In Social Networking Services (SNS), user profiles, which often consist of an image, texts, and other items, have an important role to connect with other users. However, in a preliminary study with 3,193 sample profiles on Twitter, we found that the average length of profile texts was 40 characters ( = 32.52) where the maximum length is 160. This suggests that many SNS users are missing potential opportunities to expand their social network due to the short profile texts. Therefore, we proposed a search-based interactive system to support the writing of profile texts in SNS. The proposed system was designed to dynamically search for similar profile texts while users were typing their profile so that they can get new ideas such as what to write or how to express. We evaluated the efect of the proposed system by a user study with 24 participants, and found that the proposed system enabled participants to significantly increase the length of written profile texts (+92.5% on average), compared to a baseline system with no assistance.
Social Networking Services (SNS) have exploded in popularity. According to a survey by Metaxas et al. (2014), self-expression and networking were the two main purposes for using Twitter with the proportion of 35.7% and 33.2%, respectively [1]. For the self-expression and networking on SNS, the user profile is essential to establish new connections with other users [2, 3]. For example, a study reported that longer self-description texts were Figure 1: Distribution of the number of characters in the perceived to be more trustworthy [4]. In an experimental profi=le 3te2x.t52o)f Twitter accounts ( = 3, 193, = 40, study by Counts et al. (2009), “Quotes” and “About” were found to be useful in expressing personality traits [5].
However, many SNS users do not have a rich profile text. Figure 1 shows the distribution of the number of characters in the profile text of 3,193 Twitter accounts their profile text in social networking. This observation who are deemed to associate with one of the major univer- led us to develop and evaluate a mechanism that enables sities in Japan. The accounts were manually collected by SNS users to generate richer profile texts. one of the authors from Twitter lists that were distributed The current work was inspired by the concept of Obby the associated members of the University. Figure 1 servational Learning [8] in the field of Social Psychology. shows that a large proportion of accounts uses less than In a broad sense, Observational Learning is the process half of the maximum number of characters which is 160. of learning something by observing the behaviour of othThe average length of profile texts was 40. Even though ers. We thought that the behaviour of referring to the some of them wrote limited information intentionally profile text of other users could be regarded as a form of due to some reasons like a privacy concern [6, 7], there observational learning. might be many users who can benefit from enriching
Research related to the present study includes writing support for diferent types of texts, which consist of novels, scientific paper, sentences written in a foreign language [9, 10, 11]. Roemmele, et al. developed a system to support the creation of story texts by suggesting the completion of the next sentence from a previously created text [9]. Once the user has decided on the theme and vision of the story, he or she can write specific sentences with the help of the program. Kinnunen et al. developed a system that checks for criteria such as “are keywords that occur frequently in the abstract also used in the title” to help users write readable and consistent scientific papers [10].
Ellison, et al. [2] conducted an interview study with users of an online dating service and found that many participants used other users’ profile information to find out what they should pay attention to in constructing their profiles. For example, one participant stated that she avoided using a sitting posture as her icon image because she had found that it was used by some users to make themselves look thinner.
In Information Retrieval (IR), Capra, et al. proposed the search assistance system, called “Search Guide” suing search trails [12]. Users can refer to other users’ search trails (e.g., queries issued, results clicked, pages bookmarked) as an example of the search behaviour. It was found that the Search Guide can help users’ complex search behaviour.
However, there is little work on the development and evaluation of tools to support the writing of profile text for SNS. In general, SNS profile text is much shorter than a story or scientific article. Therefore, we need a diferent approach from assisting users to maintain consistency in their writing or to write longer sentences eficiently [9, 10].
their profile text. For this aim, we proposed a system to assist users in creating SNS profile text by searching relevant profile texts created by other SNS users. More specifically, we formulated the two research questions as follows.
The first step in our work was to build a corpus which was then indexed and searched by the proposed writing assistance system. Since the idea of observational learning indicates that users can benefit from the learning of prior examples created by other users in similar contexts, we decided to collect user profile texts from the accounts that deem to have some level of association with our study participants: students at one of the major universities in Japan.
