Detection and identification of words and letters in simulated ...

Memory & Cognition 1975, Vol. 3 (2),175-182

Detection and identification of words and letters in simulated visual search of word lists*

IRA FISCHLERt

Stanford University, Stanford, California 94305

Ss were shown a rapid sequence of words and had to (a) make a speeded response to the presentation of a predefined target and (b) report a "response word" which they thought immediately followed the target in the sequence. In Experiment I presentation rate was varied orthogonally with the number of alternative targets. Detection errors and latency increased with target set size, as did the distance of the response word from the target in the list sequence. Increases in presentation rate produced greater target-response distance without affecting detection time. In Experiment II Ss were in some conditions given only the initial letter of the targets; the response could be either the whole word following the target or only its initial letter. The results indicated that Ss could, within limits, concurrently detect initial letters and identify words. Alternatives to hierarchical-type models of stimulus processing in visual

search were discussed.

In a typical visual search experiment, a person scans a list of items as rapidly as possible for the presence of a predefined target. According to Neisser (1963, 1967), target defection in such a search may be done by checking the visual features of the list items for the presence of critical target features (e.g., the hook in

"Q"). The efficiency of such a feature scan is obtained

at the price of failing to identify list items other than the target. In Neisser's (1967) description of visual search, detection and identification of stimuli are arranged hierarchically, with identification not taking place until a positive detection has been made. Since the "preattentive" detection precedes and determines identification of a stimulus, the target will typically be

the only stimulus identified by the S. In fact, Ss can rarely identify items other than the target from lists they have just scanned (Neisser, 1967, p. 70).

Shulman (1971) has pointed out that the high error rate obtained by Neisser's Ss (Neisser & Beller, 1965) probably meant that many items are not even visually registered at the faster scanning rates, making identification impossible. The unavailability of nontargets at the end of the trial could also reflect very rapid forgetting of items not needed for performance of the task. However, certain studies have suggested that, even when the problem of short-term forgetting is bypassed, having to identify each list item in turn results in poorer performance. Neisser and Beller (1965) showed that, when a person is given a specific target such as the letter "K" or the word "MONDAY," search rate is significantly faster than when the target is defined semantically, such as "any animal." A related effect is

'This research was supported in part by a NASA grant (NGRo0500Z0-507) to R. C. Atkinson and was carried out during the tenure of a Dorothy Longmire Graduate Fellowship. It is based on a doctoral thesis submitted to Stanford University in 1973. The author gratefully acknowledges the assistance of his advisor, R. C. Atkinson, and the members of his doctoral committee, E. E. Smith, H. H. Clark, and D. H. Lawrence.

tNow at the Department of Psychology, University of Florida, Gainesville, Florida 32611.

obtained in tasks which require a search for the absence

of a target. Neisser (1963), for example, showed that search for a target string which did not contain the letter "Q" was much slower than a search for its presence. Lawrence (1971) reported a similar effect regarding search for semantically defined targets (i.e., a nonanimal word embedded in a list of animal names).

Evidence against this hierarchical arrangement of a fast preattentive detection and subsequent identification has been reported in a number of studies. Sasaki (1970) found that, when the targets belonged to the same category (e.g., pieces of furniture), the number of targets simultaneously searched for had less effect on search rate than if the target words were unrelated, suggesting that semantic information could be used to decide quickly that an item was one of the targets. Similarly, Egeth, Jonides, and Wall (1972) found that, when Ss searched for "any digit" in a circular array of letters, detection time did not vary with the number ofletters in the display. Jonides and Gleitman (1972) obtained the Egeth et al (1972) context effect with the single symbol "0" depending on how it was specified prior to search. Finally, Brand (1971) obtained an effect of context for letters and digits similar to the Egeth et al (1972) result in a visual search task, and in fact found that search was actually faster for a target when it was defined categorically (i.e., "any digit") rather than specifically (cf. Ingling, 1972; Sperling, Budiansky, Spivak, & Johnson, 1971).

These findings suggest that under some circumstances categorical or semantic information can be used to

increase the efficiency of visual search. Brand (1971)

suggested that examination of each stimulus in a detection task was not so much an abbreviated form of the identification process as a selective analysis which could extract anything from visual to more complex semantic information, depending on the task and on subjective strategy. The same point is implied by Shulman's (1971) finding that recognition of nontarget

175

176 FISCHLER

items following search depended, not on the amount of time spent on a given item, but on the nature of the search task (i.e., physically vs semantically defined targets).

