What Is Machine Translation?

Machine translation (MT) is the use of software to automatically translate text from one language to another (Qun & Xiaojun, 2015, p. 105). MT has evolved substantially since its debut in the mid-20th century, with particularly rapid changes in the past decade. The latest iterations of MT rely on machine learning, making these versions faster, more efficient, and able to produce more accurate translations for certain languages (Kelleher, 2019; Lewis-Kraus, 2016; Poibeau, 2017; Wu et al., 2016). Indeed, it is increasingly difficult for language instructors to identify MT usage based on errors, one of the tell-tale signs of MT usage with past, less accurate versions of MT (Briggs, 2018; Ducar & Schocket, 2018; Stapleton & Kin, 2019).

How Does Machine Translation Work? What are the Ramifications for Language?

Machine learning relies on vast amounts of training data, composed of up to millions of source texts that are aligned at the sentence level (Pérez-Ortiz, Forcada, & Sánchez-Martínez, 2022). Much of this training data originated in 1) translation memories created by human translators (Kenny, 2022), and 2) publicly-available multilingual documents (Koehn, 2005).

While “high-resourced” languages (e.g., English, German, Spanish) have sufficient training data to ensure relatively accurate machine translations, many languages that are considered “low resource,” such as indigenous languages (see Hilleary, 2021), do not have sufficient training data and, as such, are not available on MT platforms or produce less accurate results. To respond to these problems with “low resource” languages, researchers have started to pre-train neural network models using monolingual texts (Hardach, 2021), leading, for example, to the recent inclusion of Quechua on Google Translate (Taj, 2022).

In addition to disparities in MT accuracy across languages, researchers highlight other limitations, including algorithmic biases, such as gender bias (e.g., the over-use of male forms in output and lack of non-binary options [Savoldi et al., 2021]), fluent but inaccurate translations (Kenny, 2022), and the loss of lexical diversity and variety (Vanmassenhove et al., 2019). In light of these risks, Bowker and Buitrago Ciro (2019) have advocated for the development of a new competence, that of machine translation literacy, which involves an awareness of how MT systems work, attention to how they are used, and an ability to evaluate and modify their outputs.

How Has Translation Been Understood in Language Education?

Translation—both traditional and machine—has been controversial in language education. The grammar-translation method, the dominant pedagogical approach well into the 1970s, relied on explicit explanations of discrete grammar rules in the students’ primary language and their application in the translation of single, invented, and generally decontextualized sentences. Even though this approach represents a very narrow understanding of translation, translation became all-but taboo with the rise in the 1950s and 60s of the direct method, which emphasized the development of speaking skills and the use of the target language for instruction (see Cook, 2010, for a detailed overview of this history). The recent translational turn—which emphasizes broader conceptualizations of translation as meaning making (see Vinall & Hellmich, 2022)—has resulted in a resurgence in the use of pedagogical translation, i.e., translation exercises designed to support the development of language skills. (For more on this topic, see Vinall & Hellmich, 2022, and Cook, 2010.)

How Do Language Instructors Perceive Machine Translation?

These different tides in the field’s embrace (or lack thereof) of translation have influenced instructors’ perceptions of machine translation. Indeed, research suggests that language instructors have mixed and often conflicting feelings about MT tools. Historically, one highly debated topic is whether MT constitutes academic dishonesty, including cheating and plagiarism. Most often, instructors tie the acceptability of MT with how it is used. For instance, instructors are more likely to approve of MT for shorter texts (Case, 2015; Jolley & Maimone, 2015), ungraded assignments (Clifford et al., 2013; Jolley & Maimone, 2015), and activities that do not involve writing (Hellmich & Vinall, 2021), as these uses are less likely to be seen as interfering with the development and measurement of language proficiency.

Overall, integration of MT into classroom practices has been reported to be limited (Briggs, 2018; Hellmich & Vinall, 2021; Niño, 2009). One reason for this limited integration is a lack of knowledge on the part of instructors on how to use these tools (Delorme Benites et al., 2021), suggesting a greater need to include training in teacher development programs (Zhu, 2020). That said, many instructors believe that MT can and perhaps even should be discussed in the language classroom given its increasing prevalence and ability (Case, 2015; Clifford et al., 2013; Hellmich & Vinall, 2023; Jolley & Maimone, 2015; Vinall & Hellmich, 2021).

Intuitively, the extent to which instructors integrate MT into language learning often relates to their perceptions of it. One study that looked at the metaphors used by instructors to describe MT revealed that when perceived as a “crutch,” MT is deemed destructive to learning and its use banned; when perceived as a “bridge,” MT’s role is supportive, and it is often used in tandem with the instructor to engage students in exploring meaning making processes; and when perceived as a “prosthetic,” MT can transform classroom instruction, from a focus on accuracy to one that highlights interaction and communication (Vinall & Hellmich, 2021).

How Do Students Perceive and Use Machine Translation Tools?

Overall, language learners tend to view MT tools as helpful (Garcia & Pena, 2011; Lee, 2020; Niño, 2009; Tsai, 2019) and report frequent use of MT tools (Bourdais & Guichon, 2020; Clifford et al., 2013; Jin & Deifell, 2013; Jolley & Maimone, 2015; Larson-Guenette, 2013; O’Neill, 2019; White & Heidrich, 2013). Language learners have reported relying on MT for a range of purposes, such as looking up words (e.g., vocabulary) (Clifford et al., 2013; Larson-Guenette, 2013; O’Neill, 2019; White & Heidrich, 2013) and double checking their work or instincts (Clifford et al., 2013; Jolley & Maimone, 2015).

