User:Angelaa312/SentenceProcessing7

Sentence processing is way by which people decipher and understand meaningful communication via language. Although words comprise the basic building blocks of language, they do not express ideas on their own. Furthermore, they often express different ideas in different contexts. Thus, words alone do not encompass all of language. They must be organized into larger structures, therefore in order to comprehend language we must first be able to break down sentences and extract their meanings.

The way we break down and process sentences is called the Human Sentence Parsing Mechanism (HSPM). Different models of sentence processing propose different theories on the internal algorithm of the HSPM, but the procedure generally begins with parsing the sentence, assigning the each element to a linguistic category or part of speech. For instance, when we glance through a sentence and see “the” we recognize it as a determiner and expect a noun phrase to follow (Kimball, 1973). Parsing a sentence gives us an idea of how the words relate to one another and we use that information to populate a grammatical tree structure known as a phrase marker.

How we generate a phrase marker to represent the sentence is not obvious. Language is so powerful in part because it is extremely flexible, with multiple ways of expressing the same idea, both in terms of word choice and grammar construction. At the same the time, a single expression can often be interpreted to mean more than one thing. For instance, the sentence “Bill said he tickled Mary yesterday” can be interpreted to mean either Bill did the tickling yesterday or Bill did the saying yesterday. In terms of the grammatical tree structure, we are unclear whether the adverb yesterday attaches to the verb said or the verb tickled. This particular example is a question of high versus low syntactic attachment. Late closure and minimal attachment have been posited as rules for defining attachment (Frazier, 1987); however, context has been shown to influence the attachment preferences as well (Taraband and McClelland, 1988). Ambiguity such as this and others arise in sentence processing frequently and are a major phenomenon models have to account for.

Modular vs Interactive models

Modular Models

In modular sentence processing, the overall final comprehension is a result of the work of specific and individual modules. These modules are theorized to utilize sections of overall input, process that input, and then output the processed information to a higher order process. Combinations of modules allow for information to be processed in a way that higher order functions may find usable. When higher order processes use output information taken from modules, they may create meaningful and constructive conclusions that correspond to meaning.

It is generally understood that modules may be divided into syntactic and non-syntactic processing devices. By dividing processing of syntactic information, discourse information, and semantic information, individual modules are able to render clarified yet non-definitive renditions of input information. In serial module models, only one interpretation of sentence meaning may be processed at one time (Carroll, 2008). Models such as the Garden Path Model use modular serial processing to analyze information. In models such as this, syntactic information is processed before frequency, semantics, and context (Frazier, 1987). Once a syntactic conclusion is reached, it is then “proofread” by modules that analyze non-syntactic information. If there is a discrepancy between modules, the process restarts at the syntactic stage and reassesses analysis.

The flow of information in models such as Garden Path supports Feed-Forward architecture. Because it takes a number of modules to process information, and because modular processing is a serial process, it follows the modules output their analysis into other modules in order to advance cognitive analysis. Modules serve as building blocks in that each block connects and contributes to the integrity of the next. Every transition between modules in feed forward architecture corresponds to an increase in overall processing and complexity. When errors occur in serial processing, the system must discard all conclusive information produced after the error, and it must produce a different result at the level in which the error or discrepancy occurred.

Theories of reanalysis may explain why some sentences are harder to re-asses then others. Most theories predict the discrepancies between initial syntactic parsing and later non-syntactic information result in a reanalysis from the syntactic level. Degrees of difficulty of reassessment of sentence comprehension may, however, vary with the type of ambiguity presented. Research by Pritchett shows that ambiguous sentences which require the parser to adjust hierarchical relations in the sentence tree cause the parser to have more difficulty reaching a new conclusion (1992). Further research by Ferreira & Henderson has shown that the “further the head noun is from the point of dis-ambiguation, the stronger the processor commits to a thematic analysis…” (1991) (Gaskell and Altmann, 1997). This research supports the notion that, in ambiguous sentences, attachment between more proximal phrases is more common and punctually attained than attachment between more distant phrases.

