COMPUTERS, FIELDWORK, AND THE ANALYSIS OF CULTURAL SYSTEMS

• S is transformed to S64 (GRAM#1 [2]), denoting that the piece will contain 64 units, or syllables.
• 2. 5 S64 has been randomly transformed as A32 V8 A'24 (GRAM#2 [2]). Eight units of variation will be interposed between fixed patterns of 32 and 24 units respectively.
• 3. 5 V8 is broken down into a string of Vs (GRAM#3 [3]).
• 4. 5 The first four Vs have been randomly chosen as a block V4 (GRAM#4 [4]).
• 5. 5 The second four Vs have similarly been chosen as a block V4 (again GRAM#4 [4]). Note that sub-grammar 5 will not be invoked because all Vs have not been transformed in sub-grammar 4. Had blocks been chosen that transformed only the first seven Vs, say V4 V3, then the remaining V would need to be absorbed into the preceding blocks (using GRAM#5 [3] or [4]) because a block comprising a single unit does not exist here.
• 6. 5 The first V4 is randomly transformed into tirakita (GRAM#6 [7]).
• 7. 5 The second V4 is randomly transformed into dhinagena (GRAM#6 [9]).
• 8. 5 The fixed pattern A'24 is replaced (GRAM#7 [3]).
• 9. 5 The fixed pattern A32 is replaced (GRAM#7 [4]), leaving us with the variation:

dha-ge-@tirakita@dha--dha@genadha- 
tirakita@dha--dha@genadhage@tinakena 
tirakita@dhinagena@dha--dha@genadha- 
tirakita@dha--dha@genadhage@tinakena

A number of other devices are available to the analyst should it prove necessary to elaborate the grammar to force it to generate more acceptable pieces of music. For example, all rules in sub-grammar 6 are context-free except for rule 4 which carries a negative context denoted by the hash (#) sign. Rule 4 states that a block of two units (V2) can only be transformed into gena if it is not preceded by ge. Furthermore, coefficients of likelihood, or `weights', may be attributed to rules in order to increase or decrease their chances of being invoked. In sub-grammar 2, a choice is offered between two rules that transform S64 into structures defining the positions of fixed and varied patterns. Weights of 60 and 40 have been assigned to rules 2 and 3 respectively, implying that rule 2 will be used in 60% of variations and rule 3 in 40%. Weights can be used to channel the generative process along paths more likely to be used by particular individuals.

Once the procedure for the synthesis of variations has been grasped, the machine's analytical mode can be easily understood as it is essentially the reverse of what has been described above. A variation created for a specific piece is in effect subjected to a membership test that determines whether it belongs to the `language' expressed by the grammar. Firstly, the fixed patterns are replaced by symbols representing their length and position in the piece; note that larger strings must be considered before smaller strings, because if a smaller string (GRAM#7 [3]) is a subset of the larger string (GRAM#7 [4]) and is replaced first, then there is the possibility that a number of syllables will remain untransformed, or that incorrect sequences of structural symbols will be generated. In either case, the analysis will break down. A successful analysis will ultimately terminate with the starting symbol (S).

It takes a number of minutes to synthesize and analyse simple variations by hand, and variations derived from or belonging to elaborate grammars full of complex structural information contextual rules and weights, can involve hours of calculation. The Bol Processor reduces this process to one or two seconds for any synthesis or analysis, however intricate, and has the added advantage that it can store and later recall the paths it took, so allowing for a detailed reconstruction of particular pieces that may have proved to be of special interest.

Experiments with the Bol Processor begin with the construction of a simple, initial hypothetical model for a piece agreed upon by the analyst and the informant. This model is entered into the machine and activated in order to generate variations. An important principle that must be observed is that the grammar should be capable of analysing everything it creates, thus displaying internal consistency. The new variations are then submitted to the critical ear of the informant, who may comment on them using a simple set of semantic categories (good, bad, indifferent), or instead may choose to alter or correct them. Interestingly, when asked to repeat pieces generated by the machine, informants often inadvertently altered or corrected them anyway. Armed with this new data the analyst then sets about modifying the grammar in an attempt to make it reflect the various musical values expressed by the informant. The updated grammar is then tested and remodelled until a point is reached where, by means of this constant interaction of analysis and synthesis, the grammar is capable of generating a high proportion of musically acceptable variations.