First, we manually collected 46 Twitter lists which were created by the associated members of the university. Then we collected the accounts of the members of each collected list as well as the accounts who they follow. Finally, the data for each account, including the profile text, was automatically obtained via Twitter API. As a result, a total of 785,531 profile texts were collected.
Furthermore, we had two automated steps to remove profile texts for our research aim. First, as the collected accounts included accounts of organisations and bot accounts, we excluded them from the search if they contained any of the following words.
administrator / association / booking / bot / community / closed / event / info / oficial / sales / shop / tel
were too short or did not contain enough information as observational learning. We decided to exclude profile texts with less than 40 characters from the search. In the end, 351,754 profile texts were included in the corpus.
UI consists of three main areas: Edit Area, Search Result 1.4. Paper Structure Area, and Keep Area. When a user edits a profile text on the form in Edit Area, the system retrieves relevant The rest of the paper is structured as follows. In Section other users’ profile text and displays them in the Search 2, the proposed writing assistant system is presented. In Result Area. The user can use the search results to refer Section 3, we describe the user study to evaluate the ef- to and rewrite his or her profile text. The search results fectiveness of the proposed system. In Section 4 presents were dynamically updated as the texts in Edit Area are the experimental results. In Section 5, we discuss the changed. Therefore, we provided an option to “keep” main findings and their implications on the SNS profile profile texts you find helpful in the Keep Area. Finally, there was a button at the bottom of the Edit Area to file server (Nginx), HTTP Server (RESTful API Server), indicate the completion of the writing task during the and Elasticsearch. Each component was deployed as a user study. Pod in an on-premises Kubernetes cluster. When a user
During the experiment, a controlled UI was also devel- accesses the URL of the proposed system front-end UI, oped where the overall look was identical to the proposed the static file server first serves page contents such as UI. However, the controlled UI did not have the Search HTML, CSS and JavaScript. After that, the interaction Result Area and Keep Area. events on the page and the profile text created by the user are sent to the API server and stored in the server’s file 2.3. Back-End Search System system. Finally, profile texts are retrieved from the corpus by Elasticsearch. We added some plugins to Elasticsearch Figure 3 is a sequence diagram of the proposed system. to tokenize Japanese text and configured Elasticsearch to The proposed system consists of three components: Static rank documents by Okapi BM25 [13].
Since we used a manually constructed non-standard corpus of SNS profile texts in our experiment, we tested the performance of the back-end search system to ensure that it can retrieve relevant texts. Twenty simulated queries were manually created by using partial profile texts, the top 200 documents were retrieved by the backend search system, and their relevance was assessed with graded relevance by the authors. The MAP was 0.589 and nDCG@20 was 0.557. Although this was an informal system evaluation, the results were suficiently promising for our research aim.
user study was conducted with 24 participants who used A total of 24 profile texts were generated in the experithe interactive writing assistant system which indexed ment and used for analyses. the custom collection of Twitter profile texts as described in 2.1. The user study was approved by the ethics com- 4.1. Profile Writing Experience mittee of the Faculty of Library, Information and Media Science, Univerity of Tsukuba (No. 20-7). The experiment To answer RQ1, we investigated participants’ profile writconsisted of a pre-questionnaire, a profile text writing ing experience. We prepared a questionnaire that asked: task and a post-questionnaire. Due to COVID-19, the call "Have you ever had trouble writing in the self-description for participants and all tasks were conducted online. text of your SNS profile? The breakdown of the answers (on a five-point scale from “1. Never” to “5. Every time I 3.1. Participants cannot write well”) is shown in Table. 1. Only 8% of the participants answered “1. Never”, indicating that having Of 24 participants, 14 (58%) were female and 10 (42%) a dificult experience in writing their SNS profile text is were male. All participants were undergraduate students common. The average of the answers was 3.04 and the in the age group between 18 and 22. Their academic back- standard deviation was 1.12. ground varied among Library and Information Science A follow-up question was asked only for those par(13), Computer Science (2), Medicine (2), Social Sciences ticipants who had experienced dificulties (Answer 2 to (2), Mathematics (1), Engineering (1), Disability Science 5) in writing their profile text: “What is the reason for (1), Education (1) and Media Science (1). this?” The options of answers and the results are shown in Table 2. The most common answer was “I could think 3.2. Profile Text Writing Task of many things I could write about, but I didn’t know what to write about in particular” (9 of 18 participants answered). The second most common answer was “I didn’t feel I had anything to write about” and “I wanted to remain anonymous, so I had to be careful not to include any personal information” (7 of 18). Again, these responses support our motivation that many SNS can benefit from writing assistant tools to enrich their profile texts, although some users intentionally provided limited information to keep their privacy.
