Acquiring qualitative knowledge about complex agroecosystems. Part 1: Representation as natural language

Agricultural

PII:

SO308-521X(97)00048-6

Sys~ms,

Vol. 56, No. 3, pp. 341-363, 1998 CC?1997 Elsevier Science Ltd All rights resewed. Printed in Great Britain 0308-521X/98 $19.00+0.00

ELSEVIER

Acquiring Qualitative Knowledge About Complex Agroecosystems. Part 1: Representation as Natural Language F. L. Sinclair” & D. H. Walke@ “School of Agricultural and Forest %XRO Tropical Agriculture, Private (Received

17 March

Sciences, University of Wales, Bangor, UK Mail Bag, PO Aitkenvale, Qld 4814, Australia 1997; accepted

2 1 June 1997)

ABSTRACT Quantitative and predictive scientific understanding about complex agroecosystems such as traditional agroforestry practices remains sparse. In contrast, farmers, development professionals and others have a qualitative understanding about the systems the-v have experience of which may be a useful resource in complementing scienttj?c knowledge. Participatory research approaches help to capitalise on this complementarity. However, explicit methods for recording qualitative knowledge from these various sources are requiredfor the basis of this complementaritv to be rigorously investigated. This paper reports on the development of a methodology to achieve this, designed spectfically with reference to indigenous ecological knowledge about agroforestry, but of wider applicability. Principles and requirements for such representation are discussed and an approach based on organising statements written in ‘natural language’ is described. (G 1997 Elsevier Science Ltd

INTRODUCTION There has been much recent scientific interest in the use of complex agroecosystems, especially those involving trees, as alternatives to green revolution technologies in circumstances where more intensive use of support energy is not appropriate. Such circumstances occur, for example, where capital is more scarce than labour or there is a premium on biodiversity (as in support zones around natural forest reserves). Harnessing available knowledge in 341

342

F. L. Sinclair, D. H. Walker

contributing to the development and management of these complex agroecosystems poses interesting challenges for researchers and decision makers. As these agroecosystems involve diverse mixtures of plants and animals, knowledge is required from across scientific disciplines that have traditionally developed separately-forestry, agriculture and ecology, for example. More effective approaches to interdisciplinary research are required. Furthermore, farmers who have been operating such systems for far longer than there has been a scientific interest, may be expected to have developed an understanding of their ecological functioning. Indeed, it has been previously shown that subsistence farmers often have a sophisticated ecological rationality underlying their practice (Richards, 1985; Warren et al., 1994; Thapa et al., 1995). This knowledge can be expected to be a key resource for planning and implementing research and development programmes in relation to complex agroecosystems, both because it provides an informed basis from which to proceed (Chambers and Ghildyal, 1985) and because relevant programmes are more likely to be achieved where planned with due regard for the farmers’ perspective on needs and priorities (Rocheleau, 1987). However, analysis, synthesis and interpretation of the qualitative knowledge held by farmers and professionals across a range of disciplines are conceptually and practically challenging. These issues are particularly acute in agroecosystems that comprise a complex polyculture of species where universal laws are much less likely to be available and applicable than in simplified cropping systems and where there is less scope to ameliorate the resource base using a standardised recipe of inputs. This reflects both the partial understanding of the ecology of such systems by scientists and their extreme heterogeneity over short distances. As a result, practical, contextual knowledge of how the environment varies in a particular instance becomes important in determining appropriate action, and farmers are often particularly knowledgeable at this scale of operation (Richards, 1994). The response to these challenges is seen in the development of the systems-based ethos of much rural development now manifest in participatory approaches to research, development and extension (Budelman, 1996). By contrast, the more traditional approach to dealing with complex systems, namely their abstraction as quantitative, simulation models, has had a relatively limited impact on the exploration and development of complex agroecosystems (Muetzelfeldt and Sinclair, 1993). This is perhaps not surprising given the infancy of our understanding of many of the components of these systems, let alone the myriad interactions amongst them (Anderson and Sinclair, 1993). There is also, however, a tradition of using logical rules to represent system dynamics in ecological systems (Starfield and Beloch, 1983; Robertson et al., 1991) and of combining qualitative and quantitative

Complex argroecosystems. Part 1

343

models in the development of decision support systems in agriculture (Plant, 1997). There is, therefore, a powerful case to be made for using what qualitative knowledge is available about the behaviour of complex agroecosytems from farmers, research and development workers and documented science, as a basis for modelling them. This clearly requires a means of obtaining relevant information from these sources and representing it in an appropriate way. We have demonstrated elsewhere, through analysis of knowledge flows at frontline research and development institutions, that there are opportunities for improving the targeting of research and development in support of resource-poor farmers by better understanding their existing priorities and knowledge and by marshalling the existing professional knowledge available across disciplines (Walker et al., 1997).* Developing a conceptual model of a complex agroecosystem on the basis of a synthesis of knowledge from a range of sources clearly requires efficient and effective collection and collation of knowledge from many farmers, development workers and researchers. The knowledge articulated by these sources will be largely qualitative. To be useful, this knowledge needs to be recorded in some form that allows it to be effectively stored, accessed, analysed, synthesised and, thereby, to be available for future use. The process of recording the knowledge articulated by informants is referred to here as representation. Making effective use of whatever qualitative knowledge exists about a complex agroecosystem requires methods of representation that: ??

?? ??

are sufficiently rigorous to provide a repeatable means of knowledge capture from disparate sources; result in comparable sets of knowledge from these various sources; and are sufficiently flexible to cater for a diverse and uncertain domain.

