Dissertations / Theses on the topic 'Science of language (Linguistics)'

This dissertation describes an approach to automatically detecting and correcting grammatical errors in text produced by Korean language learners. Specifically, we focus on Korean particles, which have a range of functions including case marking and indicate properties similar to English prepositions. There are two main goals for this research: to establish reliable data sources that can serve as a foundation for Korean language learning research endeavors, and to develop an accurate error detection system. The machine learning-based system is built to detect errors of particle omission and substitution, then to select the best particle to produce grammatical output. The resources and results outlined in this work should prove useful in aiding other researchers working on Korean error detection and in moving the field one step closer to robust multi-lingual methods.

I investigated linguistic patterns in middle school students' writing to understand their relevant embodied experiences for learning science. Embodied experiences are those limited by the perceptual and motor constraints of the human body. Recent research indicates student understanding of science needs embodied experiences. Recent emphases of science education researchers in the practices of science suggest that students' understanding of systems and their structure, scale, size, representations, and causality are crosscutting concepts that unify all scientific disciplinary areas. To discern the relationship between linguistic patterns and embodied experiences, I relied on Cognitive Linguistics, a field within cognitive sciences that pays attention to language organization and use assuming that language reflects the human cognitive system. Particularly, I investigated the embodied experiences that 268 middle school students learning about water brought to understanding: i) systems and system structure; ii) scale, size and representations; and iii) causality. Using content analysis, I explored students' language in search of patterns regarding linguistic phenomena described within cognitive linguistics: image schemas, conceptual metaphors, event schemas, semantical roles, and force-dynamics. I found several common embodied experiences organizing students' understanding of crosscutting concepts. Perception of boundaries and change in location and perception of spatial organization in the vertical axis are relevant embodied experiences for students' understanding of systems and system structure. Direct object manipulation and perception of size with and without locomotion are relevant for understanding scale, size and representations. Direct applications of force and consequential perception of movement or change in form are relevant for understanding of causality. I discuss implications of these findings for research and science teaching.

Most machine learning algorithms require a fixed length input to be able to perform commonly desired tasks such as classification, clustering, and regression. For natural language processing, the inherently unbounded and recursive nature of the input poses a unique challenge when deriving such fixed length representations. Although today there is a general consensus on how to generate fixed length representations of individual words which preserve their meaning, the same cannot be said for sequences of words in sentences, paragraphs, or documents. In this work, we study the encoders commonly used to generate fixed length representations of natural language sequences, and analyze their effectiveness across a variety of high and low level tasks including sentence classification and question answering. Additionally, we propose novel improvements to the existing Skip-Thought and End-to-End Memory Network architectures and study their performance on both the original and auxiliary tasks. Ultimately, we show that the setting in which the encoders are trained, and the corpus used for training, have a greater influence of the final learned representation than the underlying sequence encoders themselves.

This thesis presents an oral history of the Advocates for Indigenous California Language Survival (AICLS) and its Breath of Life Workshop. Held every other year since 1996, the workshop is designed to meet the language revitalization needs of California Indian people whose languages have no living fluent speakers. Breath of Life Workshop organizers arrange visits to four archives on the University of California, Berkeley, campus and connect participants with linguistic mentors to read and interpret archival documents in their language for the purpose of bringing their language back into use.

Through interviews with AICLS founders, Breath of Life Workshop participants, and University of California, Berkeley, linguists and archivists, this study uncovers the role archivists play in the Breath of Life Workshops and in the care of Native language collections more generally. Topics addressed include the selection and use of archival documents in the program and the changes to archival practice and policies that have resulted from archivists’ work with Breath of Life participants. The thesis also examines issues involved in the collection, arrangement, description, preservation, and access to the documentation of California Indian languages. The study concludes with recommendations for future language revitalization programs.

For every participant role filler in an utterance, speakers must choose to leave it bare (e.g., "the interviewer") or to modify it (e.g., "the interviewer on Fresh Air"). Their decision is the end result of a combination of complex factors ranging from the original message to how distracted the speaker is. When we use corpora to create language models, part of our job is understanding the observable properties in and around an event description that allow us to predict these decisions. A considerable body of work on language production and discourse pragmatics concentrates on measuring noun phrase predictability and other forms of shared knowledge that help determine the balance point between over- and under-specification of a participant role filler. Although the importance of predictability as measured by long-term probabilities has long been recognized, I present a novel quantitative analysis of participant role filler predictability, the structure of the mental lexicon, and how the interaction of these two inform a speaker's internal perception of informativity. Standard Gricean assumptions tend to be efficiency oriented. Speakers will be informative enough but not wastefully so. Using these to model corpus distributions predict that noun phrase modification rates are directly proportional to predictability in order to satisfy the speaker's obligation to always be informative. In contrast, standard Firthian models (built around the idea that "you know a word by the company it keeps") assume spreading activation—and not efficiency—is the dominant predictor of usage. Sensitivity to activation's effect predicts that noun phrase modification rates are inversely proportional to predictability. Strongly connected participant role fillers could be easily activated for production while weakly connected participant role fillers would either be mentioned less often or themselves trigger strongly connected features (not normally associated with the head verb) to be primed for production.

To distinguish between these competing assumptions, I analyze participant role filler modification rates in event descriptions with respect to three indicators: the syntactic and semantic optionality of the role filler, the general predictability of the verb's role fillers, and the predictability of individual pairs of verb/participant role fillers. First, I use insights from linguistic theory to classify verbs and their participant roles into classes of syntactic optionality and semantic optionality. Second, I quantify over a large corpus the general predictability of a verb's participant roles and the specific predictability of each pair of verb/participant role filler. Finally, I model the relationship between the three indicators and modification in order to ascertain whether speakers have a stronger tendency to modify the more predictable participant role fillers, as Grice's Maxim of Relevancy predicts, or a tendency to modify the less predictable participant role fillers, as a Firthian activation-based model predicts.

I present descriptive statistical models to chart the relationship between predictability, syntactic optionality of a participant role, and semantic optionality of a participant role. In general, verb classes with stronger mental lexicon connections to their participant role fillers according to theory also have more predictable participant role fillers in the British National Corpus. Specifically, syntactically optional direct object verbs and semantically obligatory instrument verbs have more predictable participant role fillers than the opposite, comparable verb class. I also present several linear mixed-effect models to determine how predictive of modification the independent variables of syntactic verb class, semantic verb class, and verb/participant role filler predictability are. According to these models, speakers are significantly more likely to modify the less predicted participant role fillers even when taking into account individual verb and verb class differences. I conclude that mental lexicon accessibility modulates noun phrase realization according to a Firthian activation-based model. For each factor, I discuss possible explanations for the correlations between modification, predictability, and optionality and how these correlations make sense within a larger production model.

Online reviews have become a popular method for consumers to express personal evaluation about products. Ecommerce firms have invested heavily into review systems because of the impact of product reviews on product sales and shopping behavior. However, the usage of product reviews is undermined by the increasing appearance of shill or fake reviews. As initial steps to deter and detect shill reviews, this study attempts to understand characteristics of shill reviews and influences of shill reviews on product quality and shopping behavior. To reveal the linguistic characteristics of shill reviews, this study compares shill reviews and normal reviews on informativeness, readability and subjectivity level. The results show that these features can be used as reliable indicators to separate shill reviews from normal reviews. An experiment was conducted to measure the impact of shill reviews on perceived product quality. The results showed that positive shill reviews significantly increased quality perceptions of consumers for thinly reviewed products. This finding provides strong evidence about the risks of shill reviews and emphasizes the need to develop effective detection and prevention methods.