Modeling Mental Contexts and Their Interactions
Wei Chen and Scott FahlmanCarnegie Mellon University
What is the problem?
• To represent mental states expressed in natural language– Mental states: belief, knowledge, intention,
supposition, perceived reality, etc.• Example:
“The girl was surprised to find her grandmother’s cottage door open”
1. There is a belief change2. What is true in the girls old belief and new belief3. The reality is different from the girl’s old belief
Properties of Mental States
• Nested“The police believe the thieves were trying to steal a
solar panel”
• Change with time“realize”, “remind”, “be surprised”, “forget”, etc.
• Interactions between mental states“She wants to find her father because she believes
he is still alive”
How to solve the problem?• Mental Context Network • Implemented in Scone KB • Context activation mechanism
Properties of Mental Context Network
• Mental contexts inherit from a context that holds a set of background knowledge.
• Mental contexts evolve with time
• Mental contexts are connected to events.
• Mental contexts are environment-sensitive.
Representing Dormant Memory
• “remember” and “forget”
X
X
belief
“P forgets X”
Before Context After Context
person P
dormant knowledge
context
X
X
After Context Before Context
“P remembers X”
person P
X
Inter-contextual Activities
• Default inference rules– E.g. when a conflict is detected between the
perceived reality and mental contexts, build new beliefs according to the perceived reality
LRC’s old belief LRC’s new
belief
The cottage door is open
reality
The cottage door is closed
Conflict detected
Negate
The cottage door is closed
Applications
• Subjectivity Analysis
• Question answering
• Question generation
Conclusion and Future Work
• Contributions:– A multi-mental-context network that represents
various mental states– An inter-contextual inference mechanism which
performs reasoning based on new information and a multi-modal memory
• Future Works:– Scaling up to causal relations, temporal
information, conditional statements etc.