A cost-effective approach to collecting and collating knowledge available from farmers and professionals alike is needed. Furthermore, it is evident that where a large investment is made in knowledge acquisition from a sparse and remote farming population in the context of rural development, if the knowledge acquired can be used for multiple purposes, the process will be far more cost-effective than if a series of separate acquisition exercises are required for each purpose. The approach taken to representing knowledge is of fundamental importance because it defines both how the knowledge can be subsequently used

*It should be noted that the research described approaches to participatory R,D&E. Approaches in a companion paper.

in this paper is intended to complement to integrating these methods are discussed

344

F. L. Sinclair, D. H. Walker

and how much it is transformed from the manner in which it was articulated by the source. The purpose of this paper is to explore appropriate means of representing qualitative knowledge about the ecology of complex agroecosystems with the intention of providing tools for use by researchers and development professionals engaged in the improvement of traditional agroforestry practices.

METHOD The approach described in this paper was developed as a part of a collaborative research programme investigating the use of indigenous ecological knowledge about agroforestry systems in agroforestry research and extension programmes. The approach was interactively developed in conjunction with field programmes in five countries (Nepal, Tanzania, Thailand, India and Sri Lanka). Field work in each case was done by natural science graduates working within a rural development context. This field work contributed to the development of, and made use of, knowledge-based systems software and associated methodological guidelines for the creation of knowledge bases about the ecology of agroforestry systems (Walker et al., 1994). The development cycle comprised specification of a method of acquisition followed by implementation as methodological guidelines and, where appropriate, computer software. The implementations were then field-tested and refined or redesigned on the basis of evaluating both the process of acquisition (time taken and ease of use) and the quality of the knowledge represented as compared to that articulated by informants (clarity, completeness, coherence and consistency). The work thus made use of a number of iterations of methodological implementation, application and evaluation and was grounded in considerable practical application. For the purposes of this paper we make use of the following definitions. Data are a recorded set of observations (which may be quantitative or qualitative). Knowledge is the outcome, independently of the interpreter, of the interpretation of data. Understanding is the outcome, specific to the interpreter, of the interpretation of data or knowledge: the comprehension that the interpreter achieves. So, the interpretation of observation may advance an individual’s understanding. This advance can be articulated and communicated as knowledge.* *These definitions differ from recent use in some of the farming systems literature. However, it is necessary in the present context to view knowledge as being independent of the observer because a knowledge engineering paradigm is employed, in which ‘knowledge bases’ can be created as text on a computer.

Complex argroecosystems. Part 1

345

Typically the knowledge that someone articulates will be expressed in a form of relevance to a particular discourse-in other words as a part of an argument. This does not mean, however, that the knowledge articulated could not be equally useful in some quite different context. So, when somebody-a farmer, scientist or extension worker-has articulated some knowledge, the key requirement is to be able to record it in a way that makes it as widely and advantageously available for use in other contexts. In order to record a piece of knowledge in a form that captures its generic value, it needs to be extracted from the particular argument in which it was articulated. We call this disaggregation. However, the process of disaggregation may cause the loss of important information about the way in which a piece of knowledge could be combined with others in reasoning about a new situation. As a result, contextual information about that knowledge must also be represented. Furthermore, having disaggregated knowledge into sets of statements there is also a requirement to manage these sets so that relationships between statements are maintained and a coherent whole results. Approaches to these three tasks (disaggregation, capturing context and managing sets of statements) are outlined below.

DISAGGREGATION The first fundamental decision that was made in developing an approach to the representation of indigenous ecological knowledge about agroforestry was to seek. to represent the most basic units of knowledge possible, so that these units could then be used in numerous different combinations as required. Attempts to construct knowledge-bases from previously documented studies of indigenous knowledge had found that, because the knowledge reported was generally embedded in a fixed interpretation governed by the rationale of the report, it was not possible to isolate a disaggregated set of statements that could be used to reason about the domain independently of the frame in which the research had been reported (Walker et al., 1991). As it was clear, however, that obtaining robust indigenous knowledge from farmers was going to entail extended dialogue with a dispersed group of people and thus considerable investment of resources in knowledge acquisition, it was imperative that the knowledge acquired could be used flexibly for a range of purposes, not all of which were evident prior to its elicitation. Furthermore, as the form and extent of indigenous knowledge obtainable was unclear at the outset of this research, together with the spectrum of possibilities for using what was obtainable, it was not possible to specify, in advance, either conceptual frameworks within which to represent knowledge nor precisely how the knowledge would subsequently be used.

