- The keyphrase appearance in the user’s input is a BoW-like vector.
- Transpose the matrix W_PK of keyphrase occurrences in the page and multiply by the input
- The hidden layer has a strong association with the page so far.
- This layer is Linear
- Multiply the output of the hidden layer by the mapping matrix W_PS between pages and utterances taken from those pages
- That alone is fine, but multiply by the matrix of key phrase occurrences during speech and add them together.
- Softmaxed against it will be the probability of utterance.
This page is auto-translated from /nishio/チャットボット2018-09-30 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.