Download presentation
Presentation is loading. Please wait.
Published byHilary Cummings Modified over 9 years ago
1
Interference in Short-Term Memory The Magical Number Two (or Three) in Sentence Processing `06. 11. 4 (Sat.) / Chan-hoon Park Hypernetwork Models of Learning and Memory (Artificial Neural Networks)
2
© 2006 SNU CSE Biointelligence Lab 2 Contents Introduction Definition Parsing related problems Other STM Phenomena Limitation Conclusion
3
© 2006 SNU CSE Biointelligence Lab 3 Introduction The purpose of this chapter Syntactic processing + Memory research Limited syntactic working memory capacity – 2 (or 3) Interference effects Similarity-based interference Within independently motivated architectural mechanisms and principles and their associated constants
4
© 2006 SNU CSE Biointelligence Lab 4 Definition (1/2) Problems to deal for the purpose Center embedding [The salmon that [the man that [the dog chased] smoked] fell.] Almost every parsing theories about C.E. have No psychological motivation for structure or limit ex) Stack, Look-ahead buffers, 4, 2, 1 … Interference in multifaceted working memory Retroactive interference / Within-Category similarity effects The independence of parsing from phonological STM
5
© 2006 SNU CSE Biointelligence Lab 5
6
6 Definition (2/2) An interference theory of syntactic working memory NL-Soar : computational model of sentence comprehension (from Soar : a theory of the human cognitive architecture) NL-Soar’s working memory : H/D Set – buffers partial constituents X-bar position : Surface structure
7
© 2006 SNU CSE Biointelligence Lab 7
8
8 Parsing related problems (1/5) Embedded relative clauses The boy that the man the woman hired hated cried. What the woman that John married likes is smoked salmon.
9
© 2006 SNU CSE Biointelligence Lab 9 Parsing related problems (2/5) Center embedding or Gap-filling Without Gap-filling With Gap-filling Fred is tough for the man to please. Fred is easy for the woman who the man who hired me married to please.
10
© 2006 SNU CSE Biointelligence Lab 10 Parsing related problems (3/5) Embedding or stacking Difficulty of C.E is Consecutive occurrences of NPs (= Stacking ), rather than C.E. per se. In head-final languages (Japanese, Korean...),
11
© 2006 SNU CSE Biointelligence Lab 11 Parsing related problems (4/5) Limitations on ambiguity resolution At most two nodes are available to assign the same structural relation at any given time. Amparo saw the dog under the box on the table in the room next to the library. Closure principle Keep a small set of syntactic nodes open for further attachment. Two most recent nodes
12
© 2006 SNU CSE Biointelligence Lab 12 Parsing related problems (5/5) Length effects in garden path structures The horse (raced past the barn fell). Distance to disambiguation But, Ron (believed the ugly little linguistics professor). Ron (believed the ugly little linguistics professor he had met the week before in Prague disliked him). A limited repair mechanism Length-induced garden path effect in NL-Soar The girls believe the man who believes the very strong ugly boys struck the dog killed the cats.
13
© 2006 SNU CSE Biointelligence Lab 13 In other STM phenomena Immediate verbal memory task Average word span of 2.6 The limited capacity of human memory – 2~3 similar chunks
14
© 2006 SNU CSE Biointelligence Lab 14 Limitation Possibility of Three Three NPs occupying spec-IP, but acceptable Average span is more than just 2. There are may be additional cost associated with adjuncts in general, regardless of whether they involve relativization.
15
© 2006 SNU CSE Biointelligence Lab 15 Conclusion Sentence processing theories can benefit from incorporating ideas from traditional work in short- term memory. NL-Soar’s working memory The source of memory load is open or unsatisfied syntactic relations. This leads naturally to a focus on stacking, as a source of difficulty. Increasing similarity makes things more difficult.
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.