We currently use bart-large-xsum, which is not optimal as the model is very prone to halucinations and has little input capacity. We could train a model which takes as input the transcript as well as the previous and next summary bullet points as context and produces a single bullet point.
We currently use bart-large-xsum, which is not optimal as the model is very prone to halucinations and has little input capacity. We could train a model which takes as input the transcript as well as the previous and next summary bullet points as context and produces a single bullet point.