Categorization of Translation Mistakes Made by Artificial Intelligence

Henry Whittlesey Schroeder

Contents

Introduction

1. Objective flaws

1.1 Terminology

1.2 Harmonization

1.2.1 Individual terms

1.2.2 Orthography

1.2.3 Intersegment collocation

1.3 Interpretation (obvious and concealed)

1.3.1 Segmentation

1.3.2 Abstraction

1.3.3. Blatant misinterpretation

1.3.4 Misinterpretation due to ambiguity

1.4 Detail/Precision

2. Subjective or context-based flaws

2.1 Terminology

2.2 Harmonization

2.3 Style

2.4 Tense

2.4.1 The present tense with a future meaning in German

2.4.2 The case of sollten

2.5 Omissions

2.6 Context

Conclusion


Introduction

After broaching the primary artificial intelligence issues for translation in the first article of this series, I will delve into the weaknesses at greater length here. The initial article contains a brief description of four categories of errors: style, harmonization, terminology and interpretation of abstractions/references. Those types of mistakes are critical for the stakeholders in the translation process (client/customer, language service provider, and translator). In this article, I will address these in greater depth, while also adding a few other types of flaws.

To facilitate a quick read of this article, I have broken down the types of errors into two categories, each with their own set of mistakes. While machines or algorithms produce much better translation results than computer-based iterations prior to 2017, they need to be reviewed by human translators. This review is handled more efficiently if the revisor or editor knows the types of mistakes made.

The flaws in AI translations are divided into the overarching categories of (i) objective and (ii) subjective or context-based errors. The various types of errors will then be assigned to each category. One of the fascinating aspects of AI translations and the reason for AI’s success in both translation and many other fields is the sharp reduction in the number of objective mistakes it makes. Consequently, the number of error types found under objective errors is substantially shorter than those identified as subjective or context-based. Some types of errors appear in both categories. For example, an interpretation error may be objective in one case (e.g., if the abstraction or reference is misunderstood), but can also be subjective in other instances because the context suggests that x is better than y in a given circumstance.

Errors of an objective sort consist of terminology, harmonization, interpretation (obvious and concealed), detail/precision.

Subjective flaws are found in style, tense, context errors, and the overlapping types of terminology and harmonization.

It is also important to understand in this analysis of AI translations that I am judging the results of AI translation on both the micro, i.e., single sentence, level and on the macro, i.e., full text, level. This means that I am referring to mistakes, errors and flaws in both categories and all types even if a given sentence, when viewed in isolation, is translated correctly. For example, a pronoun might potentially refer to two different words in the preceding sentence, with the AI translation picking the wrong reference. This single sentence is only incorrect when viewed in relation to the preceding one. The same applies to divergent translations of a single source word across multiple sentences. Such cases are regarded as flaws in this article.

The examples will be taken exclusively from German texts translated into English. For the most part, I do believe that these examples are representative for other language pairs. As a translator of Russian as well, I see the relevance of most points to Russian-English texts.

1. Objective flaws

Similar to what are now referred to as human translations, the algorithms and machines of artificial intelligence make objective mistakes in the generation of a translation. A comparison of the original with the translation quickly reveals the need for adjustments and touch-ups at a bare minimum.

The errors are different, however. On balance, they are less frequent and/or easier to identify for the most part.

An objective error is defined as a translation or translation-related decision that cannot be plausibly defended in the event of a complaint by the author or client of the source text.

1.1 Terminology

Objective terminological mistakes are exceedingly infrequent and found far less than in human translations. Words that resemble each other, such as Markt and Marke in German, but have entirely different meanings (market vs. brand) are not misread. Besides the harmonization of terminology and subjective terminological errors, both of which will be addressed below, AI translations have the greatest difficulty with effectively untranslatable terms in a source language.

An example of this might be Eignungsleihe in German tender agreements. It means that a company will rely on the resources and competence of another company that effectively joins the tender process and guarantees to handle the assigned duties if the contract is awarded. It differs from outsourcing in that the company to which part of a contract is outsourced is not part of the bid. In the current iteration of AI translations, such untranslatable terms appear in the target language in a haphazard manner. Some target language results are transpositions, i.e., a rough equivalent in the target language (often off-base); others look like guesses, which are generally obvious because the suggestion makes no sense in the context.

