Copyright © 1999 Massachusetts Institute of Technology
One of the reasons that reading a good novel or listening to an interesting lecture can be a pleasurable experience is because we are (blissfully) unaware of the cognitive work we do in understanding individual sentences and relating them to the discourse context. Research in sentence processing investigates the cognitive mechanism (or mechanisms) responsible for the real-time computation of the structural representation that underlies comprehension of visual or auditory language input. Sentence processing involves the rapid integration of various types of information (lexical, structural, discourse, etc.) and research in this area is necessarily interdisciplinary, drawing on work in theoretical linguistics, computer science, and experimental psychology (for reviews, see Mitchell 1994 and Fodor 1995)
A basic finding (that accords with our intuition) is that computation is "incremental." That is, we structure the words as they are perceived rather than store them as a list that is later combined when there is a pause in the input. Incremental structuring has the clear benefit of keeping short-term memory burdens to a minimum. For example, try to memorize a list of twelve random words versus a twelve-word sentence. Even a complex sentence such as "Not all of the targets were hit by some of the arrows" is easier to remember than a random array of words (e.g., "arrows by were of not of targets all the some hit the"), precisely because we can structure the sentence as we read or hear it (even in isolation, and despite the interpretive difficulties). But there is a potential cost to incremental processing if, as is quite common in human language, the input is ambiguous and processing must proceed in advance of relevant information. In fact, one of the most remarkable characteristics of the human sentence processing mechanism (commonly termed "the parser") is its ability to structure ambiguous input in an efficient manner.
One major focus of research in parsing theory, and one that serves as a good illustration of work in a wide-ranging, complex field, is the study of the incremental processing of temporarily ambiguous word sequences. A sequence is ambiguous if it is compatible with more than one well-formed structural representation. A temporary ambiguity is one that is resolved later on in the sentence. Consider the sentences in (1).
The sequence in (1a) is temporarily ambiguous -- that is, the noun phrase the schedule after the verb may function either as the nominal object of the verb (as in (1b)), or as subject of a clause (e.g., the schedule was wrong in (1c,d)). Each function has a distinct structural representation. The ambiguity is due to the following properties of English: (i) verbs such as know may take either noun-phrase or clausal objects; and (ii) the word that, which can be an indicator of a clause, is not always obligatory (as a comparison of (1c) and (1d) illustrates). Given incremental processing, the sequence in (1a) presents the parser with a choice. It may either structure the phrase "the schedule" as the object of the verb or as the subject of a new clause. If it takes the first option, then the appearance of a verb (such as was in (1c)) will prove this choice to have been the wrong one. On the other hand, if it takes the clausal option, then the failure of a verb to appear (as in (1b)) may cause processing difficulty.
To investigate the parser's operations in ambiguities such as (1a), Rayner and Frazier (1987) used an eye-tracking technique to take precise recordings of subjects' eye movements as they read temporarily ambiguous sentences such as those above, with the (d) sentences serving as an unambiguous control. What they found was a significant slow down in reading rate ("fixation durations") after the noun phrase in the (c) sentences, compared to the (b) and (d) sentences. They interpreted this comparatively slow reading rate as evidence that the parser had incrementally structured the postverbal noun phrase as a direct object in the ambiguous (b) and (c) sentences, but they had to revise this analysis in the (c) sentences when the input after the noun phrase showed it to be incorrect. The comparison of reading times after the noun phrase in the (c) and (d) sentences is particularly informative. The presence of that in (1d), in conjunction with the following the, serves as an unambiguous indicator of a new clause. The fact that a slowdown in reading rate is only observed in the (c) sentences supports the hypothesis that this slowdown is due specifically to the parser's response to the ambiguity in (c) and not to the need to process a second clause in the sentence.
These results are consistent with the hypothesis that the parser operates in accord with a Minimal Attachment principle (Frazier and Fodor 1978). Minimal Attachment states that, when faced with an ambiguity, the parser structures (or, attaches) new input to the current representation with the minimal amount of additional structure. A direct object attachment of a postverbal noun phrase is less complex than an attachment that would require the additional structure associated with a new clause. Therefore, it is the one pursued by the parser. Frazier and Fodor (1978) argued that Minimal Attachment follows from the temporal processing advantage enjoyed by minimal structures compared to their more complex competitors. Gorrell (1995) has argued that Minimal Attachment effects are due to the parser incrementally reflecting a central property of syntactic structure: economy of representations (Chomsky 1995). The preference for minimal structure has been demonstrated for numerous structural ambiguities in English, as well as German (Gorrell 1996), Italian (DeVincenzi 1991), and Japanese (Inoue and Fodor 1995).
