the productions of bilinguals.
The interested reader should consult Wulff, Lester and Martinez-Garcia (2014) and Wulff, Gries, & Lester (in press), available under the Publications tab.
My recent work on L2 use of optional relativizers (the ones (that) you can leave out) reveals that for learners, not all null-alternations are made equal. Where native speakers prefer to include the optional relativizer when complexity and disfluency increase (presumably to preserve communicative integrity), learners make the opposite choice. One possible explanation is that there is a difference in the informational load borne by the relativizer and the complementizer. The former responds to animacy (that vs. who) and for animates, case (who, whom). By contrast, the complementizer only has one overt form. The relativizer is also syntactically "heavier." The relativizer opens a long-distance syntactic dependency to a non-adjacent position within the relative clause. The complementizer merely marks a syntactic boundary. Therefore, learners avoid the "heavy" grammatical marker but expertly use the "light" grammatical marker. Another possibility is that relativizers provide the possibility to err by producing an ungrammatical utterance (e.g., *the yams who I cooked last night). To minimize chances of error, the learners may systematically avoid these structures. I am currently investigating these questions with a broader sample of L1 backgrounds to see whether transfer might play a mediating role.
The first study from this line of research has just been accepted to the International Journal of Learner Corpus Research.
Statistical analysis of small corpora for under-documented languages
In collaboration with Fermín Moscoso del Prado Martín, I have developed information-theoretic and computational tools for the analysis of small corpora. We address a radical claim about the grammatical structure of a colloquial variety of Indonesian spoken in Riau, Sumatra. This variety has been claimed to have no syntactically constrained word order. Using a corpus of less than 3,000 words, we design a robust boot-strap technique for assessing the amount of structure in word-word (bigram) transitions. With so little data, we extract reliable comparative estimates of the structure in Riau Indonesian text compared to Standard Indonesian (Bahasa Indonesia) and English baselines. I am currently interested in scaling these methods up to allow for serious comparative-typological analysis of the fine-grained statistical properties of as many languages as possible.
This research is currently under review. I
How discourse interacts with verb-argument constructions to shape argument realization
In collaboration with Jack Du Bois, I have begun to explore the effects of information flow on the lexical complexity of referring expressions in extended discourse, and how these effects relate to syntactic positioning at multiple levels of abstraction (i.e., subject position vs. subject position of Argument Structure Construction X). In one project, we examine the dative alternation (The librarian gave the man a book vs. The librarian gave a book to the man). We show that the NPs that encode the semantic roles of agent, recipient, and theme are specified for the lexical complexity they tolerate, and that these specializations vary as a mutually constrained gradient between roles, between syntactic positions, and between constructions. These patterns suggest constructional unity of the prepositional dative (i.e., the prepositional phrase is not syntactically "adjunct") as well as grammaticization at the level of lexical encoding strategies. The latter shows movement towards a "one complex argument, n simple arguments" paradigm. By this standard, the prepositional dative shows less informational grammaticization than the double-object construction.
I have performed a similar analysis of raising constructions in English, which are known to have an argument that is shared across the matrix and subordinate clause (I want him to go to the store). Using the same statistical techniques, I show that these constructions are informationally unified (the lexical complexity of any argument depends on the others) at multiple constructional levels, including the overall raising construction, as well as the matrix and complement clause.
This research is still in the initial phases.