It is also possible for a common word to be used in an odd way, e.g., as a name for a real-life phenomenon, in one language, but not in another. An example of this terminological issue was encountered recently in a story about a constellation in the sky. In German, the constellation has two names, the Greek original and its translation. In English, however, the translation of the Greek word is not commonly used for this constellation. Multiple AI translation providers rendered the German name as its English equivalent, the reptile lizard, although the constellation is almost always called by its Greek name Lacerta, not the English translation lizard:

Original: Die Eidechse ist ein unscheinbares Sternbild am Nordhimmel, das wohl nur wenige Menschen jemals zu sehen bekommen.

Wrong AI (in 2 AI systems): The lizard is an inconspicuous constellation in the northern sky that few people ever get to see.

Correct: Lacerta is an inconspicuous constellation in the northern sky that few people ever get to see.

This type of mistake can also happen in reverse. Livland in German, a region to the east of the Baltic Sea referred to as Livonia in English, was not translated by the AI algorithm. Again, these types of mistakes are uncommon, but they reveal the need for revision and proofreading.

The most common terminological mistakes made by AI translation are subjective or context-based. From an objective perspective, they are not wrong. In isolation, that is, if the sentence were viewed out of context or the text solely consisted of one individual sentence, the AI translation could be deemed correct. In the context, however, the generated translation is incorrect. There is a particularly gray zone here with regard to terminology that has an effectively official translation due to laws, norms or standards adopted by an official body like the EU, European Commission or similar. For instance, the term löschen and other terms in the context of data have basically official translations as a result of the EU General Data Protection Act. Machines repeatedly translate löschen as delete because the word delete is by far the most common way to refer to the Löschen of data electronically. However, the GDPR consistently uses the verb erase, without delete appearing even once as an alternative. As with the compound noun data protection itself, however, there are other competing norms, standards and legislation such as the California Privacy Act, which prefers the term privacy to data protection. I have included this terminological issue in the subjective category.

These subjective cases are analyzed in section 2.1.

1.2 Harmonization

Similar to the terminological issues with AI translation, the harmonization of target texts can be broken down into objective and subjective errors. The line of demarcation is fuzzy and depends on the language preferences of the author(s), but some rules of thumb seem to apply, especially regarding consistency: Critical terms in a text should be translated uniformly, the orthography of one language should be used (if applicable) and preferred practices must be applied commensurately throughout the text, especially if they have legal relevance or are to be understood in contrast to other specifically defined terms.

1.2.1 Individual terms

While some AI translation programs allow the integration of glossaries or termbases, it is more common to use pure algorithmic generation. Languages feature words that can often be rendered divergently in a given target language. The German word Kunde, for example, equates to client or customer. Geschäftsjahr is conveyed as financial year (BE – British English) or fiscal year (AE – American English).

In the translation of a contract, the machine will tend to correctly and consistently translate Kunde as, say, client for a number of sentences, especially if it repeatedly appears in most sentences of a section of text. However, when the subject matter of the contract, in our example, shifts to liability or working hours or obligations of the counterparty, without mention of the Kunde/client, the machine translation often produces the equally correct, but unharmonized alternative term, in our case, customer. The same applies in annual reports and financial statements with the term Geschäftsjahr, especially if the target language is US English. Fiscal year is used almost universally in the United States, whereas financial year is preferred for UK and international standard English, and is found in numerous EU English documents, norms and standards (IFRS, IAS, etc.). The machine translation jumps back and forth.

1.2.2 Orthography

Most AI programs allow you to choose a target language country or dialect (e.g., UK vs. US English). Accuracy is definitively above 99.9%. The same applies to spelling. Capitalization, however, can be a problem. If we continue with our example of client or customer in contracts, the term will generally be capitalized throughout a contract (e.g., “hereinafter referred to as (the) Client”) if the client or customer is a party to the contract. AI translations may capitalize the term for a while, but will eventually lose the thread similar to what happens with the harmonization of the translation of Kunde or Geschäftsjahr.

In all likelihood you can (or will soon be able to) instruct generative AI to capitalize every instance of a word. This will probably apply to most harmonization needs. For the time being, it must be done manually by a person. It is also necessary (and will be in the future) to review each instance of harmonization because a term might also appear in two (or more) senses. For example, the term customer/client might be used divergently in a contract – for a party to the contract (most of the time) and for other customers (of the contractual customer/client or the other contracting party). This will always apply, even with instructive generative AI.