346

F. L. Sinclair, D. H. Walker

On the basis of these considerations, three distinct approaches to disaggregation were developed and tested. All three approaches involved collating coherent sets of statements (knowledge bases) but took different approaches to the construction of the statements themselves. It is argued that each method is applicable under different circumstances. Approach 1: Extracts of transcripts Initially, we were concerned that any interpretation of the articulated knowledge was likely to cause distortions and, therefore, we abstracted statements directly from interview transcripts. This process was facilitated through development and use of proprietary software known as the transcript tool (Haggith et al., 1992). The approach was then applied in the analysis of 24 interviews about the ecology of complex home garden structures with 10 farmers from two villages in the Kandy district of central Sri Lanka (Southern, 1994). Interviewing was thematic and open-ended and recording was as faithful to the local idiom as possible. Interviews were tape-recorded and then transcribed. Where interviews were in a language other than English, analysis was done using a translation of the original dialogue. The extraction process simply involved highlighting the useful statements within the transcript dialogue (mark up) and constructing a knowledge base by collecting these snippets of natural language together. Information about the people who were interviewed was collected and linked to the knowledge they contributed. The basic criterion used to decide what was worth marking up was whether the marked-up text comprised a concise statement of useful ecological knowledge. Utility was recognised where a statement could conceivably be used in a reasoning process to answer some query about the domain in question. Use of the mark-up approach demonstrated that only a small proportion of interview dialogue consisted of statements that contained useful ecological knowledge. There were other fundamental drawbacks to the method in terms of knowledge representation. In particular, analysis of statements generated in this way revealed that recording the actual form of words articulated was less important than confirming their precise underlying meaning. This is perhaps not surprising given that natural language is imprecise and that a large part of the communication during interviews involved implicit elements defined by the context of the discussion. However, it challenged our perception of sources of distortion in interactions with local communities. The process of transcription was time consuming, thereby restricting the number of interviews that could be conducted with given staff resources. Furthermore, it was clear that abstracting statements from the context in which they were articulated required some form of interpretation to ensure

Complex argroecosystems. Part I

341

that their intended meaning was preserved. The mark-up methodology had much in common with a range of analytical tools for qualitative data analysis such as QSR NUD.IST (see Weitzman and Miles, 1995) and has a role to play in the process of discourse analysis but did not provide a strong basis for creating knowledge-bases. Approach 2: Template sentences

In addition to using the mark-up methodology, an attempt was made to provide some degree of predetermined structure to statements by using a template sentence approach. This was based on the hypothesis that there were a finite number of statement types that could be used to describe relationships in the agroforestry domain. The approach was implemented as a piece of computer software called the template tool (Haggith et al., 1992) which provided a set of template sentences (Table l), from which the user selected an appropriate template for each item of knowledge to be represented and filled in the blanks. The list of templates was not intended to be an exhaustive set of sentence types, but was considered to be a significant proportion of a hypothesised, finite set appropriate for the purpose of evaluating the approach. The template approach was tested using the interview transcripts from the 24 interviews described previously. The template sentence approach proved unacceptable as a practical means of representing knowledge during field testing for two reasons: 1. the template sentence options were overly restrictive. It was usually possible to squeeze a sentence into an available template type but frequently this altered the initial statement such that the sense of the informant’s original assertion was not adequately captured; 2. generating statements using a template sentence approach was time consuming. After a reasonable period for familiarisation with the task, two researchers independently creating statements from a set of trial interviews using the approach took up to ten times as long as creating the same statements in an unrestricted form. This was because a lot of time was required to select an appropriate template structure for capturing each knowledge item. That the set of templates was overly restrictive indicated that the initial list was not sufficient. A more comprehensive set of statement types increased the ability of the user to find an appropriate template type, thereby resulting in a more acceptable representation but the resulting set of template types became so large that finding a desired structural type, from the hierarchy of types took an unacceptably long time.

348

F. L. Sinclair. D. H. Walker TABLE 1

The List of Template Sentence Structures Applied in Evaluation Approach will achieve will occur asGGlresult of -. is a reason for A is a way of A is used for is important for An increase in cauSeS increase in A causes an increase in A A decrease in An increase in causes a decrease in A A decrease in causes a decrease in & A change in causes no change in A change in causes a change in causes an increase in A causes a decrease in A causes no change in A causes a change in A is a result of A causes influences ... is defined bycharacteristics and is a sort of is a part (co=nt) of & is next to A is close to A (activity) is done (place). is (distance) from A -(place). is at (place). _ happens _ happens at the same time as d happens before & . happens at time happens during . (duration). takes (frequency). happens is an attribute of (attribute) can rangefrom -to-. (attribute) has the value (attribute) is measured in-’ (units). (attribute) can take possible values . is than A -is preferred to isgreater on than on A The effect of verb ’such as ‘ AND verb type sentences ‘ shade -’

eat

of the Template Sentence

’and ’

While this experience did not cast doubt on the validity or the usefulness of improving the tractability of elicited knowledge by imposing a syntactic restriction on what could be represented, the serious practical difficulties encountered by the users indicated a limited applicability of this particular

Complex argroecosystems. Part I

349

means of restricting what was represented. It is only likely to be appropriate where knowledge acquisition is undertaken as a broad scoping exercise, particularly where interviewing and analyses are undertaken by a range of staff with limited training and expertise and where documentation of broad rather than detailed relationships is sought. Approach 3: Unitary statements in natural language Having found that neither abstracting parts of natural dialogue, nor forcing what people said into a restricted set of templates, provided a workable approach to acquiring indigenous ecological knowledge, an intermediate solution was proposed. In using this third approach, dialogue was freely interpreted to try to best capture meaning and was then represented using a set of rules relating to how knowledge was to be represented, together with support for ensuring the parsimonious use of terms. The rules could then be flexibly applied in relation to what people articulated. The concept here was to restrict the form of representation to the minimum set of structures and terms required to cope with the knowledge available without either predetermining a set of structures or allowing the largesse of recording freely articulated forms without consideration of their proximity to other knowledge already represented. Knowledge was represented as unitary statements, defined as the smallest useful units of knowledge. These are basically assertions. Utility was defined, as before, in terms of the ability to use a statement in reasoning. The ‘smallest’ criterion simply referred to the fact that a unitary statement can not be broken down into two or more statements. Thus: clover is eaten by sheep and goats is not a unitary statement because it can be broken down into the two unitary statements: clover is eaten by sheep clover is eaten by goats. Two fundamental types of statement were recognised. Statements that described a relationship between entities, such as the two above, were known as binary statements while those that simply described the properties of one thing, such as: nebharo* trees have big leaves were known as attribute statements.

