Computational Linguistics Commons^™

From Sesame Street To Beyond: Multi-Domain Discourse Relation Classification With Pretrained Bert, Isaac R. Raff

Dissertations, Theses, and Capstone Projects

Research efforts in transfer learning have gained massive popularity in recent years. Pretrained language models have demonstrated the most successful results in producing high quality neural networks capable of quality inference after training across domains via transfer learning. This study expands on the domain transfer introduced in \cite{ferracane-etal-2019-news} exploring neural methods for transfer learning of discourse parsing between a news source domain and a medical target domain. \cite{ferracane-etal-2019-news} specifically discuss transfer learning from news articles to PubMed medical journal articles. Experiments in transfer learning in the current work expand to include three domains: Wall Street Journal articles previously annotated with …

Predicting Stress In Russian Using Modern Machine-Learning Tools, John Schriner

Dissertations, Theses, and Capstone Projects

In the Russian language, stress on a word is determined via often complex patterns and rules. In this paper, after examining nearly a century of research in stress rules and methods in Russian, we turn to see if modern machine learning tools can aid in predicting stress. Using A.A. Zaliznyak’s dictionary grammar and over 300,000 word forms, we derived stress codes to aid in predicting which syllable primary stress falls on. We trained an LSTM neural network on the data and conducted eight experiments with added features such as lemma, part of speech, and morphology. While the model performed better …

A Machine Learning Approach To Text-Based Sarcasm Detection, Lara I. Novic

Dissertations, Theses, and Capstone Projects

Sarcasm and indirect language are commonplace for humans to produce and recognize but difficult for machines to detect. While artificial intelligence can accurately analyze sentiment and emotion in speech and text, it may struggle with insincere and sardonic content, although it is possible to train a machine to identify uttered and written sarcasm. This paper aims to detect sarcasm using logistic regression and a support vector machine (SVM) and compare their results to a baseline.

The models are trained on headlines from a Kaggle dataset containing headlines from the satirical news website The Onion and serious news website Huffpost (formerly …

Covert Determiners In Appalachian English Narrative Declarative Sentences, William Oliver

Dissertations, Theses, and Capstone Projects

In this thesis, I explore the syntax and semantics of covert determiners (Ds) in matrix subject determiner phrases (DPs) with definite specific interpretations. To conduct my investigation, I used the Audio-Aligned and Parsed Corpus of Appalachian English (AAPCAppE), a million-word Penn Treebank corpus, and the software CorpusSearch, a Java program that searches Penn Treebank corpora. My research shows that Appalachian English contains a linguistic phenomenon where speakers drop the D, replacing overt Ds with covert Ds, in definite specific DPs. For example, where Standard English speakers say The doctor came by horseback, Appalachian speakers may use a covert D …

From An Art To A Science: Features And Methodology In Computational Authorship Identification, Jonathan I. Manczur

Dissertations, Theses, and Capstone Projects

Nearly thirty years ago, the United States Supreme Court revaluated the criteria for accepting forensic science and expert testimony, challenging Forensic Linguistics to assert itself as a reputable science. Much work has been produced in the interim to that end, but much still needs to be accomplished to satisfy the judicial standards. Computational linguistics has the potential to provide that necessary analytical framework. This paper’s intent is two-fold. First, there are two competing theories on the proper features necessary to identify an unknown author. Four features were drawn from the syntactic computational linguistics tradition and four from computational stylometry to …

Detection And Morphological Analysis Of Novel Russian Loanwords, Yulia Spektor

Dissertations, Theses, and Capstone Projects

This paper investigates recent English loanwords in Russian and explores ways in which computational methods can help further theoretical research. The goal of the study is two-fold: to find new, previously unattested loanwords borrowed over the last decade and to examine the rate of adaptation of the new borrowings, attested by the degree to which they conform to the constraints of the Russian language. First, we train a finite-state pipeline that combines character n-gram language models, which encode phonotactic and lexical properties of loanwords, with a binary classifier to detect loanwords. The model achieves state-of-the-art performance results during evaluation, surpassing …

