Seminar- “Modeling Practical Knowledge For Natural Language Applications"
- When:
- Thursday, March 7, 2013, 10:00 am - 11:00 am
- Where:
- 6th Floor Conf. Room (#689)
- Speaker:
- Alicia Sagae
- Description:
Bio:
Alicia Sagae received a PhD in Language Technologies from Carnegie Mellon University in 2010. Her dissertation research focused on semantic similarity of natural language descriptions of images, combining symbolic and data-driven approaches. From 2006 to 2008 she was a visiting scholar at the University of Tokyo, where she developed ontologies to support text mining of biomedical data. In 2008 she joined Alelo, an ISI spin-off company specializing in serious games for culture and language training. In her four years as a research scientist at Alelo, she served as principal investigator in multiple government funded projects, developing models of culture and dialogue for characters in virtual environments.
Abstract:
Symbolic representations of domain knowledge are leveraged by many language technology systems to improve performance. Examples include WordNet, FrameNet, and Wikipedia as well as domain-specific ontologies like the GENIA ontology for biomedical text mining. These resources vary greatly in scope and structure, and choosing the right one for a given task can be a challenging problem. In this talk I will discuss how to develop, apply, and compare such resources for a variety of natural language tasks. First, ontological features that help discriminate among natural language image descriptions are used to rerank in an image retrieval setting. We compare the contribution of features from WordNet and from a domain-specific ontology that was encoded in a marker-passing logical framework called Scone. Results show improvement when using domain-specific features on previously unseen descriptions. Next, sociocultural knowledge is encoded in a common logic ontology, in order to improve the performance of a dialogue system where language learners can have practice conversations in real time with a virtual character. Capturing this sociocultural knowledge as a domain-specific resource allows the dialogue system to perform more complex and realistic behaviors. Finally, I will introduce ongoing work on applying world knowledge encoded as a common logic ontology to the task of interpreting metaphorical text.
