“Know Who” System Using iMage (Information Mixable Graph Explorer)

3.1 Joint experiment on Know Who system for sales persons

In a joint experiment between NTT Information Sharing Platform Laboratories and NTT Software, we developed a Know Who system for sales persons by using data in NTT Software and verified its usefulness by having sales persons use the service in practice.

3.1.1 Need for Know Who system for sales persons

In a company that performs system integration, such as NTT Software, there are many cases where sales persons want to know whether there are any experts on a certain technology in their company. Since many sales persons are not well informed about such matters, we developed a system called the “Technology-case matching service” that can assist in finding appropriate persons or divisions to contact. This service enables sales persons to find the right persons quickly and prevent a mismatch between a project and the people assigned to it.

3.1.2 Data sources

We use two main kinds of data for our Know Who system.

– Project planning documents

– Software utility information

NTT software handles all projects in the form of project planning documents. A project planning document gives the type of project, the name of the division in charge, the name of the project leader, the development period, the process of development, the project members and their skills, and so on, so it is suitable for the Know Who system, which can search for persons or divisions by using technical keywords. In this experiment, we used about 8000 items of data over the past four years related to about 1000 persons. In our Know Who system, the selection of search keywords is very important. Technical keywords are extracted by the system administrators and the domain experts, and we also use software utility information to extract technical keywords. Software utility information is expected to include important and proven product names for system integrators.

3.1.3 System architecture

The architecture of our Know Who system is shown in Fig. 2. iMage analyzes the structure of RDF data and extracts distinctive patterns, which are used as queries when a search is executed. To enable the Know Who system to be operated by users, we had to develop the graphical user interface shown in Fig. 2 because iMage provides only the libraries used in searches.

Fig. 2. Structure of human resources matching system.

3.1.4 Distinctive patterns

First, iMage analyzes the structure of RDF data expressed as a graph. Then, it extracts distinctive patterns from it. Finally, it retrieves knowledge or information by using the patterns as queries. Some examples of distinctive patterns automatically extracted by iMage and then selected by system administrators are shown in Fig. 3. Ordinary search engines only list persons that match the given technical keyword. However, iMage also categorizes them persons based on distinctive patterns. For example, the persons in this result are categorized as a supervisor (pattern 1), project leader (pattern 2) or technical staff member (pattern 3) of a project related to the keyword, as shown in Fig. 3.

Fig. 3. Distinctive patterns used in Know Who system.

3.1.5 Innovations to improve effectiveness

We can measure the effectiveness of the Know Who system by checking whether all of the experts for the given technical keyword are in the search result. To improve the effectiveness, we introduced two innovations to the system.

– Support for synonyms by using folksonomy [3]

– Support for different spellings

Synonym support using folksonomy means that when the given keyword is too specific to find experts in the given field, the search field is expanded by using synonyms on a higher level of abstraction. In the keyword registration form in Fig. 2, RDF and SPARQL were chosen as synonyms of Semantic Web. Therefore, when no one was retrieved using the keyword SPARQL, the system also tried searching with the keyword RDF. We tried to improve the system's effectiveness by providing a function that lets users add synonyms by themselves, just like the mechanism of folksonomy.

The spelling variant support treats alternative spellings of a word as equivalent including alphanumeric words and Japanese hiragana, katakana, and kanji. To handle this, we prepared a dictionary of representative words. We solved the problem by normalizing the words in RDF data and the search keywords to the words in the dictionary.

3.2 System evaluation

We evaluated the system by conducting a survey in the form of a questionnaire. The profiles of the subjects are shown in Fig. 4. We asked the subjects to use the system and then checked the results by asking the system administrators. The results of the questionnaire are shown in Fig. 5. The search with the underlined bold-faced keywords delivered the most effective result because all of the persons in the result were experts including some newly discovered ones who were previously unknown to the subject.

Fig. 4. Profiles of test subjects.

Fig. 5. Results of questionnaire.

We can conclude that the system is effective at supporting the jobs of system integrators for the following reasons: In 80% of the search results, experts were found. Moreover, many underlined bold-faced keywords were input by users D and E, who are sales persons, i.e., our target system users, who are not familiar with technology.