A striking finding reported by Lawrence (1971) showed that positive target detection does not necessarily lead to the identification of that particular target. Lawrence asked Ss to identify which word in a rapidly presented sequence of words was capitalized. Errors frequently occurred at the faster presentation rates; particularly, Ss most often reported the word immediately following the capitalized word as the one they saw in capitals, and they were quite confident of this identification. This could not occur if the stimulus information used in the detection task were then subjected to a more elaborate analysis in order to be identified. Lawrence (1971) suggested that, by the time a detection response had occurred, the target had been replaced by the following stimulus, and it was this stimulus which was then given the analysis appropriate for identification. Apparently, a distinction must be made between the arrangement of tasks required of an S-which often involve a detection response followed by an identification-and the processing of a particular stimulus.

In studies where detection and identification are

based on the same stimulus (the target word), examining

the relationship between the two processes is difficult: If

the target IS defined uniquely, identification is in a sense

trivial, since the S already knows the item to be identified. If the target is semantically defined, the identification is still constrained to a particular category.

The present study attempted to separate detection and identification experimentally by requiring Ss to both locate a predefined target and identify a different item on the list. This was done by using what Lawrence (personal communication) has called a "successive recognition" task, where the S is given a target for detection and asked to identify the word immediately following the target in a rapidly presented sequence of words. Using such a task, Fischler (1972) found that the "response word" given was, on the average, later in the presented sequence relative to the target when the target was a one-syllable word than when it contained two syllables. This effect was interpreted by Fischler (1972) to mean that target detection for a one-syllable word was taking longer than it was for a two-syllable word. With a slower detection response, more items will have been presented by the time the detection is made, increasing-the probability that the word identified as the one following the target will occur later in the list sequence. It was also noted that many response words had actually been presented prior to the target. Apparently, Ss did not wait until a positive detection had been made to begin identifying other list items.

The first experiment was designed to establish that longer detection latencies would produce the shift in the distribution of response words observed by Fischler

(1972) and, second, that changes in the effective rate of word identification would produce similar shifts in the distribution, independent of the detection latency. This was done by orthogonally varying the number of alternative targets and the presentation rate in a successive recognition task.

EXPERIMENT I

In his seminal experiment, Sternberg (1966) gave Ss one to six target digits on a given trial, then presented a single test digit to them. He measured the latency to decide whether or not the test digit was a member of the target set. Latency for both positive and negative trials was a linearly increasing function of the number of items in the target set. The basic result of increasing latency with larger target sets has been replicated in a great number of studies using a variety of materials.

In a study using a sequential presentation technique, Sternberg and Scarborough (1969) repeated the Sternberg (1966) experiment, varying the number of digits in the target set from one to six but, instead of a single test digit as a probe, they presented a sequence of 20 digits at the rate of 70 msec/item and measured reaction time (RT) to the presentation of a target digit in the test sequence. The RT was found to be a linearly increasing function of set size, as in Sternberg (1966),

with comparable slope and intercept (41.5 msec/item +

351 msec). In Experiment I, target set size was varied by givingSs

one, two, or three target words on a given trial. A RT response to the presence of a target word in the list was included for comparison with Sternberg and Scarborough (1969) results and to confirm that the set size manipulation was increasing the time needed for target detection. It was predicted that, as set size increased, mean RT for target detection would increase and that the response words would, on the average, be words later in the sequence relative to the target.

Presentation rate was varied across blocks of trials by using rates of 6, 9, or 12 words/sec in a block. (Previous work suggested that this range would enable Ss to detect a single target almost perfectly and choose the correct response word about half the time.) If the successive recognition task involves relating target detection to the ongoing identification of list items, then the response word distribution should shift both with increased set size-which makes the detection time longer-and with increased presentation rate-which, for a given detection time, increases the number of words that have been presented by the time a detection has been made.

It was predicted that the function relating RT to set size would be unaffected by changes in the presentation rate. Several studies have found detection performance to be independent of the rate of successive item presentation (e.g., Eriksen & Spencer, 1968; Sperling et al, 1971). Sternberg and Scarborough (1969) suggested that the analysis needed for target detection

DETECTION AND IDENTIFICATION OF WORDS AND LETTERS IN VISUAL SEARCH

177

could begin for a new item before completion of the

preceding item. They had found the slope of the RT/set

size function in their simulated search task to be about

40 msec/item, comparable to rates found in single probe

studies. This implied that, for set sizes of two or greater, the total comparison time exceeded the lSI of 70 msec.

However, the error rate did not increase substantially

with set size; nor did RT deviate from the simple linearly

increasing function, as it should if a "backlog" of list

items was building up with the larger set sizes. The

authors concluded that a serial scan of the target set was taking place in parallel for successive stimuli. Whether or

not target detection in visual search takes the form of a

memory scanning comparison, it is true that, within a

broad range, target detection is not adversely affected by increasing the rate of item input, so we are led to expect

that the change in presentation rate should have no

effect on the RT/set size function.

