Sem2 Week 7
No description yet. This curated deck is ready for daily study.
Card preview (24)
What are the two mechanisms humans use for past tense formation?
A productive rule (add -ed) for regular verbs, and memory storage/retrieval for irregular verbs
What is the regular past tense rule in English?
Add -ed to the end of the present tense form (e.g. walk → walked, jog → jogged)
Are regular verbs an open or closed set?
Open — probably tens of thousands in a native speaker’s lexicon, and new ones enter English every year
Are irregular verbs an open or closed set?
Closed — only 150–180 members, with no recent additions (sneak→snuck was the last, arriving in the 19th century)
What is the blocking principle in past tense formation?
If an irregular past tense form is stored in memory, it blocks the regular -ed rule from applying. If no form is stored, the rule applies.
What are the three stages of children’s past tense acquisition?
Stage 1: few errors; Stage 2: overgeneralisation of -ed to irregulars (e.g. holded); Stage 3: error rate slowly decreases to near zero
What is the U-shaped learning curve in language acquisition?
A pattern where performance is initially correct, then gets worse (overgeneralisation), then improves again — e.g. spoke → speaked → spoke
What does it mean that grammatical rules are implicit?
Speakers follow them correctly without being able to consciously articulate them
Give two examples of children overgeneralising the -ed past tense rule
Any two of: goed, holded, stealed, buyed (applying -ed to irregular verbs)
Besides past tense, name two other morphological rules children overgeneralise
Any two of: plural -s (mans, foots), 3rd person -s (haves), comparative -er (gooder), superlative -est (powerfullest), ordinal -th (twoth)
Who wrote On Learning the Past Tenses of English Verbs?
D. E. Rumelhart and J. L. McClelland
Why is the past tense a useful case study for language acquisition?
It has a clear productive rule (regular verbs) alongside a closed set of stored exceptions (irregular verbs), making rule vs. memory distinctions testable
What kind of errors do children make that they could never have heard adults produce?
Overregularisation errors like swang or shuck — showing children are constructing forms creatively by analogy, not just imitating
What pattern do some irregular verbs show that suggests sub-regularities?
Groups like sing→sang, ring→rang follow a vowel-change pattern, suggesting irregulars are not entirely chaotic
What is the key theoretical puzzle the Rumelhart & McClelland model addresses?
How implicit rules like the past tense rule can be represented and learned without being explicitly stored — motivating a neural network approach
What is the central theoretical claim of Rumelhart & McClelland (1986)?
That rule-governed language behaviour can emerge from a connectionist network with no explicit rules stored anywhere — rules are an emergent property, not a stored proposition
What is the ‘explicit inaccessible rule’ view that R&M argue against?
The view that grammatical rules are stored as propositions in a specialised subsystem (the LAD), just inaccessible to conscious report — R&M reject this in favour of emergent, distributed knowledge
What is a Wickelphone?
A context-sensitive phoneme unit encoding a triplet: a target phoneme plus its immediate predecessor and successor (e.g. the ‘p’ in ‘spoke’ is encoded as the triple /s-p-o/)
Why did R&M use Wickelphones rather than simple phoneme units?
Simple phonemes have no positional context — Wickelphones capture local phonological context, which is what determines most past tense changes
What is a Wickelfeature?
A distributed, more compact encoding used instead of whole Wickelphones — each unit encodes a triple of phonemic features (e.g. voiced-dental-short) rather than a triple of whole phonemes
Why did R&M use Wickelfeatures instead of Wickelphones?
The number of possible Wickelphones is ~45,000, giving over 2 billion possible connections — Wickelfeatures reduce this to a tractable size while preserving phonological context
What kind of network architecture did R&M use?
A two-layer pattern associator (no hidden units) with modifiable connections from every input Wickelfeature unit to every output Wickelfeature unit, plus a fixed decoding network to convert output back to phonology
What were the initial connection weights in the R&M model?
All set to zero — a tabula rasa with no preset input-output relations
What learning rule did R&M use to train the network?
The delta rule (perceptron convergence procedure) — weights are adjusted in proportion to the difference between the target output and the actual output