*The Nepali word for Ficus roxburghii.

350

F. L. Sinclair, D. H. Walker

This basic representational approach, together with facilities for managing sets of statements to ensure parsimony that are described below, was applied by six researchers in five countries (India, Nepal, Tanzania, Sri Lanka and Thailand, and) involving hundreds of interviews (Garde, 1992; Thapa, 1994; Kilahama, 1994; Jinadasa, 1995; Hitinayake, 1996; Preechapanya, 1996). It proved successful in providing a basis for innovative research into complex agroforestry systems in a wide range of agroecological and sociocultural circumstances. In summary, three approaches to representing ecological knowledge articulated by farmers were developed and tested. The mark-up approach tried to minimise interpretation, which, paradoxically, did not capture what people had meant very well because much of the meaning in the dialogue with farmers was implicit or contextual. Templates were then used in an attempt to categorise meaning by using a set of predetermined sentence structures. This proved unwieldy and impractical. As a result, the third approach to disaggregation relied on a free interpretation and representation of the articulated knowledge to try to best capture its meaning. This shift from literal record to interpretation is consistent with a changing view of the knowledge acquisition process in other domains, from one of extracting knowledge directly from a source to the construction of a model of the knowledge held by the source (Ford and Adams-Webber, 1992). The process of disaggregation into unitary statements allowed the capture of the sense of knowledge articulated within particular discourses, resulting in a list of statements that might be flexibly combined in addressing new issues. However, this process of disaggregation resulted in a loss of context. Context is clearly important in the use of knowledge. The approaches to capturing context without loosing flexibility that were explored are described in the next section.

CAPTURING

CONTEXT

There are many aspects of the context of an item of knowledge. These include information about the person or people that articulated the knowledge and any conditions or other caveats that they associated with its validity, as well as details of any conceptual structures that they apply in thinking about the domain. These aspects of context are valuable in interpreting and assessing knowledge and should, therefore, be incorporated into the scheme for representation. In the current research, consideration was given to contextual information about the source and conditionality of unitary statements and the hierarchical relationships between terms and concepts used in them. Approaches to

Complex argroecosystems.

capturing some other contextual practical reasons. For example, dence that the source, or others, to be very difficult. Informants statement of confidence, not least

Part I

351

information were explored and rejected for attempts were made to capture the confihad in particular knowledge but this proved generally felt uncomfortable in providing a because it appeared to question their veracity.

Source information The source of each statement was recorded both so that people could be credited with the information that they contributed and to permit retrieval and analysis of knowledge in terms of who had articulated it. Two types of source were involved: people who were interviewed and written documents. Although different information was stored in each case, a summary of name(s) and date were appended to each statement as the link between that statement and the source information pertinent to it. This was sufficient to enable knowledge collected from the same person at different interviews (possibly set in a different context and involving different questions) to be distinguished. For interviews, the minimum information recorded for each source was the name, date, gender and age. Recording subsidiary information allowed subsequent analyses of the nature and distribution within and between communities. The subsidiary information that was recorded depended on the objectives for knowledge acquisition and the context but included, for example, ethnic group, size of livestock holding and location of farm. Conditionality Few assertions hold in all circumstances and so, in general, unitary statements were conditional. For this reason the basic form of a unitary statement had two parts; the basic statement (before the ‘IF’) and then any relevant conditions as in the following example. Crops are prone to lodging IF (there is a strong wind AND crop roots are exposed) OR (there is a strong wind AND crop stems are weak). or, equally: Crops are prone to lodging IF (crops stems are weak OR crop roots are exposed) AND (there is a strong wind). Any number of conditions were permissible and were concatenated ‘AND’ and ‘OR’ as appropriate. The conditionality of statements therefore, to be explicitly considered when knowledge was represented.

with had,

F. L. Sinclair, D. H. Walker

352

Hierarchical relationships In seeking explanatory ecological knowledge, interest was restricted to statements about classes of objects rather than particular instances or individuals. Thus, while it is generally useful to know that: about 1 in 10 gini sapu* seeds germinate within seven weeks from planting IF they have passed through the digestive tract of a bird it is not usually possible to employ a single observation, such as the following statement, in useful reasoning about agroecosystems. a gini sapu seedling in Mr Hitanayake’s garden germinated after seven weeks Exactly what class of items is being referred to, however, depends upon the definition and classification of terms. In practice, many local classifications of objects, for different purposes, may operate simultaneously. For example, Nepalese farmers classified fodder trees in terms of their effects on soil fertility and associated crops (the malilo-rukho classification), in terms of the nutritive value of their fodder (the posilo-kam posilo classification) and in terms of the heating or cooling effect of the fodder (the chiso-obano classification)-when combined these produce a complex classificatory network (Fig. 1). The important point is that some statements farmers make refer to all malilo trees and, therefore, it is important to know which these are. Each line and the two items it connects in Fig. 1 equates to a unitary statement of the form: XisatypeofY where Y is above X. In practice, for a knowledge base on a particular domain, a sensible hierarchical organisation of terms results in a significant reduction in the number of statements required to represent a given amount of knowledge (Table 2). This is because the inheritance properties of the hierarchical organisation allow attributes to be propagated through the hierarchy such that a statement: leaf litter of malilo trees decays quickly applies to all tree species classified as malilo.

