CHAPTER 3. MOTIVATION APPLICATION
4.5 Candidates Generation Module
environmental, cultural and social features of a destination.
4.5 Candidates Generation Module
Candidates generation module aims to attain possible partner lists. The process is briefly described as follows. Base on the gap images identified in the last module, hereafter named “supertype”, this module uses it to generate a collection of new metaphors. These metaphors are then analyzed by the metaphor comprehension process to ensure every metaphor we generate makes sense and then each metaphor
Gap Identification Component
Step1: Take the image model of a specific SME Step2: Conduct the semantic similarity analysis
For each element of salient properties vector (goal image elements), examine its semantic similarity relative to the images of SME.
Set MaxSimilarity = the highest level of similarity of specific salient property FOR i To total number of salient properties
MaxSimilarity = 0
FOR j To total number of images of SME
TempSimilarity = Compute the semantic similarity index between and
IF MaxSimilarity < TempSimilarity THEN MaxSimilarity = TempSimilarity
END NEXT
IF Maxsimilarity > specific threshold THEN Tag with “fulfilled”
ELSE
Tag with “unfulfilled”
END NEXT
Step3: Save all of the properties tagged with “unfulfilled” as the Gap vector Figure 4.6. Gap identification algorithm
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
43
can also be projected to specific business types for possible cooperation accordingly.
Once these candidates are identified, goal fulfillment analysis is executed to ensure that cooperating with those candidates can achieve the user’s goal, attaining a series of business types for potential cooperation.
The basic idea behind the abovementioned process is to generate the new metaphors that best describe the possible partners. For example, if a new metaphor is generated like “your partner is as wonderful as flower”, the system will try to find a partner who can be best described as flower. The reason to do so is that metaphor usage in design problems has been a primary inspiration for our study of analogy. We believe analogy to the design field can give a heuristic problem solving strategy in which a problem solving method can be carried over to another new problem. In the first module, metaphor comprehension is brought into our approach for understanding the goal. In this module, metaphor generation is tfor generating possible partners who have potential to collectively achieve the goal with the initiator.
Similar to what the metaphor comprehension process does, the metaphor generation process (see Figure 4.7) uses the “gap images” as salient properties of metaphors and send the query “as gap image as *” to Google. It will return a collection of vehicles that have the “gap image” properties. This process involves the filtering work as well. For instance, if the gap properties include sweet and delicious, it will send two queries, “as sweet as *” and “as delicious as *”, to Google and then gathers collections of vehicles that can be depicted as sweet and delicious. The vehicles with those two properties in the same time are preferred. After collecting a series of vehicles, this module combines the topic and vehicles to form a complete set of metaphors and uses the metaphor comprehension process to examine the suitability of the metaphor configurations. Once this step is completed, the module then projects
‧
the vehicles to some real business types and investigates the level of fulfillment of the user’s goal.
One major limitation of the Sandonicus approach is that the sentence pattern used for generating new metaphors is improper to deal with multiple salient properties at a time. As we noted earlier, the sentence pattern we employed in metaphor generation is “as salient property as *”. Google search would return a set of words modified by the salient property. Imagine if we attempt to send a statement to Google
the number of gap image contained in the properties of a vehicle the number of gap images
Metaphor Generation Component
Step1: Take the Gap image vector
Step2 : Generate vehicles by the aid of Google API
For i To the total number of elements in the Gap vector Send the query “as Gap as *” to Google
Save the results in the Vehicle vector
For j To the total number of elements in the Vehicle vector Send the query “as Gap as Vehicle” to Google
IF the number of returned results higher than a specific threshold THEN
Step 3: call metaphor comprehension to identify the properties of each vehicle Step 4: Compute gap image coverage rate for each vehicle
Gap image coverage rate of a given vehicle =
For each Vehicle
Compute the number of gap image are contained in Vehicle
Record the maximum gap image coverage rate Next
Step 5:
For each Vehicle
If gap image coverage rate of Vehicle = maximum gap image coverage rate among Vehicle
Save the valid vehicles in Vehicle as the Candidate vehicle vector Next
Figure 4.7. Metaphor generation algorithm
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
45
like this, “as property1, property2 ,property3… as *”, it is obviously that the results would not be desired because people do not use language like that. Apparently, we’d better to try it on the other way around.
In this research, we extend the Sandonicus approach so as to generate the metaphors with multiple properties following terms of a three-stage process. At first, the each property is put into the sentence “as property as *” and separately sent to Google search. Different sets of words then are returned. In this stage, different words (i.e., vehicles) with a given property are the new metaphors generated. Second, use metaphor comprehension technique to contrarily identify the properties of vehicles.
