Fisherman
1: There were sprinkle flecks.
Fisherman 2: What do you mean? Was it
a sprinkle or were there flecks?
Fisherman
1: It was in between; when a sprinkle
becomes flecks.
Fisherman
2: I see. Strong sprinkle you mean.
(Part of a conversation among a
group of four local fishermen, and the ethnographer, regarding a sighting of
fish. The two fishermen quoted are both
captains of their respective boats; one is a pair-trawler the other is a
smaller boat which trawls for bottom fish.)
Despite the
variety of fishing techniques that are employed along the west coast of Sweden,
the majority of them are related by a few key technologies. One of these technologies, the means for
actually capturing fish, is defined largely by the properties of materials,
fish behavior, and the physics of underwater events—making nets of different sorts a common component of the fishing practices
here. As reported in a previous
chapter, this common technology creates a bridge between the different forms of
the practice, and this bridge is manifest in the nature of the language which
defines the parts of the trawl for practitioners.
A second piece
of technology common to all fishing techniques evidenced along the coast, is
the device employed to "see" fish in the water, namely sonar. As reported in a previous chapter, sonar
devices are essential elements of all fishing techniques. Sonar forms the primary means of evaluating
the productivity of a given region of water at a certain point in time. In addition to being used for specific
functions within the cycles of activity which result in hauling fish on board,
sonar devices are also generic navigational aids installed on all boats of
substantial size. Knowing something
about the water directly underneath one's boat provides important information
which can—in conjunction with a sea chart or local knowledge or general
inference—aid in navigation regardless of the nature or purpose of the
vessel. As such, talk about interpretations of sonar images—as representations of
shoals of fish—is widespread among the entire population of seagoing residents
along the coast. Nonetheless, it would
seem that sonar devices—to the extent that the information represented is
effectively employed by fishermen to catch fish—are specifically important
mediators of fishing practice, and this importance should be reflected in the
cognitive and social processes which constitute the practice.
As the
micro-analysis of the previous chapter shows, the way sonar is interpreted and
communicated among fishermen is not obvious.
The general form of this process entails each boat acting as a "probe"
in a given area of water by way of interpretations—made by driving
captains—about displayed sonar images.
Occasionally, the information provided by a single reading of the sonar
device is unambiguous. More often,
captains must assess a broad array of sightings—their own and those of other
boats—in order to fish effectively.
This process, moreover, is importantly shaped by the specific fishing
technique being employed—both in terms of what constitutes an interpretation of
"good" fish and what the constraints on communicating the sighting are. For instance, in the previous chapter we saw
how the pair-trawling fleet is composed of a semi-cooperative set of completely
cooperative communication links (the pairs themselves). This structure constrains the form and
nature of cognitive strategies employed in communication and action within the
practice of flyttrålfiske. In this chapter, three different fishing
techniques are discussed: flyttrålfiske
(mid-water pair-trawling), vadfiske
(purse seining), and trålfiske (bottom-trawling). The first two techniques are employed by
larger teams of boats for catching pelagic fish (fish which swim or drift at
varying depths of water), while the third technique is employed by smaller
economic units (individual boats of smaller size and capital investment) to
catch fish which spend most of their time on or near the sea bottom.
In the short
dialog quoted above, one fisherman employs a novel combination of two
well-defined categories of sonar representations (sprinkle, strö, and flecks, fläckar) to try and characterize a concentration of fish that he
interprets to be in between the two. A
second fisherman, who didn't understand that usage, asks for clarification and
offers his own term to cover the intended meaning. In this chapter a study of the nature of communication about
sonar representations is undertaken.
The expectation is that this communication is built upon the utilization
of language and knowledge which reflects the patterns of interaction and
cognitive experience in the specific practices of these fishermen.
It is not
uncommon for fishermen to report sightings to each other following trips across
the water. Some of these trips are a
consequence of travelling to or from fishing grounds, but reports are also
forthcoming following trips to town for repairs or supplies or on return home
from the fishing grounds. Sometimes
these reports take the form of serious information gathering or dissemination,
possibly even as part of a plan about where to go on the next fishing
departure. And of course, reports are
frequently made between boats, over radio, while fishing together in the same
area of water. As evidenced in the
micro-analysis of deciding where to set the trawl in flyttrålfiske, these reports are often part of a very specific
context of information gathering shaped by the ongoing dynamics of following
along with a fleet of pair-trawling teams, each reporting what they are
"seeing."
These reports
often entail descriptions of the fish which are grounded in the properties of
the representing devices themselves.
Thus common descriptors such as strö
(sprinkle), fläckar (flecks), prickar (dots or marks) punkter (points), klungar (clusters or groups), rand
(stripe), skuga (shadow), tunn (thin), tätt (tight), spritt (dispersed),
röd/gul/blå (red/yellow/blue), or
holding ones thumb up to indicate size or shape, are all ways of communicating
the nature of the representation generated by the sonar device. Other methods of constructing descriptions
entail using the scale of the representation to depict the size and location of
the fish in the water. Descriptions that
entertain aspects of the state (or future state) of the fish themselves might
also be employed, such as an estimate of the catch size of a shoal (e.g., "600
boxes") or the behavior of the shoal (e.g., "tight to the bottom"). Finally, the prospects for using one's
particular equipment to capture the fish might enter into the description
(e.g., "they stand well for catching" or just "it looks
productive"). The phenomenon
addressed in this chapter is: how do the representations of the sonar device
and those of the talk generated by the fishermen about the sonar device relate
to each other and the practice of fishing?
Clearly, the
basis for making such crucial decisions as deciding where to set the trawl or
the seine are dependent upon making sensible use of the information presented
by the sonar device. And yet, the sonar
images themselves are difficult to classify in any singular manner. I was on board during many draws that
appeared on the sonar (to me and my informants) to be productive, but which
turned out not to be when the trawl was hauled on board. I was also told that it was not uncommon for
the captain of one boat of a purse seining team to call the seiner over because
the former (acting as a "scout" for the mother ship) was seeing good fish—only
to be told by the captain of the seiner that "there is no fish here" once the
larger ship arrived. I was shown how
one could turn the gain up on the sonar device so that plankton shows up as
something one might consider profitable to cast the net upon! I noticed how each captain would have a
different preference for setting the gain and picture intensity of his display
screen when he was at the helm—apparently in an effort to establish a familiar
baseline against which to make judgements.
I also experienced a variety of different sonar devices on different
ships, some with different frequency of signal, most with different screen size
and display textures. The question,
then, is from this inherent variation in subjective cognitive experience, how
does a common ground for communication exist? or does it?
Sonar was first
introduced onto the boats of this area in the late 1940's and early
1950's. Invented in WWI but not made
commercially viable until after WWII, the technology was new and expensive but,
as with many technological breakthroughs since, was eagerly incorporated into
the practice of the most ambitious fishermen along this coast. One informant told me that the sonar device
on a ship his father and father's brother had built in the early 1950's
accounted for a large percentage of the entire boat's cost! Nowadays, sonar devices of much higher
quality are installed on nearly every boat large enough to support an enclosed
bridge—at costs ranging from a couple hundred to a couple thousand dollars, or
some .05 percent of a large fishing vessel.
Earlier sonar
devices employed a chemically coated paper which recorded black marks when
traversed by a mechanical stylus driven by the voltage levels generated by
sonar returns. However, Cathode Ray
Tube (CRT) color monitors have become the dominant display device in the last
decade. Many paper plotters remain
built into the instrument panels of older ships—now usually serving the
function of preventing coffee cups from dumping their contents more than
anything else. However, several
fisherman reported to me their dissatisfaction with color monitors, citing the
"trickiness" entailed in deciphering the information represented. One of the larger new boats from Vindö even
had a paper plotter built into the bridge as recently as 1988. In practice, however, even this boat seldom
employs this paper plotter, largely on account (I was told) of the expense of
paper. The complaint with color
monitors has to do with the degree of representational variation that is
possible, indeed inherent, in these devices.