Participants were asked to write a profile text in our user study. Participants were randomly assigned to one of two groups: Control and Experimental. Twelve participants in the Experimental group created a profile text using the proposed system, and twelve participants in the Control group created a profile text without using the suggestion function. They were also instructed to create a profile text under the following scenario.
“To increase online communication between students at the university, the university asked all students to create a Twitter account. What kind of profile text would you like to create for your SNS account for the campus life?”
Since our collection was focused on university-related user profiles, we designed the scenario as above to avoid zero-match results in the Search Results area during the profile writing task.
To answer the research question RQ2, we compare the distribution of the number of characters in the profile text created between the two groups. Firstly, the number of characters in the profile text created by each group in descending order is compared in Figure 4. As can be seen, the experimental group tended to have more characters
“I could think of lots of things I could write about, but didn’t know what to write about in particular” “I didn’t feel I had anything to write about” “I wanted to remain anonymous, so I had to be careful not to include any personal information” "I was concerned that I might come across as self-conscious if I wrote long sentences" “I thought it would be strange if I wrote something diferent from other users around me” Others (e.g. “I didn’t know how to introduce myself”) ing method. It is a two-tailed test and the null hypothesis is rejected at 5% level of significance. Since the sample size in this study is small (12 participants for each groups), in their profile texts. we thought it necessary to also focus on the efect size as
Next, we compared it using the boxplot and the results well as the p-value [14]. The efect size for the Mannare shown in Figure 5. It can be seen that the quartiles Whitney U test can be calculated from the test statistic and the minimum and maximum values are larger in the U and the sample size and satisfies − 1.00 ≤ ≤ 1.00 experimental group. In addition, the value of the quar- [14]. The coin package [15] (1.4-1) on the CRAN package tile range is relatively larger in the experimental group. manager for R language was used to compute p-value and This indicates that the efect of increasing the number of efect size. The results of the Mann-Whitney U test are characters by using the proposed system varied across shown in Table 4. From < 0.0.5, there is a significant participant. The descriptive statistics values are summa- diference between the two distributions. In addition, as rized in Table 3. r-value coincides with the one for Pearson’s correlation
Furthermore, a statistical test was carried out to see if coeficient [ 14], it can be considered as a moderate efect there was a statistical significance in the distribution of size (0.36 ≤ ≤ 0.67) [16]. the number of characters between the two groups. The Additionally, we analysed the time participants spent Shapiro-Wilk test was used to confirm that the distribu- during the profile writing task, and Figure 6 shows the retion did not follow a normal distribution for both groups, sult. It was found that the experimental group took 164.5 and thus, the Mann-Whitney U test was used as the test- seconds longer than the control group on the median
We asked the following questions in the post questionnaire: “What did you have in mind when you started writing your profile text?” (A) and “What was written in the final profile text?” (B). In both questions, participants selected their answers from the same list of possible answers, including “Afiliation” and “Hobbies”. We counted the number of increase between what participants “Intended to Write” (A) and what they “Actually Wrote” (B).
