Conferences, Lectures, & Seminars
Events for June
-
NL Seminar- Dirk Hovy: "Learning Whom to Trust with MACE(NAACL Practice Talk)"
Wed, Jun 05, 2013 @ 03:30 PM - 04:30 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Dirk Hovy, USC/ISI
Talk Title: "Learning Whom to Trust with MACE(NAACL Practice Talk)"
Series: Natural Language Seminar
Abstract: Non-expert annotation services like Amazon's Mechanical Turk (AMT) are cheap and fast ways to evaluate systems and provide categorical annotations for training data. Unfortunately, some annotators choose bad labels in order to maximize their pay. Manual identification is tedious, so we experiment with an item-response model. It learns in an unsupervised fashion to a) identify which annotators are trustworthy and b) predict the correct underlying labels. We match performance of more complex state-of-the-art systems and perform well even under adversarial conditions. We show considerable improvements over standard baselines, both for predicted label accuracy and trustworthiness estimates. We show that the latter can be further improved by introducing a prior on model parameters and using Variational Bayes inference. Additionally, we present a method for trading precision and recall, achieving even higher performance by focusing on the instances our model is most confident in. We provide an implementation of MACE (Multi- Annotator Competence Estimation) for download at(http://www.isi.edu/publications/licensed-sw/mace/).
Biography: Dirk Hovy is a recent PhD graduate from USC's Information Sciences Institute, working with Jerry Hobbs and Ed Hovy. He has a background in socio-linguistics. His current work includes unsupervised and semi-supervised sequential models of relation extraction and WSD, as well as annotator assessment. He has also worked on temporal relations, metaphors, and prosody. A full list of his publications can be found at(http://www.dirkhovy.com/portfolio/papers/index.php). His other interests include cooking, picking up heavy things (and putting them back down), and medieval art and literature.
Host: Qing Dou
More Info: http://nlg.isi.edu/nl-seminar/
Location: Information Science Institute (ISI) - Marina Del Rey-11th Flr Conf Rm # 1135
Audiences: Everyone Is Invited
Contact: Peter Zamar
Event Link: http://nlg.isi.edu/nl-seminar/
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
AI SEMINAR
Fri, Jun 07, 2013 @ 11:00 AM - 12:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Benjamin Snyder, Assistant Professor UW-Madison
Talk Title: Dead Languages, and Elbow Grease: Minimally Supervised NLP Models.
Series: AISeminar
Abstract: I will discuss new techniques for inducing accurate statistical models for low resource languages. In the extreme case of language decipherment, we are presented with a text with no knowledge of the language or writing system, and our goal is to identify the phonetic properties of the characters. I will present a Bayesian model that predicts whether letters are consonants or vowels with over 99% accuracy across 503 languages. The model assumes that languages are grouped into latent clusters with shared phonotactic regularities. We perform posterior inference over the identity and shared parameters of these clusters.
In a more common scenario, we are trying to build a model (e.g. for a low resource language) while minimizing our annotation effort. I will present a method based on matrix projections that allows us to quickly identify an optimal set of examples to label. This method outperforms active learning while obviating the need for incremental retraining and bootstrapping. We report error reductions of 25-40% on the tasks of pronunciation modeling and part-of-speech tagging.
Biography: Benjamin Snyder is an assistant professor of computer science at UW-Madison. He completed his PhD at MIT in 2010, receiving the ACM Dissertation Award honorable mention and the George M. Sprowls Award for best PhD thesis in computer science at MIT. He will be visiting ISI For all of June.
Host: David Chiang
Webcast: http://webcasterms1.isi.edu/mediasite/SilverlightPlayer/Default.aspx?peid=ec818feeeb5b458e87121185990150ef1dLocation: Information Science Institute (ISI) - 11th fl Large CR
WebCast Link: http://webcasterms1.isi.edu/mediasite/SilverlightPlayer/Default.aspx?peid=ec818feeeb5b458e87121185990150ef1d
Audiences: Everyone Is Invited
Contact: Alma Nava / Information Sciences Institute
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
NL Seminar- Malte Nuhn "Is Decipherment Difficult"
Fri, Jun 07, 2013 @ 03:00 PM - 04:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Malte Nuhn, Aachen University, Germany
Talk Title: "Is Decipherment Difficult"
Series: Natural Language Seminar
Abstract: Abstract: Is it possible to learn useful translations from large amounts of monolingual data to improve machine translation? The intuitive feeling is that learning a language without bilingual data is at least "more difficult than learning from example translations". In this talk, I will present recent results on decipherment: I will show that the decipherment problem is indeed difficult (NP-hard) and what approximations to the original problem can be made without hurting decipherment accuracy much.