Practice is also necessary to harmonize. To take the example of Eignungsleihe from above, a posteditor must determine whether the German term will be used in the running text (with the paraphrased explanation appearing in parenthesis afterwards) or vice-versa. If the term appears multiple times and the explanation is long, it may make more sense to use the German term throughout and only explain it in the first instance. Should the untranslatable word appear only once, the readability of the paraphrase could be preferrable. Human translators do this almost instinctually, making a decision based on experience or the first few instances and consistently repeating it throughout the translation.

This type of decision must also be reached with regard to laws. Human translators adopt a wide range of practices for the treatment of source language laws in a target language. AI translation does the same. The variables are compounded by the tendency to abbreviate laws in the source language. The type of text also influences the approach. A law in a contract, especially abbreviated multiple times, may be handled differently than one mentioned sporadically in a report on migrant assimilation, for example.

1.2.3 Intersegment collocation

Across segments, usually sentences, it may not be possible for the literal translation to be correct. Here is a marketing example where a specific German word (Beachtung) harbors another German word (achten) that is then extrapolated and built on in the final sentence:

Segment 1: Wir haben entscheidenden Einfluss darauf, welchen Dingen wir im Alltag Beachtung schenken. Segment 2: Und in dem Wort Beachtung steckt auch schon das Wort ‚achten‘ drin. Segment 3: Es gibt Dinge, auf die wir alle gemeinsam achten, weil sie uns gemeinsam wichtig sind.

AI translation: Segment 1: We have a decisive influence on what we pay attention to in our daily lives. And the word “attention” already contains the word "respect". Segment 3: There are things that we all collectively respect because they are collectively important to us.

This play on words in German with Beachtung/Achten does not work in English. The machine opts for a literal translation, in which each segment is correct in and of itself, but not as a sequential buildup (what I am calling intersegment collocation). We need creativity here.

Human translation: Segment 1: We have a decisive influence on what we pay attention to in our daily lives. Segment 2: And the word “attention” also implies “respect”. Segment 3: There are things that we all collectively pay attention to or respect because they are collectively important to us.

The problem with machine translation in this case is analogous to the harmonization weaknesses with terminology: Viewed in isolation, each sentence would be correct as each sentence with divergent translations of Kunde as client and customer is correct. Between segments or within the text as a whole, however, the translation is objectively flawed.

1.3 Interpretation (obvious and concealed)

1.3.1 Segmentation

Most translations are still processed with a CAT tool, with the machine translation integrated into the tool. The postediting is done here as well, which basically breaks up the text into sentences, called segments, generally with the original language on the left side and the translation on the right. Periods and colons are used by the tool to determine the end of a sentence. Automated segmentation on this basis causes problems when an abbreviation with periods appears in the original or when a colon introduces a long list of possibilities for continuing a sentence (e.g., the event organizer is responsible for: 1. xxx 2. yyy 3. zzz). Although CAT segments can often be joined or separated to correct the preformatted version, this is not always possible. In any case, segments fragmented as a result confuse the algorithms.

1.3.2 Abstraction

The misinterpretation of abstractions and references accounts for one of the more frequent errors made by machines. While I can only surmise what the reasons for this are, it is likely that the machines are guessing on the basis of surrounding context in some cases and, even more interestingly, adopting a position in regard to the original.

German authors like to create diversity in their texts by abstracting from the concrete case. A simple form of this is seen in the use of generic or umbrella terms preceding names or titles. For example, the descriptive term Bereich (area/segment/division, etc.) appears repeatedly for categories that would often simply be named in English with a description of their type. Here is an example:

Unseren Fokus richten wir auf das Zusammenwirken von Finanzierungs- und personellen Versorgungsengpässen bzw. -lücken in den Bereichen Rente, Pensionen, Steuereinnahmen, Pflege, Gesundheit, Bildung, öffentliche Sicherheit und Wohnen.

We have focused on the interplay between financing and staff shortages or gaps in benefits connected with the areas of retirement, pensions, tax income, nursing care, health, education, public safety and housing.

