25th MS/UW Symposium in Computational Linguistics
Location: Microsoft, Building 99, lecture hall 1919
Time: 3:30-5pm, October 28, 2011
Announcing the 25th Symposium in Computational Linguistics sponsored by the UW Departments of Linguistics, Electrical Engineering, and Computer Science, Microsoft Research, and UW alumni at Microsoft.
Come take advantage of this opportunity to connect with the computational linguistics community at Microsoft and the University of Washington. This is a regular opportunity for computational linguists at the University of Washington and at Microsoft to discuss topics in the field and to connect in a friendly informal atmosphere. We will have five short talks (see below), followed by informal mingling.
Yoav Artzi: Towards Predicting Responses in Twitter
Twitter’s open and public network allows to directly observe how messages are reaching and influencing users by following responses. Twitter provides two forms of response: replies and retweets. Responses thus serve both as a measure of distribution and as a way to increase it. Understanding this dynamic for prediction would be valuable information for any content generator. In this work, we describe methods to predict if a given tweet will elicit a response from the social network once it’s posted. To accomplish this task we exploit features derived from various sources of signal such as the language used in the tweet, the social network and the user’s history. We use these features and leverage historical data to automatically train prediction models from a stream of real tweets collected over a two weeks period. We empirically show that our models are capable of generating accurate predictions over a subset of the tweets population.
Brian Hutchinson: Tensor Deep Stacking Networks for Phonetic Classification and Recognition
We introduce a novel deep architecture, the Tensor Deep Stacking Network (T-DSN), in which multiple blocks are stacked on top of another and where a bilinear mapping from hidden representations to the output in each block is used to incorporate higher-order statistics of the input features. Using an efficient and scalable parallel learning algorithm, we train a T-DSN to classify standard three-state monophones in the TIMIT database. The T-DSN outperforms an alternative pretrained Deep Neural Network (DNN) architecture in frame-level classification (both state and phone) and in the cross-entropy measure. For continuous phonetic recognition, T-DSN performs equivalently to a DNN, without the need for a hard-to-scale fine-tuning step.
Aniruddh Nath: Generalizing Natural Language Instructions
Reinforcement learning has been successfully applied to the problem of mapping natural language instructions to actions with little or no supervision (Branavan et al., 2009). However, this mapping cannot be used to solve new problem instances unless we also receive a set of instructions for the new instance. We present an algorithm for generalizing the knowledge gained while learning to map instructions to actions, allowing us to solve new problem instances with no additional knowledge. The algorithm is a form of imitation learning using Counting-MLNs (C-MLNs), a novel statistical relational representation that
can reason about the number of objects that satisfy a formula. We present an algorithm for learning C-MLNs, and apply it to the problem of generalizing instructions for the Crossblock puzzle game. We also investigate the use of C-MLNs for standard relational reinforcement learning, without the use of natural language instructions.
(Joint work with Matthew Richardson.)
Julie Medero: NLP in Patient-Directed Medical Displays
When patients are in the Emergency Department (ED), extensive electronic records are kept about the their visit. The patient’s complaints and background are recorded, along with doctors’ notes and records of every test that is ordered and every medication that is administered. Currently, though, that information is not available to the patient, or to the
patient’s family and friends who are with them in the hospital. In this talk, I will summarize work done this summer with the NLP and CUE groups at MSR on ways that NLP technologies can be used to make electronic medical records accessible to patients in the context of a mobile, patient-friendly display. In particular, we look at normalization and splitting of the Chief Complaint records that are entered by triage nurses when patients arrive at the ED, and at extracting “patient-friendly” explanations of lab tests and medications from consumer health websites.
Amittai Axelrod: Topic Modeling for Statistical Machine Translation
Unsupervised topic models can be used effectively in language modeling and information retrieval to tailor performance on broad corpora by determining clusters of related data. We combine such a topic model based on Latent Dirichlet Allocation with our recent work on corpus sub-selection to improve machine translation system results on a variety of TED talks.
Parking: Due to other events at Microsoft Friday afternoon, parking may be busier than usual. Please allow some extra time. You can park anywhere in the garage as long as you register the car with the receptionist (if the visitor spots are full, you can take any other spot in the garage). You may also email Michael Gamon (mgamon at microsoft dot com) with “symposium parking” in the subject line and with your vehicle and license plate information in the body to be pre-registered for parking, which saves additional time.