Biography: Bio: Having studied Physics and Computer Science at RWTH Aachen University, I'm currently a PhD student at Prof. Ney's Human Language Technology and Pattern Recognition Group in Aachen. I'm particularly interested in applying decipherment techniques to improve machine translation.
Homepage:
http://www-i6.informatik.rwth-aachen.de/~nuhn/
Host: Kevin Knight and Qing Dou
More Info: http://nlg.isi.edu/nl-seminar/
Location: Information Science Institute (ISI) - Marina Del Rey-11th Flr Conf Rm # 1135
Audiences: Everyone Is Invited
Contact: Peter Zamar
Event Link: http://nlg.isi.edu/nl-seminar/
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
AI Seminar- Adam Lopez
Fri, Jun 14, 2013 @ 11:00 AM - 12:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Adam Lopez, Johns Hopkins University
Series: Artificial Intelligence Seminar
Abstract: A decade ago, students arrived at universities having been exposed to the idea of machine translation (MT) primarily through science fiction. Today, services like Google Translate have made it science fact. It is increasingly taught in courses and has its own textbook. But since it's a real application that isn't fully perfected, the best way to learn about it is still to build an MT system. Unfortunately, popular open-source toolkits for MT are mature codebases featuring tens of thousands of source code lines, making it difficult to focus on their core algorithms. Most tutorials present them as black boxes. But we want students to understand these algorithms, and black boxes are incompatible with this goal.
As a centerpiece of our MT course, we challenged students to obtain the best performance on carefully constrained instances of four key MT problems, each corresponding to key problems in AI: alignment (unsupervised learning), decoding (search), evaluation, and reranking (supervised learning). We provided concise, fully-functioning, self-contained baseline components in less than 150 lines of python. Students brought a diverse arsenal of ideas to the problems, some novel. A surprising and exciting outcome was that student solutions or their combinations fared competitively on some tasks, suggesting that even novices could help improve the state-of-the-art on hard AI problems while simultaneously learning a great deal.
Biography: Adam Lopez works on problems at the intersection of algorithms, machine learning, formal language theory, and computational linguistics with applications to problems in natural language processing, particularly machine translation. He is an assistant research professor at Johns Hopkins University. Previously he was a research fellow at the University of Edinburgh after earning a Ph.D. from the University of Maryland. This summer he is a visiting scientist at SDL Research.
Home Page:
http://www.cs.jhu.edu/~alopez/
Host: David Chiang
Webcast: http://webcasterms1.isi.edu/mediasite/Viewer/?peid=ea55185170054e13972a0ea5b932eb6c1dLocation: Information Science Institute (ISI) - Marina Del Rey-6th Flr Conf Rm # 689
WebCast Link: http://webcasterms1.isi.edu/mediasite/Viewer/?peid=ea55185170054e13972a0ea5b932eb6c1d
Audiences: Everyone Is Invited
Contact: Peter Zamar
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
AI Seminar
Fri, Jun 28, 2013 @ 11:00 AM - 12:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Christian Shelton, CS Professor, UC Riverside
Talk Title: Continuous-Time Models: Why and How
Abstract: Discrete-time models are abundant in artificial intelligence: hidden
Markov models, dynamic Bayesian networks, Markov decision processes,
and (most) auto-regressive models assume time passes in discrete jumps.
Yet, most processes modeled actually evolve in continuous time. This talk
explores the problems inherent in this dichotomy, focusing on Markovian
models.
First, I will discuss the theoretic and experimental difficulties when
modeling in discrete time. In doing so, I will present continuous-time
Markov processes, drawing analogies to their discrete-time counterparts.
Second, I will present the continuous-time analog of a dynamic Bayesian
network: a continuous-time Bayesian network (CTBN). The talk will include
an overview of the learning and inference literatures for CTBNs, showing
how continuous-time aids in the development of efficient inference
techniques. Finally, I will show some application results employing CTBNs
on real data sets.
Biography: Christian R. Shelton is an Associate Professor of Computer Science at the
University of California at Riverside. He has spent time as a visiting
researcher at Intel Research and Children's Hospital Los Angeles. He was
the Managing Editor of the Journal of Machine Learning Research and on
the editorial board of the Editorial Board of the Journal of Artificial
Intelligence Research.
Host: David Chiang
More Info: TBA
Location: Information Science Institute (ISI) -
Audiences: Everyone Is Invited
Contact: Kary LAU
Event Link: TBA
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.