Establishing the meaning of representations created by black marks on
white paper was deemed to be a simpler affair.
In spite of these objections, informants expressed a general feeling
that color monitors constitute a better technology. At any rate, the lower operating costs of color monitors appears
to resolve the question of which technology to use.
It is
interesting to note that despite the new representational medium employed
(color monitors) remnants of the old technology (paper plotters) remain not
only in the metal of the bridge but also in the language used to describe sonar
representations. Thus color is a
dominant theme in modern descriptions of sonar representations of fish. Red, yellow, blue, and white are employed as
a continuum—by these devices—to represent energy loss (from less to greater,
respectively) in the returned signal.
This color scale translates into a resource for interpreting the density
of reflecting objects, from hard/dense to soft/sparse respectively. For fishermen, the color scale provides a
resource for interpreting the quantity and quality of fish beneath their
boats. Yet a common phrase—"it's
completely white"—is often heard among (especially older) fishermen in
reference to the lack of black marks on plotter paper, and thus the lack of
fish in the region, even though the basis for their interpretation came from
evaluating the display of a color monitor.[1] Another expression I heard, "it's been about
a half paper now" quite explicitly conflates the properties of the two
different technologies in describing the fact that fish have just recently
appeared on the sonar. On the one hand,
the reference to "paper" captures quite vividly the time course of events by
evoking the image of a scrolling paper plotter. On the other hand, the "half" only makes sense in reference to
the fixed frame width of the video monitor (which captures 12 minutes of
history) and not a length of plotter
paper, which would be undefined even if the plotter were in operation.
In order to
address the cognitive and cultural properties of these phenomena, I designed an
experiment to investigate the nature of local fishermen's use of sonar images
to interpret and communicate about fish and catching fish. While it is not possible to relate
fishermen's interpretations of sonar images directly to any objective measure
of what the sonar returns represent, the interpretations can be compared among
fishermen in order to ascertain the distribution of agreement about the
meanings of these representations as instantiated in a sample of sonar screen
displays.
This is an
entirely satisfactory approach for a number of reasons. First of all, the forms employed to
communicate the meanings of sonar representations are shaped by the social
framework in which the forms are used, and therefore this distribution of
social variables should be explicitly addressed in the analysis. Second, "reality" for the fishermen is
defined by the actions of successful practitioners—there is no more pertinent
reality for them and thus any attempt to factor out this reality is both
unnecessary and unproductive. Third,
even though the "real" nature of the relation between fish and interpretations
of sonar representations is an undecidable matter, it is reasonable to assume
that there is indeed some systematic correspondence between the skills of the
practitioners and the objective reality "out there" which these practitioners
rub up against every day, and which leaves its imprint upon the forms of their
practice and the nature of their knowledge.
While on board
the fishing boats from Vindö I collected many hours of video taped images of
bridge activity, including footage of sonar displays. I constructed a set of 50 sonar images by photographing selected
video frames (re-displayed on a color video monitor). These 3 x 5 photos were then trimmed to include only the sonar
display tube image, and pasted onto solid black backgrounds. The video frames chosen included images from
6 different sonar devices on board 4 different boats taken during 2 different
kinds of fishing (pair-trawling and purse seining) for herring. In addition, the images were chosen for
their clarity. Every attempt was made
to standardize the appearance of the trimmed photos—each photo included the
vertical scale of the display along the right hand side, and all extraneous
parts of the sonar device (e.g., brand name, adjustment panel, etc) where
trimmed away. The set of trimmed photos
ranged in size from 10cm x 8cm to 5cm x 5cm, and where pasted onto the centers
of 4 x 6 black index cards (see Figure 59).
Subjects were
selected on the basis of their willingness to participate and the goal of
collecting a representative sample of fishermen from among the local
fleet. Fifteen subjects participated in
part or all of the study. They
represent fishermen from the eight largest and most active boats of the
(approx.) twelve-boat fleet and thirty active fishermen from the island of
Vindö. Eight of the subjects were
captains/owners and seven subjects were non-owning crewmen out of a total of 17
captains and 23 non-owning crewmen from these eight boats (see Table 1). The set of fishermen engaged in three types
of fishing—purse seining, pair-trawling, and bottom trawling—and were
classified according to the dominant technique employed by the boat on which
they were crew members during the year I spent on Vindö. Although these fishermen have spent time on
each others boats, there was no significant shuffling of crew between boats
during my year of residence. On the
other hand, all of the fishermen had experience fishing for herring. All of the subjects from smaller boats have
crewed on the larger boats of this study for significant periods of time in
their past (with one exception—who had himself fished quite extensively for
herring using single boat trawling).
Two different
experiments were conducted with subsets of the 15 subjects using subsets of the
50 cards as stimuli. Each subject
participated in the context of a private semi-formal interview arranged in
advance, and conducted in the ethnographer's kitchen or the subject's home.[2]
In experiment
one (here called the "Pile Sort") subjects were given an ordered stack of 50
cards and asked to place each card in one of seven possible piles. These piles were explicitly labelled with
seven name plates, each containing a number from 1 to 7, arrayed in
left-to-right order on the table at which the subject was seated. Furthermore, name plate "1" was also
labelled with the sentence "The worst I have ever seen" and name plate "7" with
the sentence "The best I have ever seen."[3] Subjects were informed that they need not
use all seven piles, but could use no more than the seven labelled piles for
sorting the cards. All 15 subjects
participated in this experiment.
Subjects were given the following written instructions and given as much
time as they desired to do the sorting:
In this part of the experiment, I would like you to sort the pictures into (at most) 7 different groups, depending on how productive they look for catching herring using the technique with which you are most familiar.
The 7 groups should be valued from 1 ("the worst I have ever seen") to 7 ("the best I have ever seen"). You are welcome to look through the pictures as much as you like.
Most subjects
found the experiment easy to perform, and placed each card one at a time from
the top of the stack. Most would adjust
a few cards' positions once they had reached the end of the stack, effectively
contextuallizing their decisions by quickly scanning through the other members
of a pile. Subject 5 would process 10
cards from the top of the stack at a time, laying them all out for comparison
and then assigning them each to one of the 7 piles, and then proceed by
spreading out 10 more cards from the top of the stack, etc. Subject 9 built piles on #2, #4, #6, and
then divided those to form piles #1, #3, #5, and #7. Subject 10 searched through the stack briefly, pulling out 7
exemplars and then used these to guide the assignments of the rest of the stack
of cards. Subject 6 searched throughout
the stack and pulled out the 10 or 12 best and divided them amongst piles #4
through #7. He then placed the rest in
piles #1 through #3, shuffling them back and forth there until he was satisfied. Subject 14 laid each pile out as a column,
in order to get a better view of every card in each pile, but in the end did
not do a lot of reshuffling.[4]
Table 2 shows
the distribution of each subjects' pile assignments. Most subjects piled more cards in the lower numbered piles
("bad") than the higher numbered piles ("good"), although two subjects (5, 9,
and 10) created quite uniform distributions of the cards with their piling.[5] One thing to note, which is pursued in depth
in the analysis below, is an observable trend for trawler fishermen (subjects
1, 3, 5, 7, 8, 9, 11, 12, 13, 15) to rate cards more favorably (Mean = 3.18)
than purse seining fishermen (subjects 2, 4, 6, 10, 14; Mean = 2.82). This difference is significant at the .01
level.[6] This finding is quite reasonably accounted
for by the fact that evaluations by trawlermen are based on an expectation of
seeing the displayed quantities of fish over some period of time while
seinermen evaluate a stand of fish based on its catching potential as displayed
in one snapshot of the sonar. Trawling
enables catching more dispersed fish due to the mobile technique, while seining
requires dense concentrations of fish to be effective and these differences in
the practice shape the ways evaluations of sonar displays are made. This theme is pursued further in the
analysis below.
Experiment two
(here called the "Matching Task") consisted of two parts: one to collect
descriptions of sonar displays (the same cards as above), the second to measure
the task of matching these cards with the elicited descriptions.