Whereas an American writer would often simply say “connected with retirement, pensions, tax income…” without a template, Germans designate the type of category before the name or specification. The text cited above then follows this passage by referring to these Bereich (areas) as Systeme (systems) for whatever reason. In this case the translation is easily understandable. While less than ideal, in my opinion, the reference is clear. The analogous terms in English can be used, with the reader understanding that retirement, pensions, tax income… are both areas and systems. Yet, such shifting of generic terms or the invention of more creative abstractions can cause the machine to confuse the reference and interpret the passage in relation to something else in the text. Humans do this as well, to be fair. It is a common issue with abstractions.

Working with machines also becomes interesting and funny when they seem to be correcting the original text. I recently encountered a text with these sentences:

Ich danke ihnen ganz herzlich dafür! Sie tragen so dazu bei, das Leben von Menschen mit sogenannter geistiger und mehrfacher Behinderung ins Bewusstsein zu rücken, Öffentlichkeit herzustellen und nachhaltig zu verbessern.

Machine translation: I thank them so much! You are helping to raise awareness, generate publicity and make a lasting difference in the lives of people with what are called intellectual and multiple disabilities.

The you (underlined) is incorrect – it should be they because it refers to them in the preceding sentence (these are organizations helping people with disabilities). It is thoroughly possible that this is an innocent mistake made by the machine. Because the second sentence begins with the pronoun referring to them, it is capitalized, which means it can be they or you in the absence of context. However, the context is readily apparent here from the preceding sentence. Neural or generative AI translation generally realizes this. It uses surrounding context to make decisions on more complex issues than the alignment of these pronouns in simple sentences. My guess is that the algorithm felt the speaker should be addressing these organizations directly in the you form. It rarely proposes something that is overtly false, like translating ihnen as you, but the moment it encounters ambiguity with Sie, it takes the opportunity to propose a better way of expressing the thanks here. Whether right or wrong, the posteditor must step in here and write they in case the line-by-line review by the customer raises questions about quality on the basis of pronoun reference mistakes such as this.

Another possibility when working outside of a CAT tool is that a chatbot will restructure paragraphs. In the preparation of a draft translation of a paper that I had pasted into chatgpt, the translation contained a different paragraph structure. Such autonomous actions can be resolved with appropriate instructions in the prompt.

1.3.3. Blatant misinterpretation

AI interpretations are phenomenally good. In terms of objective misinterpretations relative to human translators, I would estimate a 90% drop in error rates (based on my experience revising translations before artificial intelligence and postediting now).

In the example below, a client has signed a contract with a contractor. The client pays for the services of the contractor.

Soweit die von der Aussetzung betroffenen Leistungen für mehr als 10 Stunden im Zeitraum ausgesetzt sind, schuldet der Auftraggeber auch keine Vergütung für die ausgesetzte Leistung.

Wrong AI translation: If the services affected by the suspension cannot be used for more than 10 hours in the period, the Client is also not entitled to remuneration for the suspended service.

Correct: If the services affected by the suspension cannot be used for more than 10 hours in the period, the Client shall also not owe any remuneration for the suspended service.

The algorithm reversed the recipient of the payment. To be fair, this type of mistake is very rare. To my non- but near-native knowledge, the German text is not ambiguous. Ambiguity accounts for misinterpretation far more often, as we shall see in the next section on subjective flaws. In the example above, the text as a whole makes it clear who is making and who is receiving payments for the services. The sentence itself explains this as well.

An AI translation is more likely to blatantly misinterpret, understood in a broad sense, by producing an imprecise translation in a case that requires precision. Generally speaking, this falls within the realm of context-related errors. The following example shows a mistake in a report on a company’s ownership relations:

XXX GmbH gehört mehrheitlich zur YYY Gruppe.

Wrong AI: XXX GmbH is now mainly owned by the YYY Group.

Correct: XXX GmbH is now majority owned by the YYY Group.

The seemingly minor change of mainly to majority is critical in a financial context. The German text states that the YYY Group holds more than 50% of the shares in XXX GmbH. This results in numerous tax, reporting and other requirements for both companies. The word mainly is neither defined or precisely understandable in the financial sector. A reader would look at the wording mainly owned and be unsure what to make of it: Is it the same as majority owned? Does the author want to say that the YYY Group is building up a stake in XXX GmbH? It is simply unclear, while the very conventional majority owned expresses the German idea without uncertainty.