The result is summarized in Table 5. The most common answer was “Others” (“Research Interests”, “Past Afiliaprofile texts in the proposed system. profiles written by the proposed system can lead to a bet
Therefore, providing a learning opportunity based on ter extension of social networks than existing profiles or existing users with similar interests seems to be a promis- profiles written by other methods. Furthermore, searching method to help SNS users to enrich their profile texts. based observational learning seems versatile enough to Unlike the assistance system in academic writing or cre- apply to other domains, and thus, the efectiveness of the ative writing[9, 10], the profile writing seems to benefit proposed system in other writing tasks such as product from the prior examples of "similar account" since the review writing should also be an interesting research main aim of SNS is to connect with people with similar direction. Finally, how to achieve a good balance beinterests. Given that online communities are diverse, our tween privacy and efective profile is yet another imporapproach of retrieving profile texts from similar accounts tant research question to be addressed in future work. seems to have an advantage over a fixed list of items to From a system design perspective, how to integrate the include in the profile. assistance function to the operational SNS needs to be
As for the limitation of our work, first, participants of examined [18]. our user study were recruited from a single university although their academic background varied. Therefore, a similar efect might not appear in other SNS user populations (e.g., diferent age groups or occupations). Second, this study did not verify whether the profile text created by the proposed system is indeed efective at expanding social network. Third, as pointed out in previous research[17], user privacy also needs to be considered in supporting the creation of user-generated content, including profile text. Finally, this study did not investigate whether the ranking algorithm of the back-end search system afected the user experience.
Based on our preliminary observation that the average length of SNS profile texts was not nearly as long as it could be, we developed a search-based profile writing support tool. The basic idea of our proposed system was to provide users with potential ideas about what to write and how to express by ofering profile texts written by other users with similar interests. A user study was carried out with 24 participants to evaluate the efectiveness of the proposed system. The experimental results show that the proposed system can enhance the profile writing task by allowing users to include more items in the profile texts leading to longer and richer contents. The results of questionnaires also indicate that one of the reasons for short profile texts is due to the dificulty in writing profile texts, supporting the motivation of our work, while some users preferred to be anonymous with limited profile content. Nevertheless, this work suggests that the proposed system will be helpful for those who would like to efectively extend their social network by richer profile contents.
This paper demonstrated that search-based assistance was efective for SNS profile writing of a particular university student group. Future work should investigate the eefctiveness with other university student groups as well as other population groups (e.g., professionals).
Further work is also desirable to examine whether the [10] T. Kinnunen, H. Leisma, M. Machunik, T. Kakkonen,
J.-L. LeBrun, SWAN - scientific writing AssistaNt. [1] P. T. Metaxas, E. Mustafaraj, K. Wong, L. Zeng, a tool for helping scholars to write reader-friendly M. O’Keefe, S. Finn, Do Retweets indicate Inter- manuscripts, in: Proceedings of the Demonstraest, Trust, Agreement? (Extended Abstract), CoRR tions at the 13th Conference of the European Chapabs/1411.3 (2014). URL: http://arxiv.org/abs/1411. ter of the Association for Computational Linguis3555. arXiv:1411.3555. tics, Association for Computational Linguistics, Avi[2] N. Ellison, R. Heino, J. Gibbs, Managing Im- gnon, France, 2012, pp. 20–24. URL: https://www. pressions Online: Self-Presentation Processes in aclweb.org/anthology/E12-2005. the Online Dating Environment, Journal of [11] M.-H. Chen, S.-T. Huang, H.-T. Hsieh, T.-H. Kao, Computer-Mediated Communication 11 (2017) J. S. Chang, FLOW: A first-language-oriented writ415–441. URL: https://doi.org/10.1111/j.1083-6101. ing assistant system, in: Proceedings of the ACL 2006.00020.x. doi:10.1111/j.1083-6101.2006. 2012 System Demonstrations, Association for Com00020.x. putational Linguistics, Jeju Island, Korea, 2012, pp. [3] S. T. Tong, E. F. Corriero, K. A. Wibowo, T. W. Makki, 157–162. URL: https://www.aclweb.org/anthology/ R. B. Slatcher, Self-presentation and impressions of P12-3027. personality through text-based online dating pro- [12] R. Capra, J. Arguello, A. Crescenzi, E. Vardell, Dififles: A lens model analysis, New Media & Society ferences in the use of search assistance for tasks 22 (2020) 875–895. of varying complexity, in: Proceedings of the [4] X. Ma, J. T. Hancock, K. Lim Mingjie, M. Naa- 38th International ACM SIGIR Conference on Reman, Self-disclosure and perceived trustworthi- search and Development in Information Retrieval, ness of airbnb host profiles, in: Proceedings of SIGIR ’15, Association for Computing Machinery, the 2017 ACM Conference on Computer Supported New York, NY, USA, 2015, p. 23–32. URL: https: Cooperative Work and Social Computing, CSCW //doi.org/10.1145/2766462.2767741. doi:10.1145/ ’17, Association for Computing Machinery, New 2766462.2767741.
