This paper investigates how recommendation a ects tagging behaviour regarding the language adopted in tags. We conducted a study to compare the tags assigned to digital images with and without the support of a recommender system. Results pointed to have an association between users' language used to assign tags and the type of systems supporting this task.
Tag recommender systems have arisen to help users choose
tags most suited to lead to better (more accurate, more e
cient, more satisfying, etc.) content retrieval. One popular
approach used for such recommendations is to suggest tags
which co-occur [
Understanding how users perform the same task in
different environments can provide insight for designers to
decide among distinct approaches according to user and system
needs. The language used for tagging has several
implications for the dissemination of content on the Web [
In this work we report a study investigating whether the presence of recommendation on tagging system can change the users' choices of the natural language they use for tagging digital images. Participants from whom tags were collected were residents of a Portuguese-speaking country and no particular previous de nition regarding the language for tagging was given. The results show that there is di erence in the language adopted to assign tags when users are supported by recommendation. 2.
Because we had two experimental conditions to test
(tagging with and without the recommendation aid), a
counterbalancing design approach was employed in the current
study. A total of 57 participants, all of them residents in a
Portuguese-speaker country, were partitioned in two groups:
G1 had 33 participants (16 female, 17 male, with a mean
age of 27 years); G2 had 24 participants (10 female and
14 male, with a mean age of 25). Participants were
presented the same images, interfaces, and were asked to
perform the same tasks, but each group received the conditions
in reverse order. As a design platform for recommendation,
we used a model [
To process the language of tags assigned in this
experiment, we use a standalone language identi cation tool based
on a Nave Bayes classi er [
Our rst analysis assessed the e ect of using, or not using, a recommender system on the language used to assign tags. The set of tags assigned to each image by each participant was classi ed as either PT or EN. Table 1 shows the difference in the proportion of images and the main language used by each group to images that were presented in both stages of this study.
G2
NR RS 63 (65%) 25 (26%) 33 (35%) 71 (74%)
When not using the recommender system, participants tagged fewer images using EN (G1: mean = 1:54 SD = 1:60; G2: mean = 1:37, SD = 1:71). However in the RS stage, more images were tagged mainly in EN (G1: mean = 3:27; SD = 1:30; G2: mean = 2:91, SD = 1:28). A (paired) Wilcoxon signed-rank test indicated that the mean of images with tags in EN changed (p < 0:01) from one stage to another for both groups. This behaviour also was found when we looked to the language used in each image individually, before and after recommendation (McNemar p < 0:01) and also for the images that were tagged only in the RS stage.
To make sure that the results found in this study were not narrowed by a few participants' behaviour, we looked to their results individually. We classi ed users as: PT-taggers, EN-taggers or multilingual-taggers (ML); PT-taggers | had all their images classi ed mainly by tags assigned in PT; and ML-taggers had a mix of images tagged in EN and PT. Figure 1 shows the proportion of participants and their respective tagging behaviour in each stage of this study.
At the NR stage, 45% of participants were classi ed as PT taggers. However, this behaviour changed in the RS stage, only 8% of them kept tagging images mainly with tags in PT. In the RS stage, the majority of PT taggers switched their tagging language and behaved as EN- and ML-taggers. 8 .
0 % .4 0 0 . 0 ML EN
PT NR
RS
To try understand participants' behaviour, we examined the order of tags assigned in the RS stage: We noticed that at rst some images received reference tags in PT but the following reference tags were assigned in EN. We hypothesized that, as participants received tag recommendation in EN, they switched the language of reference tags. However, individual users' behaviour needs future investigation. 4.
The combination of semi-automatic recommendation and tags' co-occurrence is an interesting and important approach used for recommendation, in part because the users' reference tags are the seed for recommending other tags. This approach is useful to make rich annotation, and also intends to improve the user experience by decreasing the e ort to assign tags while still allowing users to use personal tags. In our study we have shown that the recommendation approach used a ects the language choice for tagging. Results indicate that the quantity of images assigned mainly with tags in EN changed in the RS stage (compared to the NR stage). Because co-occurrence based approaches use the link among tags to recommend other tags, the results found here have several implications for the design process of tagging recommender systems. The approach increased the language homogeneity (EN) of tags which could result in a cultural isolation of online indexed content. We are aware that the training dataset used for recommendation has a representative quantity of tags in EN and consequently many tags in EN can co-occur with tags in another languages. However, the social tag dataset we used represents the natural imbalance of the language on the Web and how tags are connected, which reinforces that the EN language functions as a hub to other languages. On the other hand, the interface used in the RS stage only recommends tags based on the reference tags assigned, so users still had the autonomy to assign their own tags and keep their vocabulary without any recommended tag. The lack of switching (or converting) languages is evident when one considers that even in the RS stage some users continued as PT-taggers. These ndings highlight the needed to investigate whether there is distinct users pro le that are more likely to accept multilingual recommendations. 5.