Found in Translation: The Use of Machine Translation Devices in Museums

Imagine you show up to a museum where you are unable to read any of the information, to listen to the docents and guides, or even determine where the bathrooms are located. Maybe it’s a museum in a foreign country. Maybe it’s the museum in your own neighborhood. Either way, that museum is inaccessible to you. Now imagine that in the palm of your hand is a device that can help you read, listen, and find what you are looking for through instant translation of text and speech. While this kind of device is still limited to sci-fi novels and movies, this ability could soon be available to us as machine translation continues to be perfected. As the technology stands now, there are still limitations to what machine translators are able to do, but they can be a helpful tool for museums to use to help expand their community outreach and better serve their multilingual visitors when used in conjunction with human translators.

Benefits of Translation in Museums

According to US Census Bureau data from 2018-2019, nearly 68 million in the United States people speak a language other than English (LOTE) at home. (US Census 2022) With this number constituting nearly triple what it was in 1980, it is clear that non-native English speakers are a growing part of the American population. In some states, 1 in 3 people speak a LOTE at home. (Zeigler and Camarota 2019) As part of the 2018 American Community Survey, which includes over 2 million households, around a third of the LOTE speakers responded that in their own opinion they speak English “less than very well.” (US Census 2022) The top five most common languages spoken across the United States were Spanish, Chinese, Tagalog, Vietnamese, and Arabic, with Spanish making up more than 60% of LOTE speakers. (US Census 2022) Museums that are looking to expand their visitation have the opportunity to target these members of their communities. By creating materials in other languages, people may feel more accepted and welcome in a museum. They will also likely be able to learn more about the museum and its collections than if they walked around an exhibit with text in a language they do not speak and be able to ask questions and engage.

Many different forms of communication exist within museums that would necessitate translation within a multilingual museum. In her Survey of Translation Practices in Museums, the Manager of Interpretive Media and Resources for the National Gallery of Art, Reema Ghazi found that about half the 26 surveyed museums regularly translated wall text, labels, maps, interpretive brochures, and resources for families. (Ghazi 2022) However, other resources like audio guides and videos were rarely translated. This is likely because interactive and audio resources are typically more expensive to translate than simple text labels. (Kushner Bishop 2022) The survey included responses from many types of museums including art museums, history museums, university museums, and children’s museums with a variety of budgets and resource availability. (Ghazi 2022, 502) While Ghazi’s survey gives some insight for how translation is used in a variety of museums, the small sample size of 26 museums included in Ghazi’s survey makes it difficult to determine if this is representative of the field at large. A larger sample size in different types of museums would give stronger evidence of the field as a whole.

A larger survey was completed in 2009 by the Association of Science-Technology Centers, Inc. (ASTC) and the Exploratorium, a science museum in San Francisco which included responses from professionals at 111 ASTC-member museums. (Garcia-Luis 2009) At that point, only 10% of American science museums that were surveyed were offering most visitor information in more than one language. While the survey is outdated in 2023, it does give some insight to the reasons for early adopters of translation in museums. Over half of the U.S. museums surveyed in 2009 said that the most common reason for multilingual offerings was an institutional mandate to make exhibits and programs accessible to a range of audiences. (Garcia-Luis 2009, 7) Around 40% also wanted to target outreach for underserved audiences. (Garcia-Luis 2009, 7) This corresponds with the reasons given by the institutions in 2021 Ghazi’s survey who cited inclusivity as the most common motivator for translation. (Ghazi 2022, 502-503)

The Limitations of Human Translation

Museums often face the challenge of limited resources. Translation services can be prohibitively costly. While the cost can vary, professional translation services for translating the wall text for an exhibit into one other language can cost thousands to tens of thousands of dollars. (Ghazi 2022, 506) The price of translation services is based on many factors including the number of words to be translated, the language that it is being translated to, the content type, turnaround time, and the expertise of the translators. (Eriksen 2023) Generally, a professional translator costs around $0.10 to $0.30 per word, so a single label with 150 words would cost between $15 and $45. (Bagińska 2023) This price can impact the ability of some museums to offer translated materials.

