Book review: "Terminology Extraction for Translation and Interpretation Made Easy"

A few months ago, I received a pre-release copy of this book as a courtesy from the author, terminologist Uwe Muegge, with a request to give a quick language check to the English used by its native German author. As I expected, there wasn't much to complain about, because he has lived in the US for a long time and taught at university there as well as been involved in important corporate roles. I was particularly pleased by his disciplined style of writing, the plain, consistent English of the text and the overall clarity of the presentation. Anyone with good basic English skills should have no difficulty understanding and applying the material.

At the time I read the draft, I was completely focused on language use and style, but I found his approach and suggestions interesting, so I looked forward to "field testing" and direct comparisons with my usual approach to terminology mining with that feature in memoQ. About a day's worth of tests shows very interesting potential for applying the ChatGPT section of the book and also made the context and relevance of the other two sections clearer, I will discuss those sections first before getting to the part that interests me the most.

Uwe presents three approaches:

• ChatGPT (https://chat.openai.com/)
• OneClick Terms (https://terms.sketchengine.eu/)
• Wordlist (https://www.webcorp.org.uk/live/wdlist.jsp)

I wouldn't really call these three approaches alternatives as the book does, because all three operate in very different ways and are fit for different purposes. That didn't register fully in my mind when I was in "editor mode", although the first part of the book made the differences, advantages and disadvantages clear enough, but as soon as I began using each of the sites, the differences were quite apparent as were the similarities to more familiar tools like memoQ's term extraction module.

Wordlist from Webcorp is functionally similar to Laurence Anthony's AntConc or memoQ's term extraction. It's essentially useful for getting frequency lists of words, but the inability to use my own stopword lists for filtering out uninteresting common vocabulary makes me prefer my accustomed desktop tools. However, the barriers to first acquaintance or use are lower than for AntConc or memoQ, so this would probably be a better classroom tool for introducing concepts of word frequencies and identifying possible useful terminology on that basis.

OneClick Terms was interesting to me mostly because friends and acquaintances in academia talk about Sketch Engine a lot. The results were similar to what I get with memoQ, including similar multiword "trash terms". I found the feature for term extraction from bilingual texts particularly interesting, and the fact that it can work well on the TMX files distributed by the Directorate-General for Translation (DGT) of the European Commission suggests that it could be an efficient tool for building glossaries to support translation referencing EU legislation, for example, though I expect only slight advantages over my usual routine with memoQ. These advantages are not worth the monthly subscription fee to me. However, for purposes of teaching and comparison, the inclusion of this platform in the book is helpful. I see more value for academic institutions and those rare large volume translation companies that do a lot of work with EU resources. 

ChatGPT was an interesting surprise. I have a very low opinion of its use as a writing tool (mediocre on its best day, clumsy and boring in nearly all its output) or for regex composition (hopelessly incompetent for what I need, and anything it does right for regex is newbie stuff for which I need no support). However, as a terminology research tool I have found excellent potential, though formatting the results can be problematic.

My testing was done with ChatGPT 3.5, not a Professional subscription with access to version 4.0. However, I am sorely tempted to try the subscription version to see if it is able to handle some formatting instructions (avoiding unnecessary capitalization) more efficiently. No matter how carefully I try to stipulate no default capitalization of the first letter of every expression, I inevitably have to repeat the instruction after a list of improperly capitalized candidate terms is created.

I keep an e-book copy of Uwe's book in the Kindle app on my laptop, so I can simply copy and paste his suggested prompts, then add whatever additional instructions I want.

The prompt

Please examine the text below carefully and list words or expressions which may be difficult to translate, but when writing the list, do not capitalize any words or expressions which don't require capitalization.

is too long, and only the part marked red is executed correctly, but this follow-up prompt will fix the capitalization in the list:

Please re-examine that text and this time when writing the list, do not capitalize any words or expressions which do not require capitalization.

Further tests involved suggesting translations for the expressions, with or without a translated text and building tables with example sentences:

Other prompt variations, for example to write terms bold in the example sentences, worked without complications.

What about the quality of the selections? Well, I used memoQ's term extraction module on the same text I submitted to ChatGPT for term extraction in order to compare something with which I am quite familiar with this new process. 

memoQ identified a few terms based on frequency, which ChatGPT ignored, but these were arguably terms that a qualified specialist would have known anyway. And ChatGPT did a superior job of selecting multi-word expressions with no "noise". It also selected some very relevant single-occurrence phrases which might be expected to arise more in later, similar texts.

