Britain was deep in recession while Germany was flourishing three years ago. France kept moving ahead steadily long after Germany had fallen into recession. But now France is plunging deeper while the German economy continues to struggle. Britain has been taking small steps toward stimulating its economy by cutting interest rates, and has finally started to emerge from recession.
Why are words/constructions of motion and manipulation such as move ahead steadily, fall, take small steps, plunge deeper, struggle, be deep, leave, sink, and start to emerge used in a narrative about international economics?
Our project provides evidence through computer simulation that a key reason for using motion words and phrases is that it allows for the deep semantics of causal narratives to be dynamic and arise from a continuous interaction between input and memory. Since knowledge of moving around or manipulating objects is essential for survival, it has to be highly compiled and readily accessible knowledge. Representations meeting these criteria must be context sensitive and allow changing input context to dramatically affect the correlation between input and memory and thereby the set of possible expectations, goals, and inferences. Speakers are able to felicitously exploit this context-sensitivity in specifying important information about abstract actions and plans that take place in complex, uncertain and dynamically changing environments.
Systematic metaphors project features of these representations onto abstract domains such as economics enabling linguistic devices to use embodied causal terms to describe features of abstract actions and processes.
Embodied metaphors project features of spatial motion and manipulation onto abstract plans and processes. This allows event descriptions to exploit the highly compiled and real-time aspects of x-schema representations to express complex, uncertain, and evaluative knowledge about abstract domains such as international economic policies.
In our theory, embodied metaphors such as the Event Structure Metaphor (Lakoff 1994) use the dense and familiar causal structure of spatial motions to permit reflex (fast, unconscious) inferences. Our x-schema based representation is able to encode theories of embodied and highly familiar domains in a manner that retains the dynamic and highly responsive , real-time nature of x-schemas and their ability to model reflex inferences. Our project models how projection of x-schema based inferences can add valuable information about abstract plans and events in real-time.
We are concerned about modeling information provided by motion and manipulation terms used in ordinary discourse (such as newspaper stories) about the domain of international economics. To this end, we have constructed an ontology and simple model of the target domain of international economic policies. Our computational model consists of multi-valued economic policy variables modeled as a Bayesian network. The inherently temporal nature of events necessitates a temporally extended network with temporal links that predicted the value of a variable at time t , given its value at time t - 1 .
Our overall computational model, shown in the figure above includes a third important piece, namely metaphor maps that project specific features obtained from x-schema executions that model the semantics of motion and manipulation terms (called the source domain onto features of the target domain of international economics. When active, metaphor maps allow the real-time inferential products of x-schema executions to influence the target domain network by setting evidence for specific features of a given temporal slice of the economic policy network to specific values. Activating a metaphor map requires that the context (domain of discourse) be identified as being about the target domain; thus shutting off projections for irrelevant contexts.
Embodied metaphor maps are thus hybrid structures connecting the x-schema based sensory motor representations to the target domain belief network that represents knowledge about international economics. Such maps project specific results of x-schema executions; adding new information about the target by asserting evidence on the temporally extended Bayes net. Belief Update algorithms (Pearl 1988, Jensen 1995) propagate the influence of this new evidence forward and backward in the temporally extended network to arrive at new posterior values for other relevant economic features at the same and different time steps.
Our model currently includes three different types of maps. One type of map corresponds to ontological maps (Lakoff and Johnson 1980) or OMAPS. OMAPS map entities and objects between embodied and abstract domains. A second type of map projects events , actions , and processes from embodied to abstract domains. Such maps are called schema maps or SMAPS. An important function of SMAP projection is to invariantly project the aspect of the described event across domains. Our event representation, called x-schemas are parameterized controllers and the third type of metaphoric map maps x-schema parameters across domains. These maps are called parameter maps or PMAPS.
The figure above shows the system interpreting the input corresponding to "European Economic Giant Falls Sick". The Metaphoric maps are shown in blue, the active source domain schemas are in pink and the target domain inferences are in cyan. The strength of the belief in a specific entitly is down by the amount of color for the specific time slot of interest. The result of processing the input enables the system to identify Germany as the referent (subject) of the input and also leads to system to conclude that the current economic state is negative and that the goal is to regain a positive economic status.
S. Narayanan (1999).
Moving Right Along: A Computational Model of Metaphoric Reasoning
about Events
(Proceedings of the National Conference on Artificial Intelligence
(AAAI '99), Orlando Florida, July 18-22, 1999. (to appear))
Fauconnier, Gilles (1997). Mappings in Thought & Language. Cambridge
University Press. 1997.
Grady, Joe (1997).
Foundations of Meaning: Primary Metaphors and Primary Scenes.
Ph.D. Dissertation, Department Of Linguistics, University of California, Berkeley, 1997.
Other Relevant References
Page created by
Srini Narayanan
Nov 20, 1997