Reading Span and Lexical Ambiguity Resolution

A Review of "Working Memory Constraints on the Resolution of Lexical Ambiguity: Maintaining Multiple Interpretations in Neutral Contexts" (Miyake, Just, & Carpenter, 1994)

Over the past few decades, researchers have continued to debate about how people process language. Much of the research has focused on how people resolve ambiguities. In 1977, Marslen-Wilson and Tyler conducted a study on syntactic ambiguity. The participants saw one of two sentence fragments: “If you walk too near the runway, landing planes…” or “If you’ve been trained as a pilot, landing planes…” In the first sentence, ‘landing planes’ is a noun phrase, whereas in the second sentence, ‘landing planes’ is a verb phrase. After hearing the fragment, either the word ‘are’ or ‘is’ appeared on a screen. ‘Are’ is an appropriate continuation of the former sentence; ‘is’ appropriately follows the second sentence. The results showed that the participants were slower to name the participle in the inappropriate context. This shows that we use context when processing syntax, which is counterevidence to the modularity theory, which posits that higher order processes cannot be influenced by lower level processes.

Swinney (1979) studied lexical ambiguity, and, contrasting Marslen-Wilson and Tyler (1977), his results supported modularity. Participants were aurally presented with a sentence such as “The man was not surprised when he found several bugs in the corner of his room.” In this sentence, ‘bugs’ is an ambiguous word – it could refer to an insect or a piece of spy technology. Immediately following the sentence, a three-letter word appeared on a screen that was either related to context one (ANT), context two (SPY), or unrelated to ‘bugs’ (SEW). While participants were slow at naming ‘SEW’ as a word, they named ‘ANT’ and ‘SPY’ as words equally quickly, suggesting that both meanings were primed. This held true even when the sentence was clear about the intended meaning (“The man was not surprised when he found several spiders, roaches, and other bugs in the corner of his room.”). This supports modularity in that context did not affect semantic processing.

Marslen-Wilson Tyler and Swinney’s findings were clearly in conflict. Therefore, Miyake, Just, and Carpenter (1994) chose to approach lexical ambiguity from a new angle. They sought to determine whether individual differences in working memory have an effect on the resolution of lexical ambiguity. In addition to supporting modularity, Swinney’s study supported the Many Meanings Theory, which says that when readers come across an ambiguous word, they hold both meanings in their head until they come across the disambiguating context that allows them to select the correct meaning. Using this as a starting point, Miyake et al. (1994) wondered if the amount of information and the length it is stored in one’s mind differs depending on their working memory capacity. Working memory, as defined by Miyake et al. (1994), “includes not only the storage but also the computational (original emphasis) component and is considered the site for both executing various language processes and storing intermediate and/or final products of comprehension” (p. 176). Therefore, those with a larger working memory span should be able to hold more items in their memory while still having resources available for computations such as resolving lexical ambiguities. Those with a smaller working memory capacity are less able to maintain information while performing computational tasks.

The capacity-constrained model of lexical ambiguity resolution has two components: (1) multiple meanings of an ambiguous word are activated simultaneously (as evidenced by Swinney), and (2) when there is no prior biasing context that informs the reader of the correct meaning, the level at which the meanings are activated is related to the frequency of the word in that language. Equibiased words are those in which both meanings of the word appear equally frequently in language. For example, ‘club’ equally means an organization/gang or a bat/weapon. Biased words have one meaning that is more prototypical of that word. ‘Boxer,’ for example, more often means an athletic fighter (dominant meaning) rather than a short-haired dog (subordinate meaning).

