Computing with Words for Direct Marketing Support System Pakizar Shamoi Atsushi Inoue Department of Computer Science, Department of Computer Science, Kazakh-British Technical University Eastern Washington University pakita883@gmail.com inoueatsushij@gmail.com Abstract imprecision. The next section considers in detail the fuzzy This paper highlights the simplicity and effectiveness of querying model implemented using Computing with Words Computing with Words (CW) in the implementation of methodology that is solely based on fuzzy mathematics. target selection. Direct marketing can be considered as one Then we provide illustrative examples of different types of of the main areas of application for this methodology. In queries and their result sets obtained from the application particular, fuzzy classification is applied in it with the developed. Finally, the last section provides the concluding purpose of choosing the best potential customers for a new remarks of this study. product or service from a client database. One of the advantages of the proposed method is that it is consistent with relational databases. Our methodology makes it The Role of Target Selection in Direct possible to form queries in natural language, such as “print Marketing the list of not very old married clients with more-or-less high income”, which is impossible using a standard query Direct marketing (DM) is a form of advertising that enables mechanism. companies to communicate directly to the customer, with various advertising techniques including email, mobile Introduction messaging, promotional letters, etc. The crucial idea there is to be able to deliver the marketing message to the clients There is one fundamental advantage of humankind that that are likely to be interested in the product, service, or needs to be inculcated into the various information offer (Mederia and Sousa, 2002). So, DM companies or systems. It is the remarkable human ability to perform a organizations try to set and maintain a direct relationship wide range of mental tasks without any measurements and with their clients in order to target them individually for any computations (Zadeh, 2002; Herrera, et al., 2009; specific product or service. Martinez, et al., 2010). That is possible due to the brain’s An important data mining problem from the world of DM crucial ability to manipulate perceptions of size, distance, is target selection (Mederia and Sousa, 2002). The main weight, speed, etc. (Zadeh, 2002). The main difference task in target selection is the determination of potential between measurements and perceptions is that the former customers for a product from a client database. Target are crisp whereas the latter are vague (fuzzy) - the selection algorithms identify the profiles of customers who transition from membership to non-membership is gradual are likely to respond to the offer for a particular product or rather than sudden. service, given different types of information, like The main purpose of using natural (linguistic) queries profession, age, purchase history, etc. In addition to the instead of numbers is that it is much closer to the way that numerical performance, model transparency is also humans express and use their knowledge. Perception-based important for evaluation by experts, obtaining confidence rational decisions in an environment of imprecision are in the model derived, and selecting an appropriate becoming highly actual (Zadeh, 2002). An important use marketing channel (Mederia and Sousa, 2002). Fuzzy of the Computing with Words (CW) methodology, which models for target selection are interesting from this angle is in the heart of fuzzy logic, is its application to decision of view, since they can be used to obtain numerically making (Zadeh, 1965; Zadeh, 1975; Zadeh, 1996; Ying, consistent models, while providing a linguistic description 2002; Herrera, et al., 2009). In fact, CW can simplify the as well. decision processes when the experts can only provide As mentioned above, in DM the selection of the target qualitative, but not quantitative information about the audience is a very important stage. Different DM evaluated alternatives (Herrera, et al., 2009). techniques benefit from accurate target selection. Take, for This paper is organized in six sections. First one is this example, a direct mail, used in the promotion of goods and introduction. Next we emphasize the critical importance of services to organizations and individuals through electronic target selection in direct marketing. Furthermore, we mail. Some DM methods using particular media, especially examine in details the how fuzzy approach was applied to email have been criticized for poor target selection