Response recommendation system

A method comprising:

• generating training data based on a corpus of conversation data including sequences of inputs corresponding to conversation sessions at the interactive conversation service, the training data including a set of context-response pairs and a label corresponding to each of the set of context-response pairs, a context of a context-response pair including a sequence of inputs during a conversation session, and a response being a subsequent input by an agent during the conversation session, the training data generated by:
• for a first conversation session in the corpus of conversation data, selecting a first set of sequential inputs of the first conversation session as a first context of a first context-response pair, selecting an input by an agent that is immediately subsequent to the first set of sequential inputs as a first response of the first context-response pair, and assigning a first value to a first label corresponding to the first context-response pair, and
• for a second conversation session in the corpus of conversation data, selecting a second set of sequential inputs of the second conversation session as a second context of a second context-response pair, selecting a random input of the corpus conversation data as a second response of the second context-response pair, and assigning a second value to a second label corresponding to the second context-response pair;
• training at least one word embedding function using the training data by:
• generating a context vector for each context of each context-response pair of the set of context-response pairs using a context word embedding function of the at least one word embedding function;
• generating a response vector for each response of each context-response pair of the set of context-response pairs using a response word embedding function of the at least one word embedding function;
• calculating a similarity score for each context-response pair of the set of context-response pairs by comparing the context vector and the response vector of each context-response pair;
• and
• determining a loss for each context-response pair of the set of context-response pairs based at least on the similarity score and the label corresponding to each context-response pair of the set of context-response pairs.

An apparatus comprising:

• a processor,
• memory in electronic communication with the processor;
• and
• instructions stored in the memory and executable by the processor to cause the apparatus to:
• receive an input comprising a string of text at an interactive conversation service, the input received during a conversation session at the interactive conversation service;
• generate, responsive to receiving the input, a context vector based at least in part on a set of inputs corresponding to the conversation session, the set of inputs including at least the received input, the context vector generated using at least one word embedding function trained on a set of training data at least partially generated by:
• for a first conversation session in a corpus of conversation data, selecting a first set of sequential inputs of the first conversation session as a first context of a first context-response pair, selecting an input by an agent that is immediately subsequent to the first set of sequential inputs as a first response of the first context-response pair, and assigning a first value to a first label corresponding to the first context-response pair;
• and
• for a second conversation session in the corpus of conversation data, selecting a second set of sequential inputs of the second conversation session as a second context of a second context-response pair, selecting a random input of the corpus conversation data as a second response of the second context-response pair, and assigning a second value to a second label corresponding to the second context-response pair;
• and
• calculate a set of similarity scores between the context vector and a set of response vectors, the set of response vectors including at least one response vector based on a custom response input by a user of the interactive conversation service;
• and
• select one or more responses corresponding to one or more of the set of response vectors for input at the interactive conversation service, the one or more responses selected based on the set of similarity scores calculated between the context vector and the set of response vectors.

An apparatus comprising:

• a processor,
• memory in electronic communication with the processor;
• and
• instructions stored in the memory and executable by the processor to cause the apparatus to:
• generate training data based on a corpus of conversation data including sequences of inputs corresponding to conversation sessions at the interactive conversation service, the training data including a set of context-response pairs and a label corresponding to each of the set of context-response pairs, a context of a context-response pair including a sequence of inputs during a conversation session, and a response being a subsequent input by an agent during the conversation session, the training data generated by:
• for a first conversation session in the corpus of conversation data, selecting a first set of sequential inputs of the first conversation session as a first context of a first context-response pair, selecting an input by an agent that is immediately subsequent to the first set of sequential inputs as a first response of the first context-response pair, and assigning a first value to a first label corresponding to the first context-response pair;
• and
• for a second conversation session in the corpus of conversation data, selecting a second set of sequential inputs of the second conversation session as a second context of a second context-response pair, selecting a random input of the corpus conversation data as a second response of the second context-response pair, and assigning a second value to a second label corresponding to the second context-response pair;
• train at least one word embedding function by:
• generating a context vector for each context of each context-response pair of the set of context-response pairs using a context word embedding function of the at least one word embedding function;
• generating a response vector for each response of each context-response pair of the set of context-response pairs using a response word embedding function of the at least one word embedding function;
• calculating a similarity score for each context-response pair of the set of context-response pairs by comparing the context vector and the response vector of each context-response pair;
• and
• determining a loss for each context-response pair of the set of context-response pairs based at least on the similarity score and the label corresponding to each context-response pair of the set of context-response pairs.