NOTE: This is the full text of the original article, as writen in 1989. Some of the names and concepts included may have changed over the years (e.g., feature is now called structure, the definition of primary identification criterion has been changed, etc.).
BUILDING A KNOWLEDGE BASE FOR PLANT-PARASITIC NEMATODES:
DESCRIPTION AND SPECIFICATION OF METADATA
Jim Diederich*, Renaud Fortuner**, and Jack Milton*
(*) University of California, Department of Mathematics, Davis, CA 95616 and (**) California Department of Food and Agriculture, 1220 N Street, Sacramento, CA 94271-0001, USA
INTRODUCTION
We are all familiar with data. For example, if Jill Rogers is a particular employee and her hourly wage is $15.00/hr., then this fact (her hourly wage) is a piece of data. However, metadata is data about data. For example, the fact that the wage is expressed in dollars rather than francs is metadata. Metadata is typically the kind of data used by a database designer to design and implement a database (Tsichritzis & Lochovsky 1982). There are other kinds of metadata, such as a requirement that every employee have an hourly wage greater than the minimum wage, which is data about the wage data of all employees. As an example for nematodes, the position of the vulva for a particular specimen is data. That the feature vulva position is given as a percentage of body length is metadata. So is the fact that the vulva is often visible under a dissecting microscope and that its position is very useful in identifying promorphs and nests (Fortuner, 1989).
It appears that a significant portion of the inference capabilities of Nemisys (Diederich & Milton, 1989) will rely on the metadata because some important aspects of a nematologist's expertise can be expressed in that form. Other aspects such as the strategies used in identification cannot easily be captured as metadata and instead will be captured in rules and heuristics (rules of thumb). In the subsequent discussion we present an initial attempt at specifying metadata for Nemisys. While it is likely that other kinds of metadata will be necessary as the project unfolds, it seems realistic that all of the metadata discussed below will prove useful in the final project. The process of developing this metadata, including various issues and their resolution, is presented in Diederich and Milton (1988).
Definitions
Some of the terms needed to discuss metadata are:
Character - a particular feature property or other way in which organisms may differ (Pankhurst, 1975).
Character state - one of two or more possible alternative expressions of a qualitative character.
Character value - the value taken by a quantitative character (real number or integer).
Feature - short for morphological feature.
Nematode - short for plant-parasitic nematode.
Morphological feature - a structural element of a nematode, for example, female body, cuticle annuli, phasmids, etc.
Nest - a group of species that share a unique set of primary identification criteria (Fortuner, 1989).
Primary identification criterion - a character easy to see, not ambiguous, non variable within a nest, and highly useful for its identification (Fortuner, 1989, and below).
Property - a characteristic of a feature; for example, the feature 'female body' has several characteristics, its length, its habitus, etc. Some properties are characters, as defined above, when they differentiate organisms. For example the properties 'body, length' and 'body, habitus' are characters, but the property 'body, color' is not a character for plant-parasitic nematodes.
Promorph - a form that can be recognized before detailed study of its morphology (Fortuner, 1989).
Instance - one from a set of elements having a common description. For example, N-scutellonema is an instance of the class Nest.
General metadata concerns
Nemisys will include a large database of nests and promorphs. A good deal of metadata will be derived from this database. For example the range can be derived by locating in the database the lowest and highest average values for a particular measurement or by finding all possible states of a qualitative character. This is called derived metadata.
Other kinds of metadata rely on the opinions of experts. For example,
the conspicuity of a feature will have to be supplied by expert nematologists.
This is called specified metadata.
METADATA OF THE FIRST KIND: INSTANCE INDEPENDENT METADATA
This kind of metadata specifies properties of the data that are independent of specific instances of promorphs and nests. For example, the character 'lateral field lines, number' is an ordinal data type. This is true for all nests. All metadata of the first kind will be supplied by the project's domain expert.
Metadata name: data type
Definition. Specifies how the data is represented.
Significance. Specified only for atomic morphological characters (those which cannot be divided into simpler characters at the level in question). The same character may have different data type at the nest and promorph levels. For example, the character 'body shape' is a nominal value, i.e., thin, fat, etc., at the promorph level, but at the nest level it is represented as the ration of length to diameter, a measurement.
