In context: AI can help multi-lingual research
Many international researchers take advantage of the speed of AI powered machine translation, which, although improving, still has limitations.
When it comes to uploading transcripts for AI analysis and reporting, the quality of the translation is critical. In the well-known phrase, ‘rubbish in, rubbish out’.
I have tested a process to prepare English transcripts (from Dutch language versions, provided by colleagues in the Netherlands) for uploading to AI, consisting of four steps.
Stage one involves importing the audio or video clips into Adobe Premiere Pro for text-based editing. This will then create a transcript in the original language (Premiere Pro offers 18 language packs for speech to text). Importantly, it will also create subtitles to view as the audio or video file is played on the timeline.
The next stage is to convert the transcripts into English, using Google translate or similar. There are often common errors in the translations, combining errors from the original Dutch speech to text and then added to from the Google translate into English.
At this stage the translation is not of sufficient accuracy to upload to AI. When we tried with this uncorrected translation, there were mistakes in proper nouns, such as place names, brand names, some slang specific to Amsterdam and technical words used to describe the products. Clearly to use AI successfully in multiple languages, the translation must be pristine.
This is where the next steps come in. It is possible to export the Dutch subtitles from Premiere, have them translated into English and then import them to replace the Dutch subtitles. We then have a collection of audio/video files in the original language, but with English subtitles.
These are then sent to the native language speaker to view, in this case hearing responses in Dutch, but seeing the subtitles in English. The next stage is for the native speaker to note the required changes on a spreadsheet and return to be corrected in Premiere Pro.
The final output is interview clips with English subtitles, together with corrected English transcripts are ready to be uploaded into Claude AI, which offers the significant advantage of a text limit of 75,000 words, meaning a good number of transcripts can be combined and uploaded in one Word file.
What are the pitfalls to be aware of when you have the output from the AI?
First, adding ‘helpful’ information can be misleading. AI can draw on data which is outside the scope of that uploaded. This is a particular risk when you ask about ‘attitudes’ rather than the opinions of your sample. In this case, ChatGPT added information about Dutch commuting which didn’t apply to the sample of respondents.
Secondly, when asking for recommendations this is also a problem, since the AI will produce a more generic set of logical steps, not necessarily related to the data. It is tempting to use the ‘creativity’ of AI to generate ideas, but beware, if these are ideas for your client’s product development, some of them will be impractical.
Over-optimism is also a function of the tendency to be helpful. The AI is inclined to interpret opinions expressed as a positive finding, irrespective of context, or small numbers.
Finally, AI cannot spot ‘the dog that did not bark’. As Sherlock Holmes noted, sometimes when something doesn’t happen it is significant. In qualitative research, we can gain critical insights from what is not said, but what we would expect to hear.
However, armed with high quality English transcripts you will be able reap the major benefits of using AI for qualitative reporting. This includes the ability to comb through the data and select, literally, everything related to a particular respondent to create respondent profiles. It can create extraordinarily good typologies (or ‘personas’). It gives you good summaries of multiple topics, and AI gives you the power to search for correlations in a quasi quantitative way. Verbatim comments, under topic headings can be generated within minutes.
I have measured the value of AI in qualitative analysis systematically, having created a report in the conventional way last year, based on thirty open-ended interviews. I have taken those thirty transcripts and uploaded them to Claude 3.5 Sonnet.
The reduction in time required is transformational for qualitative research. I don’t have a record of the time it took to do my original analysis and prepare the report, but a conservative estimate is two to three weeks. To achieve the equivalent result using AI would take two to three days.
All the detail of the AI output and the prompts used can be found in the online PDF link here, where you can see the quality and quantity of the AI output and compare with the original report.
John Habershon is director at Momentum Research

We hope you enjoyed this article.
Research Live is published by MRS.
The Market Research Society (MRS) exists to promote and protect the research sector, showcasing how research delivers impact for businesses and government.
Members of MRS enjoy many benefits including tailoured policy guidance, discounts on training and conferences, and access to member-only content.
For example, there's an archive of winning case studies from over a decade of MRS Awards.
Find out more about the benefits of joining MRS here.
0 Comments