Label Imputation For Homograph Disambiguation: Theoretical And Practical Approaches, Jennifer M. Seale

Dissertations, Theses, and Capstone Projects

This dissertation presents the first implementation of label imputation for the task of homograph disambiguation using 1) transcribed audio, and 2) parallel, or translated, corpora. For label imputation from parallel corpora, a hypothesis of interlingual alignment between homograph pronunciations and text word forms is developed and formalized. Both audio and parallel corpora label imputation techniques are tested empirically in experiments that compare homograph disambiguation model performance using: 1) hand-labeled training data, and 2) hand-labeled training data augmented with label-imputed data. Regularized, multinomial logistic regression and pre-trained ALBERT, BERT, and XLNet language models fine-tuned as token classifiers are developed for homograph …

Predicting Stock Price Movements Using Sentiment And Subjectivity Analyses, Andrew Kirby

Dissertations, Theses, and Capstone Projects

In a quick search online, one can find many tools which use information from news headlines to make predictions concerning the trajectory of a given stock. But what if we went further, looking instead into the text of the article, to extract this and other information? Here, the goal is to extract the sentence in which a stock ticker symbol is mentioned from a news article, then determine sentiment and subjectivity values from that sentence, and finally make a prediction on whether or not the value of that stock will go up or not in a 24-hour timespan. Bloomberg News …

The Public Innovations Explorer: A Geo-Spatial & Linked-Data Visualization Platform For Publicly Funded Innovation Research In The United States, Seth Schimmel

Dissertations, Theses, and Capstone Projects

The Public Innovations Explorer (https://sethsch.github.io/innovations-explorer/app/index.html) is a web-based tool created using Node.js, D3.js and Leaflet.js that can be used for investigating awards made by Federal agencies and departments participating in the Small Business Innovation Research (SBIR) and Small Business Technology Transfer (STTR) grant-making programs between 2008 and 2018. By geocoding the publicly available grants data from SBIR.gov, the Public Innovations Explorer allows users to identify companies performing publicly-funded innovative research in each congressional district and obtain dynamic district-level summaries of funding activity by agency and year. Applying spatial clustering techniques on districts' employment levels across major economic sectors provides users …

A Computational Study In The Detection Of English–Spanish Code-Switches, Yohamy C. Polanco

Dissertations, Theses, and Capstone Projects

Code-switching is the linguistic phenomenon where a multilingual person alternates between two or more languages in a conversation, whether that be spoken or written. This thesis studies the automatic detection of code-switching occurring specifically between English and Spanish in two corpora.

Twitter and other social media sites have provided an abundance of linguistic data that is available to researchers to perform countless experiments. Collecting the data is fairly easy if a study is on monolingual text, but if a study requires code-switched data, this becomes a complication as APIs only accept one language as a parameter. This thesis focuses on …

When Misclassification Is Misgendering: Gender Prediction In The Context Of Trans Identities, Sean Miller

Dissertations, Theses, and Capstone Projects

As a subdomain of author profiling, gender prediction (sometimes called gender inference) has received a substantial amount of attention—both as a task in itself, and for other downstream analyses. Throughout the existing literature various statistical and machine learning methods have been applied to extract features in order to either characterize and differentiate female and male writing styles, or simply to achieve maximum accuracy on gender prediction as a binary classification task. However, researchers often do not disclose how they conceptualize gender nor do they consider the implications that gender prediction has for non-binary and trans individuals. Along with an overview …

Mitigating Gender Bias In Neural Machine Translation Using Counterfactual Data, Alan Wong

Dissertations, Theses, and Capstone Projects

Recent advances in deep learning have greatly improved the ability of researchers to develop effective machine translation systems. In particular, the application of modern neural architectures, such as the Transformer, has achieved state-of-the-art BLEU scores in many translation tasks. However, it has been found that even state-of-the-art neural machine translation models can suffer from certain implicit biases, such as gender bias (Lu et al., 2019). In response to this issue, researchers have proposed various potential solutions: some have proposed approaches that inject missing gender information into models, while others have attempted modifying the training data itself. We focus on mitigating …