In part one,
conducted during the same interviews in which the Pile Sort data were
collected, 12 of the 15 subjects were handed the same ordered stack of 50 cards
as in the Pile Sort. Subjects were then
asked to respond (with tape recorder rolling)—using one or two sentences of
their own making—to the question "Do you see anything?" issued from a companion
boat while out fishing together. It
should be noted that the meaning of this question, as evidenced in the
transcripts of Appendix 1, is grounded in the practice of fishing and is
interpreted (in the context of that practice, at least) as a specific request
for information about the prospects for fishing an area of water.[7] All subjects were aware that my data
collection (here and in the Pile Sort) were based on an attempt to compare
responses as a way of investigating the use of this instrument among the local
practitioners of their profession.
However, subjects were not aware of the nature of part two of this
experiment (see below) which was administered one to 3 months later. The written instructions for part one of the
Matching Task were:
Imagine you are out fishing together with another boat. (For example, you are the light boat of a purse seining team, or you are one boat of a pair-trawling team.) You are searching for herring and communicate what you see to the other boat. For each sonar picture, describe what you see with one or a couple of sentences, just like you would do if the others had asked,
"Do you see anything?"
Note: The descriptions should be more or less spontaneous, but also informative.
Subjects found
this task a bit tougher to perform.
Many of the descriptions entailed a general qualitative characteristic (e.g.,
"pretty good, looks productive") often contextualized by the preceding cards
(e.g., "this looks better") or by the experimental setting (i.e., directing
responses at me rather than the imagined interlocutor, for example "this is
more [boat x's] type of fish"). Some
non-captain crewmen had a difficult time because they do not in fact have much
experience performing this task on the boats, while others seemed to enjoy the
role-playing of "captain" in the experiment.
Some subjects appeared to take what I would call a slightly "macho"
approach, finding little worth commenting on in the stimuli, which I
interpreted to be a possible experimental effect like "I had better evaluate
this quantity of fish along these lines in order to make a statement like ‘there's
nothing here worth our time'." Several commented that such descriptions
were difficult to generate based on single displays, because one generally
employs a lot of context in sonar evaluations: what else has been seen tonight?
how desperate are we to get the equipment in the water? is this representation
at the beginning or tail end of our search?
These are questions which would normally be filled in by the local
context. (See Figure 60 for a sample of
descriptions elicited during part 1 of the Matching Task.)
The responses
(50 x 12 = 600) were later transcribed and formed a pool for selecting 10
subjects' descriptions of 12 cards (12 cards x 10 speakers = 120 descriptions)
for the second part of the Matching Task.
This selection process entailed the researcher scoring each
description—in an attempt to get at the most "interesting" cards and
subjects—in the following way: 1 point was awarded to each card (for each
subject) that was placed into piles 5, 6, or 7 in the Pile Sort experiment, and
0, 1, or 2 points were (subjectively) awarded to each description based on the
extent to which it employed "descriptive content" (as opposed to, for instance,
simply stating that the fish looked "good" or "bad"). Having scored each description from 0 to 3 in this manner, the
descriptions of the highest scoring subjects (10) and cards (12) were selected
for the second part of the Matching Task.
The second part
of the Matching Task was administered between 3 and 12 weeks after part one,
under similar conditions. In part two,
the 120 descriptions (created by the 10 chosen "speakers," and now transcribed
and listed on 6 sides of paper) where given to 13 subjects ("listeners," which
included 9 of the 10 "speakers") along with a poster board containing the 12
cards referred to by the set of descriptions.
The twelve cards were arranged (by random draw) into positions in a 3
column by 4 row matrix on the board, making all 12 cards simultaneously visible
and equally salient. The task for
subjects was to pick the card best described by each description. Selections were written (by the subjects) on
a form that contained 120 places—one for each description. In each place one of the 12 cards was to be
indicated by number. The written
instructions were:
Earlier I elicited descriptions of all of these [50] sonar pictures from local fishermen. Now I have selected out a few pictures of which you should select one for each description listed.
That is, these descriptions were your own responses elicited under the following instructions:
[Repeat
instructions from part one above.]
Note: There are 120 descriptions. Just choose the picture you think was in front of the fisherman who created the description (of course, you don't know the identity of this person). There are many which are unclear, (for example, "It looks pretty good"). With many you won't be certain at all—just choose the one you think is the most probable picture.
This part of
the task took subjects anywhere from 45 minutes to one and a half hours. 13 subjects took part, although one
subject's responses were thrown out due to the fact that he had listed multiple
cards for each description (and a retest was not performed), leaving a sample
size of 12 subjects ("listeners") on the Matching Task. Subjects generally found this part of the
task difficult, yet interesting. It
seemed to constitute a kind of task which participants felt comfortable trying
their hand at—perhaps this part of the task was more "natural" from the
standpoint of being a problem-solving task that was not embarrassing to
subjects. Responses were written down
in a private, casual, atmosphere—perhaps leading to a reduction of anxiety
about being "tested." Several said they
were confident regarding about 50% of their responses, which turned out to be
in the general ball park of accurate matching with the speakers. Agreement between listeners and speakers
ranged from 40% to 58% across subjects.[8] Some claimed to recognize authors of
descriptions, and although I was unable to verify the veracity of this
perception, some identified individuals who had not been involved as a speaker
in the experiment! On the other hand,
some descriptions did seem to be
highly stylized performances that were well known to fishermen who work with
these individuals regularly. It is
possible that these styles, well learned by some who hear them often, are
incorporated into the latters' performances—a kind of imitation—leading to the
misidentification of authors. (Although
I have no data to check this speculation, support for the hypothesis does come
from the extent to which fishermen which fish together more often agree with
each other more often in their interpretations of sonar, as reported herein.)
One question to
ask of the data is what the patterns of agreement between speakers and
listeners actually were. That is, given
that speakers were describing a particular sonar image, to what extent (and
according to what distribution) were listeners able to yield the same meaning
from the description (i.e., to identify the same image by selecting it as the
picture which best fit the description).
Figures 61 through 67 show bar charts of the percentages of agreement
between speaker and listener, for each subject who participated in the Matching
Task. The first thing to notice is how much more informative (measured
as the average percent agreement with listeners) some speakers were than others
in their descriptions.[9] Subject 6, in particular, was on average
much more informative (Mean agreement with listeners = .67) than the rest of
the speakers (Mean agreement for all speakers with listeners = .49).
The fact that
listeners made accurate sense of their own descriptions no more than they
managed to do with other speakers' descriptions (see Figures 61 through 67) was
an unexpected finding. With a little
reflection, however, this finding is reasonably explained by the fact that
descriptions (as public instantiations of speakers' understandings) are no more
informative for the author (as listener) than for other listeners, at least in
the context of the Matching Task.
There was no
significant relationship found between speakers' crew status
(captain/non-captain) and informativeness, although the trend was a negative relationship, suggesting
captains are less informative. This could be an artifact of the
experimental situation since some captains displayed a tendency (I noted during
administration of part one) to under-evaluate (i.e., be less informative) in
their responses. Furthermore, many of
the non-captain fishermen (many who were younger and more comfortable in this
kind of artificial "school-like" setting as well) were more verbal and
explicitly descriptive in their responses.
Subject 6's responses, I noted at the time of administering part one,
were unusually long and detailed—they obviously carried more information which
made disambiguation easier.[10]
On the other
hand, if the two outlier speakers (6 and 8) are removed from the sample,
captains show a definite tendency to be more
informative in their descriptions (Mean agreement (captain speaker) = .505,
Mean agreement (non-captain speaker) = .451, one-factor anova F(1) = 3.042, p =
.08). Furthermore, with the two
outlying speakers removed from the data, listeners
who were non-captains become much less
agreeable as a group (Mean agreement (captain listener) = .498, Mean agreement
(non-captain listener) = .446, one-factor anova F(1) = 2.483, p = .12). This latter finding suggests that in the
absence of the most and least informative speakers (subjects 6 and 8), captain
listeners exhibited a tendency to understand speakers more than non-captain
listeners. This same relationship
obtains when the entire data set is considered, but is only weakly evidenced
(Mean agreement (captain listener) = .496, Mean agreement (non-captain
listener) = .474, one-factor anova F(1) = .874, p = .35).
