| CPC G06F 16/3329 (2019.01) [G10L 15/08 (2013.01); G10L 2015/088 (2013.01)] | 19 Claims |
|
1. A computer-implemented method comprising:
receiving, by a contact center service in a provider network, an utterance from a conversation between a user and an agent, wherein the utterance comprises one or more sentences of the conversation;
determining, by a first machine learning model of the contact center service, a probability that the utterance includes a trigger utterance, wherein the first machine learning model is a dual encoder stacked ensemble model trained to identify the probability of the utterance being a trigger utterance;
responsive to determining that the probability meets or exceeds a threshold probability value, identifying, by a second machine learning model of the contact center service, assistance information associated with the utterance by determining a similarity between a representation of the utterance and a representation of the assistance information using section-level relevance of one or more documents of the assistance information to the utterance, wherein the one or more documents are divided into sections and each of the sections includes a plurality of fragments, wherein the section-level relevance uses section embeddings that are weighted sums of fragment embeddings partitioned from corresponding sections of the one or more documents, and wherein the second machine learning model is trained to determine the representation of the utterance and the representation of the assistance information; and
transmitting, by the contact center service, the assistance information for presentation via a user interface.
|