Beyond the price of the translation itself, there are other costs associated with translated text. Languages take up different amounts of space on a wall or page even when they are direct translations. (Ghazi 2022, 506) Some languages’ orthographies also require specialized fonts to create all of the characters in that language. (Minneapolis 2019) This means additional labor to reconfigure the layout and design of translated text to fit in the same brochure or label as the initial language.

In order to cut down on the cost of translation, museums may apply for grants or raise money, but often they rely heavily on bilingual staff to perform translation services. Ghazi’s survey determined that about 75% of museums handled their translation internally by staff, volunteers, or board members. (Ghazi 2022, 502) Not a single surveyed museum included translation in their job descriptions though, so employees are being asked to use their language skills beyond what is expected of the position and what may be beyond their actual language skills. (Ghazi 2022, 502) While fluency in both initial and target languages is a huge part of converting text between two languages, translation is a skill all of its own. Professional translators and interpreters adhere to ethical guidelines for the field of interpretation, such as confidentiality and translating accurately without personal bias. (ATA 2020) Many are certified by the American Translators Association which helps to ensure that they have the translation skills necessary to complete projects successfully. (ATA 2023) Unlike these certified professionals, bilingual employees are not likely to receive any training or evaluation of their translation skills before being used to translate materials. This puts undue burden on the bilingual staff to perform beyond the scope of their knowledge and expertise. They may feel pressured to do the translations on behalf of their own communities even though those communities would best be served by professional translators.

Humans also have limitations in their knowledge of different languages. There are certainly some polyglots who know more than a couple of languages, but no one can store the hundreds of different languages that exist in our world. Each time a museum wants to offer their content in additional languages, they have to find a new translator which will double the costs discussed above. To offer live programs like tours, museums may have to hire another staff member who is able to speak that language. These human resource costs add to the overall difficulties of a museum to provide interpreted information to their visitors.

Human translation of written text also requires another limited resource in museums, time. Generally, a translator works on about 2000-2500 words per day. That is equivalent to about 15 labels of 150 words. In a large exhibit, it could take weeks to get all of the text translated. Because the wording of labels may be edited multiple times throughout development of an exhibit or program, that means that the translator will need to return to edit sections they already completed. If there are last minute edits to the original text, those changes need to be reflected in the translation and this could lead to delays in production if the translator is not given enough time to complete the changes and lead to a lesser quality of translation. (Avantpage 2018) Some translators also charge rush fees which can further increase the cost of the overall project.

With the limitations of human translators and the strain that translation can put on resources, finding solutions to these problems is imperative for museums to be able to offer materials and programs in multiple languages. Translation technology has existed for decades and continues to improve. It may offer a solution to many of the obstacles that prevent museums from presenting information in multiple languages to their guests.

Introduction to Machine Translation

While ideas of how to easily translate between two languages had existed for a long time, the first use of machine translation can be dated back to 1933 when French scientist George Artsrouni patented his idea for a mechanical translation device which he called the “mechanical brain.” (Sreelekha 2016) Artsrouni’s mechanical brain was a storage device on paper tape that was meant to help retrieve stored information quickly. One application that he pictured was producing word-for-word translations. He first successfully demonstrated the device in 1937. (Akbari 2014) Although the mechanical brain never caught on, it was an important first step in machine translation.