Metadata values.
nominal - represented by a word or phrase
ordinal - represented by an integer
measurement - represented by a real number
boolean - represented by present/absent, yes/no, +/-
Examples. The data type for 'body, shape' is nominal at the promorph level (body shape Ð thin, normal, fat, obese). The data type for 'lateral field lines, number is ordinal at the nest level (number of lateral field lines = 2, 3, 4, 5, 6, more than 6). The data type for 'body, length' is measurement at the nest level (body length = 754.5 µm). The data type for 'deirids, presence' is boolean at the nest level (deirids present or absent).
Metadata name : data unit
Definition. Unit of measurement.
Significance. Given only for measurements and only at nest level
Metadata values.
millimeter
micrometer
some measurements are given as a percentage
ratios have no unit
Examples. Data unit for body length is millimeter; data unit for stylet length is micrometer; data unit for vulva position is percentage; ratio a, ratio b, etc., have no unit.
Metadata name: derived
Definition. Data that is expressed in terms of other data values typically in the form of an arithmetic expression.
Significance. Derived data arises usually from measurements in
the course of an identification. For example, ratios are derived from measurements
as seen in the example below. The corresponding database values for each
nest will be obtained from the experts and the literature.
Example. Female body ratio a = female body length / female body
diameter
METADATA OF THE SECOND KIND: CHARACTER ORIENTED METADATA
This kind of metadata is specific to individual promorphs and nests in the database. In addition, it consists of metadata about characters and not about character values or character states. For example, it is data about the character 'body, shape' and not about the character states thin, fat, etc. The metadata in this category will have to be specified by the domain experts or taken from the default values.
Metadata name: usefulness
Definition. This indicates how useful the character is in identifying a particular nest or promorph.
Significance. A highly useful character is one which would be among the first dozen or so characters looked at by the expert in identifying at this level. For example, vulva position is highly useful at all levels while size of medium bulb valve may be of low usefulness except for some nests. Generally speaking Nemisys will consider the degree of usefulness in determining its strategy in the identification. This metadata will also be important in characterizing primary and secondary characters.
Metadata values.
high - almost always used for identification at this level
medium - used reasonably often at this level
low - sometimes or rarely used
zero - never or rarely used
Examples. Usefulness of vulva ratio v is high (if no specific promorph or nest is mentioned, this serves as a default for this character for all promorphs and all nests). Usefulness of position of excretory pore is low (this also is a default). Usefulness of position of excretory pore in N-sychnotylenchus is high (this overrides the default in the previous example for this nest; only default overrides have to be specified).
Metadata name: conspicuity
Definition. The degree to which a character is obvious to the eye given the proper equipment for observations at the promorph or nest level. For a promorph a dissecting microscope with 40x magnification is considered proper equipment. For a nest a research compound microscope with 1000x magnification using oil immersion but no special technique. For SEM face view, obviously a scanning electron microscope is required.
Significance. Conspicuity will be useful for defining primary and secondary characters. It will also serve as an indicator of the likelihood of an incorrect response from some users.
It is important to distinguish between conspicuity of a feature, and conspicuity of its various properties. The feature 'body annuli' has either high or low conspicuity depending on the nests considered. The character 'body annuli, visibility' (from invisible to conspicuous) has high conspicuity even in those nests with faint or invisible annuli, because the fact (character) that annuli are invisible is obvious to the eye.
Conspicuity must be defined for each character, and not just for each feature, because a feature may have some characters that are conspicuous and other characters that are not. For example, the feature 'outline of labial area' has the character 'shape' highly conspicuous, but the character 'shape of cross section' has low conspicuity. However, generally speaking, a feature and its properties will have the same conspicuity.
It will be possible to define default values attached to each feature that will be deemed to apply to each of its associated characters in all promorphs or nests. The domain experts will need to override this default value only in the few cases where it does not apply to a particular character, or to a particular nest.
Finally, conspicuity may differ from freshly killed specimens to old fixed specimens. For example, median bulb valve and dorsal gland opening are highly conspicuous in freshly killed specimens, but they often have low to very low conspicuity in fixed material. It may be necessary to define two values for the conspicuity of some characters at the promorph level, the first one for living specimens, the other for dead specimens, and at the nest level, the first one for temporary mount in water of freshly killed specimens, and the other for specimens fixed and mounted on permanent slides.