Interactive Models

In interactive sentence comprehension, processing draws upon and analyzes information from syntax, frequency, context, and semantics simultaneously. Because of this parallel processing, multiple renditions of sentence comprehension may be considered at one time. Much unlike in modular models, correction in interactive models does not involve re-analysis and serial processing.

Ambiguity associated with interactive models is slightly more difficult to analyze than ambiguity in serial models. Because analysis is occurring in parallel, measures other than re-analysis must be used to come to a conclusion during ambiguous sentence processing. A common explanation for conclusions reached while using interactive models to understand ambiguous sentences involves the use of frequency. Frequency in appearance of definitions associated with words that have multiple definitions has been shown to create definitional weight associated with those more common definitions (MacDonald et al., 1994). When analyzing an ambiguous sentence, definitions of words that a parcer is more familiar with tend to be used initially and more frequently. It has even been shown through saccade studies that, while analyzing ambiguous sentences, parcers spend more time looking at words with varying definitional weights (Rayner & Duffy, 1986).

Frequency may also be associated with sentences phrases or fragments. In theory, sentence fragments that are produced through written words or utterances should hold more weight than fragments that are seldom spoken. Most interactive theories assume that sentence comprehension results are increasingly primed/activated until one conclusion completes activation, much like how TRACE theory activates word and utilizes lexical competition (even though TRACE is a modular system). McRae et al. demonstrate the effect of weighted sentence fragments by showing that, on average, people are able to more quickly complete sentences that they are more familiar with than those which are relatively novel (1998).

Most current research demonstrates evidence both for and against modular and interactive theories of sentence comprehension. Discussion of modern theories leads to an unresolved consensus as to which set of processes is actually used. For most major theories of sentence processing, there are several exceptions to theorized processing. It seems that, in addition to either modular or interactive processes, there are confounding variables that contribute to our ability to understand variations of sentences.

Serial vs Parallel Comprehension

Serial versus parallel comprehension is a distinction in how the HSPM operates. Serial comprehension argues that while the HSPM processes a sentence, it holds one dominant interpretation until there is a decisive piece of evidence that forces it to understand the sentence differently. Parallel comprehension is the idea that if ambiguities occur while processing a sentence, the HSPM activates and retains multiple potential interpretations until the ambiguity can be resolved.

Work by David Swinney (1979) has shown that parallel comprehension is more likely to be the case when it comes to semantic ambiguity. In the study, a cross-modal priming task was used to measure response times at the time of and a few seconds after an ambiguous word. People responded to all the ambiguous stimuli relatively quickly, suggesting that they have all been activated as they would be for parallel comprehension

Models for Sentence Parsing

There are a number of influential models of human sentence processing that draw on different combinations of architectural choices. Parsing can be either serial or parallel, taking information one-at-a-time or all simultaneously. Both models are evidence-based, and there is still debate as to how humans parse sentences, however there are some evident preferences humans have when parsing sentences (see Human Sentence Parsing Mechanism)

Garden path model

The garden path model is a staged serial modular parsing model. It assumes that the syntactic process is encapsulated and devoid of context. A sentence is processed with the sentence structure building up incrementally. This model takes into account word function, but it ignores word meaning. This means that semantic constraints do not apply initially (for example, an object that can't act is still expected to perform an action due to the syntactic context). A word is recognized, and a syntactic modulel is then built. The semantic module is then applied, and if necessary, the sentence is revised. Contextual and semantic factors influence processing at a later stage and can induce revision of the syntactic parse. Revision requires analysing the syntactic process from the start, which can be time and energy intensive.

This experience gives support to the immediacy principle, which states that readers immediately commit to making decisions regarding the meaning and function of a word in a given sentence (Carroll 2008). Research shows that fixation times on the last words of a "garden path" sentence are longer than the times on the earlier words, implying that readers misinterpret the later word and must adjust their reading of the sentence, giving evidence to the serial processing of the garden path model (Frazier and Rayner 1982).

It may be questioned, on what grounds does the garden path model make its decisions? When the parser encounters an ambiguity, it is guided by two different parsing strategies: that of late closure and minimal attachment.

