F RAMING MODULE

國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

37

4.4 Framing module

The framing module is setting the relevant frame to communicate with our target user who is the maven on the recommender, in order to persuade them to be engaged by the focal firm. The relevant frame is the response to the knowledge information which includes the news information and some comments by user in our system summarized by article. Based on the framing theory (Dietram, 1999), our system would create the summarized article for generating knowledge information in two parts – Building the topic and setting the context.

(1) Building the topic: As our system is to generate knowledge article to communicate with mavens, we need to determine an influence agenda relevant to the focal firm to impact the maven. Our system sets a process to build the topic (See the Table 4.2).

Table 4.2 Building the topic process

Step1: Select the attribute of the firm, for example industry and image.

Step2: Collect other relevant historical content and external news of industry and image.

 Step1: Select the attribute of the firm, for example industry and image: The purpose is to find the influence agenda relevant to the focal firm. We detect the attribute of the firm to find classified maven interests, which is the same mechanism as the recommender. Moreover, based on the alliance based module, we recommend the list with the firm’s industry and image. As the result, our topic would also buildwith firm’s industry and image.

‧ 國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

38

 Step2: Collecting relevant historical content and external news: Most of the knowledge information is built from the historical behavior that are user items including comments sharing and response on recommender.

User-item ratings are preference data for user actions, for example, transaction data, explicit ratings, user comments or browse rate (Prasad, 2012), In addition, we include other information which relates with the firm’s attribute on media on step 1, for example, news or fan page post.

Moreover, we obtain the brand alliance information as well.

(2) Setting the context: After building the topic, we need to decide which of the salience attribute of content issue inside to complete our knowledge information in the article. According to the framing theory (Dietram, 1999), the framing setting step is to recognize which attribute is importance to influence the audience. We would measure the each sentence score to identify importance and select the high score sentence to complete the context in article. In order to measure our resource, the setting context process separates into two parts: Tokenize the historical content relevant with the topic with text segmentation tool. After that, choosing the important sentence in each source, we combine multiple sources into the framing context.

Table 4.2 setting the context Step1: Filtering the source relevant with the topic

Step2: Separate the words and sentence with relevant source by Jieba.

Step3: Using TF-IDF calculation to identify the critical word.

Step4: Scoring the sentence with P/N, critical word and image word.

Step5: Framing context setting with extraction source.

‧ 國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

39

 Step1: Separate the words and sentence with relevant history by Jieba: We used Jieba, a Chinese text segmentation tool, as the tokenizer for the task of text segmentation (Tomlinson, 2010), to separate the words in each sentence of relevant data which associate with the topic.

 Step2: Using TF-IDF calculation to identify the term: After collecting the numerous types of information, we need to identify which words are important terms or not. In our system, we use TF-IDF weighting statistic which is abbreviation of Term Frequency*Inverse Document Frequency (Nenkova, 2012). The advantage of using TF-IDF weighting statistic is of no need for the stop word list; on the other words, we can expect that a term which is meaningful and important wordsappear on our arbitrary text which we collected.

(3)

 Step3: Scoring the sentence with P/N, critical word and image word: In the recommender, not only the posts, but the comments contain important information and both of them can influence user. Each of the post with comment would all be extracted in this section. The sentence extraction process the representative score by total three score: First, the positive and negative words which is matching by the NTUD word based. Second, the critical word that measure on the step2, the last is the image words which represent the whole article image. the reader measurement of the comment in using the term frequency, term frequency is defined by the number of occurrence in all comments associate with a blog post, the quotation

‧ 國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

40

measurement is the number of other comments associated, and topic measurement in group comment into topic cluster using a Single-Pass clustering algorithm method, which is used to cluster comments to attain very short and incomplete messages (Shen, 2006). Finally, sentence extraction step selects the high score of representative to as a representative sentence from each post source.

Figure 4.4 Scoring the sentence process

 Step4: Framing context setting with extraction Source: In this process, we aim to combine the multiple resources (include post, news) to generate the summarize article. First, summarize article process would detect the meaningful units (MUs). A MU is composed of one or several sentence segments, and can represent a complete meaning. The system first chooses the relative to source, comparing with the topic associate. Second, clustering the all representative sentences to build the MUs that belong to the same events as an attribute of content issue. Finally, the summarize article process composed these multiple event as a context. And set these contexts to complete the article.

‧ 國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

41