*The Sinhalese

name for Michelia champaca.

Complex argroecosystems. Part I

353

(a) A classification of 11 fodder trees according to their rukho-malilo status (‘rukho’ and ‘malilo’ are Nepnli terms describing trees that depress or enhance soil fertility respectively)

I Fodder trees

I

I t’ Badahar

1

I

Rukho

Malilo / rukho

Malilo

I

I

I

I

Dudhilo

Nebharo

Siris

Utis

+I Set siris

IGogun

‘1

1

Chamlayo

I

I

Bans Nigalo

Parayang

Mel

Rat0 siris

(b) A classification of IO fodder trees according to their chiso-obano status (‘chiso’ is a

Nepali term to describe ‘cold and ye” fodder types; ‘obano’ refers to ‘warm and dry’ fodder) Fodder trees I

I Badahar

I

I

Chiso

Chiso / obano

Obano Siris

t+ Set siris

Bans

Nigalo Parayang

* Gogun Chamlayo

4 Mel

Nebharo

Dudhilo

Rato siris

(c) A classification of 11 fodder trees according to their posilo-kamposilo status (‘posilo’ is a Nepali term for highly nutritious fodders, ‘kamposilo’ fodders have a low nutirional value) I

Fodder trees Kamposilo

Posilo

f’ Badahar

I

Dudhilo

I

I Nebharo

Siris

+ Set siris

I

I

I

Bans Nigalo Parayang

1 Mel

I

I

I

Utis Gogun

I

Chamlayo

Rato siris

d) The content of trees (a), (b) and (c) combined into a single hierarchical network

Fodder trees

Badahar

Dudhilo

Nebharo

Siris Utis

I

I

Set siris

Gogun Chamlayo

Bans Nigalo

Mel

Parayang

I

Rato siris

Fig. 1. Three separate hierarchical trees (a,b,c) (from Thapa, 1994). Three separate classifications of the same set of species and the three classifications integrated into a single diagram (d).

F. L. Sinclair, D. H. Walker

354

Representation

TABLE 2 of a Set of Unitary Statements without and with Hierarchical Information

The set of statements without hierarchical information

The set of statements with hierarchical information

Crops are economically useful Legumes are economically useful Root crops are economically useful Cereals are economically useful Chick peas are economically useful Pigeon peas are economically useful Cow peas are economically useful Crops are deliberately cultivated Legumes are deliberately cultivated Cereals are deliberately cultivated Root crops are deliberately cultivated Chick peas are deliberately cultivated Pigeon peas are deliberately cultivated Cowpeas are deliberately cultivated Crops are plants Legumes are plants Cereals are plants Root crops are plants Chick peas are plants Pigeon peas are plants Cowpeas are plants Legumes photosynthesise Chick peas photosynthesise Pigeon peas photosynthesise Cowpeas photosynthesise Legumes have roots Chick peas have roots Pigeon peas have roots Cowpeas have roots Legumes have leaves Chick peas have leaves Pigeon peas have leaves Cowpeas have leaves Legumes transpire Chick peas transpire Pigeon peas transpire Cowpeas transpire Legumes are crops Root crops are crops Cereals are crops Chick peas are legumes Pigeon peas are legumes Cowpeas are legumes

Crops are economically useful Crops are deliberately cultivated Crops are plants Legumes photosynthesise Legumes have roots Legumes have leaves Legumes transpire Legumes are crops Root crops are crops Cereals are crops Chick peas are legumes Pigeon peas are legumes Cowpeas are legumes

Complex argroecosystems. Part 1

MANAGING

355

SETS OF STATEMENTS

The approach to disaggregation into unitary statements and associated representation of contextual information allowed sets of statements to be generated. However, it was also necessary to manage those sets of statements to make them as parsimonious, coherent and consistent as possible without altering their meaning. The most obvious way to achieve this objective is to define and impose clear topic boundaries in deciding what knowledge to represent within a knowledge base and what to exclude. This, however, results in a large number of small knowledge bases which then require management at the level of the knowledge base, if there is to be any compatibility amongst them, therefore, other approaches to managing large sets of statements within a single knowledge base were developed and tested. Identification of keywords

Keywords provide an indication of what statements are about and provide an opportunity for indexing them. In practice, the identification of keywords was required to implement database search functions based on Boolean combinations of ‘ands’ and ‘ors’ across sets of statements. Initially, two types of keyword were identified: words that actually occurred in the statements, known as ‘elements’ of the statement, and ‘topic keywords’ that could be appended to the statement by the person creating the knowledge base to classify statements according to subject categories. For example in the unitary statement: root worm may be transmitted from banana to coconut IF the banana and coconut are grown close together “root worm”, “transmitted” (synonymous with transmission), “banana” and “coconut” were identified as element keywords and the topic keywords ‘pests and diseases’ and ‘plant interactions’ were appended. Having collated several thousand statements resulting from interviews with farmers and extension staff in Sri Lanka, Thailand, Nepal and Tanzania, four types of element keywords were recognised: 1. objects, which were physical items e.g. trees, banana, coconut and root worm 2. processes (or events), that were changes or fluxes, e.g. germination, soil erosion and transmission; 3. attributes, which were properties of objects or processes, e.g. tree height and rate of soil erosion; and 4. values, which were the measurable state of an attribute and could be