Method Stimulus Materials. A pool of 320 five-letter words (170

one-syllable and 150 two-syllable) were selected from the Kucera and Francis (1967) norms, ranging in frequency from 10 to 100 per million. Words with homophones were avoided. An IBM 360/67 computer was programmed to generate 270 lists of 12 words each from the word pool, with the constraint that no two words on a given list begin with the same letter. For each S protocols were generated as follows: The 270 lists were randomly permuted and divided into nine blocks of 30 trials each. Within each block, set size (1, 2, or 3) and target position in Lists 3-8 were randomly varied, giving 10 trials for each set size and five trials at each serial position. Additional targets were assigned where needed, also with the first-letter constraint. One target appeared on every list.

Each list was typed on a 22 x 28 cm sheet of white paper in a single column in the center of the sheet, with triple spacing. The lists were typed with an IBM Selectric II typewriter in Gothic type font (which was considered to have a minimal amount of "noise" among the letters) with all words in lowercase. A masking stimulus was placed at either end of the list, consisting of two superimposed strings of five different consonants; this was done to prevent the first and last items from being strongly favored in perceptual analysis, as Lawrence (1971) found with unmasked lists.

Apparatus. The method of visual search has typically involved presenting an entire list to the S, measuring latency to target detection, and estimating search rate per item from the function relating the serial position of the target to response latency. This method does not control the amount of information taken in during each eye fixation. With a sequential presentation, each item is presented foveally and in the same spatial location for a brief time. This successive presentation of items at very fast rates was obtained by use of an n-channel stroboscopic tachistoscope, which has been described in detail by Lawrence and Sasaki (1970). The typed stimulus list is held by an aluminum slide that has a 5 x 23 em vertical area removed in the center. The slide moves vertically in a rectangular 30.5 x 60.1 em steel frame and is driven by a Bodine de motor with Minarek control. The front of the frame is covered by a flat-black panel which has a 3.8 x 0.5 em horizontal viewing aperture located 30.5 em from the bottom of the frame. There is an HN32 .025-cm Polaroid neutral density filter just behind the aperture. A General Radio stroboslave, Type 1539-A, drives a strobe light, which is located 7.6 cm behind the viewing aperture. At the medium setting used in these studies, the strobe has a flash duration of less Ulan

50 microscc and a peak intensity of 750,000 cp. As the slide moves vertically upward, a photocell pulses the stroboslave every .42 em. The continuous movement of the slide and list is "stopped" every .42 em, corresponding to a single typewriter line spacing, with the light passing through the list paper, then through the filter and aperture. The S, whose head was kept stationary by a chinrest, sat about 30 em in front of the aperture. At that distance, a list word subtended a visual angle of 2.4 x 0.6 deg.

Presentation rate was controlled by the speed with which the slide moved upward. For example, with the triple spacing, six words/sec corresponded to 18 flashes/sec, or 7.6 em/sec. Rated luminance was 10.4 fl. The flash was well above the 10 to 12 flashes/sec rate at which rucker is noticeable. Interitem error in the presentation rate was measured at less than 1.5%.

The trials were initiated by the S with a button placed near his left hand. This began the presentation sequence and also started a Lafayette KlocKounter. The first masking stimulus appeared 24 flashes after the initiation. A telegraph key was situated on the S's right; a keypress signaled a target detection and stopped the KlocKounter.

Procedure. At the beginning of each test session, the S was told that the experiment dealt with the ability of people to perceive words at very fast rates and that the following events would occur on each trial: (a) The E would read S a set of one, two, or three target words to look for; (b) the S was to press the start button to initiate a test sequence of rapidly presented words; (c) one of the words in the test sequence, and only one, was one of the target words and, when S saw this target, he was to press the response key at his right; (d) at the end of the trial, S was to report which target word was presented and the word that he thought immediately followed the target. The Ss were tested in l-h sessions on each of 3 consecutive days. Each day, three 30-trial blocks were presented. Three presentation rates were used (6,9, and 12 words/sec), one rate for each block, with a different order each day. Across Ss, a given rate appeared equally often in the three block positions.

Subjects. Twelve Stanford students with normal or corrected vision served as Ss.

Results Target detection latency and errors were recorded

along with the response words. The first day was

considered to be practice, and these data were discarded.

Days 2 and 3 were combined in the following analyses, since they did not differ in any essential aspects.

Detection Errors. Several types of errors were included in calculating the overall error rate for target detection, following Sasaki (1970): (a) omission errors, where no RT or verbal response was made, and (b) commission errors, where Ss reported the wrong target. The latter type of error could only occur with set sizes greater than one; there were, however, occasional trials with a single target where it was not seen. The error rates are shown in Table I. Error rate increased

both with set size [F(2,22) = 38.92, P < .001] and with

< presentation rate [F(2,22) = 18.02, P .005]. The

highest error rate, 29%, occurred with three targets at 12 words/sec. An interaction can be seen in the relatively small increase with set size for the slowest presentation

rate [F(4,44)=3.92,p ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download