make the process of target selection more efficient. strategy. This poses a problem for marketers and Particularly, it discusses the concepts of linguistic consumers alike. On the one hand, advertisers do not wish variables and hedges, fuzzification, explains the need for to waste money on communicating with consumers not interested in their products. Also, they don’t want to lose sentences in a natural or artificial language” (Zadeh, 1975). potential customers. On the other hand, people usually try So, for example, Income is a linguistic variable if its values to avoid spam. However, they want to be aware of the new are linguistic (not very low, average, more-or-less high, …) products/services that might be interesting for them. rather than numerical (100 000 tg., 150 600 tg….). As previously mentioned, in order to maximize its Following that logic, the label high is considered as a benefits direct mail requires careful selection of recipients. linguistic value of the variable Income, it plays the same So, if the selection of recipients is too liberal, it will role as some certain numerical values. However, it is less increase unnecessary spending on DM, if it is too strict – precise and conveys less information. we’ll lose some potential customers. Virtually all To clarify, in the example provided, Income is a companies that work with a database of 100 or more linguistic variable, while low, average, high are linguistic customers use email-mailing in their business (Ribeiro and values, represented in the form of fuzzy sets. The set of all Moreira, 2003). But again, this is a very delicate linguistic values of a linguistic variable is called term set. instrument, because the line between a useful message and Although a linguistic value is less precise than a number spam is very thin. Therefore, companies providing mailing it is closer to human cognitive processes, and that can be services, must constantly engage in outreach efforts, so exploited successfully in solving problems involving that due to their own ignorance, they do not lose customers uncertain or ill-defined phenomena. So, in situations where and reputation, sending spam and making other common information is not precise (which are very common in our mistakes. real life), linguistic variables can be a powerful tool that takes the human knowledge as model (Herrera and Herrera- Fuzzy Approach Viedma, 2000). Besides their primary meaning, linguistic values may involve connectives such as and, or, not and hedges such Need for imprecision as very, quite extremely, more or less, completely, fairly, Nowadays, most of the data processed in information etc. about which we will talk extensively later. systems has a precise nature. However, a query to the database, formed by a person, often tends to have some Fuzzification degree of fuzziness. For example, the result of a query in a search engine is a lot of references to documents that are It is highly important for any target selection model to ordered by the degree of relevance to the request. Another select the clients’ features that will play the role of explanatory variables in the model (Mederia and Sousa, simple example of natural query used in everyday life: "Find a listing for housing that is not very expensive and is 2002). They serve to reflect the essential characteristics of close to downtown". Statements like "not very expensive," the clients that are important for the product or service and they vary from organization to organization. Therefore, just "close" are vague, imprecise, although rent price is completely determined, and the distance from the center of for the sake of simplicity in this model we present just the apartment - up to a kilometer. The cause of all these some of the possible criteria – gender, age, status, income. So, let’s suppose we have a table “Clients”, consisting of problems is that in real life, we operate and argue using imprecise categories (Zadeh, 1965; Zadeh, 1975). 7 rows: id (primary key), name, gender (‘Male’,‘Female’), For example, a company launches an advertising age, status (‘Married’, ‘Not_married’), email, income. campaign among their clients about new services through Table 1. Structure of the sample table for the system direct mail. The Marketing Service has determined that the new service will be most interesting for middle-aged Field Type Fuzzy Comments married men, with more-or-less high income. A crisp id int auto increment query, for example, might ask for all married males aged 40 to 55 with an income of more than 150,000 tenge. But name varchar with such a request we may weed out a lot of potential gender enum ('Male', 'Female') clients: a married man aged 39, with an income of 