356

F. L. Sinclair. D. H. Walker

expressed in qualitative or quantitative terms e.g. tall, 5m, fast and 5 t ha-’ a-‘. Experience led to the topic keyword facility being replaced by a more flexible provision to save combinations of element keywords as search strings which could then be used to select all the statements relevant to a topic. Thus, instead of appending a topic keyword to every statement about the topic (‘pests’ for example), all the words that were thought to appear in statements specifically related to pests (such as: root worm, trunk worm, ants, transmission, spreading) were saved as a search string which could be easily modified in the light of new statements and repeatedly applied to a knowledge base to select the relevant statements within it as the knowledge base was developed. This allowed topics to emerge as the knowledge base matured. Managing hierarchical relationships amongst terms With knowledge represented hierarchically, as described previously, it is possible to explore a knowledge base hierarchically, looking at what knowledge there is at different levels of organisation (for example, with reference to Fig. 1, what is known about trees? ...about fodder trees? ...about malilo fodder trees? ...about malilo fodder trees that are also posilo?). Such exploration helps the person developing the knowledge base to keep a track of what knowledge has already been acquired and the language used to represent it. As with most aspects of representation, judgement is required as to whether knowledge is, in fact, most appropriately represented as a hierarchy and if so, what it is that should be hierarchically arranged. For example, is it the tree species that are being classified by the Nepali farmers as producing fodder of high nutritive value if they are posilo? Alternatively, is it the fodder itself which has an attribute ‘nutritive value’ that may have a range of values including ‘high’, so that the ‘posiloness’ of the fodder of a single tree species may vary? It is, in fact, the latter representation which captures more of the knowledge articulated by the Nepali farmers since when interviewed at different times in the year, they talk about seasonal variation of the nutritive value of fodder from a single species which may be posilo for some parts of the season and kam posilo at others. Managing definition and synonymy Words which did not have an unambiguous and ubiquitously understood meaning in the natural language used for representation were defined and records of synonymous terms kept. Thus the term nebharo is the Nepali word for, and thus a synonym of, Ficus roxburghii, while the word ‘lopping’ may require definition as ‘removal of tree foliage by cutting branches either at

Complex argroecosystems. Part 1

351

their point of attachment with the tree stem or somewhere along their length’ and may be synonymous with ‘pruning’ but not with ‘coppicing’. Recording definitions for terms was found to be of paramount importance in developing an unambiguous record of indigenous knowledge because even apparently common terms were found to have different interpretations. For example, the Nepali word rukh is generally translated into English as ‘tree’ but as used by farmers in Nepal includes species that forest scientists would distinguish as bamboos and shrubs. Useful definitions, that is those that distinguish between what is defined and other things, were not necessarily easy to derive during the process of creating a knowledge base. Definitions were, however, often implicit in a set of unitary statements. For example, some or all of the following statements could be combined to provide a definition for the term ‘jak’: jak is an edible fruit. Artocarpus heterophyllus is the binomial name for jak.

jak jak jak jak

is grown throughout the humid tropics. is an evergreen tree. grows up to 20 metres tall. is a member of the family Moraceae.

Werner and Schoepfle (1987) have classified types of definition statement (Table 3) and their categories are a useful guide in seeking to build up a definition of a term in the present context. In practice, definitions were rarely achieved by a single statement, usually demanding a more complex characterisation. The process of creating definitions can be assisted in a computer-

TABLE 3 Different Ways of Defining Terms Attributive : Term defined in terms of distinctive characteristic Functional : Term defined as a means of achieving something Spatial : Term defined through description of relation to other objects Operational : Term defined with respect to an action which is its characteristic goal or recipient Comparative : Term defined through similarity or contrast with others Class inclusion : Term defined through inclusion in a hierarchical class Synonymy : Term defined as equivalent to another term Antonomy : Term defined as opposite or the negation of another term Quantitative units : Term defined with respect to units of measurement for quantitative variables Qualitative units : Term defined in terms of the set of qualitative values it may take or as a qualitative variable for a particular attribute Adapted from Werner and Schoepfle (1987)

358

F. L. Sinclair. D. H. Walker

Fig. 2. A procedure for defining terms (start and end points are identified in bold).

ised environment and a procedure for defining terms that was used is diagrammatically stylised in Fig. 2. Representing definitions as a set of unitary statements often resulted in a more explicit articulation of the meaning of a term than the use of a description in, unrestricted natural language. Precise definition of all terms in a knowledge base was not possible, not least because ecological knowledge includes terms that are vaguely defined and others for which a consensus of understanding could not be assumed. While precise definitions, unambiguously stating what was and was not included in a term, were sometimes possible, other definitions did no more than characterise the use of a term. This may, however, be appropriate, particularly for umbrella terms such as animal, growth or competition whose utility is to cover a multitude of different items at a conceptual level and consequently may be less useful if too

359

Complex argroeco.ystems. Part I

PKXCSS

Attribute

Glossaries (Process and attributes)

Definition: Splash erosion effect caused by drops of water falling from the leaves of tree I

t

on croplernce risesQUi?i Nab Khanal. 23

Statement

I

1292)

Source information

Object kryword hierarchy

Wall between two terraces of bari land

Ficus auriculam

Fig. 3. The conceptual

structure

of a knowledge

base containing

a single statement.

rigidly defined. Spedding (1988) observes that such terms are made more precise by the use of adjectives (e.g. unicellular animal), this corresponds to a general rule for hierarchically structured sets of terms in the present context that, as one moves down the hierarchy, more precise definition is increasingly useful, and thus should be sought.

