Words provide a natural way of perceiving and manipulating information by humans. Based on that premise, a concept of linguistic description of phenomena arose, and emerged into a large and growing field. In this paper, we present the approach for derivation of meaningful linguistic summaries from databases. We make an emphasis on description of relationships between attributes of a dataset, which can find applications in a number of domains. To demonstrate that, we took a dataset obtained during a research of cervical osteochondrosis among miners and developed a prototype system, which can produce important and interesting conclusions from data, such as “Muscle strength of miners does not change significantly with an increase of work experience from about 10 years to around 15. However, when experience exceeds 20 years, it decreases dramatically.”
An extensiveness of the use of information technologies in almost all areas of human activity has grown drastically in the past few decades, enabling a collection of huge amounts of various data, which have to undergo some kind of analysis for the benefit of individuals, organizations, or nations. There are many types of data analyses, such as descriptive, exploratory, inferential, predictive, causal, mechanistic, with all of them having different goals and approaches. In this paper we will concentrate our attention on the most frequently used one, that is, descriptive data analysis.
Descriptive statistics and visualization are commonly used methods in the descriptive analysis. However, they constitute only a first step in the process of analysis: the remaining step is interpreting the statistical figures and graphs obtained during the first part. Usually, the interpretation is given in the form of a summary in natural language, such as “men are exposed to alcohol dependence much more frequently, than women”. Human-generated summaries often contain subjective and fuzzy terms, such as “much more”, “significantly”, “large”, “young”, etc., which eases perception of this information. For example, a sentence “last week, 15,000 units of the product were sold” contains less information for an analyst and sounds less natural compared to “last week, the sales were quite weak”.
Based on the fact that natural language is easier to
perceive for humans than numbers, a concept of linguistic
description, or summarization, of data arose. One of the
good examples of work in this area is [
In the next section of this paper we outline the reasons for choosing this research topic. In the part called “Overview of Existing Approaches” we familiarize the reader with modern developments in the field of linguistic summaries derivation and explain novelty of our work. Then we describe our methodology in detail, after which we present results of its implementation in “Application” section and make concluding remarks.
In this paper we demonstrate a method to derive important
and interesting conclusions from data, which were obtained
during a research of cervical osteochondrosis among miners.
We examined goals of the original study [
We believe that linguistic summarization methods can offer a lot to this field of application and can be used effectively to complement conservative data analytics tools, such as statistics and visualization. However, a methodology we have developed is quite universal and can be applied in any situation, where foregoing functionality is needed.
Emergence of Zadeh's fuzzy sets and logic theory made
research towards linguistic description of phenomena
possible, with many researchers turning their attention to
this field [
One of the commonly used methods of linguistic
summarization is Yager's approach [
• A summarizer (e.g., young) • A quantity in agreement (e.g., most) • Truth (validity) – e.g., 0.7 as, e.g., (most of employees are young) = 0.7.
In addition, a set of 's to be described may be restricted by a set of constraints, called filter. Taking into an account all of the above, typical summary looks like:
's are , e.g., “most ( ) of single ( ) employees are young ( )”.
Truth value of a summary may be determined in several
ways; the basic way is Zadeh's calculus of linguistically
quantified propositions [
This idea was implemented by Kacprzyk and Zadrozny as
an add-on to Microsoft Access, called FQUERY [
In the works mentioned above, as well as in other research
works concerning the application of linguistic summaries to
data mining, e.g., [
In our research work, we focus on using fuzzy summaries for purposes of describing pieces of data, which are particularly interesting to a user. Firstly, this includes finding a best fitting linguistic label (which may be a compound of labels) for a subset of records in a dataset, e.g., “most of young employees have middle or high income”.
In addition, we provide an advanced fuzzy comparison for identifying distinction between subsets of data. This is necessary because, while many approaches to linguistic summarization suggest a good way to assign linguistic labels to groups of records in the database (defined by different filters), it appears quite frequently that we have to directly compare two or more collections of objects' parameters and identify how significant is their difference. For instance, the proposed fuzzy summarization mechanism is able to produce summaries like “Muscle strength asymmetry coefficient of the main group is dramatically greater than that of the control group.”