Such mistakes in the interpretation of details are differentiated from “simple” objective mistakes with details. Although this error with mainly/majority can be classified as objective in the financial context, there are cases (e.g., with less technical texts) where mainly could serve as the translation. The chatbot has interpreted the given case here incorrectly. As we will see below (section 1.4), mistakes with details go well beyond misinterpretation (of context or appropriate language).

Here is one that was surprising:

Zeitlich leicht versetzt zur Pandemie begann auch in Deutschland die Inflation zu steigen.

Wrong AI: Coinciding with the pandemic, inflation also began to rise in Germany.

Correct: Shortly after the pandemic, inflation began to rise in Germany.

1.3.4 Misinterpretation due to ambiguity

As we saw with abstraction, both humans and artificial intelligence have difficulty handling ambiguity. Prior to algorithmic translation, the required efficiency of translation and the numerous considerations in each sentence left translators susceptible to misreadings. Artificial intelligence understands context much better than it did in the early- to mid-teens (2010 to 2017), but it still fails to reach the level of humans.

A sentence in German can be formulated with two different meanings, both of which could be correct in a given context. Here is a case:

Die Leistung ist so auszugestalten, dass die Daten des Auftraggebers entweder zu jeder Zeit selbstständig durch den Auftraggeber oder, soweit dies aus technischen Gründen nicht möglich ist, mit Unterstützung durch den Auftragnehmer aus der Cloud exportiert werden können.

The service must be designed in such a way that the Client can export the data from the cloud at any time independently or, if this is not technically possible, with the help of the Contractor.

Correct:

The service must be designed in such a way that the Client’s data can be exported from the cloud at any time independently by the Client or, if this is not technically possible, with the help of the Contractor.

The algorithm views the information out of context. This may in part be due to the fact that we are often still at the stage of entering fragments of entire texts into AI platforms. On the ones I know, caps are placed on the number of characters that can be entered into a machine for translation. The blocks of entered text obviously provide some context, but not the amount potentially necessary to facilitate an algorithm’s correct attribution of the data exported, as in the example above. In the event of uncertainty, algorithms guess.

Here is another example where the AI translation misinterprets the perspective of the text, confusing the addressor and the addressee:

Nur mittels transparenten Einzelanlagen wird es für die Depotinhaber/innen überhaupt möglich sein, eine stringente Darstellung der Anlagen nach Anlagekategorien, Ländern, Branchen und Währungen vorzunehmen und ihrer Kontrollfunktion nachzukommen.

Only through transparent individual investments will portfolio holders be able to provide a consistent view of investments by asset class, country, industry and currency and fulfil their oversight function.

Correct: Only transparent individual investments make it possible for portfolio holders to have a consistent view of investments by asset class, country, sector and currency, and let them fulfil their oversight function.

The critical mistake here is that AI assumes that the “portfolio holder” is the actor or addressor pursuing the “transparent individual investments.” In fact, the text was discussing how the company servicing the portfolio holder (für die Depotinhaber/innen) ensures that the investments are transparent so the portfolio holder understands the structure and can review it. The purport of the two sentences is completely contrary to each other.

A catastrophe occurs in cases like this analysis of portfolio performance:

Folglich dürfte ein Depot, das weniger als hälftig in Aktien investiert ist, rund 20% verloren haben.

Consequently, a portfolio that is invested in equities for less than half of the time is likely to have lost around 20%.

Correct: Consequently, a securities account with less than half of its assets invested in equities should have lost around 20%.

Here is another catastrophe:

Dies lohnt sich, da wir meist Fehler zu Ungunsten der Kunden und Kundinnen finden.

This is worthwhile, as we often uncover errors in our clients' favor.

Correct: This is worthwhile, as we often uncover errors to the disadvantage of our clients.

The authors of the text are trying to point out how their service benefits their clients. One aspect of this service is to check whether custodian banks are overcharging the clients whose securities portfolios the authors of the text are managing. The custodian banks overcharge clients; the investment managers discover this in an annual review and get the incorrect charges reimbursed. That is the exact reverse of what the AI translation produces.

