Reflections on policy, global affairs, and life.

Language Learning with AI

German with AI

I moved to Germany as an English speaker fluent in French, but studying German has made me feel like learning French was breezy. A main reason for this, however, was that I had studied French full-time in a language school. Now, though, between my job and my personal life, I can't make that same kind of commitment. All the same, I want my German to improve, so I needed to rethink how I was learning.

Enter Large Language Models (I'm using ChatGPT at the time of writing).

I am not just using AI to translate. I'm depending on it to coach me. At the moment, LLMs do an alright job of translating, but I think they excel as teachers.With the "Projects" feature that has popularized with some of the major AI services, I have been experimenting with custom instructions. Over time, I've found they lead to reliably helpful responses.

I prefer to focus my learning on real situations instead contrived worksheet examples. (If DuoLingo asks me how big the elephant is one more time, I'm gonna lose it).Writing emails to the people I work with (in this case, adminsitrators at the University of Bonn) has been a perfect opportunity for this! I have the time to sit and focus on what I want to say and how to get that phrasing right in a way which not only communicates my meaning but also fits the social norms of sending emails in German. More than conjugating for the right level of formality, I also want to get little details like choosing the right greetings and sign-offs, putting my commas in the right places, and use capitalization appropriately (hint: it's not the same as in America).

To do these things, I've given ChatGPT the following instructions:

Never use the internet.

Context: I am an English and French speaker living in Germany and learning German.

Follow these instructions strictly:

  1. Always answer concisely and clearly. Do not add comments or filler content.
  2. If given a phrase in English (without additional context), translate it into simple, natural German—not a literal word-for-word translation. Prioritize core meaning, common phrasing, and basic vocabulary and conjugations (A1-A2 level).
  3. After providing the German translation: a. Then, provide an idiomatic French translation (natural phrasing, not word-for-word). b. Repeat the German sentence a second time, but this time following any German words that would be hard for an A1 student with their English equivalents in brackets [like this].
  4. Provide a list of those German words you determined were difficult, including their IPA spelling. Make the IPA spelling a clickable link for the DeepL translation using the pattern: https://www.deepl.com/translator#de/en/QUERY_WORD
  5. Do not label any of the three outputs. Present them consecutively without headings or extra commentary.
  6. If given the keyword "formal" or "informal/casual/friendly", conjugate accordingly.
  7. If I give you just one word in German, I want you to give that back to me spelled using the IPA.

These rules are tailored to me, but I think they represent a good starting off point for anyone looking to use LLMs to help them learn a new language. For example, with the 3rd rule, I lean on my French to help me quickly assess if the German is being conjugated formally or informally, but this could definitely be adapted to English with a parenthetical indicating the level of formality the writing is using.

Using this collaborative approach with AI has really taken me beyond translation. Even though DeepL is best in class for most translations, it does nothing for all the times when I'm wondering why a certain word was used or whether there were rephrasings which could better convey my intended meaning. AI as a translation-partner instead of a tool has done just that for me, and maybe it can for you too.

Coded & written by James Mitofsky