Abstract: Techniques for updating a machine learning model based on user interactions are described. In particular, in some examples, user interactions with a chatbot provide aspects of a data set to be used to train or fine-tune a ML model. In some examples, this is accomplished by collecting data from a first plurality of interactions with a machine learning (ML) model; generating a variant of the ML model using the collected data by: filtering the collected data to create a first data set, training the ML model based on the first data set to generate an adapted ML model, and fine-tuning the adapted ML model on a second data set, different than the first data set to generate the variant of the ML model.

Abstract: A system receives one or more transcripts of communications between entities. The system identifies a requested action in the communications based at least in part on a mapping between the requested action and an application programming interface. The system identifies one or more statements eliciting information, based on parameters to the application programming interface. The system generates a definition of an artificial agent based, at least in part, on the requested action and the one more statements eliciting information.

Abstract: Using a first set of machine learning models, a communication from a user of a restaurant is analyzed at an order coordinator linked via a network to resources of an order management service at a provider network. A response to the communication is prepared using another set of models at the provider network and presented to the user. An order of the user for one or more restaurant menu items is fulfilled, based at least partly on analysis of a second communication received from the user after the response is presented.

Abstract: Techniques are described for a contextual natural language understanding (cNLU) framework that is able to incorporate contextual signals of variable history length to perform joint intent classification (IC) and slot labeling (SL) tasks. A user utterance provided by a user within a multi-turn chat dialog between the user and a conversational agent is received. The user utterance and contextual information associated with one or more previous turns of the multi-turn chat dialog is provided to a machine learning (ML) model. An intent classification and one or more slot labels for the user utterance are then obtained from the ML model. The cNLU framework described herein thus uses, in addition to a current utterance itself, various contextual signals as input to a model to generate IC and SL predictions for each utterance of a multi-turn chat dialog.

Abstract: Techniques for generating conversational training data are described. In some instances, a request to generate conversational training data for a goal-oriented conversation model is received, a transitional graph of intents is traversed to generate a conversation template for each intent of the transitional graph, each intent being a task to fulfill a request and comprising one or more slot to be filled by a user of the bot machine learning model, the conversation template including a path including at least one placeholder for an utterance or a slot level utterance, and at least utterances from one or more dictionaries are sampled to fill in the placeholders for the utterances of the path to generate conversational training data.

Abstract: Techniques for creating and calibrating natural-language understanding (NLU) machine learning models are described. In certain embodiments, a training service tunes parameters of a function, taking the output from an NLU machine learning model as an input of the function, to calibrate the NLU machine learning model's output to optimize the interpretability of the resulting output, e.g., confidence score(s). Embodiments herein include generating, by the NLU machine learning model, an output based at least in part on an input (e.g., utterance) from a user, and applying a tuned, output modifying function to the output from the NLU machine learning model to generate a modified output. An inference may be generated based at least in part on the modified output.