Does The Word "Chien" Bark? Representation Learning In Neural Machine Translation Encoders, Emily Campbell

Dissertations, Theses, and Capstone Projects

This thesis presents experiments with using representation learning to explore how neural networks learn. Neural networks which take text as input create internal representations of the text during their training. Recent work has found that these representations can be used to perform other downstream linguistic tasks, such as part-of-speech (POS) tagging. This demonstrates that the neural networks are learning linguistic information and storing this information in the representations. We focus on the representations created by neural machine translation (NMT) models and whether they can be used in POS tagging. We train 5 NMT models including an auto-encoder. We extract the …

Doing Away With Defaults: Motivation For A Gradient Parameter Space, Katherine Howitt

Dissertations, Theses, and Capstone Projects

In this thesis, I propose a reconceptualization of the traditional syntactic parameter space of the principles and parameters framework (Chomsky, 1981). In lieu of binary parameter settings, parameter values exist on a gradient plane where a learner’s knowledge of their language is encoded in their confidence that a particular parametric target value, and thus grammatical construction of an encountered sentence, is likely to be licensed by their target grammar. First, I discuss other learnability models in the classic parameter space which lack either psychological plausibility, theoretical consistency, or some combination of the two. Then, I argue for the Gradient Parameter …

Inferring Research Fields In Administrative Records Using Text Data, Ekaterina Levitskaya

Dissertations, Theses, and Capstone Projects

The UMETRICS database (Universities: Measuring the Effects of Research on Innovation, Competitiveness, and Science) contains rich information on grants from sponsored federal and non-federal research for 32 universities over a 15-year period. It is hosted at IRIS (Institute for Research on Innovation and Science, University of Michigan) and serves as a rich source of university administrative data; however, it does not contain information on research fields. Categorizing grants data by research field can help to measure results of investment in research and science and provide evidence for the data-driven policy-making; yet administrative data often lacks this type of categorization. In …

Genderlects In Social Media, Alina Korovatskaya

Dissertations, Theses, and Capstone Projects

Many studies have found significant differences in ways men and women use language; some argue that these differences occur as a result of culture differences, and others suggest that they are influenced by differences in social status and power between the genders. However, some of the major studies were concluded decades ago and do not reflect changes in gender relations in recent years. In this study, we analyze modern conversations using two social media platforms, Twitter and Reddit, to determine whether substantial differences between men and women’s use of language were preserved between the genders.

Phonologically-Informed Speech Coding For Automatic Speech Recognition-Based Foreign Language Pronunciation Training, Anthony J. Vicario

Dissertations, Theses, and Capstone Projects

Automatic speech recognition (ASR) and computer-assisted pronunciation training (CAPT) systems used in foreign-language educational contexts are often not developed with the specific task of second-language acquisition in mind. Systems that are built for this task are often excessively targeted to one native language (L1) or a single phonemic contrast and are therefore burdensome to train. Current algorithms have been shown to provide erroneous feedback to learners and show inconsistencies between human and computer perception. These discrepancies have thus far hindered more extensive application of ASR in educational systems.

This thesis reviews the computational models of the human perception of American …

Computational Approaches To The Syntax–Prosody Interface: Using Prosody To Improve Parsing, Hussein M. Ghaly

Dissertations, Theses, and Capstone Projects

Prosody has strong ties with syntax, since prosody can be used to resolve some syntactic ambiguities. Syntactic ambiguities have been shown to negatively impact automatic syntactic parsing, hence there is reason to believe that prosodic information can help improve parsing. This dissertation considers a number of approaches that aim to computationally examine the relationship between prosody and syntax of natural languages, while also addressing the role of syntactic phrase length, with the ultimate goal of using prosody to improve parsing.