Despite the fact that the theory of formal languages is incapable of describing certain aspects of the musical process under investigation, it can be used to simulate those aspects of music generation and evaluation that are procedural: in other words, surface structures as manifest in individual pieces of music. By controlling the Bol Processor in many more experimental situations in which musical patterns are generated and recognised, it is hoped that it will be possible to highlight the non-procedural, underlying structures of musical thought by systematically testing the power of hypothetical models to predict correct and aesthetically acceptable pieces of music; acceptable, that is, to the informants who had a crucial role in shaping the models. So, the role of the Bol Processor is to disprove very rapidly previous hypotheses and to allow for increasingly valid hypotheses of musical thought as represented through musical structure to be built, tested, and, in turn, disproved and replaced.

Unquestionably, the Bol Processor has been a great success in many respects. It has enabled me to focus at a deep level on a cultural system that has not so far lent itself to systematic investigation, where the onus is on informants themselves to make the analytical decisions. Furthermore, it has been an object of great interest to musicians, many of whom were inspired to comment that they had never before been posed the kinds of questions that the machine raised during the course of experiments, and that this forced them to consider their music and, moreover, their musical decisions in a new light. From a slightly different point of view, many months of experiments have produced dozens of floppy discs filled with musical grammars that indicate some interesting details about the music and which can perform quite well. Indeed, I can quote one informant as saying recently: `In some respects, the machine has improved on the variations I created for this piece'. But apart from showing that it is possible to replicate with considerable accuracy the products of a complex musical system, the Bol Processor has not so far thrown much light on any of the underlying cognitive processes that generate them. No doubt this situation will improve when considerably more data have been collected and tested, and I, as operator, have developed greater efficiency and expertise in both manipulating the machine and constructing formal models.

So just how far will it be necessary to go before the anticipated results become apparent? Or are these expectations futile anyway? For example, despite the fact that a number of models were developed during a series of experiments to a relatively high degree of sophistication, they were palpably inadequate to account for musical creativity as manifest in subsequent performances that were transcribed and fed into the Bol Processor for analysis. These new improvisations simply smashed the models to pieces. Therefore, it appears that to create a description of the music sufficiently general enough to encompass all possibilities is as huge a task as it is thankless. It would evidently be more useful to restrict the investigation to specific structural or combinatory problems that can be isolated and examined systematically. But this is not to say that Bel and I should refrain from considering the wider socio-cultural explanations for the variability of models designed for the same pieces of music. For example, when teaching their students, master musicians seem to be aware that their material will be memorised, practised, performed, and thus preserved for posterity. Although novices are themselves encouraged to improvise, they nevertheless retain throughout their entire lives repertoires of `fixed improvisations' dictated by their teachers. In view of the fact that a master is remembered as much by the quality of his students as by his own ability, the musical information he imparts must in these contexts be correct and uncontroversial. He therefore tends to follow step by step procedures, so adhering more to a simple, idealised model. The same would presumably be true of recordings which are likely to be replayed many times and would doubtless be subject to detailed scrutiny. Conversely, performance situations in the presence of knowledgeable listeners offer musicians greater licence to explore new channels of creativity and to stretch the limits of musical accept ability.

It has become apparent that some of the problems currently being encountered in this research may be due to the fact that I have fallen into precisely the same sort of trap to which I alerted the database programmers earlier: namely, that have failed to investigate in detail the assumptions inherent in the initial hypothetical models of musical structure I programmed into the Bol Processor before subjecting them to musicians' scrutiny. Sometimes I found that models, though obviously inadequate, nevertheless developed very little during experiments, probably because I disregarded empirical evidence and instead formalised my own intuitions (particularly regarding syllable combinations) as a type of short-cut: intuitions which were possibly several steps ahead of what might have been expected during preliminary stages. Alternatively, a number of procedural problems could perhaps be put down to incorrect assumptions on my behalf. Naturally, the more accomplished one becomes at operating the system one is studying, the more difficult it is to conduct an objective investigation. I have learned to play tabla music to a reasonably high standard, and have become a sort of evolue\*' in reverse: one who has come to grips with a sophisticated, but essentially foreign, cultural system. But, of course, I must not assume that I have the same kinds of musical ideas or that I use the same cognitive processes as my informants, in spite of the fact that I can produce similar results during performance.

