We have leveraged our existing analyses of thousands of videos to build a machine learning model to provide you with a helping hand identifying all the valuable quotes from your interviews – for you to either use (just as if created by yourself or a colleague), edit or delete.
Learning by example
The most simple way to explain how we can get a machine learning model to identify relevant quotes from a text transcription of an interview, is to say that we taught it by example – or we used supervised learning to train a neural network to predict if a sentence should be in a quote or not. We have done this by showing the model a sentence from a transcript and then told it if it should be part of a quote or not.
Does it work
To evaluate how well we have trained the model, we used a dataset with 50.000 quotes created by our specialist, that had not been used to train the model.
The result was that out of 100 potential quotes, our specialist had originally created 36 and the model created 39 – but most importantly they agree on 32 of them and also that 57 of them should not be quotes. This leaves only 11 of 100 quotes for them to ‘disagree’ on. We think that is pretty amazing.
It is important to consider that two different research specialists will never create the exact same quotes and we will therefore also never be able to train the model to 100% accuracy and therefore to replace the specialist – but we are able to provide a helping hand that should reduce the amount of work creating quotes (or coding the data) significantly and thereby also the time for you to get to the valuable insights.