Metadata values.
high - can easily be seen in most specimens
medium -sometimes obscured or blurred from one specimen to the next
depending on specimen condition
low - hard to see in most specimens but present
zero - not visible
Examples. Conspicuity of the feature 'phasmid' is medium in both temporary and permanent mounts. If no specific promorph or nest is mentioned, this serves as a default for this feature and all its associated characters for all promorphs and all nests. The conspicuity of the same feature is high in N-scutellonema. This overrides the default value in the specified nest only. The character 'phasmid, number of annuli between phasmid and anus' has low conspicuity in N-scutellonema (new default override for a particular character of the feature 'phasmids' in the nest N scutellonema).
Metadata name: ambiguity
Definition. The character, and thus its value or state, can easily be mistaken or be uncertain.
Significance. Generally characters are unambiguous to the degree they are conspicuous. Thus, the default for ambiguity will be the inverse of conspicuity. Something which is highly conspicuous will have low ambiguity unless specified otherwise. Consequently, much of the metadata for ambiguity will be derived from the metadata for conspicuity, but there may be some exceptions. For example the lateral field lines are highly conspicuous in N-trilineellus, but there is some ambiguity in the actual number of lines, which can be seen as either three or four (Fortuner & Luc, 1987).
Ambiguity is more related to the various characters of a feature while conspicuity was more attached to the feature itself. No matter what the conspicuity of a feature is, the ambiguity of a measurement character will be less than that of a corresponding nominal character. Body length measured in micrometers has very low ambiguity, but body size expressed as a nominal character (small, medium, long) depends very much on the subjective appreciation of the observer, and it has medium ambiguity.
It is assumed that a highly ambiguous character is one which is either hard to see or it is easy to see but there is a high probability that the value or the state selected for the character will be incorrect. For instance, the user may not be sure where the measurement of the spicule length should be taken, i.e., dorsal limb, ventral limb, or spicule axis.
For the same reasons as for conspicuity, it will be necessary to define two values for ambiguity for each feature and each character at promorph and nest level.
Source. Inverse of conspicuity unless otherwise specified by domain experts
Definition. This metadata determines which characters should be considered before other characters during an identification.
Significance. Primary characters will generally be used first in an identification and are used in the definition of nests. The rest of the data with usefulness not high is neither primary nor secondary. It can be used for descriptive purposes or for identification at a lower level (species). Note, the fact that primary characters are defined with variability = zero rules out all measurements as primary criteria unless overridden by the domain experts. For identification, size is taken into account primarily as a nominal criterion (small to large) rather than as an actual measurement. In exceptional circumstances, the experts can override the automatic derivation from the metadata.
Source. Derived from the metadata unless overridden by the experts
Metadata values.
primary -defined by metadata where usefulness is high, conspicuity is high, and variability is zero
secondary -defined by metadata where usefulness is high, conspicuity
is not high, or variability is not high
Examples.
Primary and secondary characters for N-pratylenchus:
Primary: outline of labial area, shape
whole stylet, aspect
gland overlap, position
vulva, position
Secondary: median bulb, shape
tail end, shape
Note: the character 'median bulb, shape' has high conspicuity, and zero
variability in N-pratylenchus but is not considered primary because of
the experts' judgment (usefulness not high).
METADATA OF THE THIRD KIND: VALUE ORIENTED METADATA
This kind of metadata is specific to character values or states in individual promorphs and nests. Much of this metadata will be derived from the database with detailed descriptions of all promorphs and nests, once the database is created by the experts.
Metadata name. range
Definition. This metadata specifies all the values taken by the data for a particular promorph or nest.
Significance. It indicates what values to expect, which allows the system to check the input from the identifier. For example, nominal data for the character 'stylet, thickness' would come from the list {thin, normal, thick}, while 'stylet, length' may be between two possible values.
Source. Derived from the database created by the experts. The experts (or the project domain expert with technical assistance) will have to check the values of each character in the descriptions of all the species in a nest with 20 or fewer species, or a representative sample for larger nests.
