In this paper we introduce Word Bucket, a mobile app that applies gami cation to the problem of learning a foreign second language (FSL). Word Bucket consumes and produces linguistic data through user interaction, which can be used to improve available datasets relying on the power of the crowd. We describe the problems around handling \live" linguistic data and how semantic technologies can help to face the problem of data integration and its consumption in this speci c scenario.
the user engagement is by combining FSL apps with gami cation. Building on these ideas this paper presents: { The use case of learning a FSL and how open linguistic data can be applied to this scenario. { A discussion on the data heterogeneity issues around language resources from di erent providers and the e orts to nd a standardized vocabulary for modeling and sharing linguistic data. { The potential bene ts of combining gami cation and crowdsourcing for facilitating the task of learning a FSL, while at the same time producing data that can be reused for improving the quality of the original data sources. { A data analysis approach based on ranking to ensure that users consume the data they expect. 2
Learning a FSL is a task that implies perseverance and continuous motivation. From all the tasks involved in the process of learning a FSL, the acquisition of vocabulary is the one which is present in all stages. From beginners to advanced students, who command the structure of the language, the chances that new vocabulary is needed are high. This is not strange, since we are used to observe this fact in the development of our mother tongue too.
Dictionaries are without doubt the resources that best suit for vocabulary acquisition. While the amount of commercial dictionary publishers is endless, the proliferation in the last years of online collaborative projects like Wikipedia has originated similar approaches for the construction of dictionaries. Wiktionary can be seen as the leading lexical resource generated by the community.
Despite the Wiktionary content can be applied in many di erent areas of Natural Language Processing and Machine learning3, there is still way to go in order to make the contribution of the Wiktionary community comparable to that of Wikipedia in terms of support and engagement. Even though it is clear that all the community would bene t from having better lexical resources, pushing the users to perform tasks like new content creation, translation or content curation is di cult as these tasks can result tedious and repetitive. A possible way of incentivizing users to accomplish this kind of tasks is by using gami cation techniques. The basic idea of gami cation is to hide the details of the real task, while generating an execution environment from which the user could directly bene t. Gami cation pursues to incentivize the user to accomplish a task that otherwise she would not do. Following this line, we have developed Word Bucket, an app that applies gami cation to the process of learning and reinforcing vocabulary in a FSL (Figure 14). 3 For an extended list of tasks where Wiktionary has been applied see \Wiktionary data in natural language processing" at http://en.wikipedia.org/wiki/Wiktionary
ulary, however this part is not shown in the gure for brevity. For more details, we encourage the reader to download the app from www.wordbucket.com. (a) Home (b) Explorer (c) Search
The relevance of Word Bucket relies on the way users interact with the data. On the rst hand, Word Bucket is a consumer of lexical data, i.e. translations, by querying information from multilingual dictionary resources. Users can save associations of words in the target language they are learning, together with the respective translations in their mother language (Figure 1b). These associations are used to build di erent kind of games or tests that proof the knowledge of the learner. On the second hand, Word Bucket is a producer of data. Users can add their own translations to the app when there are not results for their query or these are just not appropriate (Figure 1c).
So far Word Bucket extracts only content from Wiktionary, which is still a resource under construction and presents data de ciencies. This lack is deeper for some languages than others. A possible way of enriching the original datasets is by integrating the user generated content back to the dictionaries. So far, in the current Word Bucket version, there is no possibility to reuse the user interaction for the bene t of the community. Afraid of this situation, we have started the development of a solution with the aim of feeding back the original dataset with data generated on the client side.
The public Wiktionary statistics5 show that the engagement of the community has stabilized in the last 5 years. Figure 2 depicts the amount of active editors for the English Wiktionary along the time. There are not public statistics discerning between di erent kind of modi cations, so we can not know the exact amount of changes that refer to translations only. Nevertheless, if we compare the graphic with the one in Figure 3, we can see that even the amount of daily Word Bucket users (for all the o ered languages in Android) is still under the maximum number of English active editors, the potential amount of user
generated translations could be of great bene t for the community. Moreover, the associations word-sense stored by users in their apps can be used as an indicator for creating scores in the translations. This can help the user to distinguish frequent translations from uncommon ones. At the time of writing this concept does not exist in Wiktionary, but it can be appreciated in other commercial resources like Google Translate6.