A significant
relationship was found between
speakers' type of fishing (purse
seining or pair-trawling = 1, bottom trawling = 0) and Mean percent agreement
with listeners (Mean agreement (purse/pair) = .509, Mean agreement (bottom) =
.445, one-factor anova F(1) = 4.505, p =.034).
This relationship suggests that there is possibly an increasing ordering
of informativeness (relative to the
entire sample of listeners' expectations) from (smaller) bottom trawlermen to
(larger) purse seiner and pair-trawlermen.
This
relationship is reasonably accounted for in two different, but overlapping,
ways: (1) by different demands for communication in the respective types of
fishing (i.e., bottom trawlers do less communicating and less collective
fishing) and (2) by different experience with herring (i.e., bottom
trawlermen—although they have all fished herring before, and regularly witness
representations of herring shoals on their sonar displays—are generally not
evaluating sonar images of herring on a daily basis). Bottom trawlers are self-standing economic units and although
they do cooperate extensively in many matters regarding their practice, this
cooperation seldom takes the form of sharing the details of fish sightings
while in the same vicinity of water.
Although it's true that these smaller boats will often fish in the same
areas of water,[11] they
are neither economically inclined nor practically committed to elaborating upon
fish sightings as is the case with pair-trawler and purse seiner teams.
A second
question to ask of the data is what the patterns of agreement between listeners
were. It turns out that this is a more
productive relationship to measure.
This is due to 3 different factors: (1) the artificial nature of the
experimental situation is more neutralized; (2) the 66 pairs of
listener-listener agreements (each compared on 120 match items) makes for a
much larger data sample than the 120 speaker-listener pairs (each compared on
only 12 match items); and (3) the measurement of subject agreement as sensitive
to the variation in descriptions, but rather factors out the variability in
speakers descriptions by attending to listener performance in a more
constrained task. Here we ask: given a
description (regardless of its author) how often (and according to what
distribution) do listeners agree on its meaning as evidenced in their selections
of the best-matching sonar image.
Figure 68 summarizes gross inter-listener trends by showing the
distribution of average inter-listener agreements for the 120 match items. As a group, listeners found 40 items (33%)
to be "hard" (agreeing on average with only 1 to 3 others, or about 9% to 30%
of the other listeners), 20 items (17%) to be "easy" (agreeing on average with
8 to 11 others, or about 70% to 100% of the other listeners), and the remaining
60 match items (50%) to be "in between" (agreeing on average with 3 to 8
others, or about 30% to 70% of the other listeners).
In general,
listeners agreed with zero or one other on anywhere from around 20 items
(subjects 2, 9, and 10) to 30 or more
items (subjects 1, 4, 5, 13), and with everyone else on exactly 8 match items. Comparison of listener mean agreements
revealed that the average inter-captain mean agreement (.468) was significantly
larger than inter-non-captain mean agreement (.42) (one-factor anova F(1) =
6.235, p = .0185). A two-factor (Match
Item, and Inter-listener Status [captain/non-captain]) showed this relationship
to be strongest across a subset of the match items (F(116)=10.074, p = .0001,
for the interaction between the two factors).
To compare
listeners directly, we now consider not their average degree of agreement with
all other listeners but, rather, their specific levels of agreement with others
as a means of investigating variables which may govern the patterns of listener
agreement. Table 3 shows the matrix
created by using the percentages of
agreement between pairs of listeners as a measure of similarity in
performance on the second part of the Matching Task. Each entry in the similarity matrix is a measure of a pair of
listeners agreement on the 120 match items.[12]
Figure 69 shows
the results of a hierarchical clustering analysis on these data using the
Maximum Method. (Each subject joins a
cluster by virtue of the magnitude of his closest relationship—his highest
measurement of agreement—with any member of the cluster, see Johnson
1967). The first thing to notice about
the clustering diagram is the prevalence of captains in the early clusters, and
non-captains in the later clusters.
This reflects the finding discussed in the previous paragraph that
captains were more likely (on average) to agree amongst themselves than
non-captains were. This finding also
suggests that captains were more likely to agree with any other listener than non-captains were. This intuition can be checked by coding
pairs of listeners—counting the number of captains in the pair—and regressing
the agreement averages on this variable.
This regression model (R = .244, R-Squared = .059, F(66) = 4.047, p =
.0485) shows a positive relationship (significant at the .05 level) which
supports the intuition that captains are more likely to provide interpretations
of descriptions which are "acceptable" or which define the center for consensus
among the entire sample population.
Another thing
to notice about the diagram of Figure 69 is clustering of subject agreement by type of fishing. A one-factor anova revealed no significant
relationship among listeners according to their engaging in the same
practice. However, when subject 13 (the
clear "outlier" in this task) was dropped from the data set, a significant
relationship between listener agreement and
type of fishing was found (Same [Mean] = .493, Not-Same [Mean] = .456,
one-factor anova F(1) = 5.209, p = .0265).
The questions
being addressed in this experiment concern the distributions of knowledge and
language forms entailed in the use of sonar, and the cognitive and social
mechanisms responsible for these distributions. Further analysis of the data is required in order to pursue the hypothesis
that expertise with sonar (both the communicative and interpretive skills
entailed in using sonar) is constructed via the constraints and resources of
the different practices in which its use is embedded.
These
relationships between practice-based variables and experimental performances
are difficult to demonstrate within the framework of quantitative
statistics. In general, this is due to
sample size limitations—in order to capture significant trends in data which
relate to individual subjects (for example comparing captains and non-captains,
or comparing fishermen engaged in different forms of the practice), many
subjects from each category are required.
Simply comparing large numbers of performances in the tasks, with a
small sample of subjects, is insufficient.
In fact, the entire population of fishermen from Vindö would not be
sufficient for such an analysis since there was only one pair-trawling and one
purse-seining team in residence.
Finally, the variables themselves are somewhat artificial
simplifications of reality since fishermen, boats, and teams will flexibly
interact and redefine their boundaries as ecological, economic and other
cultural conditions dictate.
An alternative
analysis attempts to uncover the structure underlying the available data
without specifying, a priori, the principle components of that structure, but
rather searching algorithmically for these components as best-fitting axes of
the data. One method for doing this was
exemplified above with the hierarchical clustering analysis of inter-listener
agreement. Another well-known method is
pursued here, namely Multiple Dimensional Scaling.
The similarity
matrix of agreement among listener pairs (see Table 3) was employed as "target"
values in a Multiple Dimensional Scaling (MDS) analysis. Briefly, the analysis attempts (by means of
an error minimization technique) to find a matrix of coordinates in n dimensions (usually 2 or 3) which is
nearly equivalent to the target matrix—that is, which reflects as much of the
measured similarity structure between listeners as possible. Figure 70 shows the two- and Figure 71 the
three-dimensional solutions found by this technique, and the extents to which
they capture the similarity structure of the original data. The two-dimensional solution shows quite
clearly the structure of the hierarchical clustering analysis reported in the
last section. However, that analysis
could not reveal the multiple relationships between subjects which become more
evident in the MDS analysis here.
Looking closely at the two-dimensional MDS solution of Figure 70, we see
a division of the agreement space along the lines of type of fishing: bottom trawlermen clustering in the upper right
corner (7, 8, 9, 12), pair-trawlermen (with one exception, subject 13)
clustering in the lower left corner (1, 3, 5, 11,13), and purse seinermen lying
between the two on the diagonal from upper left to lower right corners (2, 4,
10).