As a third novelty, we provide a way to depict relations between variables in a dataset. By a relationship we mean how a variable of interest behaves in relation to the other, for example, “Muscle strength of the miners drops a little from roughly between 725 and 775 to nearly between 675 and 750 with an increase of age from around 40 or less to roughly between 45 and 50. Then it falls drastically till approximately between 600 and 625 with an increase of age till about 55 or greater.”
Also, it is of interest to note that linguistic summaries are most frequently used in the area of business (sales database is a popular example). In this paper we will try to examine what benefits this methodology can offer when applied to the field of medical research.
E-book sales drop sharply from the beginning of the week, when they are particularly strong, till midweek, when they become rather weak. However, there is a substantial increase towards the end of the week, when sales are quite strong.
- constructs sentence(s) in natural language using a set of templates How e-book sales change throughout the week?
a set of constraints appropriate linguistic labels for attributes’ values
Over the course of our research work, we have developed a system for derivation of linguistic summaries from databases with an emphasis on description of relationships between variables (fuzzy or crisp in nature), or, technically speaking, values of dataset attributes. In particular, the system is able to: • give an accurate linguistic descriptor for an attribute of subset of data, selected by applying some constraints (fuzzy or exact), for instance, “strong”, “roughly between 600 and 650”; • compare groups of objects in a dataset by a chosen parameter, e.g., “this week’s sales have significantly exceeded that of the previous week”; • discover trends in behavior of a variable in relation to the other one (see example on the Fig. 1).
Fig. 1 reflects main components of the system and their interaction. Data model, which is an abstraction of the database, is responsible for a direct interaction with it. Additionally, it contains a set of linguistic labels (fuzzy partition), associated with each attribute, and their definition (corresponding fuzzy sets). For instance, linguistic variable sales has the values: weak, average, strong. Data model is accessed by summarization engine through fuzzy querying interface, which allows obtaining sufficiently good linguistic descriptors (labels) for attributes, manifested by subsets of interest, and difference between these subsets. Summarization engine then uses these pieces of knowledge to build a summary in natural language using built-in sentence structures.
Now let us describe the most important pieces of functionality of our system.
One of the common tasks in data analysis is to characterize values of an attribute of interest of its subgroups. For instance, a researcher may be interested in values of nerve conduction velocity among stope miners with extensive work experience. Although there are standard procedures for numerical description of data, such as finding mean and standard deviation, we advocate using linguistic labels and approximate numbers and intervals, technically defined by fuzzy sets, as they provide a more human consistent way of information representation.
One of the difficulties we faced in the process of implementation of this feature is concerned with how to define fuzzy partitions for fuzzy variables so as to be able to describe and compare different subsets of the data by a certain parameter. Generally, various groups of objects (selected using different filters) might have highly diverse values of an attribute of interest, or, in contrast, highly narrow variance. On the Fig. 2 you may see a frequency diagram of muscle strength coefficient of asymmetry of miners of a main group (miners of specializations with harmful working conditions, which involve factors that influence development of cervical osteochondrosis) and a control group (miners of specializations with less harmful working conditions). As can be noticed, the main group is “wider” and one would describe it as “approximately in the interval between 1.3 and 1.7”, and “roughly between 1.1 and 1.3” for the control group, which is clearly a shorter interval. It is obvious that it is impossible to find an ideal width of a fuzzy interval and a number of fuzzy sets in partition (granularity). Therefore, we either have to define a large number of fuzzy sets which signify approximate intervals of a different length, or find a way to build fuzzy sets on demand, depending on the data itself, but without loosing their descriptive power, which means they have to be standard in a certain way.
