UNDERSTANDING LANGUAGE
Comprehension of language requires the processor to use knowledge of the language (syntax), meaning (semantics), and our knowledge of the world (scripts) and inferences about the intentions of speakers (pragmatics).
The central questions for the study of the processing system are:
How and when are these sources of information called upon?
How is the architecture of the system organized?
Is syntactic analysis carried out first, and then meaning and interpretations ascribed later? Or are they all used at any point they might be needed?
There are too many studies in this area to present a full overview here. Instead we present just two sample problems (wordsense retrieval and nonliteral meaning) to indicate how the issues may be addressed experimentally.
Word sense retrieval
When reading or listening, it is important to retrieve word meaning, and that means retrieving the right sense of a word. This is an area where the role of background knowledge is important. For instance, in understanding ‘John put his salary in the bank’, it is necessary to select the appropriate sense of ‘bank’ – i.e. a place where financial transactions take place, not the side of a river. Context usually provides the solution to this problem, but the question is when during the sequence of processing? Is just one meaning of ‘bank’ selected at the outset, or are both meanings initially recruited, and then the right one selected later? There are two main possibilities:
1. The modular view is that word meanings are stored in a way that is not context sensitive. When we encounter a string of letters that represents a word, we automatically look up and retrieve the meaning. If the string (such as ‘bank’) represents more than one word, then both meanings should be retrieved.
The modular view is attractive because it keeps the mechanisms of looking up word meaning separate from context, and so is computationally simpler (see Fodor, 1983, for a discussion of this position).
2. The interactive view suggests that word meaning information is connected to other processes of comprehension, so that which aspects of word meaning are active depends on context. This view is attractive because it implies a very adaptive organization of knowledge and word meaning, but at the cost of more computational complexity (e.g. see McClelland and Rumelhart, 1981; Morton, 1969).
An important technique for finding the answer is priming (see Meyer & Schvaneveldt, 1971). When a word is read, it becomes easier to recognize words that are associated with it. So if you read the word ‘nurse’, you will then read the word ‘doctor’ more quickly than if you had just read an unrelated word, such as ‘bread’. What will be primed after reading the word ‘bank’? If there is no biasing context, then target words relating to both senses should be primed, such as ‘river’ and ‘money’. Swinney (1979) presented participants with spoken passages like these:
(a) Mary needed to buy some presents, so she went to the bank.
(b) Mary found the river cold, so she swam to the bank.
Immediately after the presentation of the ambiguous word, he presented a single letter string on a screen. Participants had to decide whether the letter string was a word or not (a lexical decision). When the string was a word, it could either be related to the intended sense of the ambiguous word (e.g. ‘money’), related to the other sense (e.g. ‘mud’), or unrelated to either. The question was whether there would be a response time advantage for the intended-sense associate alone, or whether there would also be an advantage for the other-sense associate of the word too.
It turned out that there was equal advantage (priming) for both senses. So context did not appear to affect initial sense selection. But if there was a delay of only 300 ms between hearing the ambiguous word and reading the letter string, the priming effect remained only with the intended (contextually cued) sense. This work suggests that word meaning information is initially stored in a modular fashion, and its retrieval is uninfluenced by context. On the other hand, very shortly after a word has been processed, contextual cues inhibit the activation of word sense information that is inappropriate. This one example represents a sample of work on the problem of modularity; research in this area remains very active (see Sanford, 1999, for a fuller review).
Nonliteral meaning
How do we understand sentences? One explanation is that we assign a literal meaning to them and then integrate this into the meaning of the discourse. But the literal meaning may not make any sense, especially if the sentence conveys an indirect speech act or a metaphor. For instance, if I say ‘My job is a jail’, I mean it restricts my freedom in a way that parallels being in jail. One prevalent view is that metaphors are first interpreted literally, then, if this fails, they are interpreted as nonliteral, or figurative (Searle, 1975, 1979). As a series of processing operations, this may be formulated as follows (from Glucksberg & Keysar, 1990):