This structure
becomes more evident in the three-dimensional MDS solution shown in Figure
72. The solution is quite difficult to
interpret in three dimensional space, but by projecting this solution onto each
of the three planes constructed from the axes of the coordinate system (x-y,
x-z, y-z planes), the relationship between listener agreement and type of fishing is quite clearly seen
(Figure 72). In particular, each of the
three projections can be divided into two regions by a straight line such that
the plotted points representing each listener which lie in one of the two
regions stand for fishermen which engage in the same type of fishing. Thus the x-y plane appears to show a
division of the similarity space into "those who pair-trawl" and "those who do
not," the x-z plane into "those who bottom trawl" and "those who do not," and
the y-z plane into "those who purse seine" and "those who do not."
Following
Kruskal & Wish (1978), the coordinates of points in this solution can be
employed in a regression model (as independent variables) to quantitatively
determine how well the MDS solution's coordinate axes predict relationships between
listeners (dependent variables). In
this case, using the three MDS solution axes as three independent variables and
regressing the three variables for type of fishing (Purse, Pair, and Bottom
Trawl) over the MDS solution coordinates for each subject, we see that the MDS
solution does indeed partition the "agreement space" along lines which predict
subjects' type of fishing. Figures 73,
74 and 75 show the results of regressing each of the three dichotomous variables
Purse (1 = purse seinerman, 0 = not
purse seinerman), Pair (1 =
pair-trawlerman, 0 = not pair-trawlerman), and Bottom (1 = bottom trawlerman, 0 = not bottom trawlerman) over the
three MDS solution coordinate axes x, y, and z.
These three
independent multiple regressions support the qualitative finding, reported
above and in Figure 72; that axes y and z are the strongest predictors of Purse
seinerman location (standard Beta values =.636 and -.411, p = .026 and .116,
respectively); that axes x and y are the strongest predictors of Pair-trawlerman
location (standard Beta values = .492 and -.601, p = .042 and .018,
respectively); and that axes x and z are the strongest predictors of Bottom
trawlermen location (standard Beta values = -.507 and .692, p = .026 and .006
respectively). Collectively, the three
regression models provide very strong support for the claim that the MDS
solution has found principal components for partitioning the space of listener
agreement on the Match Task which are quite closely related to classifications
of fishermen by their specific fishing practices.
In essence, the
MDS analysis has forced the data into a coordinate system—without specifying
how that should be done. The result
entails a loss of information yet retention of the most significant structure
in the data.[13] The MDS analysis provided the most
convincing evidence for the proposition that inter-listener agreement was
importantly shaped by the patterns of interaction resulting from the boundaries
of practice which partition fishermen's activities. Despite identical grounds for making sonar judgements in this
task, pair-trawlermen, purse seinermen, and bottom trawlermen tended to agree
more within than across these boundaries.
Figure 76 shows
a scattergram of the relationship between listener (average) agreement in the
Match Task and subject correlation in the Pile Sort. Each point in the graph represents a pair of subject's agreement
in the two tasks; as listeners and as pilers.
Regressing one variable onto the other reveals a very significant
positive relationship between the two (R = .634, R-squared = .402, F(65) =
42.982, p = .0001). That is, as much as
40% of the variance in the Match Task (listener agreement) is accounted for by
performance in the Pile Sort (agreement in direct evaluation of the sonar
pictures), and vice versa.
This finding
was quite a surprise, since the two tasks do not appear to be very
similar. Assigning goodness values to
sonar representations (in the Pile Sort task) would appear, at first glance, to
be a non-linguistic task which entails the direct interpretation of visual
stimuli. On the other hand, assigning
cards (from a subset comprising less than 25% of the 50 cards used in the Pile
Sort) to descriptions clearly requires a lot of linguistic
interpretation—grounding language tokens in meanings—by listeners. If many descriptions were of the form "2 on
a scale of 1 to 7" then the similarity in subject performance would be readily
understood since this is precisely the extent to which linguistic labels were publican available (as cognitive
mediators) to subjects in the Pile Sort task.
Although many descriptions elicited in part one of the Match Task do in fact code the sonar
representations in a simple "good/bad" manner, the 120 descriptions employed in
part two (it will be recalled) were specifically selected for their
"contentfullness," thus maximizing the extent to which the linguistic nature of
the two tasks differ. A further
investigation of the language used in speakers' descriptions is conducted in
the next section. At the very least, it
is safe to say that the underlying variables which generate agreement in
subjects across the two tasks probably have much in common.
Although it was
difficult to account for additional variance in the regression of
inter-listener agreement on inter-piler agreement, some trends were
noticeable. By rank ordering the
residuals of this regression analysis it was noticed that the regression model under-predicts the extent to which
fishermen who are actively engaged in regular communication in their specific
practice agree with each other as listeners in the Match Task. Thus for 10 of the 12 listeners, the most
under-predicted other listener (by performance in the Pile Sort) was a
fisherman who engages in the same type of fishing, and 7 of these 10 others
were also captains. The strength of
this regularity throughout the inter-listener data was verified by the fact
that type of fishing (1 = same, 0 =
different) and captain count (0 = no
captains, 1 = one captain, 2 = two captains) accounts for an additional 5
percent of the variance in the residuals of the Pile Sort agreement vs. Match
Task agreement regression.
Up to this
point, the analysis has relied on subjects' performances as "listeners" (in
part two of the Match Task) as a vehicle for interpreting the distribution of
meaning-sharing among participants in the experiment. Now I will discuss an attempt to code speakers' uses of the
language as another vehicle for understanding consensus, this time among
speakers in part one of the Match Task.
Each description elicited in part one of the Match Task (12 subjects x
50 cards = 600 descriptions) was coded for the occurrence of "features." A "feature" is defined as an instance of a
language form, from among a finite set of such forms. This finite set was "discovered" in the data. That is, by making repeated passes through
the entire set of descriptions, it was determined that an overwhelming majority
of the surface content of descriptions could be captured by a (relatively
small) set of abstract forms, called "features," represented by a 3-tuple such
as {sprinkle,good,very}. Every feature
in the set (i.e., the "feature language") is generated by a simple grammar (the
"feature grammar") as shown in Table 4.
It is important
to note that no claim is being made for the ontology of the feature grammar in
fisherman thought or language at this point.
It has been derived analytically as a convenient way to characterize the
feature language by some means other than enumerating the entire set of
features. It could easily be argued,
for instance, that members of one subset of the feature language generated by
one MODIFIER in the grammar (say, thick,
generated by SIZE) really belongs in another subset generated by another
MODIFIER (say, CHARACTER, putting it in the same class with a token of
similar—but opposite—meaning, dispersed). For the moment, it is only being claimed
that the vast majority of the content of descriptions (i.e., the surface coding
of speakers' descriptions) is well-captured by the entire set of features generated by the grammar of
Table 4. (That is, nothing need be
claimed about the classes generated by the grammar, as these are—at least in
part—artifacts of the analysis. What is
meant by "well-captured"—is elaborated upon below.)
The purpose of
this coding and analysis is three-fold:
(1) To investigate the surprising result—reported in the previous
section—of the high degree of correlation between inter-subject agreement in
the Pile and Match tasks; (2) To develop an analytic tool for further
investigating the sharing of language forms and meanings across subjects; (3)
To study the phenomenon of talk about sonar by paying attention to the actual
language forms employed by speakers in the experiment and in real contexts on
the water.
Descriptions
elicited from speakers during part one of the Match Task were thus coded as
containing instances of features from the finite set of the feature language of
Table 4. This amounts to mapping
instances of actual language use into 3-tuples of the feature language, or
"features". This process was conducted
(as much as possible) at the surface
level representation of speakers' descriptions. That is, every attempt was made to make the process equivalent to
encoding the actual surface forms of language as represented in the
descriptions themselves. For instance,
when a new construct was encountered in the data, it was added to the feature
language rather than counting it as an instance of something already in the
language which might be quite similar in meaning. Of course, what counts as a "new" construct required
interpretation on my part, but every attempt was made to code descriptions
without recourse to semantic contents.
Thus, for example, we find {NOUN,good,not} and {NOUN,bad,nil} to be
distinct features even though (semantically) they clearly have a lot in
common. Similarly, {NOUN,wide,not} may
well mean the same thing as {NOUN,thin,pretty-much} but no attempt was made to
equate them in the coding of actual language forms used by speakers.
