Psycholinguistics Seminar

Jeff Lidz & Colin Phillips

Tuesday 2-5pm

MMH 1108

Spring 2006

The problem

Learning and parsing share a certain abstract structure. Learners must determine the structure of their language on the basis of partially ambiguous data. Likewise, parsers must determine the structure of a sentence on the basis of incomplete or ambiguous data. This kind of "reasoning under uncertainty" lends itself to solutions from the domain of probability theory. Looked at from another angle, learners and parsers have been shown to be sensitive to statistical information under various experimental conditions. Many have argued that the successes of probabilistic methods in NLP and the demonstrations of statistical learning in young children fundamentally change the nature of long-standing debates about the structure of the language acquisition device. However, there have been very few attempts to show how these methods can actually solve the problems that have motivated approaches such as the Principles and Parameters approach to syntax. At the same time, parameter-based learning models have also had only limited successes in explaining the kinds of problems that they were designed to solve. In this seminar we aim to explore what, if anything, distributional learning and parsing models have to offer for the problems of learning and parsing grammatically rich representations. We will start by outlining the learning and parsing challenges that we aim to address. This will include selecting a series of 'model problems' that we will bear in mind as we read relevant experimental and computational literature. We will then review first a selection of experimental studies on statistical learning and parsing, and then some of the methods that the NLP community considers to be most useful, and evaluate whether these tools provide what we need in order to solve our problems.

How to find us:

Jeff Colin

MMH 1417D MMH 1413F

301.405.8220 301.405.3082

jlidz AT umd dot edu colin AT umd dot edu

Student responsibilities

• Class Participation (10%):  Do all of the readings and come to class prepared with questions, comments, things you didn’t understand, things you think are wrong-headed, things you think are beautiful, etc. Being an active participant in class discussions will increase your understanding and will help you to develop your own research projects. As a part of class participation, you are responsible for ensuring that a learning or parsing problem that you 'adopt' gets sufficient attention during the class discussions. 

• Presentations (30%):  Each week, some number of you will be responsible for some part of the discussion. In the week(s) in which you are presenting, come prepared with either a handout or a computer presentation.  Your presentations should (a) cover the main arguments of the papers, discussing the methods used, the hypotheses tested and how the results bear on the hypotheses (30 min. max); and, (b) evaluate the conclusions reached, identify open questions and ways of testing them, integrate the conclusions with other readings or other things you know.  'Book reports' are not acceptable in this kind of class. We encourage you to discuss your presentation with at least one of the instructors beforehand.

• Commentaries (40%):  You will also be responsible for writing up two commentaries on the material that you present, expanding upon the issues in your presentation and reacting to class discussion. Your two commentaries must be based on different aspects of the class (i.e., you can't write two commentaries on experimental studies of language acquisition, although you could write on an experimental and a computational approach to language learning). 

• Short Paper (20%): In the first several weeks, you will identify a problem in either learning or parsing that you think fits into the overall problem investigated in the seminar. As we progress, you will keep track of how what we are learning bears your particular problem. At the end of the semester, you will write up some ideas concerning novel possible solutions to your problem.

Schedule of activities: This schedule will almost certainly change, depending on class interests, discussion, etc.