Experiment 1

In the present study, Miyake et al. (1994) hypothesized that people with smaller working memory spans will not be able to hold both meanings of an ambiguous word in their heads long enough (i.e., until they reach the disambiguating word), which in turn would slow down their comprehension. This result was not expected for people with larger working memories; the researchers anticipated that their comprehension would be significantly faster. More specifically, this effect was only expected when the correct meaning of the ambiguous word was the subordinate meaning. According to the capacity-constrained model, both low- and high-span readers activate both meanings upon reading the ambiguous word. The dominant meaning is activated at a much higher level than the subordinate meaning for both types of readers. Both meanings start to die out for both types of readers as time passes, but in low-span readers, the subordinate meaning dies out much more quickly than it does for the high-span readers. It is therefore believed to fall below the threshold before the reader reaches the disambiguating region. Thus, low-span readers must go back and re-retrieve the subordinate meaning, causing their comprehension to slow. Because the high-span readers still have the subordinate meaning in their working memory, they can immediately put it to use and quickly continue reading. Reading comprehension was not expected to differ between readers when the correct meaning was the dominant meaning, because they are more highly activated.

Method

Participants.
The participants were 96 college students. They were identified as low-, mid-, or high-span readers based on the Reading Span test (Daneman & Carpenter, 1980; as cited in Miyake et al., 1994). The participants heard sets of unrelated sentences and were asked to recall the last word of each sentence. The first set had two sentences, with each set adding one more sentence until the participant was unable to recall two words from the set. Scores were based on the largest set size that participants could correctly remember every word from three of five sets. Those with a score of 2.5 or less were identified as low-span readers (N 36), 3.0-3.5 were mid-span readers (N 28), and 4.0 and above were high-span readers (N = 32).

Materials.
The 64 stimulus sentences differed in three ways in each condition. The words were either biased or equibiased (referred to as Disparity), the meanings were either dominant or subordinate (Dominance), and the target words were either ambiguous or unambiguous (Ambiguity). An example of a sentence with an ambiguous, biased target with the subordinate meaning is, “Since Ken really liked the boxer, he took a bus to the nearest pet store to buy the animal.” Each sentence could be divided into four parts. The first was the opening region, which was the same for each condition (Disparity, Dominance, Ambiguity) and was neutral in regards to the intended meaning of the ambiguous word. In the example, ‘Since Ken really liked the’ was the opening region. Next was the target word, the debatable word (‘boxer’). The seven-word intervening region (‘he took a bus to the nearest’) came after the target word and before the disambiguating region. It was the same in all conditions, and did not lend itself to either meaning of the target word. The disambiguating phrase is the last piece of the sentence (‘pet store to buy the animal’). The first word of the phrase (‘pet’) indicates the intended meaning of the ambiguous word. It is always the eighth word after the ambiguous word.

Procedure.
The sentences were presented in a moving window paradigm. This means that each word of the sentence was presented on a computer screen one at a time. The participants had to click a button in order for the next word to appear. In this way, the researchers could measure the speed of comprehension. This method was believed to be an adequate measure of working memory because without the ability to go back and re-read the words that were previously presented, the participants are forced to store the information in their working memory. Additionally, a true/false question spontaneously appeared after half of the sentences. Because the participants did not know which sentences would be followed with questions, they were motivated to fully comprehend every sentence, and thus keep their working memory engaged.

Results

For biased homographs, there was no difference in the speed of comprehension based on working memory capacity when the correct meaning was the dominant meaning. However, for subordinate meanings low-span and mid-span readers were slower in processing the ambiguous sentences than the unambiguous controls. The low-span readers spent 1537ms reading the unambiguous sentence and 1746ms for the ambiguous sentence (a difference of 209ms). The mid-span readers showed the same effect, with the ambiguous sentence taking longer than the unambiguous sentence (1435ms, 1340ms, respectively; 95ms difference). The ambiguous sentences were only processed 57ms slower than unambiguous sentences in high-span readers (1402 ms, 1345 ms, respectively). The low-span readers slowed down the most upon reaching the last word of the sentence, and to a lesser extent, at the first word after the disambiguating word. The mid-span readers slowed down at the same places as the low-span readers, but did not struggle nearly as much at the end of the sentence. The high-span readers were essentially unaffected across the disambiguating region.

The equibiased ambiguities were much easier for the participants to resolve. The only significant effect was that mid-span readers slowed down at the word after the disambiguating word.

An Analysis of Variance (ANOVA) revealed that a significant interaction effect was found for Reading Span x Dominance x Ambiguity. In other words, the speed of comprehension differed based on reading span when the target word was ambiguous and the definition was the subordinate meaning. This effect was highlighted when the mid-span group was removed from the analysis.