APPL.YING THE NATURAL

LANGUAGE

APPROACH

The results of the research outlined above were synthesised and implemented as a method and software program named TEAK (Haggith, 1992). This approach to knowledge representation involved compiling a database of statements written in ‘natural language’ with subsidiary information about the meaning of key terms and the relationships amongst them. The conceptual structure of representation of knowledge within TEAK is summarised in Fig. 3. TEAK provided database functionality, with support for the user in entering and viewing information as natural language statements, specifying keywords and specifying hierarchical relationships amongst terms. As far as the user was (concerned individual unitary statements had the form: statement IF conditions (source) All keywords in the statement (objects, processes, attributes and values) were stored in glossaries. Where necessary the keywords could have synonyms and definitions associated with them. Objects could be hierarchically arranged

360

F. L. Sinclair, D. H. Walker

and search strings of keywords that would select sets of statements on particular topics could be stored. Statements could be edited, viewed either individually (with full details) or in selected sets of one line summaries and printed. Facilities were available to browse through knowledge bases hierarchically, and to edit hierarchical structures or impose any number of new hierarchical arrangements. The approach was used for the agroforestry research projects referred to earlier in Sri Lanka, Nepal, Thailand, Tanzania and India. This work resulted in eight knowledge bases ranging from 500 to 2000 statements in size. This experience demonstrated that the approach provided a viable means of undertaking a knowledge acquisition exercise. Users were not overly constrained in what they were able to record, allowing rapid progress, but were encouraged to be parsimonious in the use of terms through the identification and definition of keywords and their arrangement in hierarchies where appropriate. This allowed the knowledge base to be continuously refined as knowledge acquisition progressed and the researcher became more familiar with the domain and the way in which farmers explained it. The search facilities allowed selection of groups of related statements that were helpful when reviewing knowledge on particular topics, thereby identifying where representation could be improved or further knowledge elicitation was desirable to obtain comprehensive coverage of the domain. As such, the approach provided a core set of tools for the research projects in question. However, analysis of the resulting knowledge bases revealed a number common constraints, as follows. ??

??

??

There was significant repetition. This involved (a) statements with different wording but essentially the same meaning (for example, some statements might refer to soil nutrient content and others to soil fertility) and (b) sets of statements which could be reduced by imposition of hierarchical structure. A significant proportion of statements did not contain useful ecological knowledge. Some statements described practices with no explanatory ecological content, others were restricted in their usefulness by the use of words such as ‘can’ or ‘may’ indicating a requirement for further knowledge elicitation of the conditions in which the statement was valid. For example: banana can be grown with clove. A significant proportion of statements did not have a single unambiguous meaning, but could be interpreted in a number of different ways. This often occurred as a result of the use of unexplained value judgements (poor, good, greater, etc.) or the use of undefined or inadequately defined terms. Some attributes and values were used inconsistently. For

Complex argroecosystems. Part 1

361

example, in some statements soil fertility might take values ‘fertile’ and ‘infertile’ in others ‘low’, ‘medium’ and ‘high’. ?? Viewing long lists of statements had limitations as a means of assessing subject coverage and there were clearly areas where further knowledge elicitation was required to obtain comprehensive coverage of the domain. These problems are as much to do with the application of the methods we have described as the inherent validity of the methods themselves and suggest that users need to be forced to be more precise in their use of representational structures. Thus, in practice, these problems were significant impediments to the development of useful abstractions of knowledge and, thereby, ‘models’ of the behaviour of the systems in question. Approaches to resolving some of these issues using a formal scheme for representation are discussed in a companion paper (Walker and Sinclair, 1998). In summary, application of the methods implemented in the TEAK software system have demonstrated that simple and intuitive approaches to knowledge representation can provide a basis for generating coherent sets of statements in natural language that are precise enough in their meaning to be useful in developing an abstraction of knowledge held by a defined community about a defined subject. We believe that this approach could make a useful contribution to the transparency and repeatability of rural appraisal methods.

ACKNOWLEDGEMENTS The research reported in this paper was funded by the UK Overseas Development Administration (ODA), Forestry and Agroforestry Strategic Research Program. However, the ODA accept no responsibility for the information provided or views expressed. It involved collaboration with Robert Muetzelfeldt of the Institute of Ecology and Resource Management of the University of Edinburgh; Dave Robertson, Mandy Haggith and Gill Kendon of the Department of Artificial Intelligence of the University of Edinburgh; Barlaram Thapa of Pakhribas Agricultural Centre, Nepal (funded by ODA bilateral funding); Nishantha Jinadasa of the University of Sri Jayewardenepura, Sri Lanka (funded by ODA bilateral funding); Gamini Hitianyake of the University of Peradeniya, Sri Lanka (funded by ODA bilateral funding); Felician Kilahama of the Department of Forestry and Beekeeping, Tanzania (funded by the Ford Foundation); Pornchai Preechapanya of the Royal Forest Department and University of Chiang Mai, Thailand (funded by the Ford Foundation, Winrock International, and the Rockerfeller Foundation); and Alison Southern of the University of Wales,

362

F. L. Sinclair, D. H. Walker

Bangor (funded by the Humphrey Robert Jones Scholarship).

