portrait

Constantine Lignos

Email: c...@lignos.org

Details: ConstantineLignos on Github, @ConstantineLig, ResearchGate

My research explores the computational structure of language from a fundamentally interdisciplinary perspective. I work on projects in many domains, including traditional natural language processing, human language acquisition and processing, and language variation and change. I create simple, interpretable computational models by taking advantage of findings from theoretical linguistics, psycholinguistics, and language acquisition.

I did my graduate work in Computer Science at The University of Pennsylvania (Ph.D. 2013), advised by Mitch Marcus and Charles Yang. I then completed a post-doctoral fellowship at The Children's Hospital of Philadelphia exploring clinical applications of statisical models of language processing, working with Tim Roberts. Frequent collaborators include Jonathan Brennan, Kyle Gorman, Laurel MacKenzie, Hilary Prichard, and Vasu Raman.


Latest news

Chapter in Cambridge Handbook of Morphology

My chapter with Charles Yang, Morphology and language acquisition, will appear in The Cambridge Handbook of Morphology.

Blends research featured in TIME Magazine

My work with Hilary Prichard on word blends is featured in the July 2015 TIME Answers issue: Why Did ‘Frenemy’ Stick? Academics are unraveling the mystery behind the success–and failure–of blended words. While that article is only available to subscribers, you can read more about our work here: Quantifying cronuts: Predicting the quality of blends.

At LSA 2015

I'll be talking about my latest work on modeling word blends at the LSA 2015 Annual Meeting:
10:30-11:00, Broadway III/IV
Presentation: Quantifying cronuts: Predicting the quality of blends (with Hilary Prichard).


Projects

Unsupervised word and morpheme learning and child language acquisition

Representative publication: Infant word segmentation: An incremental, integrated model
In the first years of life, children learn to segment the speech stream into words and the morphemes they are built from. This problem is similar to the NLP task of unsupervised language learning, and my research in this area develops learning methods that are computationally efficient, cognitively plausible, and predict the ways in which infants learn over time.

Lexical and morphological processing

Representative publication: Revisiting frequency and storage in morphological processing
I use computational modeling to extend experimental work by evaluating theories of processing morphologically complex words using large scale data. My work focuses on the debate surrounding the role of frequency in facilitating processing and the nature of morphological decomposition.

Codeswitching

Representative poster: Toward web-scale analysis of codeswitching
Multilinguals are capable of mixing the languages they speak a process known as codeswitching. Codeswitching poses a problem for traditional NLP models, which typically assume that entire documents contain a single language and perform poorly when asked to identify the language of a short message such as a tweet, let alone individual words. My research in this area focuses on using linguistic insights to build simple, high-performance systems for identifying codeswitching in short forms such as social media, enabling both NLP applications and sociolinguistic research. I've used crowdsourced annotation to create a corpus of codeswitched tweets and am developing Codeswitchador, a software package for language identification and codeswitching.

Language variation and change

Representative poster: Examining extragrammatical effects on English auxiliary contraction
While we can often describe what changes in language, we are rarely able to explain how these changes occur. I develop computational models that provide and test explicit mechanisms of how successive generations of language acquisition can lead to change. I also study the language processing mechanisms that support language variation, modeling variation in linguistic corpora.

Human-robot interaction

Representative publication: Provably correct reactive control from natural language
In situations like search and rescue, we need robots to perform complicated tasks under limited supervision, a natural application for giving commands via natural language. As a part of the SUBTLE MURI project, I created SLURP, a system that allows users to specify simple tasks for a robot to perform using natural language by translating their instructions into a satisfiable linear temporal logic specification.


Software

MORSEL: a cognitively-motivated state-of-the-art unsupervised morphological analyzer I developed for Morpho Challenge 2010. It achieved state-of-the-art results in English and Finnish.

Codeswitchador: a system for identifying code-switching in social media data. This work enables the creation of large scale corpora of code-switching and identification of bilingual users. I developed this as a participant of the SCALE summer workshop at the Johns Hopkins Center of Excellence in Human Language Technology.

I'm working to release the code and data I've used for other papers. E-mail me if you need anything in particular in the meantime. Even if it isn't formally released, almost everything I work on is publicly available on GitHub.


Teaching

I teach researchers to write great Python code. The notes for the bootcamps I've done are available at Python Boot Camp for Researchers.

I maintain a list of common mistakes that programmers new to Python make: Anti-Patterns in Python Coding.

In Spring 2011 and 2012, I taught one of the CIS department's "mini-courses," Python Programming (CIS 192).

In the past I've also led some informal groups for learning Python. The slides from those groups can be found on my Python for Language Researchers Site.