Split screen review of memoQ extraction vs. ChatGPT results

The split-screenshot is an intermediate result from one of my many tests. The overlayed red box was intended to show a conversation partner the limits of ChatGPT's "alphabetizing skill", and the capitalization of the German is not correct after a prompt to correct the capitalization of adjectives misfired. It is not always trivial to get formatting exactly as I want it. However, looking at the results of each program side-by-side like this showed me that ChatGPT had in fact identified the nearly all the most relevant single words and phrases in my text. And for other texts with dates or citation formats, these were also collected by ChatGPT as "relevant terms", giving me an indication of what legislation I might want to use as reference documents and what auto-translation rules might also be helpful.

I also found that the split view as above helped me to work my way through the noise in the memoQ term candidate list much faster and make decisions about which terms to accept. The terms of interest found in memoQ but not selected by ChatGPT were few enough that I am not at all tempted to suggest people follow my traditional approach with the memoQ term extraction module and skip the work with ChatGPT.

My preferred approach would be to do a quick screening in ChatGPT, import the results into a provisional (?) term base and then, as time permits, use that resource in a memoQ term extraction to populate the target fields in the extraction grid. With those populated terms in place, I think the review of the remaining candidates would proceed much more efficiently.

All in all, I found Uwe's book to be a useful reference for teaching and for my personal work; it is one of the few texts I have seen on LLM use which is sober and modest enough in its claims that I was inspired to test them. The sale price is also well within anyone's means: about $10 for the e-book and $16 for the paperback on Amazon. For the "term curious" without access to professional grade tools, it's a great place to get started building better glossaries and for more seasoned wordworkers it offers interesting, probably useful suggestions.

The book is available HERE from Amazon.

Bring the lightning.

Quality is a slippery notion, especially when discussing it with those whose ethical approaches to providing services are even slipperier. According to one well-known figure in the trashlation sector, "Quality doesn't matter". Knowing that individual as I do, I know that this utterance was intended as a provocation, and that it is likely backed by some almost-persuasive sleight-of-hand involving differing definitions and whatnot. Given the variability in the human emotional perception of quality (as with obscenity, I cannot define quality, but I know it when I experience it), all of the attempts one sees to quantify it in language services seem all the more absurd.

All the myriad process definitions, ISO certifications, stamps and seals of sinlessness, diplomata, grants of honoris causa et cetera cannot transform the humble lightning bug into a Bolt of Zeus.

Nor are Large Language Models (LLMs) capable of such linguistic transubstantiation, but rather the opposite. The predictive practices at their core could take a training feed of all the world's great literature (and likely already have), and yet the output would be nothing more than an insipid averaging of the basest mediocrities. Only the basest of the mediocre could mistake such text for objectively good quality.

Were we to plot the degree of enthusiasm for AI as the "future" of trashlation against the degree of actual understanding and competence for good language, the graph would look something like this:

But a recent article in The Economist suggests a better way. Curiously, it is a process I resort to myself when the greatest subtlety and balance are needed in a work, for example in the translation of good poetry, or a letter of condolence occasioned by the loss of a belovèd child.

Back to pen on paper. Where the pressure of the nib is an expression in itself, as the sweeping flourish of a final letter or a well-executed ligature.

"But that's ABSURD!!!" some might protest, glancing nervously at their smartphone timers counting down to the next due delivery of linguistic sausage. Much too slow some might think. But is it? Really?

"But you need to run QA and you can't do that with a sheet of scribbles on paper!" some might suggest, more reasonably. Ah, but I can, merely dictate the text I will have read aloud already time and again as I refined the words and their rhythm, and then, in good electronic form, all the slings and arrows of outrageous regex are my quality arsenal.

We have a slow food movement. Perhaps if we want more delicious, digestible, properly communicative words in our translated lives, we should slow the fuck down and let them crystallize, with exquisite subconscious fractal creativity, to form bolts of emotion and understanding that pierce the veil between this world and others as they flash across a page.

As the morlocks cower in their caves and hovels, tapping tiny tablets in their claws, prompting their artificial gods to take this terror of meaning from their shriveled world.