A Review of: I Can’t Place the Accent: Identifying the Characteristics Traits of Computer Translation, presented by Eugenia Tietz-Sokolskaya, CT and John Riedl, CT on Friday, November 1 at ATA65

Review by Christine Pawlowski; slides from the presentation available in the Slavic Languages Presentation Archive

I do not feel threatened by AI,  and this is not because I think my language skills are superior to the machine. Perhaps it is mostly that my monthly social security check allows me to lighten my workload to a manageable volume. And maybe it is also that I am technologically challenged (being provided with a modem to do my work for FBIS in the 90’s was a huge step).

In my very limited experience with AI projects, I have enjoyed “beating” the machine, as when the AI translation of the word “wygodny” in Polish, which may be translated variously as “convenient” or “comfortable,” resulted in an English version of an apartment advertisement that sported a comfortable bedroom armoire—perhaps a magician’s prop?

Well before the October/November ATA65 Conference, we received a survey from Eugenia and John in which we were asked to select the best translation of several Russian texts, but we were not told who (or what) did the translation. When we arrived at the session, we learned that the translations had been done by Deepl, an NMT (Neural Machine Translation) service launched in 2017; ChatGPT 3.5, an LLM (Large Language Model) service launched in 2022; and a few different humans.

Given its timeliness, it is not surprising that generative AI figured prominently in many of the conference sessions. John and Eugenia’s session dovetailed beautifully with the subject of Holly Mikkelson’s Wednesday training for ATA graders: Prescriptive and Descriptive Language. In a nutshell, we investigated how we really speak.

In both Holly’s presentation and that of our Slavists, we looked for the “tell”—a clue or indication that reveals information or suggests a hidden truth. All translators—human or generative AI–have these tells. To find them, we looked to cohesion, fluency, syntax and terminology.

From the survey results and our on-site bantering about some of the linguistic conundrums, we learned that:

• ChatGPT’s renditions will be grammatical and flow deceptively well but may not be accurate.
• DeepL is easier to peg as a computer translation.
• Human translators take liberties, which can be a blessing or a curse (hence the dangers of prescriptivism and the difficulties encountered in evaluating translation).

The good news disseminated by Eugenia and John is that humans can achieve higher quality by:

• recasting or rewriting clauses
• splitting or combining sentences
• choosing subject-appropriate terms
• substituting phrases for words and vice versa

This list suggests that skills in manipulating syntax are critical. The bad news is that humans also misspell words, misuse collocations and struggle with job fatigue.

Comparatively speaking, for the three passages we studied in the session, in every case ChatGPT came out on top of DeepL, which sticks very close to the original syntax, even to the point of unreadability. In two of the three examples, the human translation won.

Discussions of AI are ubiquitous—on Linkedin, in journals and magazines. An opinion piece in the latest edition of the journal First Things offers the suggestion that society can resist the techno-tyranny trend, which is making us miserable, by demanding human-to-human businesses because “People…will pay for happiness.”  There is some nostalgia for the way things were before the modem and the ease of searching the web for the contextually right word—but not much, in my opinion.

Christine Pawlowski is a freelance Polish and Russian translator with an M.A. in Slavic Languages and Literatures from Indiana University, Bloomington. She is retired from teaching elementary-school music and delights in being Busia to her 17 grandchildren and in directing and accompanying her church choir. She is ATA-certified for Polish-into-English and an ATA Certification grader for that language pair. She may be contacted at pawlow@verizon.net.