Metadata values.
qualitative characters: a list of all the character states observed
in all the species in the nest
quantitative characters (integers or real numbers): highest and lowest
average values for measurements of all species in the nest
Examples. Range of 'body, length' is 400 to 1200 µm in
the nest N-helicotylenchus. Range of 'labial disc, shape' is {oval, square,
hexagonal} in N-scutellonema.
Metadata name: typical value
Definition. A typical value (or state) of a character is the one which is most representative for the members of a particular promorph or nest.
Significance. The typical value or state of a character in a nest is based on the experts' expectation of what the value or state for this character would be in a typical identification if the system narrowed the candidates to this particular promorph or nest. This value represents expectations external to the database, i.e., what one would expect to find in the field or lab. For example, if an expert knew that the specimen came from the nest N-scutellonema, then he/she would expect the body length to be about 700 µm.
Source. Domain experts
Examples. Typical value for 'body, length' is 700 µm for
N-scutellonema
Metadata name: frequency
Definition. Gives the frequency of occurrence of each state taken
by a character in a promorph or nest. In case of measurements, ratios,
and ordinal values, it measures the frequency of occurrence of a small
number of states (about half a dozen) as defined by the experts.
Significance. The frequency assists in focusing on the correct
promorph or nest and in evaluating possible candidates for identification.
Source. Derived from the database or calculated by the project expert with assistance
Metadata values. A percentage for each state
Examples. Frequencies of 'body, length' (6 classes) in N-scutellonema:
class (lengths) 1 (L = less than 699 µm frequency: 41 %
2 (L = 700-799 µm) 27 %
3 (L = 800-899 µm) 9 %
4 (L = 900-999 µm) 13.5%
5 (L = 1000-1099 µm) 4.5%
6 (L = more than 1100 µm) 4.5%
Frequencies of 'labial disc, shape' in N-scutellonema (estimated from SEM face views in Figure 1 of Germani et al., 1985):
class 1 square frequency: 13%
2 oval 31%
3 hexagonal 25%
4 round 31%
Metadata name: variability
Definition. This indicates the degree to which the value of a character varies within the same nest.
Significance. For nominal values the variability is estimated from a coefficient equal to:
v = (y -1)/(x - 1) * 100
where y is the number of states observed in the nest and x is the number of states of the character found in all the nests.
For measurements, the coefficient of variability Cv is used, as calculated by the usual formula:
Cv = (standard deviation/mean) * 100
The value of the coefficient will be computed based on all species in the nest up to 20, but only on a representative sample if there are over 20 species in a nest. However, variability in a nest with only one or a few species will be unrealistically low. In such cases, the variability computed by the project domain expert from the data may be overridden by the collaborating experts. For instance, in N-antarctylus (only one species known) the variability of the body length computed from the data is 0% but this would be overridden after comparison of the same character in the related nest N-helicotylenchus where Cv = 20% In the same nest however, variability of the tail shape computed from the data is 0% (only conoid-pointed tails are present) and this is accepted by the expert because tail shape is a primary criterion for this nest.
In general the variability of measurements will not be zero. Based on the definition of primary character, measurements will not be primary characters.
Source. The project domain expert with technical assistance
Metadata values.
very high coefficient = 100%
high 75-99%
medium 25-74%
low 5-24%
very low 1- 4%
zero 0%
Examples. Variability of 'body, length' is low (c.v. = 18.9%)
in N-scutellonema. Variability of 'gland overlap, type' is zero in N-scutellonema,
because only one state is present out of four possible states, which gives
v = (1-1)/(4-1) * 100 = 0. Variability of 'stylet knobs, shape' is medium
in N-scutellonema, because three states are observed (slopping, rounded,
flat indented) out of six possible states; v = (3-1)/(6-1) * 100 = 40%.
THE INFERENCE ENGINE OF NEMISYS
Metadata Entry
While it might appear that there is a great deal of specified metadata much of it can be handled through default values. For example, if it is generally the case that vulva position has high conspicuity in promorphs, then it will only be necessary to specify otherwise in those promorphs where conspicuity is not high. Some explicit examples of defaults are presented with the metadata concept usefulness above. Also, some metadata such as data type or data unit need only be specified twice for each feature - once for promorph and once for nests. The default values will be specified by the project domain expert (R. Fortuner) after consultation with the collaborating experts.