But syntactic structure is just one factor in the processing of ambiguous sentences (clearly demonstrated in a seminal paper by Bever 1970). Work within the constraint-based framework, inspired, in part, by computational models adopting a connectionist architecture, has stressed the contribution of specific lexical information rather than more general structural preferences in sentence processing (e.g., MacDonald, Pearlmutter, and Seidenberg 1994). This interest in the contribution of specific lexical information to the resolution of structural ambiguities can be illustrated with a brief description of the experimental results reported by Osterhout, Holcomb, and Swinney (1994). They recorded event-related brain potentials (ERPs) elicited while subjects read sentences similar to those in (1). ERPs are measured voltage changes in an electroencephalogram that occur within a specified time after the presentation of a particular word or other input. The particular ERP of interest here is the P600, so called because it is a positive-going waveform that peaks approximately 600 milliseconds after the event of interest. The P600 has been associated with the processing of input that represents either a syntactic violation or a disconfirmation of a prior structural choice.
Although a first experiment replicated the major findings of Rayner and Frazier (1987), the results of a second experiment, in which verb preference (for either a noun phrase or clausal object) was manipulated, reveal the important role of lexical information in processing ambiguous input. In this experiment, temporarily ambiguous sentences containing verbs which prefer noun-phrase objects (e.g., understand in The student understood the answer was easy) were contrasted with sentences containing verbs that prefer clausal objects (e.g., guess in The student guessed the answer was easy). Osterhout, Holcomb, and Swinney (1994) found that the P600 was elicited as a function of verb type rather than as a function of comparative structural complexity. Demonstrations of the parser's incremental utilization of specific verb information, which typically occurs before potential objects in English, highlights the importance of studying the processing of verb-final clauses, a common property of many of the world's languages (Schriefers, Friederici, and Kuhn 1995; Mecklinger, Schriefers, Steinhauer, and Friederici 1995).
In addition to syntactic and lexical factors, work by Crain and Steedman (1985), St. John and McClelland (1990), Altmann, Garnham, and Dennis (1992), and others have highlighted the important influence exerted by DISCOURSE context on the process of resolving structural ambiguities. For example, consider the sentences in (2).
In (2a), the phrase that he had risked his life for many people in similar fires is an assertion being told to some woman. In (2b), the phrase that he had risked his life for is a phrase specifying which woman was being addressed. As in (1), the sentences here display a temporary ambiguity between the two structures. Minimal Attachment predicts an initial preference for the assertion analysis in (2a). Altmann, Garnham, and Dennis (1992), using the eye-tracking technique, indeed found a reading-time advantage for sentences such as (2a) over (2b). But it is significant that they only found this in particular situations. That is, they found that the discourse context played an important role in determining how these sentences were read. Consider the fact that, if two women are mentioned in a conversation, then the use of distinguishing modification, as in (2b), would aid the listener in determining which woman was being referred to in a particular sentence. By testing sentences in contexts in which modification was useful in this way, Altmann, Garnham, and Dennis report that the initial reading time advantage for (2a) over (2b) disappeared. This indicates the rapid influence of discourse context on sentence processing.
Studies such as these indicate that sentence processing necessarily involves the efficient use of many different types of information. Exactly how these various information types interact in sentence processing is an important question in cognitive psychology. Sentence processing is center stage in the "modularity" debate -- that is, the extent to which information processing is accomplished by specialized subprocessors or by more general mechanisms (Fodor 1983; Marslen-Wilson and Tyler 1987; Friederici 1990). In addition to lexical, structural, and discourse information, the role of prosodic factors is also an increasing focus of experimental research (Ferreira and Anes 1994; Warren 1996).
Another important research topic, inspired by the work of Baddeley and colleagues (Baddeley 1986), is the role of short-term memory in sentence processing. As noted, the limited capacity of short-term memory is one driving force in incremental processing, and many researchers accord it a significant role in the processing of ambiguous or complex sentences (Just and Carpenter 1992; Waters and Caplan 1996).
Altmann, G., A. Garnham, and Y. Dennis. (1992). Avoiding the garden-path: Eye movements in context. Journal of Memory and Language 31:685-712.
Baddeley, A. D. (1986). Working Memory. Oxford: Oxford University Press.
Bever, T. (1970). The cognitive basis for linguistic structures. In J. R. Hayes, Ed., Cognition and the Development of Language. New York: Wiley, pp. 279-352.
Chomsky, N. (1995). The Minimalist Program. Cambridge, MA: MIT Press.
Crain, S., and M. Steedman. (1985). On not being led up the garden path: The use of context by the psychological syntax processor. In D. R. Dowty, L. Karttunen, and A. Zwicky, Eds., Natural Language Processing. Cambridge: Cambridge University Press, pp. 320-358.
DeVincenzi, M. (1991). Syntactic Parsing Strategies in Italian. Dordrecht: Kluwer.
Ferreira, F., and M. Anes. (1994). Why study spoken language. In M. A. Gernsbacher, Ed., Handbook of Psycholinguistics. New York: Academic Press, pp. 33-56.
