10 July 2023

AI is intended to facilitate communication between daycare professionals and non-German-speaking parents.

In larger cities in particular, the proportion of children who do not grow up speaking German as their mother tongue is 40 percent or more. What the report "Education in Germany 2022" proved with figures, the pedagogical specialists in kindergartens and schools know only too well from their everyday work. "Around half of all children in our facilities do not speak German as their family language," confirms Theresia Wollnitz, staff member in the staff unit Science Cooperation – International Affairs at FRÖBEL, Germany's largest national non-profit provider of daycare centers, kindergartens and after-school programs with offices in several German states. "Our employees are thus repeatedly faced with the challenge of discussing sometimes very complex matters across minor and major language barriers. This involves, for example, organizational issues, the search for and allocation of daycare places and, last but not least, always individual support for the child and the families."

In practice, conventional translation apps have proven unsuitable for use in communication with parents because they do not have terms such as "settling in" in their repertoire. And bringing in language mediators for initial or developmental discussions with parents who speak little or no German is often uncomfortable for families in view of the sensitive topics involved. For the daycare center management and the pedagogical staff, the provision of such translators is a time-consuming and organizational task that cannot be carried out indefinitely in everyday life. This situation does not do justice to FRÖBEL's claim to enable all children and parents - "regardless of whether they have a migration history or not and regardless of their social status" - to participate in everyday kindergarten life.

^{Theresia Wollnitz © FRÖBEL / Marie Baer}

(Language) barriers are there to be broken down

"A digital tool that can be used by our pedagogical staff at any time, knows pedagogical vocabulary and can easily translate would often support our employees and the families very well," Wollnitz sums up the idea.

It is true that FRÖBEL has already been working for some time on the question of how digital solutions can enrich both the administrative operations of daycare centers and the pedagogical practice. However, when it comes to technical issues, the association is happy to rely on competent cooperation partners.

"AI should serve people"

One just such was quickly found at the German Research Center for Artificial Intelligence (DFKI GmbH) with "SpeechTrans4Kita", as the project was called: "When FRÖBEL presented us with this project idea, we were immediately enthusiastic," explains Dr. Stefan Schaffer from DFKI Berlin, who has been involved since the first project phase in spring 2021. "The project combines current research in the field of machine translation with a very concrete and expected timely benefit for educational practice. AI is supposed to serve people - if we can use it to break down language barriers in special settings, it will live up to that claim."

^{Stefan Schaffer © DFKI}

Together, the partners developed the idea in summer 2020 to make AI specifically useful for pedagogical dialogue in two languages. An app will simultaneously translate conversations between daycare professionals and parents and constantly learn so that the quality of the translations keeps improving.

"In a survey of all the facilities, we were able to identify the greatest need for Arabic," Wollnitz reports of the beginnings. To feed the AI system with relevant expressions, terms and phrases, test persons - parents, students from the FRÖBEL Academy and pedagogical specialists, worked together in language tandems. "We initially oriented ourselves here to vocabulary that is used in conversations between the facility management, the pedagogical specialist and the parents at the beginning of the daycare start - the so-called familiarization conversations. At the same time, we selected important terms relating to the educational activities in the kindergarten, such as language learning diary or morning circle," she says, describing the approach. These were terms that are used in all of FRÖBEL's 220 or so facilities and that are also established technical terms.

"A glossary of about 100 company-specific and educational terms was used at FRÖBEL," adds Stefan Schaffer. "For some of these terms, different alternative formulations (synonyms) suitable specifically for Arabic were created, thus expanding the still small database to no less than 150 expressions." The creation of a larger data collection is planned for a further project phase.

How it works (so far):

The result is a translation app based on Android that is used by users via a shared tablet: There, the individual sentences of the dialog are displayed in user-specific "speech bubbles." "Using automatic speech recognition, spoken language is converted into text. The text is sent to the backend for translation. Translations into the target language are generated, as well as corresponding back translations." Stefan Schaffer describes the process. "In an additional processing step, it is checked whether special terms of the glossary have been used or parts of the utterances have already been translated by the crowd. If this is the case, the corresponding parts of the text are exchanged. Before a dialog partner receives the translation, back translations are first displayed to those who made the last input. The most suitable back translation can be selected, which then displays the translation to the dialog partners in the frontend and outputs it through speech synthesis. If a (back) translation is not good enough, it can be marked accordingly and is then sent to the crowd for translation. This step is asynchronous, which means that the response usually does not come immediately, but with a time delay. So the crowd response doesn't benefit the current dialog immediately, but can only have an effect for later uses." Instead of voice input, text can also be entered via the tablet's virtual keyboard.

It goes without saying that the - often sensitive - content of the conversation between the daycare specialist and parents will be treated confidentially by the project partners. To protect personal data and thus the interests of children, families and professionals, all speech-processing services are operated as customized open-source competencies on DFKI's own servers. "Speech recognition is an exception, as no licenses for on-premise integrations have yet been acquired for the test phase," specifies Stefan Schaffer, who is responsible for the technical implementation with his 4-person team.

Promising practical test

In November 2021, the app was put through its paces by nine participants with pedagogical backgrounds: "The test subjects were enthusiastic about the opportunities and possibilities associated with an intelligent translation app for practical use," says FRÖBEL's Wollnitz happily. "In particular, it's very helpful for building and shaping relationships with families of non-German family language." However, the cooperation partners also noted a "not inconsiderable need for development, for example in translation quality." The flow of conversation as well as usability should be more practicable and improved. "The biggest task, however, is broad application, which means that in as many conversational situations as possible, the respective language must be "learned" in order to become better and better," Wollnitz says.

Keeping an eye on world affairs: Ukrainian-German as the next step

Accordingly, the app is still a few steps away from being used in practice. After initial results for Ukrainian-German were produced in 2022 in addition to Arabic-German in view of the growing number of refugees from Ukraine, the focus this year is on further developing the prototype. "As a next step, a larger funding project is planned in which the existing demonstrator will be further developed into a viable productive solution and an operator concept will be worked out," explains Stefan Schaffer. Funding for this is currently being acquired and, according to FRÖBEL, the search is proving "relatively difficult. Nevertheless, the cooperation partners foresee a bright future for SpeechTrans4Kita. In three years, the project should "preferably have grown into a practical, tested and evaluated product," "which can translate ten other languages without any problems and is not only widely used at FRÖBEL," Theresia Wollnitz has a clear goal. The realization of this vision will certainly not fail due to the need in educational practice.