|
- Temporal Information Extraction: We implemented hybrid systems with linguistically motivated solutions and machine learning classifiers for extracting temporal information from raw text. We do deep semantic parsing and use hand-coded rules to extract events, features and temporal expressions from the logical forms produced by the parser. In parallel, we filter events, extract event features, temporal expressions, classify temporal relations using machine-learning classifiers. Our system had a competitive performance in all tasks of TempEval-2.
- Temporal Evaluation: We proposed a new metric for evaluation of temporal annotation. It uses temporal closure to reward relations that are equivalent but distinct. Our metric measures the overall performance of systems with a single score, making comparison between different systems straightforward. Our approach is intuitive and computationally inexpensive. This evaluation metric will be used in TempEval-3 to evaluate participants.
- Temporal Question Answering: We proposed a temporal QA system that performed temporal reasoning and showed how it can be used to evaluate automated temporal information understanding. With temporal reasoning, our QA system can answer list, factoid and yes/no questions. This was a joint work with Hector Llorens.
- Creating 1 million word corpus with temporal annotation and organizing TempEval-3: We teamed up with other participants of TempEval-2, Hector Llorensand Leon Derczynski, and the organizers of TempEval-2, Marc Verhagen and James Pustejovsky, to create a 1 million word corpus with temporal annotation and to organize TempEval-3. We will automatically annotate a 1-million word corpus from news domain by merging our system outputs and human-reviewing a portion of it. We will organize TempEval-3 with that corpus.
- Merging Temporal Annotations: We proposed algorithms for merging multiple temporal annotations. With such a merging algorithm, we can merge multiple systems, multiple human annotations, or a combination of both. We evaluated our algorithm over TimeBank and AQUAINT corpora considering three state-of-the-art systems. The results of merging improve over individual systems and baseline merging algorithms. We will be using this merging tool to merge automated system outputs for TempEval-3 to create the 1-million word corpus. This is a joint work with Hector Llorens.
- Multimodal Summarization: In collaboration with Jeffrey P. Bigham and James Allen, I worked on illustrating complex sentences as multimodal summaries combining pictures, simple sentence structure and summarized text. We show that pictures alone are insufficient to help people understand most sentences, especially for readers who are unfamiliar with the domain. An evaluation of our system in the Wikipedia domain illustrates both the promise and challenge of automatically creating multimodal summaries. MMS could be used to help people with cognitive disabilities, children, older people, or people whose first language is not English.
- Game Prediction with Social Media at Yahoo! Research, Barcelona (2011): We explored the use of social media for game prediction by prototyping our system on World Cup Soccer 2010 tournament with Twitter data. We extracted around 150K prediction tweets from 1.8 billion tweets by matching the syntactic structure of predictions, instead of just matching prediction words. We created a prediction retrieval exploratory search application with the high-precision extracted predictions, where all predictions for a game are shown and the predictors are ranked according to their previous prediction success. Finally, we aggregated the predictions and we predicted the most likely outcome of upcoming games with high success. A key benefit of our framework is that it does not rely on the domain dependent knowledge for prediction, rather uses crowd's predictions from the social media. As a result, we can consider crowd's analysis about the event and by aggregating the predictions, we can attempt to predict the outcome of other sport events, elections, product release dates and any other future events discussed in the social media. Find more here: TwitterPaul.
- Information Extraction on Demand at Microsoft Medical Media Lab (2009): Implemented an Information Extraction tool that can learn different patterns from user's examples and extract the pattern instances from natural language texts in medical domain, e.g. release notes, radiology reports, etc. For pattern matching, we used the medical ontology (UMLS) to understand medical patterns like diseases, medicines, etc. and considered language features for general patterns like patient names, IDs, dates, etc. My implemented tool was shared with the clinical staff at Washington Hospital Center and they found it useful.
- In-Car Dialog System at Bosch RTC (2008): Developed a syntactic and semantic grammar for a commercial in-car dialog system. It was very similar to the famous Siri, iPhone's dialog assistant. My work was in navigation, local business and MP3 domain.
|