We have identi ed the following requirements with the aim of improving the quality of the data and the user experience: { Integrate other lexical resources to compensate the potential lack of quality.
The data integration must be transparent to the client application. { Use the lexical data generated on the client side to curate and enrich the original dataset. { Model the user behavior and incorporate usage statistics that will improve the data consumption. 3
Along the di erent versions of Word Bucket we have tried to remove complexity from the device and implement richer functionality on the server side. In this way, we have modi ed how lexical data is consumed from one version to
another. In version 1.0 and version 2.0, we used a service federation approach. This means that data is consumed directly from the service providers by using their REST APIs. Figure 4a shows the original deployment, in which the device was responsible for implementing restful clients for each one of the integrated resources (in this version, Word Bucket only consumed data from Wiktionary).
As result of the REST requests, the services usually return JSON data, which needs to be parsed and converted to the internal data representation on the client side. The main drawback of this solution is that the data needs to be converted on the y for every request. On the other hand, it allows us to have the latest lexical data o ered by the resource publishers. (a) Version 1.0 (b) Version 2.0
An additional problem of the described deployment is the no possibility to share any of the user generated data. With the aim of xing this problem, we developed a backend in version 2.0 as shown in Figure 4b. While still using a service federation, with this solution we were able to move all the data integration to the server side, removing complexity from the app package. The introduction of the backend allowed us to build our dictionary provider service, which certainly behaves like a proxy delegating client requests to the di erent resource providers. A great bene t of having this centralized proxy is the implementation of a full text search layer to homogenize the way we query the di erent lexical resources that we integrate or we might want to integrate in the future. In version 1.0, the search strategy needed to be handled on the client side as part of the resource integration as well.
With the introduction of the backend services in 2.0 we could start storing
user data through a sync API, which allows the user to have a replica of their
data on the cloud. As shown in 4b, all user generated data is kept independent
of the original resources. As can be appreciated, with this approach the problem
of data heterogeneity still remains open. In order to tackle this issue, we are
currently working on the deployment of a new approach that builds on the use
of semantic technologies (Figure 5). The main idea is to use a data warehouse
solution in which the di erent lexical data is uni ed under a common format.
As we already stated in section 2, when referring to the lexical data needs, Word
Bucket focuses mainly on the consumption of translations in di erent languages.
For this purpose, we will rely on the lemon model [
In order to incorporate user feedback to the lexical resources, we have built
a module to collect usage statistics. The target of this module is to correlate
the user queries with the best option from the list of possible translations. If
the user decides to enter her own translation, this will also be considered for
computing the statistics. Collecting this data from all the users will allow us to
apply a ranking strategy and curate the dataset by removing noisy translations7.
The rankings computed in this step will be made available as part of our data
store. We will relay on the Vocabulary for Ranking (vRank) [
In the next subsections we describe further details on the implementation of the semantic backend solution. 3.1
Data consolidation
An issue that remains open in the eld of computational linguistics is the
development of knowledge artifacts and mechanisms to support the alignment of the
di erent aspects of linguistic resources in order to guarantee semantic and
conceptual interoperability in the linked open data cloud. Ontologies have proved
to be of great use in achieving this goal and big e orts have been done in the
Semantic Web community to address the conversion of datasets to RDF and its
publication as linked data [
Recent initiatives like lemon [
As stated in [
Figure 7 depicts an example of modeling a translation in DBnary and Figure 9 the respective SPARQL query to retrieve all Spanish translations associated to the term \bank".
An important part of our project focuses on incorporating user feedback to the data model. Applying similar mechanisms to those of Web Information Retrieval, we rank the available translations depending on the usage. As a rule of thumb, we consider that a translation has more chances to be right if many users keep it in their local Word Buckets. Following this principle, the scores associated to the translations will start converging in the datastore after heavy user interaction.