1.4 Detail/Precision

In section 1.3.3 we looked at the impression of mainly/majority owned as an example of a blatant interpretive mistake. Besides making objective mistakes with details, AI algorithms also simply err from time to time. As I have said and will repeat, these cases are rare, but you cannot have official documents and reports containing objective errors, especially in critical details:

Durch die Bp. werden im Schätzungswege der hälftige Aufwand aus den Rechnungen vom 17.12.2018 Notariat YYY und vom 31.12.2018 XXX GmbH als zusätzliche Kosten des Vermögensübergangs berücksichtigt, und dem Einkommen hinzugerechnet.

Wrong AI: The fiscal tax audit estimates half of the expenses from the invoices dated December 17, 2018 from Notariat YYY and December 31, 2018 from XXX GmbH as additional costs of the asset transfer and adds them to the profit.

Correct: The fiscal tax audit considers as an estimate half of the expenses from the invoices dated December 17, 2018 from Notariat YYY and December 31, 2018 from XXX GmbH as additional costs of the asset transfer and adds them to the profit.

The German formulation is an elliptical or highfalutin way of saying that too many costs have been deducted in German. The fiscal tax audit (like an IRS audit) has made an estimate and determined that half of the expenses are to be considered additional costs of the asset transfer (resulting in a higher tax bill than previously paid). This detail is obviously critical. Firstly, the German text formulates this act of determining the allocation of expenses as a statement of fact (berücksichtigt/considers). It also includes the means in which this is done (im Schätzungsweg/as an estimate). By transforming the “means” to the verb, the AI translation attenuates the factual character of the determination. It also creates an (at best) murky correspondence to the original: The fiscal tax audit estimates half of the expenses. Does that mean it did not estimate the other half? What is the definitive nature of this estimate that adds costs to asset transfer and profit. To a certain extent, the end of the sentence clears this up, but the estimate of expenses can be interpreted widely and wrongly.

Another example of this can be seen in the following translation where the “neutral” German formulation is translated by AI in a commensurately neutral manner. However, even if the neutral translation is possible in English, it is very uncommon:

Anfang des Jahres gingen die Märkte davon aus, dass die amerikanische Notenbank Fed im Jahr 2024 insgesamt sechs Zinsschritte verkünden würde.

Wrong AI: At the beginning of the year, markets assumed that the American central bank, the Federal Reserve (Fed), would announce a total of six interest rate steps in 2024.

Correct: At the beginning of the year, markets assumed that the American central bank, the Federal Reserve (Fed), would announce a total of six interest rate cuts in 2024.

This case is particularly problematic for AI because it involves context that it apparently does not (yet?) know. A layperson could also make this mistake. To translate “Zinsschritte” correctly in English, that is, in a formulation conventional in US English, you must specify whether the interest rate “step” or “act” is a “hike” (going up) or a “cut” (going down). An informed translator will obviously know this – we are currently (forecast to be) in a rate cutting cycle, so the neutral German term is translated as “cut.” Interestingly, the term “Zinssenkung” appeared two sentences later, so it is at least conceivable that AI could have determined the context of the term “Zinsschritt.” It is also possible that its frame of reference is the UK and that such a formulation is more common (a quick google search suggests that it is not).

2. Subjective or context-based flaws

As is apparent from reading solely the target text of a translation, the mistakes in AI translations tend to be more subjective or context-based than objective. AI translations read smoothly and are coherent as long as the source text on which they are based is defined by these characteristics. German texts created in a professional setting almost always exhibit the necessary clarity and wording to allow for easily read translations. In quantitative terms, the majority of the corrections I make in AI postediting can be attributed to subjective or context-based considerations.

We will begin the analysis by examining error types that also fall under objective flaws. These consist primarily of terminology and harmonization, but in a different sense than the objective cases. These elucidations will be followed by the highly subjective types of style, tense (the latter possibly somewhat peculiar to German) and omissions. Finally, we will look at context-based flaws that I have included in the subjective category, but staddle the line between the delineation.

2.1 Terminology

Specialists or experts in a field easily find flaws in algorithmic terminology, as we have seen here with mainly/majority owned. Aside from misinterpretations of specific terms in a given context, artificial intelligence must know standards, norms and other documentation. These reference works establish certain terms in a specific field for continental European use of English.

Algorithms often spit out terms that may be common in the UK or United States, but the fundamental equivalent is not used in the European context. Bilanzstichtag in German is an example. Reporting date would be far more common in American English, but this term is almost always translated as balance sheet date, a formulation that sounds quite stilted in the United States. Another is Finanzamt – the equivalent of the Internal Revenue Service (IRS) in the United States. You would never translate Finanzamt in a German context as IRS, although a machine might.

