We recently switched over to the
tensorflow_embedding pipeline (mostly due to multiple intent support and less memory consumption).
However (due to the architecture of the pipeline), the model doesn’t generalize over new examples at all.
In many cases the models returns an incorrect intent with very high confidence and the correct intent (along with many others) is returned as zero.
Is it possible to have a structured way of solving this problem?
I tired the pipeline specified in this thread but it didn’t work and I got the same results.