250,000 age int √ tenge does not fall into the query result, although he is a status enum ('Married', 'Not_married') potential customer of the new service. email varchar income int √ Linguistic variables One of the main hindrances of modern computing is that a By the way, in practice, a certain threshold of concept cannot be well understood until it is expressed membership value is given in excess of which records are quantitatively. This is where linguistic variables come in. included in the result of a fuzzy query. Usually it is a The main motivation to prefer linguistic variables rather number between 0 and 1 and can be represented to the user than numbers is that a linguistic description is usually less in the form of percentage. So, an expert can manoeuvre specific than a numerical one (Zadeh, 1975). with it to make the query more or less strict. One of the According to Zadeh , “By a linguistic variable we mean situations in which a threshold can be very efficient is a variable whose values are not numbers but words or when expert receives a long list of clients as a response to a query. Then, he can decide to be stricter and make the The last thing left to do - to build certain membership threshold higher in order to be more confident in the functions belonging to each linguistic term – fuzzy set We buying power of the clients. define the membership functions for the young, middle- Usually, in real-world decision making processes there aged, and old fuzzy sets with the following parameters are experts - decision makers who choose the appropriate [a,b,c] = [18,35,65]. In general form they look like: initial parameters to define the fuzzy variables (Martinez, et al., 2010; Herrera and Herrera-Viedma, 2000). So, because of different cultural reasons or different points of view and knowledge about the problem it seems reasonable to give possibility to decision makers to provide their preferences about the problem on their own. That is why a, b, and c parameters for each of the fuzzy variables should be input to the system by the expert. Again, this is done, since it seems difficult to accept that all experts should agree to the same membership functions associated with primary linguistic terms. Many decision problems need to be solved under Figure 1. Fuzzy sets for young, middle-aged, and old uncertain environments with blurred and imprecise information. The use of linguistic information in decision making involves processes of CW (discussed a bit later). Fuzzy sets and logic play a major role in this project. Fuzzy mathematics allows us to use the imprecision in a positive way. It is very efficient in complex problems that can’t be handled using standard mathematics, like processing human elements – natural language, perception, emotion, etc. The term “fuzzy” can be defined as “not clear, blurred, or vague.” For example, the word “tall” is fuzzy, since it is subjective term. For some people, man Figure 2. Membership functions for young, middle-aged, and old with the height 190 cm (6.2 feet) is tall, whereas for others 170 cm (5.7 feet) is enough to call the person “tall”. As Now we can, for example, calculate the degree of Zadeh said, “Fuzzy logic is determined as a set of membership of a 30-year-old client in each of the fuzzy mathematical principles for knowledge representation sets: based on degrees of membership rather than on the crisp membership of classical binary logic” (Zadeh, 1996). According to traditional boolean logic, people can be either tall or not tall. However, in fuzzy logic in the case of the fuzzy term “tall,” the value 170 can be partially true and partially false. Fuzzy logic deals with degree of membership with a value in the interval [0, 1]. In this paper fuzzy sets are used to describe the clients’ age and income in linguistic terms which are fuzzy variables. A computationally efficient way to represent a fuzzy number is to use the approach based on parameters of its membership function. Linear trapezoidal or triangular membership functions are good enough to catch the ambiguity of the linguistic assessments (Herrera and Figure 3. Membership of a 30-year-old client in young, middle- Herrera-Viedma, 2000). It is not necessary to obtain more aged, and old. (μ [Young] (30) = 0,294, μ [Middle-aged] (30) = accurate values. The proposed parametric representation is 0,71, μ [Old] (30) = 0). achieved by the 3-tuple (a; b; c) for each fuzzy variable, it is enough for 3 fuzzy sets, since we applied a fuzzy Another fuzzy variable in the system is client’s income. partition. We define it for the domain X = [0, 1000 000], so, the Now let’s try to formalize the fuzzy concept of the universal set U = {0 ,1,…, 250 000, …,1 000 000}. The client’s age. This will be the name of the respective term set consists of 3 fuzzy sets – {"Low", "Average“, linguistic variable. We define it for the domain X = [0, 90], “High“}. The membership functions for income variable so, the universal set U = {0 ,1, 2, …..,89, 90}. The term set term set are totally similar to the ones discussed above. The consists of 3 fuzzy sets – {"Young", "Middle-aged“, parameters are the following [a,b,c] = [40 000, 100 000, “Old“}. 