Another problem is the translation of terms that require a judgement about the priority of the author. Gemeinde is a good example of this. It can be reproduced interchangeably as community or municipality in almost any context. However, if the author is evidently using Gemeinde as a type of politically defined location such as a city, town and Gemeinde, then the machine may not pick this up in an isolated sentence (without city and town for reference), but a translator processing the entire text and knowing the big picture can judge the appropriate translation as municipality.

2.2 Harmonization

The subjective side of harmonization becomes a consideration when artificial intelligence decides to translate individual non-critical terms, especially verbs, divergently. As I discussed in harmonization under objective errors, critical terminology must be translated consistently. This rule does not necessarily apply to preferred phrasing or verbs. A target-language translation of a verb might have a close equivalent, but sound awkward (e.g., sich auszeichnen/is characterized by). Using the near equivalent might be acceptable if it appears only once, but, for the sake of readability, the algorithm (and posteditor) might prefer a smoother and less precise translation that also eschews harmonization.

The German language, like all others, is littered with popular verb formulations that are problematic in English. This is particularly the case in marketing texts. Sich auszeichnen, Erscheinungsbild, erfolgen are just a few popular means of expression that come to mind spontaneously.

2.3 Style

The biggest weakness in machine-based translation is the style, which, at least as produced by DeepL, is somewhat stilted from an American perspective. It should be noted that even when DeepL is set to American English, the formulations and style are influenced by British English on balance. British English also has more of a clausal structure than thoroughly streamlined American English (SPO ad nauseum), so I may not be the best judge of the results produced by it. Nonetheless, I have found that moderately-complex-to-complex sentences require a lot of rewriting. This is the most time-consuming aspect of working with machine translation. It is the reason why, as I wrote above, the price difference between a machine-based and human translation is “only” 20-30% (I am aware that this is a huge amount in economics and business administration). If you take an average sentence of 10-15 words (in German) and need to make 3 adjustments in the pre-translated English version, that is roughly equivalent to the amount of time needed to translate the sentence by yourself without machine support based on my estimates.

To achieve a fluid English text with at least a reduction in the clausal structure of German, while preserving the correspondence between the original and the translation, you must reconfigure the sentences produced by algorithms. Generative AI combines sentences, splits sentences and can even remove superfluous or redundant information, but at least the current iteration often shies away from the subjective decisions on sentence restructuring. My guess is that this is partially due to the fact that these sentences in and of themselves are not objectionable as clausal sentences, but when they appear in a series of multiple consecutive sentences, it becomes too cumbersome to read. This is another iteration of the context-based weaknesses in AI translation.

2.4 Tense

Another problematic issue is tense. There are two primary reasons for this: (i) the tense of the verb is not indicative of the actual tense, and (ii) the tense of the verb is ambiguous.

2.4.1 The present tense with a future meaning in German

Germans often use the present tense in texts with a future meaning. There are not two present tenses in German (as in English and Russian, for example) and there is a future tense for the future, but, for whatever reason (presumably the repetition of the future-indicating verb werden (will/shall)), German authors often resort to the present tense for the future. This does not work in English with the same frequency. And again, in the absence of sufficient context or clarity in the context, algorithms have difficulty recognizing this idiosyncratic use of tense. It should also be noted that Germans also have an affinity for the present tense when speaking about events in the past, especially in historical works, to make them more immediate or emphatically imply the parallels between the recounted history and its interpretation on the one hand and current events on the other. AI translations (and also human translators) often shift this to the past. This point is controversial and can be debated. Both are definitely possible. But the present tense for future events is not common in English. It must generally be corrected.

An example of the present tense being used for a future event can be seen in this marketing fragment on a sports competition:

Wir erwarten über 10.000 Athletinnen und Athleten aus der ganzen Welt. Hinzukommen 20.000 Begleitpersonen, 3.000 Offizielle, sowie jeweils etwa 15.000 Angehörige und geladene Ehrengäste.

AI translation: We expect over 10,000 athletes from all over the world. In addition, there are 20,000 accompanying persons, 3,000 officials, and about 15,000 family members and invited guests of honor each.