Griffiths

Travel

Scholarship

and John

REFERENCES Anderson, L. S. and Sinclair, F. L. (1993) Ecological interactions in agroforestry systems. Agroforestry Abstracts 6(2), 57-9 1. Budelman, A. (ed.) (1996) Agricultural R and D at the Crossroads. Merging Systems Research and Social Actor Approaches. Royal Tropical Institute, Amsterdam, The Netherlands, 247 pp. Chambers, R. and Ghildyal, B. (1985) Agricultural research for resource poor farmers-the farmer first and last model. Agricultural Administration 20, l-30. Ford, K. M. and Adams-Webber, J. R. (1992) Knowledge acquisition and constructivist epistemology. In The Psychology of Expertise: Cognitive Research and Empirical AZ, ed. R. R. Hoffman, pp. 121-136. Springer-Verlag, New York. Garde, L. J. (1992) A study of indigenous ecological knowledge about a livestock/ forest production system in the Uttar Pradesh Hills of India. MSc. thesis, University of Wales, Bangor. Haggith, M. (1992) TEAK, Tools for Eliciting Agroforestry Knowledge, A Guide for Users. Version 2. School of Agriculture and Forest Sciences, University of Wales, Bangor. Haggith, M., Robertson, D., Walker, D. H., Sinclair, F. L. and Muetzelfedlt, R. I. (1992) TEAK: tools for eliciting agroforestry knowledge. Proceedings of the British Computer Society Symposium on IT-enables change in developing countries. Hitinayake, H. M. G. S. B. (1996) Farmers’ knowledge and the development of complex agroforestry practices in Sri Lanka. PhD thesis, University of Wales, Bangor. Jinadasa, P. N. M. (1995) Indigenous ecological knowledge about mother plant selection and plant siting in Kandy home gardens of Sri Lanka. M.Phil. thesis, University of Wales, Bangor. Kilahama, F. B. (1994) Trees and indigenous ecological knowledge about agroforestry practices in the rangelands of Shinyanga region, Tanzania. Ph.D. thesis, University of Wales, Bangor. Muetzefeldt, R. I. and Sinclair, F. L. (1993) Ecological modelling of agroforestry systems. Agroforestry Abstracts 6(4), 207-247. Plant, R. E. (1997) A methodology for qualitative modelling of crop production systems. Agricultural Systems 53(4), 325348. Preechapanya, P. (1996) Indigenous ecological knowledge about the sustainability of tea gardens in the hill evergreen forest of Northern Thailand. Ph.D. thesis, University of Wales, Bangor. Richards, P. (1985) Indigenous Agricultural Revolution. Hutchinson and Co Ltd, London. Richards, P. (1994) Local knowledge formation and validation: the case of rice production in Sierra Leone. In Beyond Farmer First, eds. I. Scoones and J. Thompson, pp. 165-170. Intermediate Technology Publications, London.

Complex argroecosystems. Part I

363

Robertson, D. H., Bundy, R., Muetzelfeldt, R. I., Haggith, M. and Uschold, M. (199 1) Eco-logic, Logic-based Approaches to Ecological Modelling. MIT Press, Cambridge, MA. Rocheleau, D. E. (1987) The user perspective and the agroforestry research and action agenda. In Agroforestry: Realities, Possibilities and Potentials, ed. H. L. Gholz, pp. 59-87. Martinus Nijhoff, Dordrecht. Southern, A. J. (1994) Acquisition of indigenous ecological knowledge about forest gardens in Kandy district, Sri Lanka. M.Phil. thesis, University of Wales, Bangor. Spedding C. R. W. (1988) An Introduction to Agricultural Systems. Applied Science Publishers, London, U.K. Starfield, A. M. and Beloch, A. L. (1983) Expert systems: an approach to problems in ecological management that are difficult to quantify. Journal of Environmental Management

16, 261-268.

Thapa, B. (1994) Farmers’ ecological knowledge about the management and use of farmland tree fodder resources in the mid-hills of eastern Nepal. Ph.D. thesis, University of Wales, Bangor. Thapa, B., Sinclair, F. L. and Walker, D. H. (1995) Incorporation of indigenous knowledge and perspectives in agroforestry development. Part 2: case study on the impact of explicit representation of farmers’ knowledge. Agroforestr)! Systems 30, 249-26 I.

Walker, D. H. and Sinclair, F. L. (1998) Acquiring qualitative knowledge about complex agroecosystems. Part 2. Formal representation. Agricultural S_vstems 56, 365-386.

Walker, D. H., Sinclair, F. L., Muetzelfeldt, R. I. (1991). Formal representation and use of indigenous ecological knowledge about agroforestry: pilot phase report. School of Agricultural and Forest Sciences, University of Wales, Bangor. Walker, D. H., Sinclair, F. L., Kendon, G., Robertson, D., Muetzelfeldt, R. I., Haggith, M. and Turner, G. S. (1994) Agroforestry knowledge toolkit: methodological guidelines, computer software and manual for AKTl and AKT2, supporting the use of a knowledge-based systems approach in agroforestry research and extension. School of Agricultural and Forest Sciences, University of Wales, Bangor. Walker, D. H., Sinclair, F. L, Joshi, L. and Ambrose, B. (1997) Prospects for the use of corporate knowledge bases in the generation, management and communication of knowledge at a frontline agricultural research centre. Agricultural Systems 54(3), 291-312.

Warren, D. M., Brokensha, D. and Slikerver, L. J. (eds) (1994) Indigenous Knowledge Systems.. The Cultural Dimension of Development. IT Publications, London. Weitzman, E., and Miles, M. (1995) A Software Source Book: Computer Programs for Qualitative Data Analysis. Sage Publications, Thousand Oaks, CA. Werner, 0. and Schoepfle, G. M. (1987) Systematic Fieldwork Volume 1: Foundations qf Ethnograph.y and Interviewing. Sage Publications, Thousand Oaks, CA.