Late Closure

The principle of late closure states that if two phrases are of equal complexity, then the phrase should be incorporated into the current constituent. For example, "John said he would leave yesterday" would be parsed as John said (he would leave yesterday), and not as John said (he would leave) yesterday (i.e., he spoke yesterday).

Minimal Attachment

Minimal attachment proposes that phrases should be attached into a sentence creating the simplest structure possible (that is, the one with the fewest phrasal nodes). For example, if given the sentence "He loved the girl and her dog...", we are more likely to interpret the sentence meaning that his love extends towards both the dog and the girl, as opposed to "her dog" being the beginning of a new and separate clause. Common examples of garden path sentences are found in headlines.

Constraint-based model

The constraint-based model, unlike the garden path model, is a parallel and interactive parsing model. All available information is used to contribute to the parsing of the sentence simultaneously. When a word is recognized, it is used to build both a syntactic and a semantic module, which are separate processes that interact with each other. Semantic influence impacts the sentence processing. This allows for "multiple possible interpretations", as opposed to garden path's required investment in one interpretation at a time. Context and sources available will inhibit or activate interpretations until one correct meaning remains.

Evidence for this model can be found in the application of lexical knowledge, where readers may prefer a more complexly structured sentence because it is more familiar. The more common and frequently encountered interpretations are more quickly activated than less common constructions. When given "A thief stole paintings in the night" and "A thief stole the paintings in the museum", a reader would prefer the latter, as paintings and museums are contextually closely related words. Likewise, sentence forms that are less frequently used (for example, using a passive particle with a verb that rarely uses passive particles) will take longer to be processed. Research also shows that listeners are quick to "direct their eyes to the referents of what they hear" (Allopenna et al. 1998).

Reading v. Hearing

Reading is an automatic process that takes a lot of cognitive resources to understand what we are reading and to be able to draw information from it. There are different systems for what we read; in the English language we use an alphabetic system in which each letter corresponds to a sound, but in other logographic languages (such as Chinese) every word has a different symbol or combination of symbols associated with it. We know that reading is an automatic process due to many studies, but one of the most famous studies demonstrating the Stroop Effect (Stroop, 1935). In this study, participants were asked to name the color of ink that a word was printed in. When the word was printed in the same color that the word spelled, participants were faster to respond than if the word/color pairings were incongruent (Stroop, 1935). This shows that it is not easy to separate meaning from words because the process is automatic, if it was a more effortful process, we would easily be able to change the meaning of what we are reading depending on the task at hand. In reading, there are three different types of words: pseudo words (pronounceable non-words), regular words (each letter uses its most common pronunciations), or irregular words (letter pairs do not have their most common pronunciations) (Harley, 2010). An example of each is “smeat”, “beef”, and “island,” respectively.

Two different models for reading are the dual-route model of reading and the triangle model of reading. According to the dual-route model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) there are two different routes that we can access to get from print to sound for comprehension. The direct route (lexical route) is used when the print activates an entry in our mental dictionary (lexicon) without a lot of effort; this route is faster and more effective for skilled readers and well-learned words. The second route in this model is the indirect route (non-lexical route) where the letter-sound correspondences are used for the grapheme-phoneme conversion (GPC) (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001). GPC is the process that takes place when reading new words in which the reader takes the word letter-by-letter, turns each letter into a sound for pronunciation. However, GPC does not work for irregular words because each letter in an irregular word does not correspond to the most readily accessible pronunciation (Harley, 2010). The dual-route model is effective because when we do not know a word, we would not be able to go straight to the lexicon; instead, we would have to use GPC. According to the triangle model of reading there is no lexicon, and there are not two separate routes (as shown in the dual-route model). The triangle model asserts that the sounds of words are patterns of activation across phonological units, the meanings of words are patterns of activation across the semantic units, and the print form of words are patterns of activation across the orthographic units (Seidenberg & McClelland, 1989). This triangle view of reading can explain how the pronunciations of regular, irregular, and pseudo-words are possible by using three separate types of units within the brain that account for the different types of word input (sounds, meanings, print).