Fodor, J. A. (1983). The Modularity of Mind. Cambridge, MA: MIT Press.
Fodor, J. D. (1995). Comprehending sentence structure. In L. R. Gleitman and M. Liberman, Eds., An Invitation to Cognitive Science, vol. 1: Language. 2nd ed. Cambridge, MA: MIT Press, pp. 209-246.
Frazier, L., and J. D. Fodor. (1978). The sausage machine: A new two-stage parsing model. Cognition 6:291-325.
Friederici, A. (1990). On the properties of cognitive modules. Psychological Research 52:175-180.
Gorrell, P. (1995). Syntax and Parsing. Cambridge: Cambridge University Press.
Gorrell, P. (1996). Parsing theory and phrase-order variation in German V2 clauses. Journal of Psycholinguistic Research 25:135-156.
Inoue, A., and J. D. Fodor. (1995). Information-paced parsing of Japanese. In R. Mazuka and N. Nagai, Eds., Japanese Sentence Processing. Hillsdale, NJ: Erlbaum.
Just, M., and P. Carpenter. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review 99:122-149.
MacDonald, M., N. Pearlmutter, and M. Seidenberg. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review 101:676-703.
Marslen-Wilson, W., and L. Tyler. (1987). Against modularity. In J. Garfield, Ed., Modularity in Knowledge Representation and Natural-Language Processing. Cambridge, MA: MIT Press, pp. 37-62.
Mecklinger, A., H. Schriefers, K. Steinhauer, and A. Friederici. (1995). Processing relative clauses varying on syntactic and semantic dimensions: An analysis with event related brain potentials. Memory and Cognition 23:477-497.
Mitchell, D. (1994). Sentence parsing. In M. A. Gernsbacher, Ed., Handbook of Psycholinguistics. New York: Academic Press, pp. 375-409.
Osterhout, L., P. Holcomb, and D. Swinney. (1994). Brain potentials elicited by garden-path sentences: Evidence of the application of verb information during parsing. Journal of Experimental Psychology: Learning, Memory, and Cognition 20:786-803.
Rayner, K., and L. Frazier. (1987). Parsing temporarily ambiguous complements. The Quarterly Journal of Experimental Psychology 39:657-673.
Schriefers, H., A. Friederici, and K. Kuhn. (1995). The processing of locally ambiguous relative clauses in German. Journal of Memory and Language 34:227-246.
St. John, M., and J. McClelland. (1990). Learning and applying contextual constraints in sentence comprehension. Artificial Intelligence 46:217-257.
Warren, P. (1996). Prosody and parsing: An introduction. Language and Cognitive Processes 11:1-16.
Waters, G. S., and D. Caplan. (1996), Processing resource capacity and the comprehension of garden path sentences. Memory and Cognition 24:342-355.
Altmann, G. T. M., Ed. (1990). Cognitive Models of Speech Processing: Psycholinguistics and Computational Perspectives. Cambridge, MA: MIT Press.
Baddeley, A. D., and S. E. Gathercole. (1993). Working Memory and Language. Hillsdale, NJ: Erlbaum.
Balota, D. A., G. B. Flores d'Arcais, and K. Rayner, Eds. (1990). Comprehension Processes in Reading. Hillsdale, NJ: Erlbaum.
Carlson, G. N., and M. K. Tanenhaus, Eds. (1989). Linguistic Structure in Language Processing. Dordrecht: Kluwer.
Carpenter, P. A., A. Miyake, and M. A. Just. (1994). Working memory constraints in comprehension: Evidence from individual differences, aphasia and aging. In M. A. Gernsbacher, Ed., Handbook of Psycholinguistics. New York: Academic Press, pp. 375-409.
Clifton, C., L. Frazier, and K. Rayner, Eds. (1994). Perspectives on Sentence Processing. Hillsdale, NJ: Erlbaum.
Garfield, J. L., Ed. (1987). Modularity in Knowledge Representation and Natural Language Understanding. Cambridge, MA: MIT Press.
Garrett, M. (1990). Sentence processing. In D. N. Osherson and H. Lasnik, Eds., Language: An Invitation to Cognitive Science, vol. 1. 1st ed. Cambridge, MA: MIT Press, pp. 133-175.
Garman, M. (1990). Understanding utterances. In Psycholinguistics. Cambridge: Cambridge University Press.
Mitchell, D. (1994). Sentence parsing. In M. A. Gernsbacher, Ed., Handbook of Psycholinguistics. New York: Academic Press, pp. 375-409.
Sells, P., S. M. Shieber, and T. Wasow, Eds. (1991). Foundational Issues in Natural Language Processing. Cambridge, MA: MIT Press.
Simpson, G. B., Ed. (1991). Understanding Word and Sentence. Amsterdam: Elsevier .
Copyright © 1999 Massachusetts Institute of Technology