1. Derive a literal interpretation of the utterance.
2. Assess the interpretability of that interpretation against the context of that utterance.
3. If that literal meaning cannot be interpreted, then and only then derive an alternative nonliteral interpretation.
The sequence above suggests that in order to make an appropriate interpretation of a statement, we need to know whether it is meant to be literally true or not. But it also makes strong assumptions about the processes underlying comprehension that subsequent work has suggested may be incorrect. The account has been examined for both indirect speech acts and metaphor comprehension. Gibbs (1979) showed that people take no longer to process indirect requests such as ‘Must you open the window?’ – meaning ‘Don’t open the window’ – than to understand literal uses of the same expressions (in the present case, meaning ‘Need you open the window?’). These data suggest that people do not need to obtain a literal meaning of an expression first in order to comprehend an indirect speech act. This runs against the traditional model (Glucksberg & Keysar, 1990). Gibbs (1983) claimed, more strongly, that participants do not always derive a literal meaning at any point. To establish this would be another blow to the traditional odel, since this model specifies that literal meanings are necessarily established. Gibbs had participants read stories that ended with critical sentences such as ‘Can’t you be friendly?’ In different stories, the sentence was given a literal meaning (‘Are you unable to be friendly?’) or an indirect interpretation (‘Please be friendly’). After reading a passage, participants had to decide whether a string of words was a grammatically correct sentence. Some of the strings were either the literal or the nonliteral interpretation of the critical sentence. Gibbs predicted that the literal context would prime the literal interpretation, and the nonliteral context would prime the nonliteral interpretation. These results should be reflected in a priming effect on the subsequent sentence judgement task. In two experiments, the results confirmed these expectations. In particular, when the context biased the interpretation of the critical sentence towards a nonliteral interpretation, there was no priming of the literal interpretation.
These findings show that the applicability of the standard comprehension model (Glucksberg & Keysar, 1990) is at best limited, although it is worth noting that the comprehension of sentences in stories (such as have been used in most of the studies reported here) and actual interactions in dialogue are very different situations, so we must guard against simplistic conclusions. Nevertheless, work on indirect speech act comprehension reinforces the view that literal interpretation is not always necessary. Similar findings have been obtained for metaphor comprehension. For example, Glucksberg, Gildea and Bookin (1982) asked participants to decide whether simple statements were literally true or false. For example, consider the statement ‘Some desks are junkyards.’ This is literally false, and so (according to the conventional model) the obvious metaphorical interpretation should not interfere with processing and the production of a ‘no’ response. Yet it does. A statement with an obvious figurative interpretation akes longer to reject as literally false than does a sentence with no obvious figurative meaning, such as ‘Some desks are roads.’ So, in the case of ‘some desks are junkyards’ it seems that the metaphorical meaning is computed automatically even though it is not needed, which indicates that testing for literal meaning cannot represent the previous, modular processing stage that the classic position would claim (see also Glucksberg and Keysar, 1990). Our sample of work on the comprehension of metaphors shows how simple response time studies can be used to evaluate the sequence of language processing events. The conclusions suggest that the straightforward classical view that literal interpretation takes place first, and then nonliteral interpretation takes place later if needed, is wrong.
Comprehension of language requires the processor to use knowledge of the language (syntax), meaning (semantics), and our knowledge of the world (scripts) and inferences about the intentions of speakers (pragmatics).
The central questions for the study of the processing system are:
How and when are these sources of information called upon?
How is the architecture of the system organized?
Is syntactic analysis carried out first, and then meaning and interpretations ascribed later? Or are they all used at any point they might be needed?
There are too many studies in this area to present a full overview here. Instead we present just two sample problems (wordsense retrieval and nonliteral meaning) to indicate how the issues may be addressed experimentally.
Word sense retrieval
When reading or listening, it is important to retrieve word meaning, and that means retrieving the right sense of a word. This is an area where the role of background knowledge is important. For instance, in understanding ‘John put his salary in the bank’, it is necessary to select the appropriate sense of ‘bank’ – i.e. a place where financial transactions take place, not the side of a river. Context usually provides the solution to this problem, but the question is when during the sequence of processing? Is just one meaning of ‘bank’ selected at the outset, or are both meanings initially recruited, and then the right one selected later? There are two main possibilities:
1. The modular view is that word meanings are stored in a way that is not context sensitive. When we encounter a string of letters that represents a word, we automatically look up and retrieve the meaning. If the string (such as ‘bank’) represents more than one word, then both meanings should be retrieved.
The modular view is attractive because it keeps the mechanisms of looking up word meaning separate from context, and so is computationally simpler (see Fodor, 1983, for a discussion of this position).
2. The interactive view suggests that word meaning information is connected to other processes of comprehension, so that which aspects of word meaning are active depends on context. This view is attractive because it implies a very adaptive organization of knowledge and word meaning, but at the cost of more computational complexity (e.g. see McClelland and Rumelhart, 1981; Morton, 1969).
An important technique for finding the answer is priming (see Meyer & Schvaneveldt, 1971). When a word is read, it becomes easier to recognize words that are associated with it. So if you read the word ‘nurse’, you will then read the word ‘doctor’ more quickly than if you had just read an unrelated word, such as ‘bread’. What will be primed after reading the word ‘bank’? If there is no biasing context, then target words relating to both senses should be primed, such as ‘river’ and ‘money’. Swinney (1979) presented participants with spoken passages like these:
(a) Mary needed to buy some presents, so she went to the bank.
(b) Mary found the river cold, so she swam to the bank.