We decided to follow the second path, and designed an
algorithm for finding a fuzzy set, combined from predefined
fuzzy sets in a partition using disjunction, which fits a subset
of data best, with a degree of truth being higher than some
threshold value (we took 0.7). In this algorithm we used
Zadeh's formula [
Let us demonstrate the results of running this algorithm on some examples beneath. See Fig. 2 and Fig. 3 for comparison (fuzzy intervals identified for the example groups approximately correspond to intervals of high density on the frequency diagrams).
Muscle strength asymmetry coefficient of the main group is approximately between 1.4 and 1.6.
Muscle strength asymmetry coefficient of the control group is about 1.3 or less.
Muscle strength of stope miners is approximately between 675 and 775.
Muscle strength of timbermen is roughly between 650 and 750.
Muscle strength of loader operators is nearly between 625 and 750.
This is important to note that the method described above can be equally efficiently used for crisp and fuzzy subsets of data (selected using fuzzy filter, e.g., “middle-aged” for age).
Another main component of the system is a mechanism of fuzzy comparison, which is used to identify a degree to which two groups of objects are different. Comparison of groups is very useful for many applications. Specifically, in our application it is necessary for such tasks as: determining distinction between the main and control group of workers, as well as between different professions of miners (also with varying labor conditions) to prove negative impact of certain workplace factors on the development of cervical osteochondrosis and overall health, as well as find core influencing factors; compare medical test results before and after treatment, and results of several different therapies.
A standard way to assess difference between two samples is t-test. However, we did not use it for a number of reasons. Firstly, numbers are perceived and interpreted by human worse, than words, and, secondly, there is a problem designing weighted t-test (for cases when subsets are fuzzy).
Instead, we invented our own approach for finding a coefficient of distinction based on linguistic summaries. In brief, we take the best fit fuzzy sets (let’s say, 1 and 2) corresponding to two subsets of interest (lets call them and ), and find an average of validity of two summaries: “ is NOT 2 ” and “ is NOT 1” (“not” is in traditional Zadeh’s sense). This method also identifies a direction of this difference (less or greater).
A linguistic label for a coefficient of distinction is found as a fuzzy set in the partition (defined subjectively) having the largest degree of membership of this coefficient.
Some of the summaries, generated by the program, that implements fuzzy comparison (refer Fig. 2 and 3 for visual comparison) can be seen beneath:
The simple kinds of summaries given in the previous subsections do not worth much attention per se. However, if combined into more complex structures, they can become a really powerful tool, that enables to depict not only single subgroup or relation between two subgroups' parameters, but a relationship between variables, i.e. attributes of a dataset, which can be fuzzy or crisp in nature. For instance, we are interested in a relationship between age of the workers and muscle strength, which can be represented by a couple of sentences in a natural language (see Fig. 4. for comparison; red dots are means of muscle strength of workers with age intervals loosely corresponding to fuzzy sets in the partition):
Muscle strength of miners drops a little from roughly between 725 and 775 to nearly between 675 and 750 with an increase of age from around 40 or less to roughly between 45 and 50. Then it falls drastically from approximately between 675 and 750 to approximately between 600 and 625 with an increase of age from roughly between 45 and 50 to about 55 or greater.
of miners from 1 and 2 with an increase of from 1 to 2, where: • and are an independent and dependent variable respectively, age and muscle strength according to the example;
• 1 and 2 are two linguistic labels associated with domain, that are neighboring if to put all labels in ascending order (approximately 40 or less and between 45 and 50, between 45 and 50 and 55 or greater in the sentence); • 1 and 2 are their respective linguistic descriptors, that were calculated by a method explained earlier in the text (e.g., roughly between 725 and 775 and nearly between 675 and 750);
• is a relation between subsets 1 and 2 (increase, decrease or no change), exemplified by drop and fall (synonyms) in the above summary;
• D is a linguistic label for a coefficient of distinction (slightly, dramatically, etc.), found using a method described in the previous subsection.
In order to make our summaries more realistic and humangenerated like, we introduced synonyms to the functional parts of a summary and the linguistic labels (refer to Tab. 1). So, for instance, to specify that is significantly greater than , system may also take any one of its synonyms, such as substantially, noticeably.