Chapter 2 examines the effect of syntactic phrase length on prosody in double center embedded sentences in French. Data collected …

Ghost Peppers: Using Ensemble Models To Detect Professor Attractiveness Commentary On Ratemyprofessors.Com, Angie Waller

Dissertations, Theses, and Capstone Projects

In June 2018, RateMyProfessors.com (RMP), a popular website for students to leave professor reviews, removed a controversial feature known as the “chili pepper” which allowed students to rate their professors as “hot” or “not hot.” Though past research has rigorously analyzed the correlation of the chili pepper with higher ratings in other categories (Felton, Mitchell, and Stinson, 2004; Felton et al., 2008), none has measured the effect of the removal of the chili pepper on the text content submitted by students. While it is a positive step that the chili pepper has been removed, text commentary on teacher attractiveness persists …

Demographic Factors As Domains For Adaptation In Linguistic Preprocessing, Sara Morini

Dissertations, Theses, and Capstone Projects

Classic natural language processing resources such as the Penn Treebank (Marcus et al. 1993) have long been used both as evaluation data for many linguistic tasks and as training data for a variety of off-the-shelf language processing tools. Recent work has highlighted a gender imbalance in the authors of this text data (Garimella et al. 2019) and hypothesized that tools created with such resources will privilege users from particular demographic groups (Hovy and Søgaard 2015). Domain adaptation is typically employed as a strategy in machine learning to adjust models trained and evaluated with data from different genres. However, the present …

Do It Like A Syntactician: Using Binary Gramaticality Judgements To Train Sentence Encoders And Assess Their Sensitivity To Syntactic Structure, Pablo Gonzalez Martinez

Dissertations, Theses, and Capstone Projects

The binary nature of grammaticality judgments and their use to access the structure of syntax are a staple of modern linguistics. However, computational models of natural language rarely make use of grammaticality in their training or application. Furthermore, developments in modern neural NLP have produced a myriad of methods that push the baselines in many complex tasks, but those methods are typically not evaluated from a linguistic perspective. In this dissertation I use grammaticality judgements with artificially generated ungrammatical sentences to assess the performance of several neural encoders and propose them as a suitable training target to make models learn …

Analyzing Prosody With Legendre Polynomial Coefficients, Rachel Rakov

Dissertations, Theses, and Capstone Projects

This investigation demonstrates the effectiveness of Legendre polynomial coefficients representing prosodic contours within the context of two different tasks: nativeness classification and sarcasm detection. By making use of accurate representations of prosodic contours to answer fundamental linguistic questions, we contribute significantly to the body of research focused on analyzing prosody in linguistics as well as modeling prosody for machine learning tasks. Using Legendre polynomial coefficient representations of prosodic contours, we answer prosodic questions about differences in prosody between native English speakers and non-native English speakers whose first language is Mandarin. We also learn more about prosodic qualities of sarcastic speech. …

The Perception Of Mandarin Tones In "Bubble" Noise By Native And L2 Listeners, Mengxuan Zhao

Dissertations, Theses, and Capstone Projects

Previous studies have revealed the complexity of Mandarin Tones. For example, similarities in the pitch contours of tones 2 and 3 and tones 3 and 4 cause confusion for listeners. The realization of a tone's contour is highly dependent on its context, especially the preceding pitch. This is known as the coarticulation effect. Researchers have demonstrated the robustness of tone perception by both native and non-native listeners, even with incomplete acoustic information or in noisy environment. However, non-native listeners were observed to behave differently from native listeners in their use of contextual information. For example, the disagreement between the end …

Obfuscating Authorship: Results Of A User Study On Nondescript, A Digital Privacy Tool, Robin Camille Davis

Publications and Research

For those who write anonymously, particularly for safety reasons, authorship attribution poses a threat. Nondescript, my web app, guides writers in achieving stylometric obfuscation in order to preserve anonymity. The app runs simulations of authorship attribution scenarios by analyzing the user’s linguistic features. In this paper, I will describe the conception of the Nondescript app; discuss related work; and present the results of a user study. Most users in the study were able to anonymize their writing in at least 5 out of 10 authorship attribution scenarios. Users rated the anonymization process an average of 3.6 out of 5 in …