200 000]. Income fuzzy variable, so as Age, is partitioned by three fuzzy sets associated with linguistic labels. Each fuzzy set corresponds to perception agents – low, average, Figure 5 demonstrates that very hedge steepens the curve. or high salary. As it can be seen from the graph, there are no sharp boundaries between low, average, and high. Figure 5. Visualizing the hedge very Furthermore, the modifier that can weaken the statement - more-or-less X (i.e. more-or-less middle-aged) would be given as a square root of the initial function (Zadeh, 2002): Figure 4. Fuzzy Sets for low, average, and high income. A fuzzy μ F MORE − OR − LESS ( X ) = μF ( X ) partition. Figure 6 illustrates that more-or-less hedge makes the curve less steep. It is highly important to remember that decision making is an inherent human ability which is not necessarily based on explicit assumptions or precise measurements. For example, typical decision making problem is to choose the best car to buy. Therefore, fuzzy sets theory can be applied to system to model the uncertainty of decision processes. Linguistic Hedges Knowledge representation through linguistic variables characterized by means of linguistic modifiers – hedges Figure 6. Visualizing the hedge more-or-less makes the query more natural, so their main advantage is the ability to be expresses in natural language. Hedges can Finally, not X (i.e. not young) which is a complement change the statement in various ways – intensify, weaken, fuzzy set, can be expressed by subtracting the membership complement. Their meaning implicitly involves fuzziness, function of X (middle-aged) from 1: so their primary job is to make things fuzzier or less fuzzy. Let’s consider the most common ways of generating new μ F NOT ( X ) = 1 − μ F ( X ) fuzzy sets based on the initial fuzzy set using various hedges. This is useful for constructing various semantic structures -composite words - from atomic words (i.e. young) that reinforce or weaken the statements (Zadeh, 1996) such as very high salary, more-or-less old, etc. Again, the main motivation is to strengthen or weaken the statement (Zadeh, 2002). For reinforcing there is the modifier very, to weaken - more-or-less or almost, approximately. Fuzzy sets for them are described by certain membership functions. Hedges can be treated as operators which modify the meaning in a context- Figure 7. Visualizing the modifier not independent way. For example, let’s suppose that the meaning of X Let’s enjoy calculating the membership of 30-year-old (middle-aged) is defined by some membership function. If client to each of the fuzzy sets: middle-aged, not middle- we want to strengthen the statement, we use very aged, very middle-aged, and more-or-less middle-aged. intensifier (Zadeh, 2002). Then the meaning of very X (i.e. very middle-aged) could be obtained by squaring this function: μFVERY ( X ) = ( μF ( X )) 2 database in response to a query expressed in a natural language. That is our result - terminal data set (TDS). So, our goal is to derive TDS from IDS. In our model, we process the natural query step by step, by constraining the values of variables, this process will be considered in details later. Our aim is to make explicit the implicit fuzzy constraints which are resident in a query. So, how can we make explicit the fuzzy constraints that are given in natural language and so as are implicit? In linguistic features of variables, words play the role of the values of variables and serve as fuzzy constraints at the same time (Zadeh, 1996). For example, the fuzzy set young plays the role of a fuzzy constraint on the age of clients. Figure 8. μ [middle-Aged] (30) = 0,71; μ [not middle-aged] (30) Young takes the values with respect to certain membership = 0,29; μ [very middle-aged] (30) = 0,5; function. In a very general case query consists from a set of μ [more-or-less middle-aged] (30) = 0,84. criteria and set of constraints put on those criteria. So far, we have primary terms for income and age - high, average, As we have seen, hedges intrinsically convey the low, young, middle-aged, old; hedges - not, very, more-or- imprecision in themselves. The main flexibility they less; connectives - and, but, or. provide is that they