Chapter Five
Sentence processing
Outline
Introduction
Structural properties of sentences
- Statistical Approximations to English
- Where Do People Pause When They Speak?
Syntactic processing
- Syntactic Resolution Is Needed for Comprehension
- Surface Structure versus Deep Structure
- Competence versus Performance
- Syntactic Structure of Sentences
- Clausal Processing
Sentence parsing & syntactic ambiguity
- Local Ambiguity versus Standing Ambiguity
Models of sentence parsing
- Garden Path Model of Sentence Processing
- Constraint Model of Sentence Parsing
Meaning: The goal of sentence processing
Is syntax processed separately from meaning?
The role of prosody in sentence processing
On-line interactive models of sentence processing
- Shadowing and Gating Studies
- How On-line Is Gating?
Where does context operate?
Comprehension of nonliteral meaning
The role of memory in language processing
- Speech Perception and Lexical Identification
- Syntactic Parsing and Retention of Phrases
- Retention of Semantic Propositions
- How Specialized Is the Memory System Used for Sentence Processing?
A processing model of sentence comprehension
Summary
Key Concepts From the Book
I. The study of sentence processing examines how people are able to decode and understand speech at the sentence level. This is a difficult task because of the rapidity of connected speech and the fact that in fluent speech words run together and are not as clearly articulated as they are in isolation. (Page 228)
II. The predictable structure of ordinary sentences, coupled with the relatively limited number of words that are in common use, facilitates our ability to rapidly process speech. The average college-educated adult may have a speaking vocabulary of 75,000-100,000 words. However, the 50 most common words in the English language make up about 60% of all the words we speak and 45% of all the words we write. (Pages 229-230 )
III. Moray and Taylor (1960) demonstrated the predictive quality of English with an experiment in which they gave people several words of a sentence and asked them to fill in the next most likely word. The more words the subjects were allowed to see of the sentence, the more English-like the resulting sentence. Sixth order approximations, in which subjects were given five words and asked to guess the sixth, produce more grammatical results than second order approximations, in which the subjects were only allowed to see one word. (Pages 230-231)
IV. Intonation patterns and rhythms also help people decode the spoken word. Pauses in fluent speech may take up as much as 40-50% of speaking time. These pauses tend to occur before words of low probability in context, suggesting that the pauses are determined by linguistic content, not by the need to take a breath. (Pages 231-232)
V. Noam Chomsky's theory of transformational grammar attempts to explain the structure of language, positing that every sentence is composed of a surface structure which is the actual expression of its deep structure or meaning. Sentences can have the same deep structure and different surface structures, or different deep structures and the same surface structure. (Pages 232-233)
VI. A complete theory of sentence processing must take into account both competence and performance. The average speaker of English has the knowledge to form complete and grammatical sentences (competence), but most actual spoken sentences consist of incomplete and ungrammatical fragments (performance). A theory of performance must explain how we can understand speech, however incomplete and fragmentary. (Page 233)
VII. In order to understand a sentence, a listener must parse it to determine its syntactic structure. The words of a sentence can be grouped into phrases that form separate units. Trace theory says that parts of a sentence may move from one position to another, leaving behind a "trace" in the surface structure of the sentence. In order to understand the sentence, the listener must detect this trace, thus reactivating the semantics of the correct lexical antecedent. (Pages 233-235)
VIII. Analysis of the speed at which subjects understand sentences has shown that people tend to process sentences in smaller, more manageable units or chunks. (Pages 235-237)
IX. Models of sentence parsing address how the syntactic functions of individual words determine the overall syntactic structure of clauses and sentences. Syntactic ambiguity gives clues about general processing principles in language comprehension. Local ambiguity refers to cases where the syntactic function of a word is later clarified as we hear more of the sentence. Standing ambiguity refers to cases where the sentence remain ambiguous even when all of the lexical information has been received. (Page 238)
X. The garden path model of sentence processing says when we encounter ambiguous sentences, we process only one meaning. If this meaning does not work, we must go back and reparse the sentence. The constraint satisfaction model says that when we listen to ambiguous sentences, we unconsciously process both meanings. (Page 239)
XI. Two important principles in the garden path model of sentence processing are the late closure principle and the minimal attachment principle. According to the late closure principle, listeners and readers tend to hold off as long as possible in determining when they have reached a major clause boundary. This principle is a special case of the more general minimal attachment principle, which states that listeners attempt to interpret sentences in terms of the simplest syntactic structure that is consistent with output. (Pages 239-240)
XII. By studying reading time, theorists have been able to find support for both the garden path model and the constraint satisfaction model of sentence processing. In spoken language, speakers usually reduce or eliminate syntactic ambiguity through prosodic cues. (Pages 240-241)
XIII. Under ordinary circumstances, people strive to understand a sentence as quickly as possible, with the ultimate goal of developing a propositional representation of the utterance. While the surface structure of the sentence may be retained in short term memory, only the deep structure, or meaning, is retained in long term memory. In some cases (such as in jokes and insults) listeners do accurately remember the surface forms of sentences. (Pages 241-243)
XIV. Early theories of sentence processing were based upon the premise that language is processed serially, with phonology, lexical processing, syntactic processing and semantic processing following each other in strict sequence, and syntactic processing being conducted independently from semantic analysis (syntactic autonomy). (Page 244)
