PSY 365: Psychology of Language - Midterm 2 Study Guide

Flashcards covering syntax acquisition, spoken and visual word recognition, syntactic parsing theories, speech planning and production errors, and computational language models.

What is the difference between syntax and semantics?

Syntax refers to the rules for how words combine into sentences, whereas semantics refers to the meaning of those sentences.

What does the principle of compositionality state in syntax?

The meaning of the whole sentence can be derived from the meaning of its parts plus a set of rules for combining them together (e.g., Adjective-Noun combinations like "Red House").

What is a syntactic constituent, and how can differing parses affect sentence meaning?

A constituent is a group of words treated by the parser as a single unit (such as a Noun Phrase like "The red hat"). Parsing the same group of words into different constituents alters meaning (e.g., "[I shot] [an elephant in my pyjamas]" vs "[I shot] [an elephant] [in my pyjamas]").

What are the generative, hierarchical, and recursive properties of syntax?

Generative means being able to produce and understand novel sentences never heard before. Hierarchical means structures exist at multiple levels (words group into constituents, which group into higher-level constituents). Recursive means a constituent can sit inside another constituent of the same kind (e.g., an NP containing another NP).

What is semantic bootstrapping (Pinker, 1987)?

The theory that children learn syntax more easily once they know the meaning of words, effectively "pulling themselves up by their bootstraps" using word meaning.

What is distributional learning (Mintz, 2003), and what did Mintz (2006) demonstrate in $12\text{-month-olds}$?

Children use statistical patterns in word distributions to learn syntactic categories; frequent frames (like "the ___ is") predict parts of speech with $90\%$ accuracy. Mintz (2006) showed $12\text{-month-olds}$ looked longer at a novel word placed in an incorrect frame, showing they use distributional cues.

What are Tomasello's (1992) "verb islands"?

Verb-specific constructions where children use new verbs in familiar frames before generalizing them to novel structures (e.g., generalizing "Daddy cuts cheese" to "Mommy cuts cheese", but failing to generalize to "Can I cut it?").

What did MacWhinney (1975) and Brooks & Tomasello (1999) find regarding children's early syntactic generalization?

MacWhinney found $85\text{--}96\%$ of what two children said derived from just $40$ templates. Brooks & Tomasello found children exposed to a new verb solely in the passive voice produced it in active voice only $28\%$ of the time.

Why do overgeneralization errors (Bowerman) support abstract rule learning?

Children produce novel expressions not heard in adult speech (e.g., "Button me the rest" or "Why did you fall me off?"), demonstrating that they are generalizing abstract grammatical rules rather than merely repeating adult utterances.

What did Wonnacott (2011) discover about children's generalization using an artificial miniature language?

When grammatical particles ("dow" vs "tay") appeared freely with any noun, children generalized them broadly; however, when each noun was restricted to only one specific particle, children stopped generalizing.

What is semantic priming (Meyer & Schvaneveldt, 1971), and how did Yee & Sedivy (2006) demonstrate it in spoken language?

Semantic priming is faster recognition of a word when preceded by a related word (e.g., "doctor" after "nurse"). Yee & Sedivy demonstrated real-time priming by showing participants who heard "hammer" looked more frequently at a picture of a nail.

What is neighborhood density, and how does it affect word recognition (Goldinger et al., 1989)?

Neighborhood density refers to the number of phonological neighbors (words differing by one sound). Words from dense neighborhoods (e.g., "sling") are recognized more slowly because neighboring candidates compete during recognition.

What is mediated priming?

Indirect semantic priming where a prime speeds recognition of a target via an unstated intermediate associate (e.g., "lion" primes "stripes" through intermediate activation of "tiger").

How do homophones, homographs, homonyms, and polysemous words differ?

Homophones share sound but differ in meaning and spelling (flea/flee). Homographs share spelling but differ in meaning and sound (wooden bow/low bow). Homonyms share sound and spelling with unrelated meanings (dog's bark/tree bark). Polysemy involves related senses of a single word (run a race/run a business).