York, NY, USA, 2017, p. 2397–2409. URL: https: [13] S. Robertson, S. Walker, S. Jones, M. M. Hancock//doi.org/10.1145/2998181.2998269. doi:10.1145/ Beaulieu, M. Gatford, Okapi at trec-3, in: Overview 2998181.2998269. of the Third Text REtrieval Conference (TREC[5] S. Counts, K. Stecher, Self-presentation of person- 3), Gaithersburg, MD: NIST, 1995, pp. 109–126. ality during online profile creation, in: Conference: URL: https://www.microsoft.com/en-us/research/ Proceedings of the Third International Conference publication/okapi-at-trec-3/.
on Weblogs and Social Media, ICWSM 2009, 2009. [14] M. Tomczak, E. Tomczak, The need to report efect [6] H. Krasnova, S. Spiekermann, K. Koroleva, T. Hilde- size estimates revisited. an overview of some recbrand, Online social networks: why we disclose, ommended measures of efect size 21 (2014) 19–25. Journal of Information Technology 25 (2010) 109– [15] H. Torsten, H. Kurt, M. Van de Wiel, Z. Achim, 125. URL: https://doi.org/10.1057/jit.2010.6. doi:10. Implementing a class of permutation tests: The coin 1057/jit.2010.6. package, Journal of Statistical Software 28 (2008). [7] F. Stutzman, J. Kramer-Dufield, Friends only: doi:10.18637/jss.v028.i08.
Examining a privacy-enhancing behavior in face- [16] R. Taylor, Interpretation of the correlation coefibook, in: Proceedings of the SIGCHI Conference cient: A basic review, Journal of Diagnostic Medical on Human Factors in Computing Systems, CHI Sonography 6 (1990) 35 – 39. ’10, Association for Computing Machinery, New [17] Y. Fang, A. Godavarthy, H. Lu, A utility maximizaYork, NY, USA, 2010, p. 1553–1562. URL: https: tion framework for privacy preservation of user //doi.org/10.1145/1753326.1753559. doi:10.1145/ generated content, in: Proceedings of the 2016 1753326.1753559. ACM International Conference on the Theory of [8] A. Bandura, Observational Learning, 2 Information Retrieval, ICTIR ’16, Association for ed., Macmillan Reference USA, New Computing Machinery, New York, NY, USA, 2016, York, NY, 2004, pp. 482–484. URL: https: p. 281–290. URL: https://doi.org/10.1145/2970398. //link.gale.com/apps/doc/CX3407100173/GVRL? 2970417. doi:10.1145/2970398.2970417. u=cuny_hunter&sid=GVRL&xid=369c6e06, topic [18] G. Dworman, S. Rosenbaum, Helping users to overview. use help: Improving interaction with help sys[9] M. Roemmele, A. S. Gordon, Creative help: A story tems, in: CHI ’04 Extended Abstracts on Human writing assistant, in: International Conference on Factors in Computing Systems, CHI EA ’04, 2004, Interactive Digital Storytelling, Springer, 2015, pp. p. 1717–1718. URL: https://doi.org/10.1145/985921. 81–92. 986198. doi:10.1145/985921.986198.