can make a fuzzy natural query even In outline, a query q in a natural language can be more natural. considered as a network of fuzzy constraints. After processing procedure we get a number of overall fuzzy Computing with Words constraints, which can be represented in the form X is R, Y is S…, where X is a constrained criterion variable (i.e. age) Computing with words (CW), originally developed by which is not explicit in q, and R is a constraint on that Zadeh, provides a much more expressive language for criterion. So the explicitation process can be defined as: knowledge representation. In it, words are used in place of q → X is R, Y is S…, etc. numbers for computing and reasoning, with the fuzzy set As a simple illustration, let’s consider the simple query: playing the role of a fuzzy constraint on a variable (Zadeh, not very young males with more-or-less high income. As 2002). CW is a necessity when the available information is we can observe, some of the variables in a query are crisp, too imprecise to justify the use of numbers, and when there while some have fuzzy constraints. Let’s assume that the is a tolerance for imprecision that can be exploited to user chose the threshold value μ Total as a sufficient level achieve tractability, robustness, low solution cost, and of precision. So, we obtain: better rapport with reality (Zadeh, 1996). YOUNG[Age; not, very; μ Total ] ∩ HIGH[Income; more-or- A basic premise in CW is that the meaning of a proposition, p, may be expressed as a generalized less; μ Total ] ∩ MALE[Gender; ; μ Total = 1] constraint in which the constrained variable and the Notice that not very young and very not young are constraining relation are, in general, implicit in p (Zadeh, different things. Therefore, the order is important. Another 1996). In the system proposed here the CW methodology main point is that μTotal is a membership value reflecting is slightly adapted and modified in order to be able to the degree of membership to not very young and more-or- process natural queries, not propositions. less high, not to young and high. We need to pay special attention to it. To obtain the answer we need the membership value that corresponds to young and high, of course. That is why, the process is reversed: before we presented the formulas to shift to very young from young. Now, instead, we want to define young using very young. So, if we squared the threshold for young to get the threshold for very young, now we apply the inverse operation – square root. Furthermore, we get: YOUNG[Age; very; 1- μ Total ] ∩ HIGH[Income; ; μ Total 2] ∩ MALE[Gender; ; μ Total = 1] = YOUNG[Age; ; Figure 9. CW approach to target selection 1 − μ Total ] ∩ HIGH[Income; ; μ Total 2] CW methodology is changed a little bit to correspond to the proposed system. In particular, the initial data set ∩ MALE[Gender; ; μ Total = 1] (IDS) is a database with clients’ information. From the IDS we desire to find the subset of clients from the Let’s consider the translation rules that can be applied nowadays it is becoming increasingly important to access singly or in combination (Zadeh, 1996). These translation information in a more human-oriented way – using natural rules are: language. a) Modification rules. Example: ‘very old’; We presented a fuzzy querying model capable of b) Composition rules. Example: ‘young and more-or-less handling various types of requests in a natural language middle-aged’ ; form. The interface developed allows experts to express Constraint Modification Rules. X is mA → X is f (A) questions in natural language and to obtain answers in a where m is a modifier – hedge or negation (very, not, readable style, while modifying neither the structure of the more-or-less), and f (A) defines the way m modifies A. database nor the database management system (DBMS) It should be stressed that the rule represented is a query language. convention and shouldn’t be considered as the exact The main advantage of our model is that the query to the reflection of how very, not or more-or-less function in a system is done in a natural language. Besides, existing natural language (Zadeh, 1996). For example, negation not clients databases do not have to be modified and is the operation of complementation, while the intensifier developers do not have to learn a new query language. very is a squaring operation: Basically, we just have a fuzzy interface that is used as a if m = not then f (A) = A’ top layer, on an existing relational database, so, no if m = very then f (A) = A 2 modifications on its DBMS were done. Constraint Propagation Rules. Constraint propagation The main problem in developing human-oriented query plays crucial role in CW. It is great that all the stuff with interfaces was how to allow users to query databases in a numbers takes plays outside of a user’s vision. natural way. The motivation for that is that usually users The rule governing fuzzy constraint propagation: If A don’t wish to define the clear bounds of acceptance or and B are fuzzy relations, then disjunction – or (union) and rejection for a condition, that is, they want to be allowed conjunction – and (intersection) are defined, respectively, some imprecision in the query (Ribeiro and Moreira, as max and min (Zadeh, 1996). 