How do bottom-up and top-down activation theories differ regarding lexical ambiguity resolution?

Bottom-up activation proposes that acoustic input activates all meanings of a word initially, and context selects the correct one later. Top-down activation proposes that context pre-activates the correct meaning so alternate meanings are never fully activated.

What did Swinney (1979) conclude about lexical ambiguity resolution using the word "bugs"?

Immediately after hearing "bugs", participants were faster to recognize both "spy" and "ant"; several words later, only the context-appropriate meaning ("ant") remained facilitated, supporting the bottom-up activation model.

How does the Cohort model (Marslen-Wilson, 1987) explain spoken word recognition?

All words matching the initial acoustic onset form an active cohort (e.g., "bea-" activates beaker, beetle, beacon). As successive sounds arrive, mismatching candidates drop out until reaching the uniqueness point where one word remains.

How does the TRACE model (McClelland & Elman, 1986) differ from the Cohort model?

TRACE proposes that incoming sounds activate matching words regardless of where the match occurs, giving no special status to word onsets, which allows rhyming words to compete (e.g., "beaker" activates "speaker").

What did Allopenna et al. (1998) observe when participants heard "Pick up the beaker"?

Eye gaze increased to both the onset competitor ("beetle") and the rhyming competitor ("speaker"), supporting TRACE's rhyme prediction, though more fixations went to "beetle", showing early onset preference.

What did Spivey & Marian (1999) find regarding bilingual lexical access?

Russian-English bilinguals hearing the English word "marker" looked at a stamp (Russian "marka"), indicating that lexical candidates across both languages remain active in parallel.

What do mondegreens reveal about the vulnerability of the Cohort model?

Misheard song lyrics show that if word segmentation boundaries are misidentified, the Cohort model uses an incorrect word-onset anchor, causing downstream word recognition to fail.

What evidence from shadowing tasks indicates that sentence parsing is incremental and eager?

People can shadow normal connected speech within $\approx 250\text{ ms}$, but are much slower when repeating nonsense strings, demonstrating that real-time structural parsing enables rapid processing.

What is a garden path sentence?

A sentence containing a temporary ambiguity that leads the parser to initially build an incorrect syntactic structure (e.g., reading "raced" as a main verb in "The horse raced past the barn fell" before crashing at "fell").

What do N400 and P600 ERP components reflect during sentence comprehension?

The N400 amplitude increases in response to semantically unexpected words, while the P600 amplitude increases when encountering syntactic anomalies or words that are difficult to integrate into the grammatical structure.

How does Garden Path Theory (Frazier & Fodor, 1978) explain sentence parsing?

The parser is modular: it initially builds the simplest possible structure using purely syntactic information, ignoring context, meaning, and frequency until a subsequent reanalysis stage if the initial parse fails.

How does the Constraint-Based Approach (MacDonald et al., 1994) explain sentence parsing?

Multiple syntactic structures are activated in parallel, with the parser immediately evaluating multiple informational sources (syntax, meaning, frequency, context, thematic fit, visual scene). Garden paths occur when constraints collectively support an incorrect parse.

How did Trueswell et al. (1994) show that thematic relations influence syntactic parsing?

"The patient rushed to the hospital died" produced garden path difficulty because patients can perform the action of rushing, whereas "The blood sample rushed to the hospital was contaminated" caused less difficulty because blood samples cannot be the agent of rushing.

What did Tanenhaus et al. (1995) find regarding visual context and syntactic parsing of "Put the apple on the towel into the box"?

In a one-apple context, listeners initially misparsed "on the towel" as the intended goal location; in a two-apple context, listeners immediately interpreted "on the towel" as a modifier distinguishing between the apples, avoiding the garden path.

How did DeLong et al. (2005) demonstrate phonological prediction during reading?

When participants read sentences like "The day was breezy so the boy went outside to fly…", seeing the unexpected indefinite article "an" instead of "a" elicited an enlarged N400 before the word "kite" appeared, showing that the specific upcoming word was anticipated.