XV. Prosody, which consists of intonation patterns, word stress, pauses, and the lengthening of the final vowel in words prior to a clause boundary, plays an important part in language processing. (Page 245-246)
XVI. An interactive model of language processing posits that information from all levels of analysis can interact freely. An autonomous model would run independently at its own rate, slowing down only if it had to cross-check its operation against possible semantic interpretations derived from prior context. (Pages 246-247)
XVII. An interactive view begins with bottom-up processing, describing the way the listener analyzes the acoustic signal, processing it upwards into phonemes, words, sentence structure, and finally into meaningful utterances. Top-down processing describes how what we already know about what a sentence contains helps us to assemble meaning from bottom-up sources. (Page 247)
XVIII. On-line interactive models of sentence processing differ from serial models by saying that all levels of language processing interact freely and continuously with one another. Some interactive models say that top-down sources influence the recognizability of likely words even before acoustic signals are heard. Other models say that top-down processes only begin to operate after sensory information activates a list of potential words. (Pages 247-248)
XIX. Studies of the interaction of top-down and bottom up processes include both shadowing and gating. In shadowing studies, subjects are asked to repeat what they hear on a tape as it is being said. In gating studies, subjects are asked to guess a word both in context and out of context when parts of it are presented to them acoustically. (Pages 248-249)
XX. Shadowing studies show that people tend to spontaneously correct errors in pronunciation and grammar, often making these corrections within 200 milliseconds of the word's onset. Gating studies reveal that people are able to guess a word within 175 to 200 milliseconds of onset when presented in context. Out-of-context words take around 333 milliseconds to recognize. (Page 249)
XXI. Gating studies have also revealed that syntactic constraints have a smaller effect on supplying the correct word than do semantic constraints. (Pages 250-252)
XXII. Modularity theorists such as Forster, believe that operations such as lexical activation are performed automatically and are informationally encapsulated. Early modularity theorists proposed three separate processing systems devoted to language:
A. The lexical processor, which activates particular words based on phonological or orthographic input. The output of this process is passed along to
B. The syntactic processor, which conducts an analysis of the sentence to create a surface structure representation. The output then moves to C. The message processor, which extracts meaning from the surface structure. (Page 253)
XXIII. In contrast with interactive models, in Forster's conceptualization, no processor has information from operations conducted by higher-level processors. Whereas interactive models account for the speed of speech processing by presuming continuous interaction between levels of analysis, the autonomous model explains this speed by saying lexical analysis is not slowed by the need to integrate context as each word is processed. (Page 253)
XXIV. Cross-modal lexical priming studies have shown that even in a highly constrained lexical-semantic context, all meanings of ambiguous words appear to be activated. These findings have suggested to researchers that sentence comprehension systems consist of autonomous subsystems or modules that act automatically, uninfluenced by higher level processes. (Pages 254-256)
XXV. Some sentences, such as sarcastic statements, sentences employing idioms, metaphors and indirect requests, have non-literal meanings. Researchers have come up with two accounts of the processing of sentences with non-literal meaning. The first assumes a three stage process:
A. The individual determines the literal meaning of the sentence
B. The individual determines whether the literal meaning seems appropriate in the context
C. If the statement does not fit the context, the individual rejects the literal meaning and searches for a non-literal one. (Page 257)
XXVI. A second model for the comprehension of figurative language says that individuals do not need to process the literal meaning of a sentence before understanding the non-literal meaning. This theory is supported by cross-modal lexical priming studies that have revealed that the non-literal meanings of sentences are often as available to listeners as literal meanings. (Pages 257-258)
XXVII. Although most sentence processing occurs quite rapidly, some sort of short term memory buffer which stores as-yet unprocessed words and phrases seems to be a necessary component of sentence comprehension. These needs have been divided into three categories:
A. Cases of ambiguous acoustic input in connected speech or poorly articulated words would require the retention of phonological information.
B. Cases of local ambiguity would require a listener to retain a representation of the surface form of the sentence.
C. Cases in which the initial misinterpretation of an ambiguous word leads to errors in the formation of semantic propositions would require the listener to have a memory buffer. This buffer would operate whenever on-line sentence processing lags behind the input. (Pages 258-259)
XXVIII. Studies of brain-injured subjects with poor short-term memory have shown that despite severe short term memory deficits, many have relatively good sentence comprehension. The short-term memory required for sentence processing, then, appears to be quite complex, and is not easily measured by standard tests of memory capacity. (Pages 259-260)
XXIX. Other studies, such as those done with college students, have suggested that the memory structures used for storage and recall of such things as word lists are not the same memory structures used for on-line sentence processing. This may point to a special form of memory used for the semantic representations of sentences. (Pages 261-262)
XXX. A recent processing model of sentence comprehension proposes that in active speech perception, linguistic input is processed in cycles on a segment-by-segment basis. The phonological stream is rapidly recoded into propositions. Then, connections among these propositions are established. This relationship is referred to as a coherence graph. Finally, the propositions most important to the message structure are selected. (Page 262)
|
 |
|
|