A few studies have used direct observation methods to explore exactly how learners use MT tools (Deifell, 2018; Garcia & Pena, 2011; Hellmich, 2021; Hellmich & Vinall, 2023; Tight, 2017; Vinall et al., 2023). Hellmich (2021), Hellmich and Vinall (2023), and Vinall et al. (2023) observed 74 beginning learners of Chinese, French, and Spanish as they completed a writing task in the target language. The learners frequently used MT tools, but they also used these tools in complex ways that surpass common instructor assumptions. For instance, student participants analyzed the results of the MT tools more than half of the time, rather than simply copy-pasting them into their writing document. They also demonstrated advanced analysis strategies, such as using multiple tools to triangulate MT tool search results, drawing on their own metalinguistic knowledge, and analyzing examples to identify the most appropriate translation for their needs.

MT, when used critically, has been shown to support student writing, although some studies have suggested that these benefits can vary depending on proficiency level (e.g., Chung, 2020; Chung & Ahn, 2021; Lee, 2022). When trained on how to use MT thoughtfully, learners’ written texts can improve–for instance, fewer errors and more linguistic diversity, among other things (Fredholm, 2019; Tsai, 2019; Lee, 2020, 2022; O’Neill 2019), although these benefits disappear if MT training is not refreshed (Fredholm, 2019; O’Neill, 2019). In addition, work done by Thue Vold (2018) suggests that using MT can help learners develop metalinguistic awareness.

How Can MT be Addressed in the Language Classroom?

As a result of the often negative perceptions of MT, language instructors often enact policies that restrict its usage. Common policies include banning MT altogether or limiting MT to the word-level (“using MT like a dictionary”) (Hellmich & Vinall, 2023). Unfortunately, these policies can have unintended negative consequences. In the case of banning MT, research demonstrates that students are already using MT. Banning a ubiquitous tool, rather than teaching students how to use it effectively, can create distance between teacher and students as well as reduce learning opportunities (Vinall & Hellmich, 2021; Hellmich & Vinall, 2023).

In the case of encouraging students to use MT at the word-level, such policies can lead students to more inaccuracies: current machine learning algorithms require longer strings of input in order to produce more accurate translations, meaning that using MT “as a dictionary” does not offer MT tools the input necessary to work at its best. For instance, an MT program will not be able to distinguish which version of the English word “fly” is appropriate for the intended situation unless more input is provided (e.g., “the fly,” “to fly,” “a fly on the wall.”) (Hellmich & Vinall, 2023).

Some pedagogical approaches to MT have focused on pre/post editing. The lineage of these approaches is in translation studies, when translators pre-edit texts to be translated to best suit translation software and/or post-edit texts that are produced by MT. For language learners and writers, pre-editing involves identifying and simplifying/removing potential ambiguities (e.g., cultural references, linguistic structures) in the source text in order to reduce the possibilities of MT-produced errors (Sánchez-Gijón & Kenny, 2022). Post-editing involves identifying and correcting raw MT output (see Niño, 2008). Much of the recent work in this area has focused on language learners’ error analysis (e.g., Chung, 2020; Kol et al., 2018). That said, the pre/post distinction does not always capture the dynamic process of composing with the support of online tools, where these distinctions are often fuzzy and overlapping (Vinall et al., 2023).

Another pedagogical approach looks to teach students to critically engage with MT as a resource in making meaning. This approach begins at the level of the tool itself, making space for collective conversations and analysis of MT tools as well as other online tools available to learners (Hellmich & Vinall, 2023; Klekovkina & Denié-Higney, 2022; Vinall & Hellmich, 2021). In a classroom context, for example, instructors might ask students to investigate different tools (e.g., online dictionaries, machine translation platforms, parallel corpora) and discuss collectively what each tool has to offer (see Activity 1 below).

Supporting students to engage with MT tools critically also involves teaching them to think about what goes into MT tools (input) and how they assess what comes out (output). For instance, instructors can encourage students to collect examples of how different inputs render different results (see Activity 2 below). They can also teach students to analyze the output using specific strategies, such as triangulating the results of MT tools with other online resources or, as Ryu and colleagues (2022) demonstrate, with resources such as a Google Image search.

Finally, much of the consternation caused by MT tools can be avoided when the objectives of language learning and language learning tasks shift from accuracy to meaning-making. That is to say, when students are encouraged to be creative, agentive users of language, MT is repositioned as a means to an end, rather than an end itself. For instance, Klekovina and Denié-Higney (2022) restructured a unit on 18th century French theater to focus on a debate wherein students argued their views, making MT a tool but not a central component of the learning activity.

What’s Next?

There is a lot we do not yet know about machine translation and language learning. First and foremost, more research is needed on the impact of different approaches to using MT in language education, particularly teaching students to use it critically. This kind of research includes how these approaches impact tool usage, text creation and interpretation, language learning, and how to best support instructors in navigating MT tools in the language classroom.

In addition, more research is needed on MT and language education outside writing. To date, research on MT and language education has focused on its use to support the development of writing skills. However, technological advances, such as screen readers that can instantly translate text and speech recognition used in voice-based machine translation, suggest the need for further inquiry into the potential of MT use to support listening and reading comprehension and speaking skills. There are also potential bridges to be built between MT/language education and broader fields and movements in education, such as with translation studies or digital literacies.

While an exact path is hard to predict, it is clear that continued advances in technology will make MT an important topic of research and practice in language education in the years to come. In particular, the importance of MT in world language education research and practice will only continue, given the emergence of genAI (for more on genAI and MT, see Hellmich & Vinall, 2024).