Eleven groups of experts dealing respectively with tylenchids, anguinids, dolicho/belonolaimids, tylenchorhynchids, hoplolaimids, pratylenchids, heteroderids, criconematids, aphelenchids, longidorids, and triochodorids, will propose default values for the metadata usefulness, conspicuity, and ambiguity in each of these groups of nematodes.
Later, experts will study one promorph or one nest at a time. They will verify the relevance of the default values and eventually modify them for this particular nest or promorph. They also will propose values for the metadata priority, range, typical value, frequency, and variability. These values will be given by each expert depending on his/her knowledge of the group and subjective appreciation.
Parallel to this effort, the description of up to twenty species will be gathered for each nest. The metadata range, frequency, and variability will be calculated from the species data. These more objective values will be compared to the values proposed by the experts to obtain a measure of the fuzziness of expert knowledge.
Using Metadata in Nemisys
To give an indication of how metadata will be used consider the following. Each nest is defined by a set of primary identification criteria. Consequently, when an expert suspects that the specimen being identified is from a particular nest and asks for confirmation, the primary characters will be displayed prior to the display of secondary characters and of other characters that are neither primary nor secondary. A primary character would be one whose metadata specified its usefulness as high, its conspicuity as high, and its value variability as zero. So a rule to distinguish primary characters might read something like:
if the level is nest and
if the character usefulness is high and
if the character conspicuity is high
if the character value variability is zero,
then the character is primary
Likewise, secondary characters could be defined as usefulness is high, but conspicuity is not high, or ambiguity is high, or value variability is not zero.
Another important consideration for accurate identification is the reliability of the observations entered by the user. This reliability depends on the expertise level of the user, on the number of specimens studied, and on several metadata concepts. For example data entry for a particular character could be:
If conspicuity is low and
if ambiguity is high and
if variability is high and
if number of specimens observed is low and
if observer is beginner,
then reliability is very low (almost useless)
Reliability would increase when the number of specimens observed increase. It would be higher for characters with higher conspicuity or lower ambiguity and variability. The system would search for such characters and ask the user to confirm the observation with low reliability with other, more reliable characters.
Rules and Heuristics
Other factors of the identification process cannot easily be captured as metadata, and they will rely on rules and rules of thumb (heuristics). For example, reliability is improved when several observations agree. The default values for the character 'median bulb valve, size' are: conspicuity is low, ambiguity is high, and variability is medium. A beginner entering size = large with number of specimen observed = one, would have an observation with very low reliability. However, if 'stylet, size' is given as elongate, this would increase dramatically the reliability of valve size because the two characters states are related.
Rules can be used to define relatedness among characters:
number of genital branches is related to position of vulva (with exceptions),
and they can also be used to define relations between character values:
if 'stylet, size' = elongate
then 'procorpus, shape' = swelled
'junction procorpus/median bulb, type' = fused
'median bulb valve, size' = large/very large
Heuristics are rules of thumb that will be proposed by the domain experts to direct and verify the identification. An example of an heuristic would be:
If specimen resembles P-melo
If host = citrus
then check whether it is N-tylenchulus
REFERENCES
Diederich, J. & Milton, J., 1988. Creating domain specific metadata for data and knowledge base. Submitted. [Published as: Diederich, J. & Milton, J., 1991. Creating domain specific metadata for scientific data and knowledge bases. IEEE Transactions on Knowledge and Data Engineering, Vol.3 No. 4, pp. 421-434.
Diederich, J. & Milton, J., 1989. NEMISYS, An expert system for nematode identification. In: Fortuner, R., 1989 (Editor). Nematode identification and expert system technology. New York, Plenum Publishing Corp., pp. 45-63.
Fortuner, R., 1989. A new description of the process of identification of plant parasitic nematode genera. In: Fortuner, R., 1989 (Editor). Nematode identification and expert-system technology. New York, Plenum Publishing Corp., pp. 35-44.
Fortuner, R. & Luc, M., 1987. A reappraisal of Tylenchina (Nemata). 6. The family Belonolaimidae Whitehead, 1960. Revue Nématol. 10: 183-202.
