Loading...
Please wait, while we are loading the content...
Similar Documents
Seeking coherent explanations -- a fusion of structured connectionism, temporal synchrony, and evident reasoning - eScholarship
| Content Provider | Semantic Scholar |
|---|---|
| Author | Shastri, Lokendra Wendelken, Carter |
| Copyright Year | 2000 |
| Abstract | Seeking coherent explanations — a fusion of structured connectionism, temporal synchrony, and evidential reasoning Lokendra Shastri and Carter Wendelken International Computer Science Institute 1947 Center Street, Suite 600 Berkeley, CA 94704 Abstract A connectionist model capable of performing rapid inferences to establish explanatory and referential coherence is described. The model’s ability to perform such inferences arises from (i) its structure, (ii) its use of mutual inhibition among “sibling” types, entities, and rules, (iii) the use of temporal synchrony for representing dynamic bindings, and (iv) its ability to rapidly modify weights in response to convergent activity. Introduction Consider the following simple narrative: “John fell in the hallway. Tom had cleaned it. He got hurt.” Upon hearing the above narrative most of us would infer that Tom had cleaned the hallway, John fell because he slipped on the wet hallway floor, and John got hurt because of the fall. These inferences allow us to establish causal and referential coherence among the events and entities involved in the narrative. They help us explain John’s fall by making plausible inferences that the hallway floor was wet as a result of the cleaning and John fell because he slipped on the wet floor. They help us causally link John’s hurt to his fall. They help us determine that “it” in the second sentence refers to the hallway, and “He” in the third sentence refers to John, and not to Tom. Empirical data strongly suggests that inferences required to establish referen- tial and causal coherence occur automatically during language understanding (see e.g., Just & Carpenter 1977; Keenan, Bail- let, and Brown 1984; Kintsch 1988; McKoon & Ratcliff 1980, 1992; Potts, Keenan, & Golding, 1988). Any system that attempts to explain our ability to establish causal coherence during language understanding must pos- sess a number of properties: First, such a system must be representationally adequate. It must be capable of encoding specific facts and events and expressing general regularities (aka rules) that capture the causal structure of the environ- ment. In particular, the system should be capable of encoding context-dependent and evidential cause-effect relationships. Second, the system should be inferentially adequate, that is, it should be capable of drawing a range of explanatory inferences by combining evidence and arriving at coherent interpretations that are quasi-optimal with reference to a cost- function (Hobbs et. al, 1993). Third, the system should be capable of establishing referential coherence. In particular, it should be able to unify entities and events by recognizing that multiple designations might refer to the same entity or event. Fourth, the system should be capable of learning and fine- tuning its causal model based on experience, instruction, and exploration. Finally, the system should be scalable and com- putationally effective. The causal model underlying human language understanding would be extremely large. Yet we understand language at the rate of several hundred words per minute (Just & Carpenter 1977). Hence, a system for estab- lishing causal coherence should also be capable of encoding a large causal model and rapidly performing the requisite in- ferences within fractions of a second. This paper describes several key extensions to the con- nectionist model SHRUTI that enable it to draw the sorts of inferences described above. SHRUTI is a neurally plausible system capable of expressing causal knowledge involving n- place relations, limited quantification, and type restrictions. It encodes specific events as well as context-sensitive priors over events. It expresses dynamic bindings via the synchronous fir- ing of appropriate node clusters and performs inferences via the propagation of rhythmic activity over node clusters. This propagation amounts to a parallel breadth first activation of the underlying causal graph, and hence, the reasoning in SHRUTI is extremely fast. The use of weighted links and activation combination functions at nodes allow SHRUTI to encode soft rules and perform evidential inference. SHRUTI supports su- pervised learning which allows it to fine-tune its causal model in a data-driven manner (Shastri & Ajjanagadde, 1993; Shastri & Grannes, 1996; Shastri, 1999; Shastri & Wendelken, 1999; Wendelken & Shastri, 2000). In order to carry out inferences for establishing referential and causal coherence, however, SHRUTI ’s core functionality had to be extended in a number of ways. These include the ability to (i) unify entities and relational instances (events) (ii) posit the existence of entities that are left implicit in the utter- ance, and (iii) favor interpretations that are more plausible and more likely over others that are less so. These functional ex- tensions were realized in part by introducing mutual-exclusion clusters in the encoding of types and entities and by modifying the behavior of node-types. But more importantly, SHRUTI ’s inferential behavior was modified by (i) introducinginhibitory interactions among rules sharing a common consequent (ef- fect) and (ii) modeling short-term-potentiation, a biological phenomena whereby synaptic strengths (link weights) un- dergo rapid but short-lived changes in response to convergent activity. Both these changes play a critical role in favoring coherent and more-likely interpretations over less coherent and less likely ones. The rest of the paper is organized as follows. The next section presents SHRUTI ’s basic representational machinery. This is followed by an elaboration of evidential reasoning in SHRUTI . Next we discuss mechanisms particularly aimed at the problem of establishing coherence and illustrate the functioning of the model with the help of an example. SHRUTI ’s representational machinery Figure 1 illustrates the encoding of the following fragment of knowledge (expressed in SHRUTI ’s input syntax): (1) 8 x:Agent, y:Location [slip(x,y) ) fall(x,y) (600,900)]; (2) 8 x:Agent, y:Location [trip(x,y) ) fall(x,y) (800,900)]; (3) *TF: trip(Person, Location) 100; (4) *TF: slip(Person, Location) 50; |
| File Format | PDF HTM / HTML |
| Volume Number | 22 |
| Alternate Webpage(s) | http://www.icsi.berkeley.edu/pubs/ai/seekingcoherent00.pdf |
| Alternate Webpage(s) | http://www1.icsi.berkeley.edu/~shastri/psfiles/cogsci00.pdf |
| Alternate Webpage(s) | http://www.icsi.berkeley.edu/~carterw/papers/cogsci2000.pdf |
| Alternate Webpage(s) | http://www.icsi.berkeley.edu/~shastri/psfiles/cogsci00.pdf |
| Alternate Webpage(s) | http://www.icsi.berkeley.edu/~carterw/papers/cogsci2000.ps |
| Alternate Webpage(s) | http://www.cis.upenn.edu/~ircs/cogsci2000/PRCDNGS/SPRCDNGS/PAPERS/SHA-WEN.PDF |
| Alternate Webpage(s) | http://www.icsi.berkeley.edu/~shastri/psfiles/cogsci00.ps |
| Alternate Webpage(s) | https://cloudfront.escholarship.org/dist/prd/content/qt74f7n8vx/qt74f7n8vx.pdf?t=op352h |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