In case a translation does not exist, users have the possibility to create it according to their personal knowledge. These personal translations can be in6542173 s e l e c t d????i wwwss::beatooonirrr<<ndddsechh<<tatt<tthh<pp?httwtth::t//ppttop//t::rww//pd: ww///:ww//ww?www/s.. wwwwelleewwn..mmwlls.eeooel.mmennlmeoo?mnnommdnoeoonmmfddmieeoonmllddoi..deeotnnlldeiee..leonntt.ln//n.eellntteee//wtemm/lltheeoo/lemmennlrme##ooemnnowv##nofar#ncsli#uteadtennLee>soefnenixR>n?iicecidapta?eli>lsFfoEeionnnn"r>tbsimrteay>ino>k.:nb" @:.a..en . . 8 g order by ?word ? sense corporated to the system and o ered to the rest of users as part of the lexical dataset when they reach certain \credibility" threshold, i.e., many users have created the same translation in their buckets.
As stated previously, we need some kind of data model in order to make the
translation scores persistent. For this purpose we have decided to use the
Vocabulary for Ranking9 (vRank) introduced in [
Previous works have been performed trying to improve the quality of linked
data by using human contribution to achieve certain data related tasks. In [
which implements a Web adaptive learning solution based on RDF data models.
MAPLE uses a reasoner to match tailored educational content with user
proles, in order to provide a custom learning experience. For this purpose, authors
rely on the use of an extended RDF version of the LOM standard [
Closer to the idea of using games for generating data are the works described
in [
In this paper we have described the use case of learning vocabulary in a FSL by reusing lexical data from online providers. We have provided an overview of Word Bucket, a mobile app that combines gami cation and crowdsourcing in this context, and discussed how the adoption of semantic technologies can help to solve the problem of lexical data heterogeneity and content enrichment.
Future work will focus on two main directions, namely, extending the o er
of available languages and increasing the engagement of users. The rst issue
is an ongoing task since the beginning of the Word Bucket project. To increase
the amount of languages supported, we need to incorporate speci c resources
targeting those languages. Due to the multilingual nature of Word Bucket,
nding this kind of resources is not easy. The main problem resides in getting data
containing bilingual translations, one for each pair of languages we would like
to o er. A potential step towards a solution could be the addition of
commercial dictionaries within Word Bucket. The integration of private and commercial
dictionary data into a global dataspace can open new business models for
dictionary editors and service providers [
A requirement of this new approach is the need for providing provenance
information within the data model, so that authorization mechanisms can be
implemented. This problem has been already addressed by other authors in the
context of pharmacological data [
Regarding the issue of engagement, a possible strategy would involve implementing new tests and minigames within the app. Engagement is directly related to the way users utilize the app. From a social perspective, mobile platforms have revolutionized how we interact with the information. Users carry their mobile devices most of the time. This factor facilitates the online presence of the user in comparison to using other devices like laptops or desktops. While mobile devices can be used everywhere, the second group of devices is only used in places like homes or o ces. By using mobile apps, this fact can be properly exploited to generate content that otherwise could be hard for the user, not only because of the tedious of the task, but mostly because of nding the right time to accomplish it. Applied to the use case described in this work, lexical data can be enriched with user generated content (UGC) like notes, audio and images by using the right mechanisms to incentivize the user interaction. These annotations add an extra value to the original dataset and most importantly, they can be reused as part of the learning process.
Acknowledgments. The work described in this paper has been prepared in a close collaboration with English Bubble Ltd. Word Bucket and all the technical infrastructure around it is property of English Bubble Ltd. The authors acknowledge all the support provided by English Bubble team during the preparation of this paper. Word Bucket is the result of joining the e ort and young spirit of very passionate people from di erent part of the globe. Special thanks go to Danny Smits and Robert Hanley.
We also thank Dieter Fensel and Miguel-Angel Sicilia for the support and feedback regarding this work.