Sentence Processing

Major source: Mitchell, D. C. (1994). Sentence parsing. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 375-409). San Diego: Academic Press.

Major Issue: autonomy of language subsystems

Autonomous systems (modularity; e.g., syntactic parsing free (initially) of pragmatic, semantic, etc. effects) VS.

Interactive systems (systems not modular; syntactic parsing influenced (initially) by pragmatic, semantic, etc. effects)

Methodological Issues:

Products of processing (e.g., memory) may reflect

retrieval or representational effects rather than initial processing (e.g., Sacks effect, Bransford effects)

On-line measures: eye-tracking (word fixation duration), priming, reading time (moving window)

Current on-line measures are rough measures of processes that are extremely brief; difficult to determine when systems are in play (e.g., initial access vs. post-access), so difficult to evaluate modularity debate

Off-line measures: recognition memory, recall, question answering

Research emphasis on ambiguous sentences (unusual) in English language; Do results generalize to other unambiguous sentences and in other languages?

Lexical Ambiguity

Issue: are all meanings of polysemous words accessed at comprehension?

Early study: Foss (1970) - phoneme monitoring; longer following ambiguous words suggesting all meanings activated. Effect is very brief; disappears after two syllables.

E.g., The men started to drill before they were ordered to do so.

The sergeant ordered the men to drill before they were ready to do so.

Much evidence for activation of multiple meanings.

Even a biasing context can fail to restrict multiple activation.

E.g., Swinney (1979) cross-modal priming study:

Hear: Rumor had it that, for years, the government building has been plagued with problems. The man was not surprised when he found several spiders, roaches, and other bugs in the corner of his room.

Lexical decision: bug, spy, sew. Facilitation for bug and spy immediately, but not after 4 syllables.

Meaning frequency and context may interact:

Duffy et al (1988) eye gaze technique

For balanced (e.g., right) words, both meanings activated when context occurs after word; for polarized (e.g., yarn) only dominant meaning is activated.

When context occurs before word and favors subordinate meaning, both meanings of a polarized term are activated.


Multiple meanings appear to be activated in many situations

Appears to be automatic process

Ambiguities appear to be resolved during post-access checking.

Syntactic Ambiguity


The cop told the motorist that he had noticed .....

to what is "that he had noticed" attached (complement structure (VP) or relative clause (modifies motorist)) when it is encountered?

Parsing Possibilities (note similarity with lexical ambiguity):

1. Single structure activated; Garden path model or effect;

Serial processing models (Bever, Fodor); Modularity assumed.

Serial analysis without annotation

Serial analysis with annotation; options are tagged as possibilities; facilitates later reanalysis if initial choice is incorrect

2. Multiple structures activated;

Parallel processing

resource free; no processing costs for multiple structures

resource limited; processing costs

unbiased; all analyses equivalent

weighted; certain analyses are favored

3. Partial analysis; minimal commitment; wait and see;

Assume parsing temporarily suspended until encountering information that resolves ambiguity, at which point analysis runs and completes itself.

While suspended, low-level processing (e.g., NP and PP assembled) may continue but links (e.g., between NP and PP) are delayed

4. Hybrid models; different strategies in different circumstances.

E.g.: Just & Carpenter: parallel processing if resources available; serial strategy if overloaded (see Carpenter, Miyake, & Just, 1995, ARP)


Processing load in ambiguous region; predictions:

Tagged serial = predict slowing

Parallel; resource limited = predict slowing

Serial (no tags) = no difference

Parallel; resource free = no difference

Minimal commitment = faster processing

Results: no processing difference between ambiguous and matched ambiguous; supports no-tag serial and resource free parallel

Processing immediately after ambiguous region; predictions:

Sample stimuli:

The cop told the motorist that he had noticed to avoid overtaking the school bus.

*italicized = region after ambiguous region

Serial models = processing difficulty if initial analysis not supported; no difficulty if initial analysis continues

Parallel models = generally no effect predicted (multiple analyses have been activated)

If weighted, some difficulty if the analysis supported is not the one initially weighted

Minimal commitment = depends on the model's assumptions about what's maintained in ambiguous region; varying degrees of difficulty expected as some structure will need to be instantiated.


Strong support for processing increment in disambiguation region across a number of different domains. For example:

Preposed adverbial clauses (with vs. without commas):

After the young Londoner had visited(,) his parents prepared to celebrate their anniversary.

Reduced vs. Unreduced relatives:

The defendant (that was) examined by the lawyer turned out to be unreliable.

Data fail to support unbiased parallel models

Data support untagged serial,


Best support for untagged serial and biased parallel (no difficulty in ambiguous region, but difficulty in disambiguating region)

Mechanisms of Initial Choice

Weighted parallel, serial, and some minimal commitment models must predict initial parsing choices:

1. Strategies based on tree structure analyses

Serial, garden-path analyses; e.g., Frazier (1978)

Late closure: Where possible, attach incoming material to phrase or clause currently being processed

E.g.: Mike had a date with the sister of the professor who was named in the newspaper last week.