Generative Adversarial Networks And Word Embeddings For Natural Language Generation, Robert D. Schultz Jr

Dissertations, Theses, and Capstone Projects

We explore using image generation techniques to generate natural language. Generative Adversarial Networks (GANs), normally used for image generation, were used for this task. To avoid using discrete data such as one-hot encoded vectors, with dimensions corresponding to vocabulary size, we instead use word embeddings as training data. The main motivation for this is the fact that a sentence translated into a sequence of word embeddings (a “word matrix”) is an analogue to a matrix of pixel values in an image. These word matrices can then be used to train a generative adversarial model. The output of the model’s generator …

Recursive Neural Networks For Semantic Sentence Representation, Liam S. Geron

Dissertations, Theses, and Capstone Projects

Semantic representation has a rich history rife with both complex linguistic theory and computational models. Though this history stretches back almost 50 years (Salton, 1971), recently the field has undergone an unexpected shift in paradigm thanks to the work of Mikolov et al., 2013(a & b) which has proven that vector-space semantic models can capture large amounts of semantic information. As of yet, these semantic representations are computed at the word level, and finding a semantic representation of a phrase is a much more difficult challenge. Mikolov et al., 2013(a&b) proved that their word vectors can be composed arithmetically to …

Multimodal Depression Detection: An Investigation Of Features And Fusion Techniques For Automated Systems, Michelle Renee Morales

Dissertations, Theses, and Capstone Projects

Depression is a serious illness that affects a large portion of the world’s population. Given the large effect it has on society, it is evident that depression is a serious health issue. This thesis evaluates, at length, how technology may aid in assessing depression. We present an in-depth investigation of features and fusion techniques for depression detection systems. We also present OpenMM: a novel tool for multimodal feature extraction. Lastly, we present novel techniques for multimodal fusion. The contributions of this work add considerably to our knowledge of depression detection systems and have the potential to improve future systems by …

Intergroup Variability In Personality Recognition, Arundhati Sengupta

Dissertations, Theses, and Capstone Projects

Automatic Identification of personality in conversational speech has many applications in natural language processing such as leader identification in a meeting, adaptive dialogue systems, and dating websites. However, the widespread acceptance of automatic personality recognition through lexical and vocal characteristics is limited by the variability of error rate in a general purpose model among speakers from different demographic groups. While other work reports accuracy, we explored error rates of automatic personality recognition task using classification models for different genders and native language groups (L1). We also present a statistical experiment showing the influence of gender and L1 on the relation …

Describing Doggo-Speak: Features Of Doggo Meme Language, Jennifer Bivens

Dissertations, Theses, and Capstone Projects

Doggo-speak is a specialized way of writing most commonly associated with captions on Doggo memes, humorous images of dogs shared in online communities. This paper will explore linguistic features of Doggo-speak through analysis of social media posts by Doggo fan pages. It will use the discussed features as inputs to five machine learning classifiers and will show, through this classification task, that the discussed features are sufficient for distinguishing between Doggo-speak and more general English text.

Speech Perception In “Bubble” Noise: Korean Fricatives And Affricates By Native And Non-Native Korean Listeners, Jiyoung Choi

Dissertations, Theses, and Capstone Projects

The current study examines acoustic cues used by second language learners of Korean to discriminate between Korean fricatives and affricates in noise and how these cues relate to those used by native Korean listeners. Stimuli consist of naturally-spoken consonant-vowel-consonant-vowel (CVCV) syllables: /sɑdɑ/, /s*ɑdɑ/, /tʃɑdɑ/, /tʃ^hɑdɑ/, and /tʃ*ɑdɑ/. In this experiment, the “bubble noise” methodology of Mandel at al. (2016) was used to identify the time-frequency locations of important cues in each utterance, i.e., where audibility of the location is significantly correlated with correct identification of the utterance in noise. Results show that non-native Korean listeners can discriminate between …