Accordingly, a set of vehicles are generated and the properties of each vehicle are identified as well. The primary principle to find the most appropriate vehicles is that the properties of the vehicle should contain gap images as many as possible. Finally, the vehicle contained the most gap properties will be chosen in order to form the new metaphor.
After a set of metaphors are generated, candidate discovery component then attempts to match the metaphors to real businesses. More specifically, if a new metaphor “your partner is just like a flower (vehicle)” is created, the component will try to find out which business entity is just like a flower. To do so, the image model similarity analysis will be executed. The design logic of image model similarity analysis is very similar to gap image analysis introduced in section 4.4. Table 4.3 presents an example of image model similarity analysis. In the previous step, the properties of the vehicles have been identified and these properties are also adjectives.
Therefore, the first column in the table shows the properties of the vehicle and the second column presents the major image elements of a given business model. The only difference between gap image analysis and image model similarity analysis is
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
46
that gap image analysis tries to understand which image element cannot be fulfilled but image model similarity analysis is focusing on the image elements that can be fulfilled. Figure 4.8 shows a simple formula is developed for indicating the level of similarity between two different image element sets. In general, a vehicle will be matched to the business model with the highest similarity score.
Table 4.3 An example of image model similarity analysis Adjectives from analyzing
the vehicle
Adjectives from the business image model
Fulfilled or not
wonderful pleasant No
beautiful charming Yes
gorgeous gorgeous Yes
Image model similarity score
= the number of image element that can be ful illed the number of image element in a given image model Figure 4.8 the formula of image model similarity score
The computation complexity would be high if the system tries to perform image similarity analysis for every image business model in the system. For example, if there are 1000 business image models in our system, and then it will be 1000 times of image similarity analysis to be conducted in order to find the best match for a vehicle.
For the purpose of reducing computation complexity, the image classification module is developed and will be elaborated in a later section. Put simply, if these 1000 image models can be grouped into several classes (e.g., 100 classes) according to their image configuration, then the image similarity analysis only need to perform 100 times.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
47
Once the best match of image class is identified, the image models belonging to this class are all matched because we assume that the image models in the same class have similar image configurations. For more details, please reference Figure 4.9 and 4.10.
Figure 4.9.Candidate Discovery algorithm (1)
Candidate DiscoveryComponent
Step1: Match vehicles to real businesses
For each Candidate vehicle, examine its similarity relative to the image models of business classes.
Set MaxSimilarity = the highest level of similarity between a specific vehicle property and a specific image of a specific business class
Set CountFulfilledProperty = the number of fulfilled properties (the initial value of CountFulfilledProperty = 0) Set MaxFulfilledProperty = the maximum number of fulfilled properties.
Set Flag = the index of Business image class
Set SimilarityLevel = the level of similarity of specific vehicle and real business image class
FOR i To total number of 0 Candidate vehicles MaxFulfilledProperty = 0
FOR j To total number of Business image class CountFulfilledProperty = 0
Set P = Read the properties of a specific candidate vehicle from CandiVehicleProperites
Set BI = Read the images of a specific business image class FOR k To total number of P
Maxsimilarity = 0
‧
Figure 4.10.Candidate Discovery algorithm (2)
For each matched real businesses, compute its level of goal fulfillment. Since the goal of the SME is to convey a specific image to public through the aid of cooperation, it’s essential to have the anticipated effect of cooperation forecasted when different partner compositions are formed. To this end, the image model mixing API is accordingly adopted. The image model mixing API belongs to color mixing
TempSimilarity = Compute the semantic similarity index between
IF Maxsimilarity > specific threshold THEN Tag Gap with “fulfilled”
CountFulfilledProperty = CountFulfilledProperty+1 END
NEXT
IF MaxFulfilledProperty < CountFulfilledProperty THEN MaxFulfilledProperty = CountFulfilledProperty
Flag = j END
NEXT
Gap
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
49
image model similarity analysis, the fulfilled images can then be identified. Then, goal fulfilment score can be computed. The formula is showed in Figure 4.11. The denominator is the number of goal images and the numerator is the number of goal images which are fulfilled. The higher score indicates the higher level of goal fulfilment and its value will between 0 and 1. For more details of goal fulfilment analysis, reference Figure 4.12. Finally, the output of candidate generation module will be the candidates list with the highest goal fulfilment score.
Goal ful illment score = the number of goal images that can be ful illed the number of goal images
Figure 4.11 The formula of goal fulfilment score