While this
strategy seemed like a reasonable way to avoid assuming the semantic structure
that the study was meant to uncover, it was not without problems. The primary problem comes in the coding
itself. When is an instance of language
use the same as another (already taken from a noisy speech stream filtered by
means of transcription)?[14] Clearly, I had to draw on my knowledge of
the domain of discourse just to disambiguate the referents of speech. And even then, many subtle ambiguities
remained unresolved or conflated in the coding. For instance, many expressions were contextualized by the
preceding sentence or the preceding stimulus, making coding difficult. The expression "it's beginning" could mean
that the current sonar display shows fish beginning (a change in the display
from left to right) or that the subject was taking the artificial experimental
context seriously and comparing the current stimulus to the previous one.[15] Sentences which used "it" as the head noun
were notoriously difficult to code because the referent could be an object on
the display, or the entire display, or the fish referred to by the
display. Sometimes the surrounding
sentences provided cues about the intended referent, but many times it remained
ambiguous.[16]
Furthermore,
surface coding of language forms breaks down when expressions are
polysemous. Thus "thin" (as an
adjective) was used to modify a noun which identifies a solid representation of
fish (thus meaning something like "not wide") but was also used to modify a
noun which identifies a scattered representation of fish (thus meaning
something like "not concentrated"). In
the current coding, these two meanings of the adjective are conflated as
identical features. Other expressions
that were difficult to code were those constructed by negating some other
expression, for instance "not one of those super tight sprinkles." Should this be coded as {sprinkle,tight,not}
or {sprinkle,tight,nil}? (Generally, if
no other sentence was available in the description to help resolve this issue
the latter coding was employed as a kind of default.)
Another difficulty
was introduced by expressions which employed contextualizations within the
displayed image as a means for modification.
Thus a few subjects (particularly, the most "prolific" author, speaker
6) would refer to parts of referents
(of which generally only the latter were identified by most subjects), and
describe these (e.g., the "edges" of the sprinkle, or the layer above or below
some other identified area in the representation). These "sub-descriptions" were not coded unless they could stand
on their own as self-referential features.
For example, "high up [near the surface] looks good" could reasonably be
coded by the pair of features {nil,high-up,nil}, {nil,good,nil}, while
"underneath that looks better" was not coded since its components refer to other
aspects of the display and do not stand on their own as self-referential
features. Notice, however, that even
the pair of features in the former case introduces "noise" into the meaning of
the description since the pair are not tied together, but become independent
components of the entire list of coded features which stand for the speakers'
description in this analysis. Thus
although the speaker in this case meant only that the upper layer was "good,"
the coding now simply records that an evaluation of "good" was made for the
sonar display in question.
The most
coarsely-coded features were those in the class "ACTION," because the coding
ignored many aspects of the specific techniques of fishing which were commented
upon by speakers. Thus, "if we could
keep it on the bottom here it could be something" was coded simply as
{nil,draw,maybe}. This particular issue
was not that troublesome because I felt that many of these specifics were
offered up to inform me, the
experimenter, and thus were an artifact of the experimental situation. In fact, I can not recall that this type of
content was ever employed in real
situations. In real searches out on the
water, the history of observations serves to resolve the ambiguities which
emerge in the experiment as a product of reasoning with uncertainty. Furthermore, captains on the water do not
regularly engage in giving each other lessons about the problems of
fishing. Instead, they make judgements
based on what they have seen and offer these up as statements of what they are
seeing.
The purpose of
coding speakers' descriptions, as reported above, was to generate a means for
comparing descriptions—at the level of regularity in choice of language forms
used to communicate about sonar displays.
The notion here is that speakers employ words, and constructs built from
words, which serve to model what they are seeing on their sonar displays,
enabling communication of this to others—so long as the constructs, visual
representations, and experience with each of these are shared by interlocutors.
A simple
measure of description similarity was thus devised to attempt to measure the
extent to which language use is similar for referring to similar sonar
representations. The measure performs a
simple count of the components shared between all pairs of features of two
descriptions, in the following way:
Given two features f1={a1,b1,c1}
and f2={a2,b2,c2},
F(f1,f2) =
1 point is
awarded if a1=a2 (the same NOUN is employed),
2 more points
are awarded if b1=b2 and c1=c2 (the
same MODIFIER is employed).[17]
The language
similarity between two descriptions d1 and d2 (call it, L(d1,d2)) is defined as
the sum of F(fi,fj) for all fi in d1 and all fj in d2. The language similarity, across speakers
{s1, s2, ...sn}=S, for some pair of sonar images (card kl and card km) is then L(S,kl,km) given by the sum obtained by
applying L to all unique pairs of descriptions generated by members of S about
card kl or card km (excluding all terms L(di,di) generated by comparing
descriptions with themselves).
L then, gives a measure of the
similarity in language use, among a set of speakers, regarding a pair of sonar
displays. Notice that this measure is
strictly at the level of shared language forms. Although there is wide-spread agreement about the meanings of
sonar images (in terms of the potential for catching fish at the represented
location) it was shown in a previous section that this varies significantly
among fishermen. In order to ascertain
the extent of language sharing at the level of "meaning" we need a way to
determine if the sharing of language forms is systematically related to
speakers understandings about what the sonar images stand for as
representations of fish and prospects for catching fish.
Define P((si,kl),
(sj,km)) to be the similarity between card kl (as judged by speaker si) and
card km (as judged by speaker sj), obtained by a simple distance metric on
these speakers' pile assignments for kl and km during the Pile Sort task. Call those assignments—that is, the value
subject si gave card kl and the value subject sj gave card km in the Pile
Task—p(si,kl) and p(sj,km), respectively.
Then:
P((si,kl),
(sj,km)) = 1.0 - |p(si,kl) - p(sj,km)| / maxp
Since pile
assignment values ranged from 1 to 7, maxp = 6. The pile similarity, across subjects {s1,s2,....sn} = S, for some
pair of sonar images (card kl and km) is then P(S,kl,km) given by the average
obtained by applying P to all unique pairs of pile assignments made by members
of S to cards kl or km (excluding
all terms P((si,kj),(si,kj))—that is, terms generated by comparing pile
assignments given by the same subject for the same card). Thus—for n subjects—P (and L) involves the
average (resp. sum) of n^2 terms of P (resp. L) if kl≠km, and n^2 - n
terms of P (resp. L) if kl=km.
In summary, for
the 50 cards under consideration, there are 1275 terms [(50^2 - 50)/2 + 50]
each of P and L, one for each unique pair of cards. The values of P and L for each pair of cards, across all 12
speakers,[18] are shown plotted against each other in the
scattergram of Figure 77.
The most
significant thing to notice about the scattergram of Figure 77 is that large values of L (high consensus in the use of language forms for pairs of
cards) quite successfully predict large
values of P (similar evaluations
of pairs of cards in terms of "goodness").
In other terms, the language used by subjects to communicate about cards
reflects the meanings of cards in terms of understandings about the
representations of fish believed to be shown in each sonar image. Pairs of cards which share many language
features, also share properties which lead subjects to believe they represent
similar prospects for catching fish. Furthermore,
small values of P (dissimilar evaluations of pairs of cards in terms of
"goodness") quite successfully predict
small values of L (low consensus
in the use of language forms for pairs of cards). Pairs of cards which subjects believe to represent very different
prospects for catching fish (or for which there is little agreement about those
prospects) do not engender shared language features in the descriptions of
those cards by speakers.
The fact that
large numbers of points lie in the region of the plot marked by large values of
P and small values of L can be explained by three different
factors. First of all, there were many
cards (nearly a third of the 50 it will be recalled) which were summarily
regarded as "bad" by pilers, with little consideration given. Pairs of these cards will generate large P scores because they all landed in
pile one or two for all subjects.
