Topic 1 -- Lexical Access and Content-Addressability
References:
Forster, K. I. (1976). Accessing the mental lexicon. In R.J.Wales & E.Walker (Eds.), New Approaches to Language Mechanisms. (pp. 257-287). Amsterdam: North-Holland. (available on ERes)
Associationism, direct connections - how?
- two methods - software (hash-coding, decision trees) or hardware (associative memory)
- decision trees
- one branch for every legal combination => very large tree, largely empty
- but it is content-addressable (given content, can determine address)
- pruning branches that do not lead to words reduces size, but predicts that nonwords would be detected faster than words of the same length; also predicts length effects in word recognition time
- no way to explain frequency effects
- hash coding techniques
- collisions avoided only with very large address space
- necessity for search
Hardware solutions
- word detector models (logogen)
- neural nets
- Kanerva's sparse distributed memory -- ref: Kanerva, P. (1988). Sparse Distributed Memory. Cambridge, Mass.: MIT Press.
Difficulties with word-detector accounts
- 'first-past-the post' selection of best match (logogen model)
- the blight-bright problem suggests that the correct detector may not be the first to reach threshold
- evidence that nonwords must excite word detectors
- conclusion is that the input activates a range of candidates, not just one
- predicts interaction between frequency and visual quality (not found)
- 'survival of the fittest' selection of best match (neural nets)
- competition from neighbors should have an inhibitory effect (but it doesn't)