2003). Users can express the intersection in 3 ways The past and new conceptual structure of the model is distinguished by the connective type – and, but, or no schematized and illustrated below, in figure 10. connective at all. As it was previously stated, for this operation, we take the minimum of two memberships to get the resultant membership value: μA( x) ∩ B ( x) = min[μA( x), μB( x)] Union is represented solely by or connective. The resultant membership value is equal to the maximum of two values provided: μA( x) ∪ B( x) = max[μA( x), μB( x)] The threshold used in the system serves as the α-cut (Alpha cut), which is a crisp set that includes all the members of the given fuzzy subset f whose values are not less than α for 0< α ≤ 1: fα = {x : μ f ( x) ≥ α } We also know how to connect α-cuts and set operations (let A and B be fuzzy sets): ( A ∪ B ) α = Aα ∪ Bα ( A ∩ B ) α = Aα ∩ Bα , So, using the formulas provided above, in order to find the result of a query with a certain threshold – α, containing or or and operations, we first find the α-cuts and then take the crisp or / and operation. In dealing with real-world problems there is much to be gained by exploiting the tolerance for imprecision, Figure 10. a) Traditional approach. b) CW approach uncertainty and partial truth. This is the primary motivation for the methodology of CW (Zadeh, 2002). Let’s look at examples of natural language queries given with the purpose of demonstrating the capabilities of this human-oriented interface. Application and Examples The application interface is very friendly. If some user More and more employees are depending on information experiences problems forming a natural query, he can use from databases to fulfill everyday tasks. That is why another menu provided. In it the criteria are listed on the screen, and user can just pick and choose which ones he wants. Furthermore, the parameters he chooses appear and YOUNG[Age; not, very; μ Total =0.7] ∩ he needs to choose needed values (“very young”, “not old”, etc.) on the respective pull-down menu. Moreover, in order MARRIED[STATUS; ; μ Total =1] ∩ (AVERAGE[Income; ; to adapt this model to other databases we won’t need to change the logic, because it is context-independent. We μ Total =0.7] ∪ HIGH[Income; more-or-less ; μ Total =0.7]) = will just need to change the list of fuzzy variables. YOUNG[Age; ; μ Total ≈ 0.55] ∩ MARRIED[STATUS; ; Example query 1. not old married males with very high μ Total =1] ∩ (AVERAGE[Income; ; μ Total =0.7] ∪ income. [Threshold value: 0.5] HIGH[Income; ; μ Total =0.49]) Here we have two crisp criteria - status is married, gender is male. Furthermore, there are two fuzzy criteria – The result set is the following: age is not old and income is very high. So, we have: OLD[Age; not; μ Total =0.5] ∩ HIGH[Income; very; id name gender age status email income 5 Ernar M. Male 32 Married era@gmail.... 890 009 μ Total =0.5] ∩ MALE[Gender; ; μ Total = 1] 8 Karl L. Male 50 Married karl@hotm... 200 300 ∩ MARRIED[Status; ; μ Total = 1] = OLD[Age;; 9 Amina L. Female 74 Married amina@ya... 120 000 13 Alfi A. Male 67 Married alfi@gmail... 88 000 μ Total =0.5] ∩ HIGH[Income; very; μ Total ≈ 0.7 ] Last thing to note, the hedges can be applied infinitely in ∩ MALE[Gender; ; μ Total = 1] ∩ MARRIED[Status;; any order! In order to demonstrate that in practice, consider the following example. μ Total = 1] Next our system finds the values of age and income that Example query 4. very very very old or very very very correspond to the thresholds obtained. For the age, the young. [Threshold value: 0.5] constraining relation will be “ ≤ 50 ”, for the income – OLD[Age; very, very, very ; μ Total =0.5] ∪ YOUNG[Age; “ ≥ 170 710 tg.”. very, very, very; μ Total =0.5] = OLD[Age; ; μ Total ≈ 0.92] Now, having a look at our sample table, we can find 2 clients, fully satisfying the query. The system gives us the ∪ YOUNG[Age; ; μ Total ≈ 0.92] same result: The targeted clients are: id name genderage status email income id name gender age status email income 5 Ernar M. Male 32 Married era@gmail.... 