Experiments on Processing Lexical Blends

Experiments on Processing Lexical Blends*

Adrienne Lehrer and Csaba Veres

University of Arizona University of Bergen

Lexical blends have existed in English for a long time, but in recent years their creation has accelerated greatly, so much so that, at least in the US, it is hard to avoid new ones. Yet mainstream morphologists have treated them as marginal in spite of their increasing popularity. When they appear in context, however, they seem easy to process, so we decided to carry out some experiments to tap into the mechanisms of processing them. In 1997 and 1998 we carried out several experiments on how subjects process lexical blends under time pressure. Although this work was never published, the International Conference on Lexical Blending has provided us with an opportunity to present our results and to suggest ways that we might have continued with our project.

Our goal was to gain some insight into processing through psycholinguistic experiments. There is a large literature on how much speakers of a language decompose complex words, including compounds (the basis of lexical blends), and our hypothesis was that similar decomposition may take place with blends. The work built on earlier work of Lehrer (1996), in which subjects were presented with blends and asked to identify the targets (the two-word compounds underlying the blends) and to provide meanings. There were no time constraints in those studies.

The point of these new experiments was to try to find evidence for rapid, automatic processing in the decomposition of the blends. Thus, while the conscious task of overtly identifying the components of meaning and assigning an interpretation requires time and effort, the process might involve an initial, possibly preconscious stage where candidate words are automatically activated for further consideration. Since these are not reported through conscious introspection, more subtle experimental techniques are required. After establishing some basic properties of timed blend identification, we present two experiments. The first is a variation of the stem completion task and the second a lexical decision task.

1. Timed Responses and Stem Completions

Experiment 1 consisted of two parts: (1) identifying the underlying words in a blend and (2) a stem completion task. The first task that experiment we carried out was basically a replication of Lehrer's earlier experiment with the addition of time constraints. We based our method oon lexical decision tasks although our task was not a lexical decision task but rather a lexical interpretation task..

Experiment 1: First task. Pilot.

Method

Some blends were novel and others not. Subjects were 38 undergraduate students in linguistics courses, and they were given extra credit for their voluntary participation A pilot test was carried out with four subjects who were told that the purpose of the experiment was to identify the target words and interpret the blends. They were given several examples and also had some practice before the experimental stimuli were presented. Subjects sat in front of a computer screen and pushed a foot pedal to bring up an item on the screen. They were told to press the yes button as soon as they could identify and understand the blend, which ended the trial and recorded the elapsed time from the presentation of the blend. Although the apparatus also had a no button, subjects were told not to use it for this experiment. A tape recorder was placed beside the computer, into which subjects spoke the two target words and a gloss which they felt was appropriate for the blend. The four subjects needed a long time to respond, an average time of 2222ms. (Null responses were never included in our calculations.) In addition, the response time was quite variable. We hypothesized that much of the time might involve finding a plausible interpretation of the blend itself, so we eliminated this step from the instructions in the main experiment and told subjects only to identify the target words.

Experiment 1 Identifying the targets.

Subjects were 38 undergraduate students in linguistics courses, and they were given extra credit for their voluntary participation Subjects saw one of two lists of 40 blends with different blends on each list. The purpose of the two lists was for the stem-completion task, described below. Each list each contained four types of blends: word + splinter (oildralic < oil + hydraulic) splinter + word (narcoma < narcotic + coma), two splinters (dramedy < drama + comedy), and complete overlap (slanguage < slang + language). Some blends were novel, such as bombphlet and others not. As before, subjects brought up an item onto the screen by pushing a foot pedal, After pushing the yes button, the computer recorded the time, and the subjects spoke the target words into a microphone.

Table 1 presents the time that subjects took, which was even longer than response times for the pilot. The mean response time forThe 21 subjects who saw List 1 and the 17 who saw List 2 required almost three seconds.

Table 1 Response times for identification of y blend targets.

Pilot NS = 4 / List 1 NS = 21 / List 2 NS = 17
Mean response time / 2222 ms / 2986 ms / 2994 ms
Mean standard deviation / 680 ms / 828 ms / 595 ms

An ANOVA was carried out to compare the response time for different blend types, but the F value of .69 was not significant.

A conjecture for why reaction times are so long is that speakers automatically wait for an interpretation before responding, so the reaction times overestimate the amount of time it takes simply to recognize the component words Whether the two processes of identifying and then interpreting the two targets are done sequentially, overlapping, or in parallel is something for future research. In the subsequent experiments we tried to find out if the component words were actually available before the response was recorded.