Discussion

The results are consistent with the capacity-constrained model of lexical ambiguity. The participants with a high-reading span were able to maintain both meanings of the ambiguous word in their working memory, so the speed of processing was unaffected by the dominance of the word. However, low-span readers do not have the capacity to store both meanings in working memory for as long as the high-span readers can. They have a finite number of resources at their disposal, and therefore, to free up some extra space, one meaning rapidly dies out. Because the dominant meaning is initially activated at a higher level, and because prior knowledge tells us that the dominant meaning is more likely to be the appropriate meaning, it is the subordinate meaning that dies out. If this was in fact the intended meaning in that context, it would require a low-span reader to reevaluate the word and reactivate the subordinate meaning, thus causing the rate of comprehension to slow.

There are other possible explanations for the results of this study. One alternative is that neither low- nor high-span readers activate both meanings simultaneously, but that they only activate the dominant meaning initially. What distinguishes the high-span readers from the low-span readers is that they are better able to mentally backtrack and reassess the sentence and choose a new, appropriate meaning. However, this explanation is not entirely plausible because it fails to explain why low- and high-span readers performed equally quickly in the case of equibiased words. A second possibility is that high-span readers are also faster readers. This explanation can be refuted by the fact that high- and mid-span readers have equal reading rates, but high-span readers still had more rapid rates of comprehension in the biased homograph, subordinate meaning condition.

Experiment 2

The second study is essentially an extension of the first. Rather than comparing low- and high-span readers with each other, a within-subjects design was employed to compare each participant with him/herself. In this study, working memory was more directly assessed by varying the length of the neutral region. The shortened sentence, for example, was “Since Ken liked the boxer, he went to the pet store to buy the animal.” The neutral region is four words in the length (‘he went to the’). In the extended version, the sentence read, “Since Ken like the boxer very much, he went to the nearest pet store to buy the animal.” The neutral region in this sentence is seven words long. The model asserts that the shortened sentences will enhance the speed of comprehension. Not only do additional words consume more cognitive resources, but longer sentences leave more time for the subordinate meaning of a biased homograph to die out.

Method

Participants.
The participants were 22 college students. They were all identified as mid-span readers, according to the Reading Span test (Daneman & Carpenter, 1980; as cited in Miyake et al., 1994).

Materials.
Twenty-four sentences with biased homographs and 24 with equibiased homographs were used in this experiment. For each type, 12 sentences had long neutral regions and the remaining 12 had shortened neutral regions. Only subordinate interpretations were used in this experiment. True/false statements followed one-third of the sentences.

Procedure.
The same procedure that was followed in Experiment 1 was utilized here.

Results
Once again, the results largely supported the hypotheses predicted by the model. In the case of equibiased homographs, the length of the neutral region did not affect processing. However, the mid-span readers experienced difficulty comprehending the subordinate biased homographs in the long sentences. An ANOVA revealed that in the short sentence in the biased condition, reaction time did not differ as a function of ambiguity; in the long sentence condition, the ambiguous sentences took longer to process (1737ms) than the unambiguous sentences (1552ms).

Discussion
That the longer sentences took longer to process is further evidence in favor of the lexical ambiguity resolution model. As the neutral region lengthens, so does the amount of time required to maintain both interpretations in working memory. Thus, it makes sense with the model that the shorter sentences had faster comprehension at the homograph – both interpretations were still being activated at that point.

The two experiments presented above shed light on the way we process information and how we resolve lexical ambiguities. The method is the same for all types of readers (high-, mid-, and low-span), but the level of success differs. When we come across an ambiguous homograph with two or more possible meanings, both meanings are activated, with the dominant meaning being activated more strongly than the subordinate meaning. Someone with a high-reading span is able to maintain both interpretations in working memory until he/she reaches the disambiguating context and can choose the appropriate meaning. One with a low-reading span, however, does not have the cognitive capacities to maintain both meanings, and thus the subordinate meaning rapidly dies out, often before reaching the disambiguating context. This is just one example of how individual differences influence the language comprehension process.

Main Menu

See Also

Main Menu