Actually, these sentences were generated by a computer program, that compared subsets of data corresponding to all linguistic labels associated with an independent variable (age in the example) by the value of the second parameter (muscle strength), after which it putted the results of each comparison into a simple template:
Synonyms about X, around X, roughly greater than X, approximately less than X, nearly between A and B decrease, fall, drop, decline increase, rise, climb, grow quite significantly, slightly, a little significantly, substantially, noticeably very significantly, dramatically, drastically a) not significant difference almost the same, identical
While the kind of summaries demonstrated above is quite human consistent, it is possible to introduce a number of refinements in the future:
a) Add linking words. Words like but, furthermore, nevertheless, likewise, again would allow to increase quality of summaries by highlighting a contrast in trends or their repetition. Examples:
Electromyography test results for the groups of workers with less than 10 years of experience and with 10 -15 years of experience do not differ significantly. However, that for the group of workers with more than 20 years of experience fall sharply from 700-750 to 600-625.
Nerve conduction velocity falls gradually from the age of 35-40 till the age of 40-45. Then it decreases sharply till the age of 45-50 and after that it starts to decrease gradually again.
b) Creating a structure and organization for sentences, summarizing distinctions between groups subsetted using a parameter which is crisp in nature and cannot undergo qualitative comparison, for instance, profession or a method of treatment a worker undergone:
Nerve conduction velocity (NCV) of workers, who have undergone basic therapy treatment, has significantly increased due to the therapy. NCV of workers, who have undergone DENS treatment, has also improved, but to the lesser degree.
In order to enliven the developed methodology and identify its advantages and disadvantages, we implemented some of the aforementioned functionality of the proposed system in Ruby programming language, together with Ruby on Rails web framework, that we utilized for easier database manipulation and user interface design. Let us now demonstrate our results.
Example 1. Age and muscle strength coefficient of asymmetry (refer to Fig. 5 for comparison).
Muscle strength asymmetry of miners rose quite significantly from approximately 1.5 or less to roughly between 1.4 and 1.6 with an increase of age from approximately 40 or less to approximately between 45 and 50. Then it rose drastically from approximately between 1.4 and 1.6 to roughly 1.6 or greater with an increase of age from roughly between 45 and 50 to roughly 55 or greater.
Example 2. Age and muscle strength coefficient of asymmetry with a finer scale for age (see Fig. 6 for comparison).
Muscle strength asymmetry of miners climbed quite significantly from roughly less than 1.5 to approximately between 1.4 and 1.6 with an increase of age from approximately less than 40 to roughly 45. Then it climbed a little from approximately between 1.4 and 1.6 to roughly between 1.5 and 1.6 with an increase of age from about 45 to roughly 50. Then it climbed very significantly from roughly between 1.5 and 1.6 to approximately greater than 1.6 with an increase of age from around 50 to around 55. Then it was almost the same with an increase of age from roughly 55 to approximately greater than 55.
Earlier in this paper, we mentioned several use cases of
our system. Let us provide a more comprehensive list of its
potential applications with accordance to the goals of the
research [
2. Find relationships between various parameters of patients, such as age and work experience, and results of medical tests, so as to prove negative impact of miner's workplace factors on the development of cervical osteochondrosis.
3. Analyze a relationship of potentially unfavorable workplace factors that influence development of cervical osteochondrosis, like heavy physical labor, vibration (coming from drilling machine, loader, etc.), microclimate, noise, and degradation of functional parameters of the body.
In this paper we presented a system for derivation of linguistic summaries from databases, which possesses functionality of linguistically describing relationships between attributes of a dataset and distinction in properties of subsets of data. Main challenges we have faced were construction of a good descriptor for a subset, deciding on the metric for significance of difference between subsets, and designing the summary structure. In our future works we plan to enrich summary structure with new templates and components like linking words. Also, we plan to examine a potential of our methodology for application in areas other than medical research, as we believe that it is universal and can be applied in any context where linguistic summaries can aid in decision support or other purposes, involving data analysis.