Research
Full list of publicationsRecent Projects & Topics |
This is an ongoing project with my advisor, Markus Dickinson. I plan to extend this work as my dissertation. The work involves a set of stimuli and a small corpus of native and non-native English speaker sentences, and I am working on getting the proper permissions to make the corpus available online. Shallow Semantic Analysis of Interactive Learner Sentences. Focusing on applications for analyzing learner language which evaluate semantic appropriateness and accuracy, we collect data from a task which models some aspects of interaction, namely a picture description task (PDT). We parse responses to the PDT into dependency graphs with an an off-the-shelf parser, then use a decision tree to classify sentences into syntactic types and extract the logical subject, verb, and object, finding 92% accuracy in such extraction. We also process native speaker responses to the task in this same way, and use the resulting semantic triples as a gold standard for the automatic evaluation of the learner responses, with some degree of success. The specific goal of the initial paper is to examine the challenges involved in extracting these simple semantic representations from interactive learner sentences. Leveraging Known Semantics for Spelling Correction. Focusing on applications for analyzing learner language which evaluate semantic appropriateness and accuracy, we build from previous work which modeled some aspects of interaction, namely a picture description task (PDT), with the goal of integrating a spelling correction component in this context. After parsing a sentence and extracting semantic relations, a surprising number of analysis failures stem from misspellings, deviating from expected input in ways that can be modeled when the content of the interaction is known. We thus explore the use of spelling correction tools and language modeling to correct misspellings that often lead to errors in obtaining semantic forms, and we show that such tools can significantly reduce the number of unanalyzable cases. The work is useful for any context where image descriptions or some expected content is available, but not necessarily expected linguistic forms. Word-level language identification in The Chymistry of Isaac Newton (with Sandra Kübler and Wallace Hooper). In this paper, we introduce the task of word-based language identification in multilingual texts, in which every word needs to be classified with regard to its language. This task is necessary for multilingual texts in which language switches can occur within sentences, often more than once, as is the case in the texts in The Chymistry of Isaac Newton collection. We present a novel method based on character n-grams in combination with a weighting scheme that allows us to model the probability of language switches at different points in sentences. This method reaches the highest accuracy of 89.94% when 5-grams are used. The IUCL+ System: Word-Level Language Identification via Extended Markov Models (with Eric Baucom, Timur Gilmanov, Sandra Kübler and Daniel Whyatt). We describe the IUCL+ system for the shared task of the First Workshop on Computational Approaches to Code Switching, in which participants were challenged to label each word in Twitter texts as a named entity or one of two candidate languages. Our system combines character n-gram probabilities, lexical probabilities, word label transition probabilities and existing named entity recognition tools within a Markov model framework that weights these components and assigns a label. Our approach is language-independent, and we submitted results for all data sets (five test sets and three "surprise" sets, covering four language pairs), earning the highest accuracy score on the tweet level on two language pairs (Mandarin-English, Arabic-dialects 1 & 2) and one of the surprise sets (Arabic-dialects). With Eric Baucom and Sandra Kübler, I've been exploring methods for taking parsers trained on newspaper text and improving their performance on (transcribed) dialogs, when dialog training data is not available. In the first paper, we used constituency parsers, and we are currently performing similar experiments with dependency parsers. Note that my portion of this work mainly consists of the co-training experiments. Domain Adaptation for Parsing. We compare two different methods in domain adaptation applied to constituent parsing: parser combination and co- training, each used to transfer information from the source domain of news to the tar- get domain of natural dialogs, in a set- ting without annotated data. Both methods outperform the baselines and reach similar results. Parser combination profits most from the large amounts of training data combined with a robust probability model. Co-training, in contrast, relies on a small set of higher quality data. |
Programs |
I don't have much on GitHub yet, but I'm in the process of moving more of my current and completed work there. |