Speech Recognition Engineer (previously Computational Linguist)
- Integrated class-based and ngram language modelling in a WFST architecture, for improved speech recognition of classes of English words
- Maintained code to domain-adapt (at the word level) a generic English speech recognition model to companies on-the-fly based on metadata
- Expanded the scope of this domain adaptation to the user level, increasing the recall of name transcription by 15% relative, without negatively impacting WER.
- Optimized the training algorithm for this domain adaptation, enabling a 288x speedup (including parallelizing with Kubeflow).
- Maintained multi-dialectal English lexica through manual review of entries; wrote a syllabifier to auto-flag entries that violate English syllable structure
- Scoped speech recognition and scaling for languages other than English and made a long-term plan for the company, weighing linguistic considerations, technical and data limitations, and financial constraints
- Spearheaded the creation of a policy about inclusive language at the company, both technical and otherwise
- Organized a computational linguistics olympiad for the data science teams
- Maintained linting, pre-commit, CI/CD code for automated deploys of packages and Docker images.
- Vowels (vowels in phonetics, phonology, acoustics and speech recognition)
- Ethics in NLP (introduction to concepts in ethical NLP)
- Catchy Title Goes Here: A metapresentation (how to make better presentations)
- Flagging Mispronunciations (introduction to phonology and syllabification)
- Intermediate RegEx Practice (writing effective regular expressions in multiple languages)
- Technical Changes and User Impact (how to communicate better about technical changes)
- How To: ARPABET (introduction to phonemic transcription using ARPABET)
- Git (introduction and intermediate topics to version control with git)
- RST workshop (writing good documentation in reStructured Text)
- Intro to observability
This project seeks to quantify gender bias in the media by counting the proportion of people quoted in news text who are women. To do so, we developed a system that analyzes articles published online by certain English-language Canadian news outlets, using parts we built in combination with off-the-shelf NLP tools.
Nota bene: This project uses gender analysis that has some notable flaws. Although we have three categories of gender—male, female and other—the 'other' category lumps together cases where the speaker's gender is unknown as well as where the speaker's gender is known to fall outside the binary. Furthermore, our system depends on databases that encode binary gender (which erases nonbinary people) and on other NLP systems, e.g., for coreference analysis, that are known to have significant performance gaps for resolving instances of singular they and for English neopronouns (ze/hir, xe/xyr, etc.). We have manual overrides for (famous) people whom we know are systematically misgendered by these systems we depend on, but for most cases, we make probabilistic guesses at a person's gender based on their name - a problematic and cissexist practice. This impacts the accuracy of all reported results and we discuss these ethical considerations and more in our paper.
Gender is fluid and personal, and therefore "gender recognition" as a concept is impossible to do. At the same time, equitable representation in the media is important, and all previous manual attempts at quantifying this gender gap in media quotes have required orders of magnitude more time and labour, while still suffering from many of the same assumptions (e.g., correlating certain names with a gender, assuming pronouns unambiguously tell you someone's gender, etc.). Self-reported gender for speakers quoted in newspaper articles is not information that reporters collect, though this would be how to do this research in a perfect world.
If you are considering doing similar work or just generally examining gender as a variable, I encourage you to look at the following references: this blog post on trans-inclusive AI, this excellent talk by Kirby Conrod on How To Do Things With Gender, the Cao-Daumé paper on gender-inclusive coreference resolution which includes some great examples and new datasets, and this paper by Brian Larson on gender as a variable in NLP, informed by the Belmont Report. Look up Kirby Conrod, Os Keyes, Lauren Ackerman and Ártemis López. For gender and facial recognition, read this paper evaluating commercial facial recognition technologies, and this Os Keyes paper about the impact of gender recognition using facial analysis on trans folks. Also follow this guide for HCI researchers on how to do better with gender on surveys.
- Paper: The Gender Gap Tracker: Using Natural Language Processing to measure gender bias in media (2021). PLoS ONE 16(1). Fatemeh Torabi Asr, Mohammad Mazraeh, Alexandre Lopes, Vasundhara Gautam, Junette Gonzales, Prashanth Rao, and Maite Taboada.
- Website: Gender Gap Tracker
- Article: Tracking the gender gap in Canadian media (2019). Maite Taboada, Fatemeh Torabi Asr.
- Press: The future of women's voices in media: an interview with Gender Gap Tracker's lead SFU researcher (2019).
- Talk and panel discussion: Tracking the Gender Gap in Canadian Media - Taking a Big Data Approach (2019). SFU Big Data Hub. I was a panelist.
- Code: Gender Gap Tracker GitHub repository
This project analyzed millions of online comments on English news articles to examine toxicity, constructiveness, and content in this genre and their interplay. Our results paint a more positive picture of the internet than our lay perceptions because it turns out that moderated comment sections are quite constructive. Our findings yield insights for news outlets to moderate comments and advise commenters, as well as for internet citizens looking to have more meaningful interactions with others online.
- Technical report (PDF): Constructiveness and Toxicity in Online News Comments (2019). Vasundhara Gautam and Maite Taboada.
- Article: Tyee Commenters Assessed (2019). Vasundhara Gautam and Maite Taboada.
- Radio: Recent Study Finds Online Comments Surprisingly Constructive (2019). Redeye, Vancouver Cooperative Radio, CFRO 100.5FM. Vasundhara Gautam.
- Poster (PDF): Automatic comment moderation: topics and toxicity in online news (2019). SFU Linguistics Poster Session. Vasundhara Gautam and Maite Taboada.
- Talk: Content moderation in social media (2018). SFU Public Discourse and Data Science Symposium. Maite Taboada and Vasundhara Gautam.
- Talk: Automating Comment Moderation: Topics and Toxicity in Online News (2019). Undergraduate Research Symposium. Vasundhara Gautam.
- Code: TACT GitHub repository
This project aimed to create a semantic classification of the [adverb-ly adjective] construction - an adjective modified by an adverb ending in 'ly'. I used R and Python to computationally extract, count and classify instances of these constructions in large corpora (COCA, CORE, SOCC, CMU Movie$, Cornell movie reviews, NYT reviews)
- Paper: Hilariously ridiculous and other adverb-adjective combinations: Classification and frequency distribution across registers (2018). 4th Meeting of the American Pragmatics Association. Cliff Goddard, Maite Taboada and Rada Trnavac.
- Poster: The English [adverb-ly adjective] construction: Classification and distribution across corpora and registers (2019). Canadian Linguistics Association. Cliff Goddard, Maite Taboada and Rada Trnavac.
- Code: Adverbly Adjectives GitHub repository
Software Engineer (Intern)
- Developed software for Predix App Engine, a platform for the Industrial Internet of Things - added features, resolved bugs, wrote automated tests, and maintained documentation.
- Led the effort to write a comprehensive technical Installation Guide for on-premise installation of our multi-node, multi-service platform, which required broad knowledge of the platform services, networking, system administration and security.
- Helped overhaul the software architecture of authentication and access control for the product using Cloud Foundry User Account and Authentication (UAA).
- Worked in a group led by Dr Ashley Farris-Trimble on several projects aimed at studying phonological processing by using eye-tracking experiments in the visual world paradigm
- Curated visual and auditory stimuli, created word lists, programmed the experiments with SR Research Experiment Builder and MATLAB
- Used an eye-tracker and an audiometer to run participants in experiments
- Organized and ran linguistics-themed day camps to recruit children for studies