Human translation: We expect over 10,000 athletes from all over the world. In addition, there will be 20,000 accompanying persons, 3,000 officials, and about 15,000 family members and invited guests of honor each.

The two sentences in German contain the verbs erwarten and hinzukommen, both in the present tense. The event will take place in the near future. In English, verbs such as expect and anticipate allow for future meaning, so no difference is seen between the AI and human translation here. The second sentence would sound very awkward, however, if hinzukommen is translated as there are for an event in the future. The reader would think the two sentences are referring to different events – one in the future (expect) and one now (there are).

2.4.2 The case of sollten

The modal verb sollten (should/supposed to) is ambiguous in German. It can refer to the present tense in the meaning of should or be supposed to on the one hand or the past tense in the meaning of was/were supposed to. Only the context can dictate the tense and thus meaning. This seems to be a problem for AI based on extensive empirical findings. Here is an example:

Die Auswahl der Veranstaltungsstätten sollte es Berlinerinnen und Berlinern aus allen Bezirken ermöglichen, die Weltspiele ohne weite Anfahrtswege besuchen zu können.

This passage was translated in the present tense as should. However, it appeared in a report about an event that had already happened. The selection took place in the past. This entails that the ambiguous sollte must be translated in the past tense (as was supposed to).

A mishmash of tenses, especially in the case of the present tense being used for past events, is the most common outcome generated by artificial intelligence translation. In such cases, the mistake(s) effectively fall under objective errors. Consistency across sentences or segments remains the job of the human translator/posteditor.

2.5 Omissions

If I had categorized these translation mistakes in the days of human translation, the omission of text – a very common problem with human translation – would have certainly been categorized as an objective error. Especially in long sentences, it is common to oversee a sentence fragment, say, an independent clause or a dative or genitive phrase embedded in all the other information. Today, it is very different with machines.

Experience and familiarity with algorithmic translation leaves you very comfortable with regards to omissions. An AI translation does not omit critical information. Again, this statement applies to 99.9% of automated translation. That is not to say, however, that there are no omissions at all. One of the most fascinating aspects of working with machine translation is that the algorithms also analyze the source text and sometimes decide that it contains superfluous or redundant information or can be phrased in a more concise way with far fewer words. This causes translation to drift toward what I would call transposition: The function and meaning of the original are retained, i.e., roughly, the form, while the content is altered. These altered sentences lead to omissions. For the most part, algorithms do not engage in this. And in the case of generative AI, you can instruct the provider not to combine sentences or leave out information. Based on my experience, it is fairly obvious, not common, and does not tend to crop up in longer sentences.

2.6 Context

Context, as alluded to or directly mentioned in various parts of this article, remains one of the primary stumbling blocks in AI translation. One particularly problematic subarea of this issue is titles or headings of books, chapters, sections, etc.

Original: Erklärung zum Bezug bzw. zur Überprüfung orts- und familienbezogener Besoldungsbestandteile (OFZ-Erklärung)

AI translation: Explanation regarding the reference or verification of location- and family-related salary components (OFZ Declaration)

Correct: Declaration on the Receipt of Benefits or Verification of Location- and Family-related Salary Components (OFZ Declaration)

At the point when artificial intelligence reads this heading, it does not know what the heading is referring to. Explanation and declaration are both common translations for Erklärung. In part, as already mentioned, this is due to the block-like nature of artificial intelligence, as the platforms can still only handle a limited number of characters. When entire texts can be prepared, this weakness will certainly improve, e.g., by having the title or heading effectively translated after the entire translation has been generated. In the above case, the algorithm should have matched the first instance of Erklärung with the second, which it translated correctly as declaration. Generally, however, such duplicates are not present and the reference to a regulation may not be realized.

Conclusion

Although artificial intelligence has evolved and automated translation to a degree unanticipated a decade ago, the human translator or editor still occupies an indispensable place in the process. The algorithm produces a better first draft than a human, but its results must be revised and proofread just as those of a human translator.

As we have seen above, the revision stage has changed slightly due to the types of errors slipping into translation now. The frequency of objective or hard errors has dropped with automation, while some subjective or context-based mistakes, soft errors, have remained and in some cases increased due to the abstract preparation of the draft.

In the following article, we will explore some of the fun and experimental possibilities facilitated by this new method of preparing translations.