Late closure - attach relative clause (italicized) to professor rather than sister

Minimal attachment: attach incoming material into phrase marker being constructed using the fewest nodes consistent with the well-formedness rules of the language under analysis

E.g.: The criminal saw the cop with binoculars but the cop didn't see him.

An extra node is required if the PP (with binoculars) is attached to the NP (the cop with binoculars) rather than to the VP (the criminal used binoculars to see the cop).

2. Lexical frame-driven strategies;

No processing tree; rather, lexical properties influence assignment

E.g., Ford, Bresnan, & Kaplan, (1982): Verb frames specify the type of arguments they can take;

E.g., 'positioned' requires object NP and optional PP

"The woman positioned the dress on the rack"

3. Thematic assignment strategies (e.g., Pritchett, 1988; based on Chomsky);

Every NP has thematic role (e.g., agent, patient, goal, etc.) and parsing is based on the linking of these roles:

e.g., I gave my money to the bookie.

4. Discourse based strategies;

Initial choice based on extent to which interpretation fits with prior discourse

E.g; I saw the man with the telescope. (Prior discourse might constrain attachment to VP rather than NP)

5. Exposure based strategies;

initial choice based on statistical frequency of comparable structures. For example, in an English corpus most 'that' continuations are complement rather than relative structures; as a result the complement interpretation will usually be the initial choice

E.g: The cop informed the motorist that he had noticed ....

6. General processing cost strategies;

initial choice subject to multiple influences

E.g., Carpenter & Just (1992); high capacity readers take into account a broad range of potential constraints (lexical, syntactic, discourse, etc.); while low capacity readers are restricted to most immediate constraints (e.g., syntactic).


1. Tree structure; consistent support for late closure; RT for ambiguous NP when new phrase is not connected to current phrase.

As soon as he had phoned his wife started to pack

As soon as he had phoned his wife she started to pack.

But, punctuation and prosody (e.g., comma or pause after phoned in above) can have a large effect (but shouldn't according to strict tree structure parser)

Also, some cross-linguistic differences: English displays late closure tendency but Spanish does not (Cuetos & Mitchell, 1988)

E.g., Someone shot the son of the actress who was on the balcony with her husband.

English (late closure) attach relative clause to actress; Spanish speakers tend to attach it to son.

Less support for minimal attachment; sentences requiring nonminimal attachment take longer than controls requiring minimal attachment:

E.g.: Nonminimal attachment: The spy saw the cop with the revolver but the cop didn't see him.

Minimal attachment: The spy saw the cop with the binoculars but the cop didn't see him.

But: Semantic expectations (spy saw........binoculars) decrease RT rather than minimal attachment. Nonminimal attachment conforming to semantic expectations can be faster than minimal attachment that doesn't conform (Taraban & McClelland, 1988):

The couple admired the house with a friend but knew that it was overpriced (minimal attachment)

The couple admired the house with the garden but know that it was overpriced.

2. Lexical frame strategies;

Many results are ambiguous due to difficulty in separating out initial lexical/thematic effects vs lexical/thematic filtering (revising) effects.

However, some studies have demonstrated instances when lexical information did not have an immediate effect (lexical effects ignored)

E.g., effects of transitive (saluted) vs. optionally transitive (fainted)if lexical effects, it should be easier to process material when lexical item constrains what can follow (saluted), but this does not happen. Longer Rts for saluted than fainted:

After the private had saluted/fainted the sergeant decided to end the military drill,

3 & 4. Thematic role assignment and discourse-driven;

Like #2, difficult to determine at what point thematic roles and prior discourse come into play.

5. Exposure-based strategies;

Some evidence that statistical frequency does influence initial parsing

E.g., Corpus examination:

English - late closure

Spanish - 1st NP attachment

On-line studies support above

However, some disconfirming evidence:

Corpus - NP attachments

On-line - VP attachment (a la minimal attachment)

"I saw the man with the telescope"

Overall: Least ambiguous support for tree structure parsers and past experience models.

Checking Processes

Little known about checking of final sentence structure; there are numerous constraints that must be checked:

number matching between verbs and subjects

thematic role for noun phrase (only one possible)

compatability with discourse

Relevant questions:

Do checking processes operate without delay?

Are they applied in a consistent order?

Does failure result in starting from scratch?

Summary and Evaluation

Emphasis on ambiguous stimuli; ambiguous stimuli make relevant variables measurable; may not be major problem

Emphasis on English; more of a problem. Want to know about language processing, not English processing.

Cross-linguistic differences can serve as natural experiment (e.g., English-Spanish exposure based differences)

Time-frame problem: current methodology makes it difficult to determine time course unambiguously;