Events - NLP Talks

Talks for Fall 2007

Our approach follows a word order monotonization strategy making use of syntax information (dependency parse trees) of the source language to build a set of automatically extracted reordering rules. The input sentence is extended to a graph built with reordering hypotheses, hence, allowing for a constrained search on the syntactically motivated reorderings.

I will also discuss the work carried out at Columbia University with Nizar Habash, which basically focuses on the Arabic-English language pair and introduces chunking information.

In this talk, I'll present WordsEye, an experimental web-based system for creating 3D scenes using natural language input. WordsEye relies on a large library of 3D models and 2D images to generate its scenes and uses an open source raytracer to perform final renders with reflections and shadows. I'll discuss the overall software architecture and some of the linguistic, semantic, and representational issues raised by this type of application.

This work has been done in collaboration with Richard Sproat.

In this talk, I'll discuss a system we developed that analyzes raw news text and predicts its impact on public opinion of political candidates. To measure public opinion, we use Prediction Markets -- online future markets that allow participants to invest money in predicting future events. Using NLP techniques we show that news data can be used to predict future directions in these markets, which serve as a proxy for public perception about the candidates.

This work was done at the University of Pennsylvania, with Ari Gilder, Mark Dredze, and Dr. Fernando Pereira.

In this talk I will describe some experiments using a spoken language dialog system for collecting information about patients in the ICU. Speech in the hospital has always been a no-no (usually due to difficult acoustic environment), we are re-visiting this to see if advances in IT, recognition and devices technology are sufficient to cross that barrier. During this talk I will briefly describe the dialog system we used, talk about the results from our experiments (both speech and dialog). I will also discuss the results of a survey we have completed recently which tries to distinguish between two different approaches for collecting information with varying levels of initiative. This work is part of a collaboration between our CS department, Columbia Presbyterian and IBM.

The field of opinion (or sentiment) analysis has received much recent research attention motivated by its practical appeal and the interesting computational problems that it presents. Sentiment analysis is concerned with automatically extracting attitudes, opinions, evaluations, and sentiment from text. More precisely, we are interested in the area of fine-grained sentiment analysis, which is concerned with opinions at the level of sentences, clauses, or individual expressions of opinions.

Our work assumes that we have access to automatically extracted fine-grained expressions of opinions. We argue that to be fully useful fine-grained expressions of opinions need to be augmented with additional information (e.g. opinions need to be grouped by their opinion holder and/or their topic). We refer to this task as opinion summarization. Our work focuses on combining fine-grained opinion expressions into higher-level opinion summaries, which are easy to browse, explore and manipulate.

In my talk I will define our notion of opinion summary and give some motivating examples of why opinion summaries are needed. I will then focus on the research challenges that have to be addressed in order to create opinion summaries. I will discuss each of the challenges and talk about our approaches.