1.1.2 Identifying the targets.

The spoken responses were transcribed and classified as correct. "Correct" responses were determined by their identity to the ads, magazine headings, and brand names which provided the blends we used, or they were conventional items, like infomercial. There were a small number of additional responses which were equally good (as judged by AL). For example, the target words for biologue were biography and dialogue, but biology was classified as plausible. However, when the additional items were added to the "correct" one, the results were not affected much. Table 2 summarizes the results.

Table 2: Percentage correct for each types of blend

Word + splinter / Splinter + word / Two splinters / Overlap
List 1 / 52% / 50% / 34% / 69%
List 2 / 79% / 77% / 54% / 59%

Subjects who saw List 2 did better than those who saw List 1, and this might beis due to the fact that List 2 was easier. Note that in spite of this, the reaction times did not differ in the previous analysis. Perhaps this is because even though the component words in List 2 were easier to recognize than in List 1, the reaction time is controlled by the time it takes to generate a gloss, as we have already suggested.

There was no attempt to match the two lists, and in retrospect we should have done this to explore more variables. The number of items with complete overlap was small, but the items were easy: On List 1: guestimate, slanguage, and clandestiny. On List 2 were palimony, globaloney, and selectric. Palimony pal + alimony was easily identified because at the time that word, coined by journalists, was in the news to describe a scandal. The actor Lee Marvin, who had a house in Tucson, had broken up with his partner. She was suing him because she claimed that he gave her a verbal promise to share his wealth as he would a wife in case of separation. She said he had broken his promise.

Experiment 1 Second task 1.2.1 Stem completions and implicit memory

The stem completion task attempted to determine if subjects remembered the parts of the blends, even in the absence of explicit recognition of those parts. That is, can we find evidence for the activation of a lexical item involved in a blend that is never explicitly recognized? The experiment was suggested by work on implicit memory. Graf and Mandler (1982), Graf and Schacter (1985), and Schacter (1987) summarize data showing evidence for two memory systems: implicit memory and explicit memory. Implicit memory is defined by Graf and Schacter (1985: 501) as facilitation on a task "in the absence of conscious recollection," whereas explicit memory is revealed when performance requires conscious recollection of previous experiences.

Graf and Mandler (1982) devised a study-task on memory. Subjects were given a stack of cards with one word on each card. There were two conditions: a semantic one, in which subjects were to decide how much they liked or disliked the words, and a non-semantic one where they subjects had to decide if the letters on one card were similar to those on the previous card. Then all subjects were given a list of stems, three-letter words that could be completed by an English word. Subjects were told to complete the stems with the first word that came to mind (implicit memory instructions). They were then asked to recall as many of the words from the cards as they could (explicit memory). Only those who performed the semantic task could recall some of the words; the other subjects recalled few. However, on the stem completions, subjects from both groups did equally well, and both completed the stems with the words they saw five times as often as a control group which had not seen the words.

Returning to our own experiment, immediately after the computer work was finished, subjects performed the stem completion task. They were given a list of stems, a sequence of two or thee letters that could be the beginning of possible English words. Most stems were three letters; two were used only when three would certainly trigger the target. Half the stems were from blends on List 1 and half from those on List 2. In addition, half of each of these could be the beginning of the first part of the blend and half the beginning of the second. Table 3 gives examples of the stems we used.

Table 3: Examples of stems

Stem / Blend / Target
COM__ / dramedy / comedy
BRA__ / branchizing / branch
PAM__ / bombphlet / pamphlet
DES__ / deskercize / desk
INC__ / coaccidental / incidental
HUR__ / hurricoon / hurricane

Subjects were instructed to complete the stem with the first word that came to mind beginning with those letters.

1.2.2 Stem completion results

There were significantly more stem completions based on the blends seen by subjects than chance, where chance was defined by the responses of those who had not seen the blend. (Each group served as the control for the other group.)

Table 4: Stems completions with blend target.

Saw / List 1 # = 21 / List 2 # = 17
L1 Target / 252 / 109
L2 Target / 77 / 122

Х2 = 52 p. < .001

In addition we calculated responses that used a morphologically related word. For example, in the blend successories, success + accessories, if a subject completed the stem SUC with succeed, that counted as related. However, most related words were plurals of singular nouns or inflectional variants of verbs.