Immediately after the presentation of the ambiguous word, he presented a single letter string on a screen. Participants had to decide whether the letter string was a word or not (a lexical decision). When the string was a word, it could either be related to the intended sense of the ambiguous word (e.g. ‘money’), related to the other sense (e.g. ‘mud’), or unrelated to either. The question was whether there would be a response time advantage for the intended-sense associate alone, or whether there would also be an advantage for the other-sense associate of the word too.
It turned out that there was equal advantage (priming) for both senses. So context did not appear to affect initial sense selection. But if there was a delay of only 300 ms between hearing the ambiguous word and reading the letter string, the priming effect remained only with the intended (contextually cued) sense. This work suggests that word meaning information is initially stored in a modular fashion, and its retrieval is uninfluenced by context. On the other hand, very shortly after a word has been processed, contextual cues inhibit the activation of word sense information that is inappropriate. This one example represents a sample of work on the problem of modularity; research in this area remains very active (see Sanford, 1999, for a fuller review).
Nonliteral meaning
How do we understand sentences? One explanation is that we assign a literal meaning to them and then integrate this into the meaning of the discourse. But the literal meaning may not make any sense, especially if the sentence conveys an indirect speech act or a metaphor. For instance, if I say ‘My job is a jail’, I mean it restricts my freedom in a way that parallels being in jail. One prevalent view is that metaphors are first interpreted literally, then, if this fails, they are interpreted as nonliteral, or figurative (Searle, 1975, 1979). As a series of processing operations, this may be formulated as follows (from Glucksberg & Keysar, 1990):
1. Derive a literal interpretation of the utterance.
2. Assess the interpretability of that interpretation against the context of that utterance.
3. If that literal meaning cannot be interpreted, then and only then derive an alternative nonliteral interpretation.
The sequence above suggests that in order to make an appropriate interpretation of a statement, we need to know whether it is meant to be literally true or not. But it also makes strong assumptions about the processes underlying comprehension that subsequent work has suggested may be incorrect. The account has been examined for both indirect speech acts and metaphor comprehension. Gibbs (1979) showed that people take no longer to process indirect requests such as ‘Must you open the window?’ – meaning ‘Don’t open the window’ – than to understand literal uses of the same expressions (in the present case, meaning ‘Need you open the window?’). These data suggest that people do not need to obtain a literal meaning of an expression first in order to comprehend an indirect speech act. This runs against the traditional model (Glucksberg & Keysar, 1990). Gibbs (1983) claimed, more strongly, that participants do not always derive a literal meaning at any point. To establish this would be another blow to the traditional odel, since this model specifies that literal meanings are necessarily established. Gibbs had participants read stories that ended with critical sentences such as ‘Can’t you be friendly?’ In different stories, the sentence was given a literal meaning (‘Are you unable to be friendly?’) or an indirect interpretation (‘Please be friendly’). After reading a passage, participants had to decide whether a string of words was a grammatically correct sentence. Some of the strings were either the literal or the nonliteral interpretation of the critical sentence. Gibbs predicted that the literal context would prime the literal interpretation, and the nonliteral context would prime the nonliteral interpretation. These results should be reflected in a priming effect on the subsequent sentence judgement task. In two experiments, the results confirmed these expectations. In particular, when the context biased the interpretation of the critical sentence towards a nonliteral interpretation, there was no priming of the literal interpretation.
These findings show that the applicability of the standard comprehension model (Glucksberg & Keysar, 1990) is at best limited, although it is worth noting that the comprehension of sentences in stories (such as have been used in most of the studies reported here) and actual interactions in dialogue are very different situations, so we must guard against simplistic conclusions. Nevertheless, work on indirect speech act comprehension reinforces the view that literal interpretation is not always necessary. Similar findings have been obtained for metaphor comprehension. For example, Glucksberg, Gildea and Bookin (1982) asked participants to decide whether simple statements were literally true or false. For example, consider the statement ‘Some desks are junkyards.’ This is literally false, and so (according to the conventional model) the obvious metaphorical interpretation should not interfere with processing and the production of a ‘no’ response. Yet it does. A statement with an obvious figurative interpretation akes longer to reject as literally false than does a sentence with no obvious figurative meaning, such as ‘Some desks are roads.’ So, in the case of ‘some desks are junkyards’ it seems that the metaphorical meaning is computed automatically even though it is not needed, which indicates that testing for literal meaning cannot represent the previous, modular processing stage that the classic position would claim (see also Glucksberg and Keysar, 1990). Our sample of work on the comprehension of metaphors shows how simple response time studies can be used to evaluate the sequence of language processing events. The conclusions suggest that the straightforward classical view that literal interpretation takes place first, and then nonliteral interpretation takes place later if needed, is wrong.
No comments:
Post a Comment