Germani, G., Baldwin, J.G., Bell, A.H. & Wu, X.Y., 1985. Revision of the genus Scutellonema Andrassy, 1958 (Nematoda: Tylenchida). Revue Nématol. 8: 289-320.
Pankhurst, R.J., 1975. Biological Identification with Computers. London, The Systematic Association, Special Volume n°7, Academic Press, x + 333 p.
Tsichritzis, D.C. & Lochovsky, F.H., 1982. Data models. Prentice-Hall, Englewood Cliffs, New Jersey.
DISCUSSION
Tarjan: Suddenly I feel like the boy in the old Dutch story sticking his finger in the hole in the dike with all this stuff coming at me. How does variability enter into all of this?
Milton: Variability will be used, first of all, to help distinguish primary characters for a given nest, i.e., variability will have to be zero. It will also play a role in judging the certainty of a particular value, but exactly how that will be done depends on input from the experts. For example, it may be unwise to ask a non-expert to enter the value of a character with high variability in a particular nest, or to consider characters with high variability across all nests. Certainly, variability along with other metadata will be available to the user of the system in the same way the data will be available.
Fortuner: As far as I can see, there are two ways to include
metadata about the variability. One way would be to ask the experts to
give their opinion on the variability of each character in a particular
nest. An expert on Ditylenchus, for example, would know that stylet
length is not very variable, whereas body length can vary by a factor two
in some species. The other way is to include in the database a description
of all the species in a nest. The system would then compute the coefficient
of variability from the measurements or make a comparison of frequencies
of each character state for qualitative characters. In fact, it would be
very interesting to compare the subjective appreciation of variability
by the experts to the objective calculation of the variability made from
hard data.
Valdivia: I have three questions. First, how do you classify the different
set of rules in Nemisys? Second, who controls meta rules and in what order
are they applied? And third, how do you plan to handle uncertainly?
Milton: I am not exactly sure what you mean by classifying the different sets of rules, but perhaps I can indicate some relevant facts. We have not really decided what inference mechanism to use yet. Much of the expertise can be captured in the metadata, and we will look to rules to handle such things as exceptions, special cases, and context dependent actions. Initial planning of the architecture of the system involves doing something like setting up a scoring function, as in the Internist system, which does not involve rules at all. This scoring function could be used as a basic mechanism to discriminate between the various taxa, and to do such things as to identify the leading candidates at a given time during an identification. Then as for rules, an example of a context dependent use of a rule would be to assist with the display of a given nematode. Rarely would we wish to display all characters at once, and what we display depends on what we are doing at the time. Since all contexts in which rules might help are not yet clear, and an inference mechanism is not set, the order of application of the rules is also not clear. Actually the rule base may be fairly small, so that order will not be critical. There is at least one commercially available shell that runs on top of Smalltalk, Humble, that supports both forward and backward chaining, and we may make use of it. If by "control" you mean "specification", the experts will have to tell us what the important rules are, but also we will build in ways for the identifier to override facts and/or rules. As for uncertaincy we will probably take a Bayesian approach, but we have also talked about the concept of fuzzy sets. The initial work of the system has been primarily in terms of specifying the metadata, characterizing the identification process, developing tools for the identification workbench, and setting up a reasonable first cut at a computer interface. We cannot really make some basic architectural decisions until we have a reasonable sense of what the expert does in the identification process and the context is clear, so we will be looking to you for guidance on these matters.
Valdivia: How will you handle nominal characters that can be defined in terms of size, such as very large, large, small, very small.
Fortuner: I understand the question is about the fuzziness of the characters, e.g., what is a large stylet compared to a small stylet. Again, there are two ways to decide. One is to use subjective appreciations from the expert. You can just ask each expert what he/she thinks qualifies as a small, medium, and large stylet. The second way would use hard data to compare such a nominal character to the actual measurement. You would find that experts in a particular group of nematodes qualify as large stylets that are 60 µm or longer, while the threshold value would be for example 50 µm for taxonomists familiar with another group of nematodes. With enough input from various experts, you can construct a curve with the cumulative frequencies for the threshold values for "large"tylench stylets. When a user defines a stylet as large, the system will know that its length at least equal to the smallest threshold proposed by the experts , and you can calculate the likelihood that its value is equal to the various values attached by the various experts to the concept of a "large" stylet.