When we are hearing someone speak to us, the processing of their spoken information is different than when we are reading the words. Word recognition is primarily a bottom-up (or data driven) process in which we use information that we are given to help us decipher and understand information. This “bottom up” information eliminates incompatible semantics that do not fit into the input so that we can determine what the speaker is trying to say. The Cohort model (Marslen- Wilson & Welsh, 1978) explains how syntax provides some of the constraints when we are hearing information. Prior context in the sentence along with the new input help the listener determine what words are being said. All of the spoken words activate a “cohort” of possible words that could be next in the sentence, and as the speaker continues to provide new input, the size of the cohort steadily decreases with more input. All of these spoken words have a point which they become unique and the listener has received enough information so that there is only one semantic option remaining; this is referred to as the “uniqueness point.”

The Cohort model is an example of an interactive model because higher-level processing effects lower-level processing in a “top-down” effect, which is not supported by the modular models of comprehension. The phonemic restoration effect (Samuel, 1997) supports this idea of interactive models because in his experiments, even though certain phonemes were omitted from sentences and replaced by other noise, the listeners were still able to attend to the sentence without any issues with the removed phonemes. Listeners were able to continue to hear the sentences without any interruption from the missing phoneme because they were able to use lexical knowledge (higher-level processing) to help with these missing phonemes (Samuel, 1997).

Thus far into psycholinguistics, the TRACE model has become the predominant model that represents how auditory input is processed into meaningful sentences. According to this model, there are three levels of processing units; feature level, phoneme level, and the word level. Each level is bi-directional and every node on each level is connected to the level above or below it. These connections can be inhibitory or excitatory; connections between levels are excitatory where as connections within levels are inhibitory. This model relies upon other activation-based models because the more information that we receive from the auditory input, the spreading activation takes place so that only the correct nodes are active and help determine the target word (McClelland & Elman, 1986).

Neuropsychology and Sentence Comprehension

Current research in neuroscience has identified several anatomical factors behind language and sentence comprehension. It has been well documented, since 1861 after Broca autopsied an aphasic’s brain, that the brain contains sectional areas associated with specific language attributes. It has also been shown that atrophy to these sectional areas have somewhat consistent effects on language function. Broca’s and Wernicke’s aphasics, who have developed atrophy in the infero-frontal cortex area and posterior middle temporal gyrus, respectively, display specific patterns of dysfunction. By examining this dysfunction, we are able to deduct the roles that these neural areas play in language functioning.

Anatomical Division

Both Broca’s and Wernike’s areas have been associated with language comprehension and production. These areas, however, interact via the articulate fasciculus, and play slightly different roles in language functioning. It is generally understood that Wernicke’s area plays a role in word recognition and lexicon storage, and that Broca’s area is more closely related to speech production and transmittal of language information to the prefrontal gyrus and motor network (Lichtheim, 1885). Research with aphasics strengthens the argument that Broca’s and Wernicke’s areas play somewhat specified roles in language. Broca’s aphasics tend to display dysprosodic and anomic symptoms, and they have difficulty producing written and spoken words. Their comprehension, however, is rather intact and functional. Wernicke’s aphasics, on the other hand, often display poor short-term memory and comprehension. Their speech is fluent but is marked by substitution errors, neologisms, and repetition difficulty. Strengthening the argument that Broca’s and Wernicke’s areas are connected and interact, damage to the articulate fasciculus results in dysfunction associated with both areas.

Interplay of Language Areas

By testing for syntactic processing and observing the brain and its activity though both functional and structural MRI imaging, Tyler et al. have shown that Broca’s and Wernicke’s aphasics are insensitive to syntactic preference, compared to a control (2011). These patients displayed deficiencies in both verbal communication and syntactical comprehension on their given task. This study has interpreted these results to demonstrate that the connections between Broca’s and Wernike’s area “… [play] an essential role in the neural network that carries out syntactic computations (Tyler et al., 2011).” Another study by Swabb et al. demonstrates a similar effect by observing sentence comprehension in aphasics (1997). This study used an EEG to analyze the reaction time of aphasics in the recognition of an ambiguous word. Results from the study demonstrate significantly longer reaction time in aphasics via a delayed N400 spike compared to a control group. Swabb et al. have determined that the results from this study demonstrate dysfunctional comprehension in aphasics related to “…integration of lexical information into a higher order representation of the preceding sentence context (1997).” Because the dysfunction of the aphasics from this study originates with the atrophy of Broca’s and Wernicke’s areas, Swabb et al. have demonstrated that these areas play a crucial role in sentence comprehension (1997).