Likewise, these cards were not of particular interest to speakers, as
witnessed by the short descriptions employed—such as "it's nothing"—which (for
pairs of such descriptions) will generate small L scores even though the language used is completely shared. Secondly, the measure of language sharing in
the experiment is unidirectional in the sense that speakers utterances are
scored for their positive contribution to consensus (the extent they match
other usage) but there is no penalty exacted for failing to be
informative. Speakers could choose to
(and often appeared to) under represent the contents of the displays in their
speech, leading to many instances where descriptions failed to score well due
to lack of data. Thirdly, the coding
and scoring of features (as described above) was very conservative in that
every attempt was made to keep the analysis at the level of language forms in
order to avoid making semantic judgements which would bias the results. This clearly resulted in a very "noisy"
measure of language consensus. These
three factors together contribute to the large population of points in the
upper left-hand corner of Figure 77
The systematic
relationship found, between language use about and evaluations of sonar images,
provides an explanation for the surprising result which showed large positive
correlation between inter-subject agreement in the non-linguistic Pile and
linguistic Match tasks. The
Pile-Similarity/Language-Similarity plot (Figure 77) suggests that many
descriptions entail constructs which effectively model the significant features
of sonar representations as fishermen understand these. Furthermore, this understanding of sonar
representations—both linguistic and visual—is likely to be grounded in the
actual practices which generate shared sightings of sonar images, talk about
these, and experience with the catching of fish which they represent. To the extent that this is true, we should expect subjects who agree on the
Pile task to also agree on the Match task, which is what we found to be the
case.
Figure 78 shows
a scattergram of the values of P and
L generated for only those cards,
descriptions, and evaluations involved in part two of the Match Task. The plot contains 78 points [(12^2 - 12)/2 +
12], one for each card-pair of the 12 cards used in this part of the Match
Task. The L values were generated from a feature scoring of the 120
descriptions which constituted the items to be matched to the 12 cards in this
part of the experiment. The P values were generated from the pile
values assigned to the 12 cards in question by the 12 listeners. The plot provides even more convincing
evidence for the proposition that similar inter-subject agreement in the Pile
and Match tasks is a likely outcome.
Again, we see that cards which are rated similarly for their "goodness"
(measured by subject consensus in the Pile task) tend to share language
features. For the data involved in part
two of the Match Task, this relationship is so strong that it seems to
approximate a linear relationship exhibiting uniform variance (see Figure
78). A regression model of this
relationship reveals a correlation (R) of .52 which accounts for almost 27
percent (R-squared) of the variance in the data.
While the
correlation between individual performances in the Pile Sort and Match Task
might lead some to the conclusion that "individual competence" of an
unspecified sort accounts for the data, this is clearly an inadequate
explanation. There is strong evidence
for the convergence of shared practice-based variables accounting for the
patterns of agreement across the tasks of this experiment. Factors such as the type of fishing pursued
and the amount of time one has spent interpreting and evaluating sonar images
as a captain or a regular crew member appear to play significant roles in
determining the distribution of agreement in both the Pile Task (the assignment
of private evaluations to sonar images) and the Match Task (the assignment of
publicly generated descriptions to sonar images). Inter-listener agreement in the Match Task appeared to depend
even more significantly than the Pile Task upon the channels through which
communication about sonar images takes place.
This should be an expected finding since the Match Task entails explicit
employment of the terms which propagate through these communication channels in
the various practices of the fishermen.
Finally, the positive correlation in subject agreement across the two
tasks, although surprisingly high, is reasonably accounted for by the
systematic use of language to reflect (and possibly construct) fishermen's
understandings about sonar representations.
Throughout the
analysis of this chapter, relatively little has been said about the phenomena entailed in generating
interpretations of sonar images and communicating these to others. Instead, a more formal treatment of these
phenomena—made possible by factoring them out via measures of subject
performance in experimental tasks—was attempted in order to yield structural
and quantitative comparisons which support the principles of practice-based
cognitive and social coordination elaborated upon in other chapters in more
phenomenological terms. Nonetheless, it
is important to return to those phenomena—the details of the situations,
processes, and contents entailed in evaluating sonar images, seeing them as
representations of fish, and communicating these images and what they represent
to others.
One pervasive
feature of the language used by fishermen to describe sonar images is the
multi-referential framework employed.
As with the discourse entailed in the activity of disambiguating the
identities of radar echoes (see Chapter 4), the talk about sonar images takes
on a complex referential structure due to the fundamental roles played by the
representations generated by the device which mediates actors interpretations. Captains of a team of pair of trawlers
negotiating shared understanding of boat locations mediated by different
representations (as shown in Chapter 4) will regularly take the perspective of
their interlocutor when speaking, in order to facilitate mutual understanding. Occasionally, this process will break down
under the introduction of foreign objects into the representational field of
one interlocutor (e.g., real or false products of the speaker's local radar
device) which do not make sense, or are unavailable, to the other
interlocutor. When this happens,
additional repairs are made to the speech which attempt to eliminate its
multi-referential nature. And yet, in
general, the discourse affords very robust generation of shared (or at least,
collective) knowledge among the driving captains of the pair.
In the current
chapter, the analysis of language used by fishermen to talk about sonar
displays makes it clear that a similar kind of perspective-taking is inherent
in the sharing of forms and meanings which make communication possible. This socially constructed intersubjectivity
is only partially constrained by the differences between, on the one hand,
structures displayed on the sonar screen and, on the other hand, structures
underneath the surface of the water which are represented by the sonar
display. Thus, language is free to
drift (and regularly does) with respect to the domains of reference—display, or
underwater objects. For instance,
although "red," "yellow," "blue," and "white" refer to properties of the display
screen, they simultaneously refer to the properties of the fish which the
display is representing. Common
examples evidenced in the descriptions elicited in the experiment included,
"red groups which lie on the bottom," and "productive sprinkle with a little
red up there [near the surface]" In
both cases, red is a property used to modify nouns which are somehow
simultaneously on the display and located in the water.
This
multi-referentiality of the language does not lead to confusion because the
device effectively translates differences in the intensity and timing of sonar
returns into differences in display characteristics which are salient to sonar
users—namely, color and depth scaling.
Lacking any additional constraints on communication which would require
more precise locating of the referents of speech, the system of linguistic
codes can (for fishermen) transparently denote the represented fish via the
mediating properties of the device by which fishermen regularly experience the
resource which is the object of their labors.
Clearly, it is not just the device which makes this system work but is
also the demands for shared understanding—functional coordination among
participants of the practice—engendered by the specifics of the activity
itself. Thus a phrase such as "there is
no color" unproblematically (yet paradoxically) refers both to a sonar
image—which in fact does show some color—and represented fish which is
interpreted to be a poor prospect for fishing.
In a similar manner, constructs such as "a little bit which disappears
off the screen," attempt to simultaneously quantify the fish (that which is
represented) and the display (that which is representing)—and appear to do so
quite successfully.
Although
descriptive terms such as "tight," "collected," "hard," "dispersed," "empty,"
and "concentrated" clearly reflect properties of the display image, these
descriptors are also very salient (often more salient) in terms of the
properties of the state and behavior of fish which the display represents. Clearly the display can represent something that's "hard" (again, via the relationship
between intensity of returned signal and color) but cannot be "hard" or even show
"hardness" directly. Similarly, patches
of solid color on the display can be "collected," "dispersed," "tight
(together)," and "concentrated," and the entire display could perhaps be called
"empty" when it is solid black, but these descriptions are more apt (I claim)
for characterizing the density of discrete entities in three-dimensional
space—yielding evidence that theses phrases may simultaneously evoke properties
of both represented and representational objects for practitioners.
Of course, the
flexibility in this communication system has limits, and many of these are
placed by the constraints on shareability which operate over time (and in the
ongoing efforts to coordinate actions within the activity and which take place
in time) to shape the form of language consensus. In the data of my experiment, descriptors such as "juicy,"
"grainy," "showers," "toothpick," "groups," and "sparse" were examples of
idiosyncratic terms which attempt to capture something deemed important by
taking advantage of displayed structure in the image.[19] These terms were idiosyncratic because they
were employed only once, or many times by only one speaker, although many of
them are understandable either because of the distinct image they invoke or
because the term has a history of use (even though it may currently be
unpopular).
