CHAPTER 3. MOTIVATION APPLICATION
4.4 Goal Comprehension Module
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
37
Purpose: evaluate the market niche potential for different choices
Major Input: a series of partner composition choices
Components: attractiveness analysis and uniqueness analysis
Major Output: attractiveness score and unique score (4) Image classification module:
Purpose: process image data in advance to reduce the computation complexity of niche assessment
Major Input: customers’ image preferences and business images
Components: customers’ preferences classification and business images classification
Major Output: customer image preference clusters and business image clusters
We will describe each of the modules in additional process details and provide the algorithm and formula of these modules.
4.4 Goal Comprehension Module
An assumption of our partner recommendation system is that the users have an ability to clearly define their image they would like to build. For instance, a SME may want to make its customers think they’re so happy and just like in paradise when they visit.
Then, a SME user needs to have an ability to think of the word “paradise” and input it into our system. After that, the system would start from analyzing the goal “paradise”
and identify its latent meanings in order to find the most appropriate partners for collectively achieving the goal. The goal comprehension module is designed to perform this task.
In order to comprehend metaphors, e.g., a SME (target) is just like a paradise (vehicle), we adopt a web-driven, case-based approach called the Sandonicus approach (Veale & Hao 2007), which leverages the text of web as a plentiful knowledge source to identify what properties are most contextually appropriate to
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
38
apply to both sides of target and vehicle. This approach employs Google search engine as a retrieval mechanism for finding properties of words by using Google supported APIs, which allow the search of wildcard term * as any possible words. For example, if you send a query “as * as paradise” to Google, you may get a series of words, such as beautiful, gorgeous, wonderful. That implies paradise can be beautiful, gorgeous and wonderful. More specifically, these words can be considered as the properties of paradise. We treat these properties as the meaning of paradise.
However, there are few points that need to be addressed for benefiting from the Sandonicus approach. First, this system will operate in Chinese environment and the Sandonicus approach is basically designed for English environment. To employ the Sandonicus approach, the words need to be translated from Chinese into English. We adopt unofficial Google dictionary API to do so (source:
http://code.google.com/p/google-api-translate-java/ ). Second, given a user is assumed to input a phrase to describe the goal, the phrase should be decomposed into processable lexical units. This task is performed by using Chinese word segmentation API from Sinica (source: http://ckipsvr.iis.sinica.edu.tw/). For example, if the user input (in Chinese form:電影巨片) is “blockbuster movie” instead of “paradise”, then this phrase is first decomposed to two lexical units (i.e., blockbuster and movie) and then they are separately sent to Google dictionary to translate them from Chinese to English. After translation, the system then starts to send “as * as blockbuster movie”
to Google Web query and gets a series of adjectives. Finally, it’s unavoidable to attain undesired results when we employ the sentence pattern “as * as vehicle” in Google.
To settle this issue, we develop a two-step process to ensure we can get the quality results. The first step is to establish an exception word list. Any word in this list will be filtered out from the search results. For example, when we send a query “as * as
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
39
paradise”, we may get the result like “as well as paradise”. The word “well” is obviously not a result we would like to get. Then, all we need to do is add the word
“well” to the exception word list, and the program would automatically filter it out.
We create the list during the test runs. The second step is to leverage SentiWordNet to keep the positive words in the results. SentWordNet is a lexical resource of opining mining (Baccianella et al. 2010;Esuli & Sebastiani 2006). One of its ability is to determine a word whether has positive or negative meaning A word determined positive means that the word demonstrates good side of a thing. For example, happy is a positive word and sad is a negative word. Only the positive words are the words that we care about because the words are used to describe goal and only positive aspects of the goal are important in building good images.
All the details mentioned above are the design logic of the component “metaphor comprehension”. The input of this component is a goal in the form of metaphorical statement and the output is a set of adjectives, which are the meanings of the goal. For more details of metaphor comprehension process, see the algorithm in Figure 4.5.
Thereafter, the gap identification component attempts to catch the missing part of existing images of the SME for achieving the goal. As describe in section 4.2, our system also collected the images of SMEs (i.e., business image model). The collected images are basically adjectives (i.e. image elements) used for describing a SME.
Through comparing the collected image elements of SMEs and identified images elements from the goal, the image gap then can be identified.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
40
During the comparison process, the semantic similarity analysis is conducted to evaluate how close the meanings of two words are, given that the images elementsare adjectives. We use DISCO (extracting DIstributionally related words using CO-occurrences) is a Java class that allows to retrieve the semantic similarity between arbitrary words (Kolb 2008, 2009). It will output a semantic similarity score. Higher score indicates higher semantic similarity. If any of wanted images cannot be found in the existing images of SME or cannot be found in the existing image with a high level of similarity, that image would be considered as one of the gap images. In other words, gap images are those which are not able to be fulfilled by existing images.
Table 4.2 presents an example of gap image analysis. Assume that there are three adjectives (wonderful, beautiful and gorgeous) identified by metaphor comprehension process. This means the SME would like to deliver a service experience that makes people think it’s wonderful, beautiful and gorgeous. In order to understand which
Metaphor Comprehension Component
Step1: Decompose the metaphorical statement into processable lexical unit Step2: Translate every word from Chinese to English
Step3: Set vehicle = the phrase used to define the goal (e.g., paradise) Send the query “as * as vehicle” to Google
Step4: Get a series of properties of the vehicle Save the results of query as the image vector
FOR i TO the total number of image elements in the image vector Remove Image if Image exists in exception list or Image is determined as a non-positive word
NEXT
Step6: Save the rest of image elements in the image vector as the salient properties vector
Figure 4.5. Metaphor comprehension algorithm
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
41
image element is missing, the system would start to compare the goal images (wonderful, beautiful and gorgeous) with the image model of the SME. The table 4.2 shows that the SME give people charming, pleasant and gorgeous images. If any goal image element can be found an exactly same adjective or can be found an adjective with similar meaning in business image model, then this goal image element is determined “fulfilled”. In this example, the goal image “wonderful” cannot be fulfilled because there are no words with similar semantic meaning. On the other hand, the goal image “beautiful” can be fulfilled owning to similar word “charming” and the image “gorgeous” can be fulfilled because of the exactly same word in the image model of the SME. Those unfulfilled adjectives are the gap images. The algorithm of gap identification is specified in Figure 4.6. In sum, the identified gap images will entail what are the elements that should be complemented by others for achieving the goal so that they can serve as the good starting point for partner candidate generation.
Table 3.2 An example of gap image analysis Adjectives from analyzing
the goal
Adjectives from the business image model
Fulfilled or not
wonderful(gap image) pleasant No
beautiful charming Yes
gorgeous gorgeous Yes
However, there is one important thing needed to be noticed. Setting an appropriate goal is relevant to success. If the goal setting is either undesirable or the users do not take features (i.e., environmental, cultural, social aspect) of the destination where the service is offered into account, the users may not get the optimal results. Therefore, in this study, we assume the users can come up with an appropriate