References

Carroll, D. W. (2008). Psychology of language, 5th ed. California: Thompson West Inc.

Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance: The psychology of reading. Hillsdale, NJ: Erlbaum.

Pritchett, B. L. (1992). Grammatical competence and parsing performance. Chicago, University of Chicago Press.

Ferreira, F., & Henderson, J. M. (1991). Recovery from misanalysis of garden-path sentences. Journal of Memory and Language, 30: 725-725.

Gaskell, G., & Altmann, G. (2007). The oxford handbook of psycholinguistics. New York: Oxford University Press.

McRae, K., Spivey-Knowlton, M. J., Tanenhaus, M. K. (1998). Modeling the influence of thematic fit (and other constraints) in on-line sentence comprehension. Journal of Memory and Language, 38: 283-312.

Macdonald, M. C., Pearlmutter, N. J., Seidenburg, M. S. (1994). Lexical nature of syntactic ambiguity resolution. Psychological Review, 101: 676-703.

Rayner, K., Duffy, S. A. (1986). Lexical complexity and fixation times in reading: Effects of word frequency, verb complexity, and lexical ambiguity. Memory and Cognition, 14(3): 191-201.

Frazier, L. & Rayner, K. (1982) Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology 14, 178-210.

Treiman, R., Clifton, C., Jr, Meyer, A. S., & Wurm, L. H. (2003). Language comprehension and production. Comprehensive Handbook of Psychology, Volume 4: Experimental Psychology. Pages 527-548. New York: John Wiley & Sons, Inc.

Jens-Max Hopf, Markus Bader , Michael Meng , Josef Bayer (2003): Is human sentence parsing serial or parallel? Evidence from event-related brain potentials. Brain Research (2003) Volume: 15, Issue: 2, Pages: 165-177

Allopenna, P.D. Magnuson, J.S., & Tanenhaus, M.K. (1998). Tracking the time course of spoken word recognition using eye movements: Evidence for continuous mapping models. Journal of Memory and Language, 38, 419-439

Swinney, D. 1979. Lexical access during sentence comprehension: (Re) consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645-659

http://robotpedagogue.com/human-parsing.php

Swaab, T., Brown, C., Hagoort, P. (1997). Spoken sentence comprehension in aphasia: event related potential evidence for a lexical integration deficit. Journal of Cognitive Neuroscience, 9(1):39-66.

Tyler, L. K., Marslen-Wilson, W. D., Randall, B., Wright, P., Devereux, B. J., Zhuang, J., Papoutsi, M., Stamatakis, E. A. (2011). Left inferior frontal cortex and syntax: function, structure and behavior in patients with left hemisphere damage. Brain, 134: 415-431.

Lichtheim, L. (1885). On aphasia. Brain, 7:433-484.

Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Ziegler, J. (2001). DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review. 108, 204-258.

Seidenberg, M.S., & McClelland, J.L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523-568.

Marslen-Wilson, W., & Welsh, A. (1978). Processing interactions and lexical access during word recognition in continuous speech. Cognitive Psychology, 10, 29-63.

Samuel, A.G. (1997). Lexical activation produced potent phonemic percepts. Cognitive Psychology, 32, 97-127.

McClelland, J.L., & Elman, J.L. (1986). The TRACE model of speech perception. Cognitive Psychology, 18, 1-86.

Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journal Of Experimental Psychology, 18(6), 643-662. doi:10.1037/h0054651

Kimball, J. (1973). Seven principles of surface structure parsing in natural language. Cognition 2, 15-47.

Harley, T. A. (2010). Talking the talk: Language, psychology and science. New York, NY: Psychology Press.