890 009 9 Amina L. Female 74 Married amina@ya... 120 000 8 Karl L. Male 50 Married karl@hotm... 200 300 10 Alan D. Male 18 Not_mar alan@gmai... 35 000 13 Alfi A. Male 67 Married alfi@gmail... 88 000 Example query 2. middle-aged but not more-or-less old clients. [Threshold value: 0.5] For sure, such type of human oriented interfaces can be Here we need to make the conjunction of two constraints very useful for all companies that face the problem of on one fuzzy variable – age: efficient target selection of clients. MIDDLE-AGED[Age; ; μ Total =0.5] ∩ OLD[Age; not, An Issue of Hedges more-or-less; μ Total =0.5] = MIDDLE-AGED[Age; ; There is one thing that disorients me in our Direct μ Total =0.5] ∩ OLD[Age; ; μ Total =0.25] Marketing System. Using Zadeh's definition of the very We obtain the following result (note, that if we queried intensifier it follows that the curve for very young, must hit just for middle-aged, then 50-yeared client Karl L. would the values 0 and 1 at exactly the same places as the curve be included to the result set): for young. It is counterintuitive in this particular application (as well as others), since it can be absolutely id name gender age status email income possible that someone is young without it being absolutely true that he is very young. This contradiction, no doubts, 4 Iliyas T. Male 28 Not_mar.. iliyas@gma... 305 000 gets even worse with very very young, very very very 5 Ernar M. Male 32 Married era@gmail.... 890 009 young, etc. According to Zadeh's, they all hit the values 1 6 Kamin... Female 40 Married kaminari@... 55 000 and 0 at the same place as young. 11Madina.. Female 34 Not_mar… madina_@... 30 000 A different model for the hedges may likely be necessary for a future improvement, that narrows down the range of Example query 3. not very young married clients with ‘very young’ whose membership values are 1 when average or more-or-less high salary. [Threshold value: 0.7] comparing with that of ‘young’ for example. We obtain the following: Conclusion (Ribeiro and Moreira, 2003). R. A. Ribeiro, Ana M. Moreira. Fuzzy Query Interface for a Business Database. The main goal of this research was to demonstrate the International Journal of Human-Computer Studies, Vol 58 effectiveness of Computing with Words approach in (2003) 363-391. natural query processing. In a nutshell, it allows us to form queries in natural language, which is impossible (Ying, 2002) M. Ying. A formal model of computing with using a standard query mechanism, thus simplifying the words. IEEE Transactions on Fuzzy Systems, 10(5):640– life for an expert. 652, 2002. In certain areas, like Direct Marketing, target selection of information from databases has very blurred conditions. (Zadeh, 1965) L. A. Zadeh. Fuzzy sets. Information and Fuzzy queries can be very efficient there. Similarly, fuzzy Control, 8:338–353, 1965. queries can be used in the variety of other fields. Namely, in selecting tourist services, real estate, etc. (Zadeh, 1975) L. A. Zadeh. The concept of a linguistic To conclude, the use of natural language in decision variable and its application to approximate reasoning. Part problems is highly beneficial when the values cannot be i. Information Sciences, 8(3):199–249, 197 expressed by means of numerical values. That happens quite often, since in natural language, truth is a matter of (Zadeh, 1996) L. A. Zadeh. Fuzzy Logic = Computing with degree, not an absolute. Words. Life Fellow, IEEE Transactions on Fuzzy Systems, There are future improvements. However, those are Vol.4,1996. mostly some minor technicality such as the matter of linguistic hedges being counterintuitive and some (Zadeh, 2002) L. A. Zadeh. From Computing with auxiliary, cosmetic functionality such as a parser and GUI Numbers to Computing with Words – from Manipulation when considering some system development. of Measurements to Manipulation of Perceptions. Int. J. Appl. Math. Comput. Sci., Vol. 12, No. 3, 307-324, 2002. References (Herrera and Herrera-Viedma, 2000) F. Herrera, E. Appendix I Herrera-Viedma. Linguistic decision analysis: steps for Sample table data solving decision problems under linguistic information. Fuzzy Sets and Systems 115 (2000) (Herrera, et al., 2009) F. Herrera, S. Alonso, F. Chiclana, E. Herrera-Viedma. Computing with words in decision making: foundations, trends and prospects. Fuzzy Optim Decis Making (2009) 8:337–364 DOI 10.1007/s10700- 009-9065-2 (Mederia and Sousa, 2002) S. Maderia, Joao M. Sousa. Comparison of Target Selection Methods in Direct Marketing. Eunite, 2002 (Martinez, et al., 2010) L. Martinez, D. Ruan, F. Herrera. Computing with Words in Decision support Systems: An overview on Models and Applications. International Journal of Computational Intelligence Systems, Vol.3, No.4, 382-395, 2010.