midterm 2 Flashcards
(38 cards)
dense phonological neighborhood, effect on word recognition in noise?
- a group of words that sound very similar to one another (e.g., “cat” has neighbors like “cap,” “cad,” “mat,” etc.). These similar-sounding words are called phonological neighbors.
Effect on word recognition?
-Words from dense neighborhoods are harder to identify.
- High-frequency words in sparse neighborhoods are the easiest to recognize.
- lexical competition
Homophones vs. Homographs
Homophones: Same sound, different meanings (e.g., “watch” = a timepiece or to observe).
Homographs: Same spelling, different meanings/pronunciations (e.g., “bass” as a fish vs. a musical tone).
Parallel Access Effect for Homophones (Tanenhaus et al., 1979)
Cross-modal priming task: Participants hear a sentence ending in a homophone (e.g., “watch”) and quickly see a word on a screen (e.g., “TIME” or “DART”) to decide if it’s real.
Result: Both meanings (noun and verb) are activated initially — showing parallel access.
🕰 Context filters meaning later (~250 ms), suggesting bottom-up activation followed by contextual selection.
Cohort Model: What words might be activated by “element”?
The Cohort Model (Marslen-Wilson):
Words activated early on share the same initial phonemes.
For “element”: Possible cohort includes “elephant,” “elegant,” “elevate,” etc.
These are narrowed as more sounds are heard, until the uniqueness point (e.g., “elementa…” rules out others).
TRACE Model: Words activated by “element”? The TRACE Model (McClelland & Elman):
Words are activated based on any overlapping phonemes, not just the beginning.
So for “element,” words like “cement,” “ornament,” or even “elephant” could be weakly activated.
TRACE allows bidirectional activation (phoneme ↔ word levels).
Cohort vs. TRACE Models
Activation:
Cohort: Words sharing initial sounds
TRACE: Words sharing any overlapping sounds
Processing:
Cohort: Strictly feedforward
TRACE: Interactive (bottom-up & top-down)
Handling of noise:
Cohort: Struggles with noisy input
TRACE: More robust to ambiguity/noise
Allopenna et al. (1998): Visual World Eyetracking & the TRACE Model
Task: “Put the beaker above the square.”
Objects on screen: beaker, beetle, etc.
👀 Eyetracking showed: Participants looked at both “beaker” and “beetle” early on — consistent with TRACE’s prediction of overlapping activation.
🧪 Cohort would predict no fixation on “beetle” (not a cohort member).
📌 Thus, TRACE is better supported.
What does it mean to say that sentences are comprehended in “real time”?
Sentence comprehension is incremental and dynamic — we interpret meaning as each word is heard, not after the sentence ends.
This is evident from garden-path sentences, where we’re misled by initial interpretations.
What is a garden-path?
A garden-path sentence leads the reader/listener toward one interpretation, which later proves incorrect:
“The horse raced past the barn fell.”
The parser assumes “raced” is main verb, then has to backtrack when “fell” appears.
syntactic ambiguity
A sentence has syntactic ambiguity when it can be parsed in more than one way.
Example:
“The boy put the book on the shelf into his backpack.”
Did the boy put the book that’s on the shelf into his backpack, or put the book onto the shelf that’s in the backpack?
Modular Syntactic Parser vs. Cascading Interactive Parser
Modular Parser:
- Uses only syntax at first
- Fast, automatic, rigid
- Explains garden-paths via structure
Cascading Interactive Parser
- Integrates syntax, semantics, pragmatics
- Flexible, context-sensitive
- Predicts garden-paths can be avoided
Minimal Attachment & Parsing Ambiguity
Minimal Attachment: The parser defaults to the syntactically simplest structure.
For example, in:
“The athlete accepted the prize would not go to him.” We parse “the prize” as the object, not the subject of a clause — leading to confusion.
Tanenhaus Visual World Study: “Put the shoe on the towel…”
Sentence: “Put the shoe on the towel into the bucket.”
Ambiguous unless there’s visual context.
🧠 Participants misinterpret if there’s only one shoe (assume “on the towel” is destination).
In unambiguous control (e.g., two shoes: one on towel), they resolve correctly.
📷 Key finding: Visual context affects parsing in real time — evidence for interactive comprehension.
Kamide & Altmann: Predictive Eye Movements
Sentences: “The man/girl will ride/taste the…” (e.g., beer, candy)
Participants look at the appropriate object before it’s mentioned.
⏱ Shows real-time prediction using syntax + semantics.
How Neural Networks Learn Output Patterns
Start: Random output.
Learning: Adjust weights (connections) using error correction (backpropagation).
Gradually: Patterns become closer to target output.
Why Feedforward Word Predictors Fail with Context
Simple model: Predicts based on just the previous word.
Fails in:
“The man will ride the…” vs. “The girl will ride the…”
Can’t distinguish because it lacks memory — produces same prediction regardless of earlier context.
Can a Feedforward Digit Predictor Learn a Next-Digit Task?
Yes, it can eventually learn the pattern (e.g., 0→1, 9→0), because the mapping is:
Consistent
Deterministic But: It lacks flexibility — e.g., it can’t generalize to “if input is X, output X+1” rule without retraining.
Simple Recurrent Network (SRN)
Like a feedforward net but stores memory via a context layer.
This context allows it to:
-Keep track of prior inputs.
-Learn sequential structure.
-Predict better over time.
SRN vs. Large Language Model (LLM)
Architecture:
SRN: Small-scale, few layers
LLM: Deep Transformer models
Memory:
SRN: Fixed, short-term
LLM: long context via attention
Learning:
SRN: Trained on small datasets
LLM: trained on billions of tokens
Are LLMs Good Models of Human Language?
Not quite.
LLMs:
Use surface-level statistical regularities.
Don’t “understand” meaning or have grounded experience.
Don’t acquire language like children (no embodied learning or social cues).
Humans:
Learn through interaction, attention, grounding in perception/action.
Readings (Blank, 2023; Cai et al., 2024) agree: LLMs are impressive but cognitively impoverished compared to human language use.
Classic View — Language as a Mirror of Thought
Supported by Chomsky and Fodor
Language labels pre-existing concepts
Thought = universal; Language = expression
🧠 Example: Infants and non-linguistic primates have similar conceptual structures
Linguistic Relativity (Whorfian View)
Language may alter concepts, The idea that language shapes thought
Labels change how we see, think, or remember
🧪 Evidence:
Words may cause conceptual merging or splitting
New labels → new mental categories
Ex: “Wug”, “dax”, “zif” might lead to category distortions over time
Linguistic Relativity:
“Observers are not led by the same evidence to the same conclusions unless their language backgrounds are the same.” (Whorf)
Examples from Event & Spatial Language:
Manner languages (English) → verbs describe how
Path languages (Spanish, Greek) → verbs describe where
Visual World Experiments:
Speakers of different languages attend to different parts of a scene depending on their verb bias (manner vs. path).
Space Language
Egocentric (left/right) vs. Allocentric (north/south)
Tzeltal (Mayan): uses only allocentric
Early studies: Tzeltal speakers failed egocentric tasks
BUT…
👁 Li & Gleitman showed that Penn students behaved similarly depending on context — suggesting environmental factors matter, not just language
Language shapes what we attend to, especially during communication — not necessarily how we can think.
How Common is Bilingualism? is one language turned off?
60%+ of world population speaks more than one language
Monolingualism is rare globally
❌ NO.
Bilingual minds show co-activation of both languages
Even in single-language contexts, non-target language activates subtly
