Support for Creating SNS Profile Text Through Interactive Information Retrieval
Kousuke Nagase (AY 2020)
Profile text is an important piece of information that impresses users on social networking sites. However, according to a preliminary survey conducted by the author (n=2,376), the average number of characters in a Twitter profile (maximum 140 characters) is 40 (median 29), indicating that not many users create a well-developed profile text. Against this background, we developed and evaluated a system to support the creation of profile text on social networking sites.
The proposed system implements a feature that dynamically searches for profile texts so that users can refer to the content and writing style of profiles of other users who are similar to themselves. The idea of referring to similar profiles was inspired by the theory of observational learning in the field of social psychology. The proposed system has been implemented as a general web service. Specifically, when a user enters part of a profile text in a page form, the system automatically searches for related profiles of other users and displays them next to the text. The user can use the search results to rewrite his or her own profile text, or fix a particularly helpful profile text to be displayed on the screen.
To evaluate the proposed system, we evaluated the ranking performance and the usefulness of the system through human subject experiments. For the ranking performance evaluation, since there was no suitable existing dataset for evaluation, we constructed our own dataset using the following procedure. First, 20 fictitious SNS user personas were considered, and a profile statement was created for each. This was used as a query, and for each query, the suitability of the profile sentences for the search target was manually judged by the author. Using this evaluation dataset, the MAP and nDCG@k scores for the search function of the proposed system were obtained. Note that the ranking algorithm used was a combination of Okapi BM25 and the cosine similarity between query and document of the document distributed representation. The results showed that the values were all above 0.5, indicating that the proposed system has a certain level of retrieval performance. In the subject experiment, the proposed system was evaluated on 16 students of Tsukuba University. The subjects were divided into two groups of 8 students each, one using the automatic search function and the other not using it, and were asked to create a profile statement with the setting of an account for university interaction. The results of the analysis showed that the group that created profile texts using the suggestion system had more information in their profile texts than the group that did not use the system. It was also confirmed that the subjects in the group that used the suggestion system tended to write about more topics in their profile texts than they had expected.
As a future direction, we would like to verify whether the profile text created using the suggested system is effective when actually used on social networking sites. In addition, it is necessary to further diversify the test subjects and test collections to obtain more accurate evaluation results.
(Translated by DeepL)