World War II spurred on translation technology research by the military and by the 1950s, multiple universities had begun research on mechanical translation including Massachusetts Institute of Technology (MIT), University of Washington Seattle, and University of California Los Angeles. MIT hosted the first conference on machine translation in 1952. The field continued to grow slowly from there over the next few decades as it was difficult and costly to research. In the late 1980s, IBM researchers published an article about statistical approaches that led to increasing the power of computers and the capacity of their memory as well as decreasing costs. This helped researchers develop a new “example based” approach to translation. This new statistical technique “translates the phrase and words based on earlier translated texts.” (Sreelekha 2016, 234-235) In the 1990s, the first online translators were introduced. (Sreelekha 2016, 235)

During the 21^st century, translation tools have continued to improve rapidly. The early online translation tools used the phrase-based statistical methods developed in the 1980s. Today, many of the most common machine translation tools–including the most popular tool Google Translate (Elliot 2022)–use an artificial intelligence system called neural machine translation (NMT). NMT has faster, more accurate translations than the previous types of machine translations. Rather than translating a word or phrase individually, NMT produces translations of whole sentences by pulling from data sets and language modeling to better understand the context of the whole text. This creates a more human-like quality to the translations.

Machine Translation, a Fix?

Machine translation tools are able to address some of the biggest limitations of human translators. For museums that have limited budgets or resources, machine translators are a cost saving measure. There are many free text machine translators such as the most popular one, Google Translate. Machine Translators can translate into hundreds of different languages and dialects, so museums are able to provide materials in many languages without raising the cost of translation.

Machine translation tools are also much quicker than human translators. Machine translation only takes a matter of seconds rather than days or weeks. Machine translation tools are available all day, every day which allows for last minute changes to be made to the text in exhibits or promotional materials without a delay.

With all of the benefits of machine translation, it would seem to be the ideal tool for museums, but one of the biggest limitations of machine translation is accuracy. While NMT has greatly improved the quality of translation compared with previous translation tools, it still does not provide the same understanding of language as a human translator. (Zhang 2022) NMT struggles with polysemy, or words that have multiple meanings or different parts of speech such as “present” as both a noun and a verb. (Zhang 2022, 179) NMT is trained through a limited corpus of language, so its understanding of the different variations of words is limited to those included in the training. It also struggles with expressions. (Zhang 2022, 179) Human translators generally summarize the meaning of the figure of speech rather than the individual words. NMT is unsuccessful in understanding and transforming the expression into an equivalent meaning in the target language. NMT is attempting to copy the human thinking process, but still has a long way to go. (Zhang 2022, 180) Presenting a poor-quality translation to visitors may give them an overall worse impression of the museum than if no translation was provided at all.

In addition to its inaccuracy, machine translation is only as good as the data it was trained on. NMT falls victim to a problem with many forms of artificial intelligence with limited data sets: bias. Nearly 30% of text on the internet is in English, so the tools have plenty of material to pull from for English translations. (Che Wai Peh 2023) Languages that are not used as frequently provide less data for machines to be able to pull from which can make the translations even more inaccurate. The translations may even be inappropriate or offensive.

These two issues balance out the benefits of machine translation. An inaccurate translation can have detrimental results. Without someone who speaks the target language to review the translations, museums risk insulting their visitors and making them feel unwelcome or unwanted which goes against the goals of introducing translated materials. However, this does not mean that machine translation should be totally eliminated. Museums can hire human translation editors to review the machine translated materials for less than the cost of a full translation which would help to address the major concerns of NMT. Another form of machine translation that is gaining popularity is computer-aided translation (CAT). CAT is a type of software that reduces the work of human translators. They include different tools that speed up the process like translation memory. Essentially translation memory pulls from the work that a human translator has already completed. Say a sentence or phrase is used multiple times in a project, the computer will save the translation and apply it to the later uses in the project. With time saved by the translator, money and resources are also saved by the museum. The discount can range from around 5% to 25% depending on the company and the repetitions used in the project. (Tarasiewicz 2023) At this point, machine translation alone is not the ideal solution, but it can help to reduce the overall costs for museums.