One solution may lie in the further automation of the process of analysis. This would present two major advantages: firstly, a machine cannot assume that which it does not already know, and therefore must act only on empirical evidence; and secondly, it would promote a more direct inter-action between informant and computer, so reducing the researcher's role as `interpreter'. This last point is important because, as time goes by, I am faced more and more with credibility problems: the musicians know that I know most of the answers to the questions put to them, and thus they steadily become less cooperative. Software programs that facilitate greater automation now exist, or at least are at present being developed. Bernard Bel has created Segmentation, a program that will chunk syllables into compound blocks solely on the evidence of the ways in which they have been combined and juxtaposed in examples provided by informants. Another, as yet prototypical, program, called Formal, will incorporate Segmentation in an attempt to build its own initial hypothetical grammar. Both have been written in Prolog for the Apple Macintosh computer system, and are consequently not as yet transportable to small, portable systems like the Apple IIc. That may be only matter of time, because rapid technological advances in computer design ensure that more powerful and compact systems will become available in the next few years, if not months. But in the meantime, there is a strong argument for conducting a series of experiments in computer laboratories, a move I have until now resisted and argued against strongly on the grounds that such situations are culturally insensitive. However, it should not be thought that I am proposing to pluck informants from low-technology societies and dump them in totally bewildering high technology environments; my informants belong to large, North Indian conurbations, shave with electric razors, use automatic cameras, and occasionally listen to Glenn Miller on Sony Walkmans. Many have even travelled to Europe and the Middle East on concert tours. Moreover, they have become habituated over the last seven or so years to eccentric researchers wielding sophisticated tape recorders and portable computer systems. Thus, in order to develop more adeqate research tools for the field in the course of what is, after all, an experimental project, the Western laboratory seems to be the most likely scenario for the next stage in Bol Processor research.

Although I have been describing the analysis of a musical system, it should be realised that, like other `expert systems', the Bol Processor is not restricted to work within one specific domain of knowledge but may be used to process other kinds of cultural data as well. Experiments with expert systems have been conducted in what may be termed the more mainstream areas of anthropology: for example, in Fischer's study of Punjabi arranged marriages, a machine digested a body of information relating to a number of variables such as caste, consanguinity, education, beauty, size of dowry, and the location of families before it pronounced on the suitability of a possible marriage [(.fischer 1986.)]. Studies of kinship systems is another domain that stands to benefit from computer-aided research methods, though conversely belief systems might prove less amenable to codification and formal representation. The results of further experiments with machines like the Bol Processor will determine the extent to which computer-aided research can contribute constructively to techniques and methodologies for the analysis of cognitive systems as representations of cultural systems. Both anthropology and artificial intelligence stand to benefit enormously from mutual interaction in the ways in which they attempt to delineate broad representational and operational models of knowledge and reasoning.

Bibliography (in Refer Format)

%A H. Russell Bernard 
%T The construction of primary data in cultural anthropology 
%J Current Anthropology 
%V 27 
%N 4 
%P 382-96 
%D 1986

%A Rada Dyson-Hudson 
%A Nigel Dyson-Hudson 
%T Computers for anthropological fieldwork 
%J Current Anthropology 
%V 27 
%N 5 
%P 530-1 
%D 1986

%A Robert S. Gottlieb 
%T The major traditions of North Indian tabla drumming 
%M Vol. 1 
%C Munich 
%I Musik Verlag Emil Katzbichler 
%D 1977

%A Jim Kippen 
%T The dialectical approach: 
a methodology for the analysis of tabla music 
%J International Council for Traditional Music (UK Chapter) Bulletin 
%V 12 
%P 4-12 
%D 1985

%A Jim Kippen 
%T Computational techniques in musical analysis 
%J Bulletin of Information on Computing and Anthropology 
%V 4 
%P 1-4 
%D 1986

%A Jim Kippen 
%T The tabla of Lucknow: 
a cultural analysis of a musical tradition 
%S Cambridge Studies in Ethnomusicology 
%C Cambridge 
%I Cambridge University Press 
%D 1988

%A Jim Kippen 
%T Computing culture: 
some reflections on the uses of computers in anthropological research 
%J Current Anthropology 
%V x 
%P x 
%D Forthcoming