"Fleck" and
"sprinkle" were the most popular nouns, and they were used in a consistent
manner to identify well-bounded regions of solid color (schools of fish) and
fine-grained, dispersed color (fish) respectively. In fact, "fleck" is so consistently utilized to denote a
self-standing, solid area of color on the display (almost always red in color),
that I would argue it is particularly salient in terms of it's representation
as a shoal of fish.[20] As such, a "fleck" takes on properties (in
speech) which are normally attributed to schooling behavior of herring as this
is known to the fishermen. For example,
a fleck can "take hold of the bottom" and "lie on the bottom" or can "stand in
the middle of the water" and can be "cast upon" with the seine "drawn upon"
with a trawl and some instances of "fleck" can be modified with a term to
denote the fact that it contains a particular kind of herring or would probably
yield a particular size catch. It seems
clear that this term has particular salience for fishermen as a description of
fish—in particular, of a shoal of herring which is tightly packed together in
the water as it typically will be when the sea is cold and calm and under
certain conditions of water luminance, current, and the behavioral and life
cycles of fish.
However, a less
popular usage of "fleck" was evidenced in some speakers' efforts to employ it
more generically, or compositionally, to denote smaller—usually
multiple—patches of solid color on the display screen. This usage was clearly marked by use of the
plural form of the noun, "flecks," and often referred to solid color objects
located within some other referenced field of the display image. This usage seldom entailed content which
could be said to evoke reference to fish in the water, but instead focused upon
aspects of the display image itself.[21]
In the quoted
conversation which opens this chapter, the friendly disagreement between two
fishermen over the use of "flecks" and "sprinkle" provides evidence that some
kind of boundary between the two terms must be maintained in the current
linguistic system to afford intelligibility.
In that conversation Fisherman 1 attempts to combine these two popular
terms as a way of denoting a state of the display which lies in between the
prototypical states represented by the two terms.[22] Fisherman 2 complains about this usage and
offers a construction based on the modification of "sprinkle" (a "strong
sprinkle") which covers the intended meaning of a "sprinkle" that is showing
patches of solid color in places. Based
on the observed use of "fleck" in the experimental setting, it might be
conjectured that it's salience as a term which is simultaneously a state of the
display and a school of fish inhibits acceptance of using the term as a
productive modifier of other terms which focus primarily upon the state of the
display—a use which would conflict with the term's salience as a representation
of a shoal of fish.
[1]Interestingly, this language is also consistent with the ways color monitors represent fish, white being the end of the scale used to represent low amplitude sonar returns and thus low density concentrations of fish.
[2]One subject performed part two of the Matching Task (see below) on board during a fishing outing.
[3]Here and throughout this experiment all interaction was conducted in Swedish.
[4]NOTE: Because one subject's stimuli differed from the other subjects' by 5 cards (these 5 cards were replaced with higher quality sonar images for subsequent subjects), comparisons between subjects which rely on identical stimuli entails the 15 subjects responses on only 45 of the 50 cards. Alternatively, in order to consider responses to all 50 cards, data from 1 subject would have to have been dropped from the sample. A test of the two alternatives did not reveal enough difference to justify dropping this subject from the sample, particularly in light of a need to compare subject performance in this task with the results from a smaller sample in the Matching Task.
[5]Note that these are the same individuals which used more "algorithmic" techniques in their sorting. These individuals were all younger fishermen with significantly less experience and more recent (and, in many cases, more extensive) formal schooling.
[6]The test employed was a comparison of u1 and u2, the means of independent samples from an assumed normally distributed population. z = (u1 - u2) / Sqroot(SE1^2 + SE2^2) = 2.79. A two-factor (Type of Fishing, and Card) analysis of variance of the pile values assigned to the set of (45) cards by all subjects, shows this difference between trawling fishermen and seiners to be concentrated to a subset of the cards, namely those which were not considered "bad" by everyone.
[7]That is, the question posed to subjects should not be interpreted as a more general one of "Describe to me (the researcher) what you see in front of you." Of course, there is no guarantee that this latter speech context was not also salient for subjects, and there is some evidence (see below) that it often was.
[8]By comparison, chance agreement on this task would be 1 in 12, or 8% with a standard error of 9.2%, making even the lowest level of agreement (40%) significant at the .001 level (one-tailed t-test).
[9]A one factor (Speaker) analysis of variance showed differences between speakers as a set was very significant (F(9) = 5.048, p = .0001).
[10]It should be noted here that this kind of verbosity was explicitly devalued in the local culture. As reported in other chapters, fishermen learn by doing and are valued by working hard. This is often explicitly contrasted with those who "talk a lot," make claims for personal knowledge or abilities, and are thus antithetical to the ideal personality of being a captain. The point is that this social fact may have worked against the objectives of the experiment (to elicit representative samples of communication about sonar) due to the artificial nature of the dialog for speakers in the experimental context.
[11]One form of this collective work is exhibited in the form of kräftfiske, where the activity of many boats dragging their trawls over the same area of bottom actually stirs the small lobsters up, making them more catchable for all who are fishing in the area.
[12]There were actually two instances of illegible responses, one for subject 3 and one for subject 4, which had to be thrown out. Note also that we expect, by chance, any two subjects to agree on 1 in 12 items, or 8% with a standard error of 2.5% on the entire set of 120 match items. Thus every entry in the matrix is significant at the .001 level.
[13]It should be noted that all of the MDS solutions reported here are "robust." That is, they were found repeatedly (under different hill-climbing parameters of the error minimization algorithm) and consistently, and therefore represent either a very common local minimum of the solution space or a true global minimum. Furthermore, similar solutions were also found at other dimensionalities in order to check for the possibility of local minima in the reported solutions.
[14]The entire analysis was performed in the original language to avoid translation biasing, but is reported here mostly in English glosses of the originals.
[15]It will be recalled that the speech context was defined to be during a search for herring out on the water. Several subjects took this to mean that the previous cards could be employed as referents of speech. This was not something I had anticipated and yet (in the real setting) this time-course contextualization of sightings is a fundamental component of the joint activity of interpreting sonar representations (see Chapter 4).
[16]This multi-referentiality of the language is an interesting phenomenon in its own right and is discussed further below.
[17]Modifiers involving integer values, as opposed to labels, were said to match if the values were within 5 units of each other.
[18]Although only 10 of the original 12 subjects for whom descriptions were elicited in part one of the Match Task were "speakers" in part two of the Match Task, here we are considering the entire original sample of descriptions elicited from participants in part one as "speakers" (yielding 12x50=600 descriptions). Actually, 2 speakers' descriptions included 5 cards each which were later exchanged in the set of 50 sonar representations. These 10 descriptions were thrown out, leaving 590 descriptions as the data corpus under consideration in this section.
[19]"Juicy" was used (only once and by one subject) to characterize a "stripe" which was not well bounded (a solid, horizontal, red area of the display which had ragged edges which gave the appearance of a liquid dripping in both vertical directions). Interestingly, this particular image (and the term used to describe it) also connote a very "delicious" find as this sonar image revealed a very lucrative shoal of fish.
[20]"Fleck" or "flecks" was used in 71 of the 590 descriptions in reference to 20 of the 50 cards in part one of the Match Task. Of these, 56 instances involved only 9 of the 50 cards. 11 of the 12 speakers of part one of the Match Task used "fleck" to describe the sonar image shown in Figure 59.
[21]All speakers used "fleck," in the singular, in at least two different descriptions. The majority of speakers used it in this form in four or five different descriptions. Only two subjects used "flecks," in the plural, more than once and four others used the term this way just once.
[22]In fact, Fisherman 1 was one of the two speakers in the experiment who used "flecks" in the plural form in more than two of his descriptions. Fisherman 2, on the other hand, was a speaker in the experiment who did not use "flecks" in the plural form in any of his descriptions.