In Issue 1/2024

In the era of AI, all that once looked extremely difficult and requiring skilled talents to be accomplished is now at a click’s distance. Or at least this is what it seems. In fact, AI does not always manage to meet our expectations. The reason for this is something that has always caused problems in communication among humans: understanding. In technical terms this is called prompting, but fundamentally it is the capacity of humans to speak a language that the machine can understand.

This is nothing new. Think of all the misunderstandings in everyday conversations among humans. These can be fixed. When talking, we can contextualise information, correct a word or rely on body language and intonation. In written communication this is a little more cumbersome as we can only rely on characters lining up on screen one after the other. This makes the difference between written and spoken communication worth further investigation.

In linguistics, Paul Grice (1975) clearly stresses the importance for speakers to follow four maxims for an effective conversation:

  • provide enough information (maxim of quantity),
  • say the truth (maxim of quality),
  • be pertinent (maxim of relevance),
  • be clear and avoid obscure speech (maxim of manner).

Grice soon realised that actual conversations rarely stick to these rules, because speakers fail to provide correct, pertinent, clear and enough information, or simply because they don’t want to. Think of slogans or journal titles sometimes using partial truths, violating the maxim of quantity; of irony, lies, fake news, violating the maxim of quality; of when you abruptly move to a new topic because you think the current one may annoy or embarrass someone, violating the maxim of relevance; and of political demagogy, violating the maxim of manner. In diamesic translation, even if you go for verbatim transcription, what is said and how it is said can make the task very hard for live captioners, subtitlers, court and parliamentary reporters, and secretaries, as shown by the experiment I carried out for British Institute of Verbatim Reporters, demonstrating how 10 professionals transcribed the same input in 10 different manners (Eugeni, 2021).

Furthermore, communication doesn’t only depend on collaborative speakers, it also requests collaborative listeners. When dealing with comprehension, van Dijk and Kintsch (1983) develop the “strategic model of discourse comprehension”, according to which six strategies need to be applied by collaborative listeners when processing the speaker’s input to turn it into meanings:

  • sounds and grammar into comprehension of the surface level of a sentence (propositional strategies);
  • sentences into comprehension of the links between ideas (local coherence strategies);
  • main communication moves into comprehension of the reasoning (macrostrategies);
  • text type into inference and predictability of what comes or may come next (schematic strategies);
  • shared knowledge of the world into overall comprehension of the contextualised text (production strategies); and
  • other strategies that allow for understanding nuances, as delivered by the speaker’s style, register, and non-verbal and conversational elements.

When asked to turn speech into a specific type of text, diamesic translators apply both understanding strategies and production strategies. Adapting Kohn and Kalina’s model (1996), I suggest that diamesic translators adopt the following strategies to produce higher quality texts:

  1. elaborative inferencing: to anticipate grammatical, lexical and conceptual elements;
  2. memorising: to postpone grammatical, lexical and conceptual elements;
  3. monitoring strategies: to detect possible grammatical, lexical and conceptual mistakes;
  4. adaptation strategies: to bridge possible grammatical, lexical and conceptual gaps between the spoken and the written language, via explicitation of intonation, punctuation, formatting, standardising spelling and grammar, disambiguating homophones, etc;
  5. evasion strategies: to produce text that does not commit too much in case of different possible interpretations of the same output, as in the case of the French MP discussed in my last column which reports on how French parliamentary reporters had to deal with a case of homophony that could lead to interpretations with political consequences (Eugeni 2023).

Back to AI, all these strategies are something that the machine cannot apply as these are mental activities that require human intelligence. However, we all know from experience that AI can produce pretty good results at times. How come? The reason is that a lot of understanding and production strategies can be automated. And as long as collaborative speakers stick to Grice’s maxims, the machine can replace human skills. What AI cannot replace (yet!) is the capacity of humans to understand speakers who are not collaborative and to go beyond surface to get the speaker’s intended meaning. It is not, in this sense, a collaborative listener. This means there is still a need for diamesic translators, as long as speakers fail to be collaborative.

Allow me to conclude with a personal consideration: if we – diamesic translators – once hated bad speakers because they were hard to transcribe, we are now turning into their best supporters because we want to keep our profession alive, valued and decently paid!

Carlo Eugeni is Tiro’s Scientific Adviser.


Van Dijk, T. A. & W. Kintsch (1983). Strategies of Discourse Comprehension. Academic Press. URL:

Eugeni, C. (2021). What does ‘verbatim’ mean? – Tiro 2/2. URL:

Eugeni, C. (2023). Of Homophones, referents, and prejudices. – Tiro 4/2. URL:

Grice, P. (1975). Logic and conversation. In Cole, P. & J. Morgan. (eds) Syntax and semantics, vol. 3, pp. 41–58. Academic Press. URL:

Kohn K. & K. Sylvia (1996). The Strategic Dimension of Interpreting. – Meta 41 (1), pp. 118-138. URL:

    pingbacks / trackbacks
    • […] Carlo Eugeni:Strategies in Diamesic Translation […]

    Leave a Comment

    This site uses Akismet to reduce spam. Learn how your comment data is processed.