Beyond Text in Museums

While this paper has focused so far primarily on text translations because these are the most common forms of translations in museum, I want to address machine voice translation in live educational programs because this is a developing area for museums. (Ghazi 2022, 502) Text translations only address part the problem for LOTE-speaking visitors in the United States. If they want to be able to ask questions and engage with museum staff or take a live tour with a docent, they need to be able to listen and speak the language of the staff member or the staff member needs to speak the visitor’s language. For live programs, there are a number of emerging devices that promise simultaneous translation. (Tenement Museum 2023) While these devices may have a high initial cost, once a museum has invested in them, they do not have to pay the cost of hiring additional multilingual staff. In theory, any staff member can use the devices to speak to guests in a variety of languages.

The Tenement Museum in New York is currently utilizing the Ambassador Interpreter from Waverly Labs for this purpose. Their website describes the machine as “wearable, over-the-ear interpreter device that captures speech and provides both an audio translation and a written transcription in real time.” Visitors who wish to use the device can either borrow a Tenement Museum tablet and earpiece or download the Ambassador mobile app to use their own phones to connect via Bluetooth to the earpiece used by the educator. The device is able to interpret into 20 languages and 42 dialects. (Sieber 2021)

Unfortunately, the device does not seem to live up to all that is promised. According to reviews on the App Store, the translations are not entirely accurate and the differences in the dialects are negligible. (App Store 2023) The device requires slow, short sentences and clear enunciation for accurate translations which creates a stunted, unnatural speech pattern. The devices are only able to be used in a visitor’s right ear and cannot comfortably be used on the left which could be a problem for visitors with limited hearing in one ear.

The United States Capitol in Washington, DC is providing tours in Spanish and Mandarin using a Travis the Translator device. The Capitol offers the language tours once a day in the morning. The Capitol Visitor Center’s website does not provide as much information about their device as the Tenement Museum. It only says that “guides lead the tour with a translation device.”(US Capitol 2023) However, the Capitol website does include the information about the tours in English, Spanish, and Mandarin while the Tenement Museum only provides the information on their tours on their website in English. Unlike the Ambassador, it does not require a separate user device like a phone or tablet to be usable. The device is able to be used on its own and can be played out loud for visitors to listen or can be with the microphone and assisted listening devices the Capitol uses for its standard English tours. The device allows for translation to and from both the language of the guide and the language of the visitor, so visitors are able to ask questions while on tour.

In my experience as a tour guide at the Capitol Visitor Center, the tours I have led at the Capitol have had many of the same problems as the Tenement Museum. The Travis device is unable to keep up with human speech in real time unless the speaker uses short, simple sentences. The Travis device can take a few seconds to load and produce the target language which means the speaker must wait for the processing to complete before being able to continue speaking. The translations are not always accurate or lack nuance which leads to confusion among visitors. For example, when speaking about a statue of the Ponca leader Chief Standing Bear recently, the device translated his name as “jefe oso de pie” or “boss bear standing.” Visitors then asked where the bear was held in the Capitol. I do not speak fluent Spanish and have no knowledge of Mandarin, so I am unable to recognize when the translations are not correct unless a visitor communicated confusion. A Spanish speaking tour guide would have been able to better explain that the phrase was his name and title rather than referring to a literal bear.

Devices like the Ambassador Interpreter and Travis the Translator rely on the same NMT systems as text translations, so they are subject to the same problems and more. Additionally, voice translation devices require speech recognition and speech synthesis software to work. This leads to more systems that can produce errors. For example, the Travis device registered me saying “that man” as “Batman” on one of my tours and translated it as such. With text translations, museums are able to edit materials before presenting them to the public, but with live translations there is no quality assurance testing. Until the technology improves, hiring presenters who can speak the target language still seems to be the best option for live programs.

Conclusion

Translation technology can address some of the issues that acts as a barrier for museums, but the technology still needs to improve before it can replace the human element that language requires. The translation field continues to be improved every day, so it may not be long until it is able to equal human translations, but in the meantime, it can still be used as a tool to assist people. Computer Assisted Translation can help reduce the cost of text translations. Although voice translation devices may not be ideal for translating entire programs, they could be utilized for visitors to ask questions as they navigate museums. Until the technology improves, museums should continue to invest in making their content accessible to all.