What is surprisal (Hale, 2001; Levy, 2008), and what does it predict in reading?

Surprisal measures how unpredictable a word is given its preceding context; words with higher surprisal values cause longer reading times and elicit larger N400 amplitudes.

Why are object relative clauses harder to process than subject relative clauses according to Gibson's (1998) memory-based account?

Object relatives (e.g., "The senator who the reporter spotted shouted") require holding the initial noun phrase in memory across intervening material until reaching the main verb, placing higher demands on working memory.

What is the kindergarten path effect (Trueswell et al., 1999)?

Five-year-old children told to "Put the frog on the napkin into the box" frequently hopped the frog to the napkin first due to immature prefrontal cortex development, which limits their cognitive control to suppress their initial parse.

What did Smith & Wheeldon (1999) demonstrate regarding the scope of speech planning?

Speakers took longer to initiate speech when the subject phrase was complex ("The dog and the foot move up") than when the object was complex ("The dog moves above the foot and the kite"), indicating that speakers plan the initial phrase before starting articulation.

What is a lemma in models of language production?

An abstract mental representation of a word that contains its semantic meaning and syntactic category, but does not include its phonological sounds.

What evidence do word exchanges and sound exchanges provide for speech production stages?

Word exchanges respect syntactic categories across distances ("The tree fell from the apple"), whereas sound exchanges occur between adjacent words irrespective of category ("Shoving leopard"), demonstrating that lemma selection precedes sound retrieval.

How do picture-word interference results (Meyer, 1996) differ between semantically and phonologically related distractors?

A semantically related distractor slows down picture naming due to lemma competition, whereas a phonologically related distractor speeds up naming because the target word's sounds are pre-activated.

What does syntactic priming (Bock, 1986a) show about sentence generation?

Repeating a passive sentence increased the probability of producing a passive structure for an unrelated scene by $\approx 8\%$, demonstrating that abstract syntactic frames are primed independently of specific lexical items.

What is lexical bias in speech errors, and what are its two main theoretical explanations?

The tendency for phonological speech errors to form real words more often than expected by chance. It is explained either by bidirectional activation feeding back from sounds to words, or by an internal pre-articulatory monitor filtering non-words.

What did Motley, Camden & Baars (1982) find when spoonerisms would result in taboo words using the SLIP technique?

Participants made significantly fewer speech errors, exhibited longer response latencies, and showed elevated galvanic skin conductance, confirming that an internal monitor detects taboo errors prior to articulation.

What is the distributional hypothesis (Firth, 1957)?

The principle summarized by "You shall know a word by the company it keeps," stating that word meaning can be derived from the statistical patterns of surrounding words in text.

What are three major limitations of n-gram language models?

Sparsity (most possible combinations never occur), inability to generalize (unseen sequences receive zero probability), and failure to capture long-range linguistic dependencies as context tables explode in size.

What is a key architectural limitation of Word2Vec (Mikolov et al., 2013) that Transformers resolved?

Word2Vec assigns a single static vector to each word form (meaning river bank and financial bank share the same vector), whereas Transformers compute context-dependent representations using dynamic attention mechanisms.

How do Query, Key, and Value vectors operate within the Transformer attention mechanism (Vaswani et al., 2017)?

A current word's query vector is compared against preceding key vectors to compute attention scores; these scores then weight the corresponding value vectors to dynamically incorporate relevant contextual information.

How does the training data volume of modern Large Language Models compare to human linguistic exposure?

A $10\text{-year-old}$ child hears $\approx 100\text{ million}$ words, whereas GPT-3 was trained on $300\text{ billion}$ words and recent frontier models are trained on more than $100\text{ trillion}$ tokens.

What did Trott et al. (2023) find when testing Theory of Mind in GPT-3 using false belief scenarios?

GPT-3 tracked character knowledge and correlated with human responses, but unlike humans, it was significantly less sensitive to whether a character held a true versus false belief.