Since 21 subjects saw List 1 and only 17 saw List 2, the numbers in Table 4 are adjusted for a better comparison by multiplying the responses to List 2 by 1.24 to show the percentages in Table 5. As can be seen, subjects who had been primed by seeing a blend on the computer screen were twice as likely to use the underlying constituentthat word to complete a stemet as the unprimed subjects.

Table 5: Percentage of primed and unprimed stem completions.

Primed / Unprimed
% / %
Target / 68 / 34
Target + Related / 65 / 35

In retrospect, we should have looked at the various subclasses of responses, especially to seelearn if there were significant differences between stems that could be completed by the first word of the target as opposed to the second word in the target.

1.2.3 Stem completions and identification of targets.

So far, we have shown that stem completions can be used to measure the effects of previously presented items in this task. But the really interesting question is still unanswered. That is, what happens to items that were not

fully processed to the point where subjects could give a coherent gloss? To answer the question we analyzed the oral response of subjects and correlated them with stem completions. Responses were categorized into four classes: (1) subjects identified the target and completed the stem with that word; (2) they identified the word but did not choose it for the stem; (3) they did not identify the word but completed the stem with it anyway,; and (4) they neither identified the word nor completed the stem with it. The fourth class contained a subclass where the subjects misidentified the target word and used that in the stem completion. There were five instances with subjects who saw List 1 in this subclass. Table 6 shows these results.1

Table 6: Correlation of stem completions and correct identification of targets.

Identified word
and chose for
stem / Identified word
but did not
choose for stem / Did not identify
word but chose
for stem / Did not identify
word or choose
for stem
No. % / No. % / No. % / No. %
Saw List 1 / 216 27 / 294 37 / 64 8 / 226 28
Saw List 2 / 101 18 / 271 50 / 28 5 / 147 27

The relevant case here is the items in the third column, cases where the subjects did not identify the word but used it on the stem completion anyhow. The total number for List 1 is 64, which is 8% of all responses, and for List 2, 28, which is 5% of all responses. Since the numbers are small, we feel these data are preliminary.2

1.3 Familiarity

Finally subjects were given a list of all the blends and asked which ones they had seen or heard before the experiment. Items were divided into those which over half of the responses were correct and those in which fewer than half were correct. Then the items in the first group wereas further divided into those in which over half the subjects reported having seen or heard the blend previously and those who had not. T-tests showed a significant difference in response times for List 1 (p < .01) but not significant for List 2. This mixed result suggests more investigation is needed, but even if there are significant differences, the phenomenon is compatible with two explanations: (1) blends are stored as parsed with pointers to target words, or (2) subjects had previously decomposed them and could do so again, but more quickly than the first time.

Unfortunately we left the project before completing more detailed analyses of our data and/or replicating the experiment by controlling for more variables, such as the frequency and neighbors of the target words. We could have also looked at each subject's protocol, such as familiarity instead of pooling the data. Also in retrospect we did not focus on the interesting category 3 results in which subjects did not identify the targets but chose that word for the stem anyway. Finally, we should have matched the two lists in the ways described at the end of the paper.

Experiment 2. Lexical decisions tasks with masked priming.

The second set of experiments was intended to explore rapid, automatic decomposition. Although the research showing that morphological decomposition for complex words and compounds is mixed, it still suggestsed that we might find evidence for automatic decomposition of blends.

This set of experiments used the lexical decision task with masked priming. In a lexical decision experiment, subjects are presented with a string of letters and they have to decide as quickly as possible whether the letters spell a word in their language or not. They then press a yes or no button, which records the time. Usually half the stimuli are words and the other half nonwords.

One variation uses a prime, which is an item preceding the target item which is either identical to the target or related, either phonologically, morphologically, orthographically or semantically. related. For example, if the target word is operate, then presenting operate again or operation or doctor earlier in the list will speed up a subject's response to operate. Forster (1985) has shown that an identical or related word presented can serve as a prime for up to 10 following words.

A modification is the use of a masked prime, which is a repetition effect that occurs when the same lexical item is accessed twice in rapid succession, usually for a very short time, e.g., 50 milliseconds. This effect is independent of word frequency and word type., Although difficult to detect, a masked prime produces a reliable effect in speeding up the response time for words, but not for nonwords (Forster, 1985).

Our hypothesis was that when words are primed by blends, subjects would correctly respond faster to words than would subjects who were not primed.