Abstract: A disambiguation dialog may be generated by determining candidate responses based on an intent of a user's message. A utility value and a relevance probability may be determined for each of the candidate responses. An intermediate ranking may be computed for each of the candidate responses based on the utility value and the relevance probability. Candidate dialogs may be formed with the top candidate response, the top two candidate responses, and so on. Additional candidate dialogs may be generated by varying a presentation format of the candidate responses. Discoverability probabilities may be associated with each of the candidate responses within a candidate dialog. A joint metric for each candidate dialog may be computed as a function of the utility value, relevance probability and discoverability probability associated with each of the candidate responses included in the candidate dialog. The highest ranked candidate dialog may be selected as the disambiguation dialog.

Abstract: An automated agent may communicate with a user via a chat channel to proactively alert the user of an L1 IT support issue. The L1 IT support issue may be determined based on monitoring indications of human-initiated activities maintained by a system of record, and may, prior to the automated agent's alert, be unknown to the user. In some instances, a natural language understanding (NLU) module may be used to identify an entity and intent from the indications of human-initiated activities, and the L1 IT support issue may be determined based on the determined entity and intent. After alerting the user of the L1 IT support issue, the automated agent may inform, via the chat channel, the user of a remediation step available to address the L1 IT support issue. Upon obtaining the user's permission, the automated agent may perform the remediation step to address the L1 IT support issue.

Abstract: The content of a knowledge datastore is evaluated and improved. In a first aspect, the content effectiveness of individual snippets is evaluated and a content creator is requested to improve snippets with a low content effectiveness. In a second aspect, the supply of and demand for content in each content topic is evaluated, and a content creator is requested to create articles for content topics for which the demand exceeds the supply. In a third aspect, the message responsiveness and content effectiveness of content topics is evaluated and a content creator is requested to create articles for content topics with a low message responsiveness and/or content effectiveness. In a fourth aspect, the content utilization and content effectiveness of individual snippets is monitored and snippets with a high content effectiveness and a low content utilization are promoted, whereas snippets with a low content effectiveness and a low content utilization are deprecated.

Abstract: An automated conversation is facilitated between a user and a virtual agent. A system receives an input message from the user and analyzes an intent of the input message. Based on the intent of the input message, the system generates a plurality of bids for responding to the input message, and assigns an intent confidence score to each bid from the plurality of bids based on a confidence level of each bid from the plurality of bids. The system determines a winning bid from the plurality of bids based on the intent confidence score associated with each bid from the plurality of bids, and generates a response based on the winning bid.

Abstract: A form filling application is configured to minimize the form filling effort of a user. The configuration follows concepts from reinforcement learning, and includes optimizing a policy for selecting agent actions in a manner that maximizes a reward signal. In the context of the form filling application, an agent action may specify one or more slots of the form for the user to fill, and further specify one or more user interfaces for filling the specified one or more slots. The reward signal may be defined as an inverse function of the user effort, so that maximizing the reward signal has the desired effect of minimizing the user effort.

Abstract: An automated agent may communicate with a user via a chat channel to proactively alert the user of an L1 IT support issue. The L1 IT support issue may be determined based on monitoring indications of human-initiated activities maintained by a system of record, and may, prior to the automated agent's alert, be unknown to the user. In some instances, a natural language understanding (NLU) module may be used to identify an entity and intent from the indications of human-initiated activities, and the L1 IT support issue may be determined based on the determined entity and intent. After alerting the user of the L1 IT support issue, the automated agent may inform, via the chat channel, the user of a remediation step available to address the L1 IT support issue. Upon obtaining the user's permission, the automated agent may perform the remediation step to address the L1 IT support issue.

Abstract: A form filling application is configured to minimize the form filling effort of a user. The configuration follows concepts from reinforcement learning, and includes optimizing a policy for selecting agent actions in a manner that maximizes a reward signal. In the context of the form filling application, an agent action may specify one or more slots of the form for the user to fill, and further specify one or more user interfaces for filling the specified one or more slots. The reward signal may be defined as an inverse function of the user effort, so that maximizing the reward signal has the desired effect of minimizing the user effort.

Abstract: A multi-domain chatbot is used to service a message of a user. An automated agent of the multi-domain chatbot may act as an intermediary between the user and a plurality of domain-specific modules of the multi-domain chatbot. The automated agent may receive the message from the user, determine an intent of the message, and based on the intent, determine a group of the domain-specific modules that should be investigated. The automated agent may then investigate the group of domain-specific modules by sending the user message to and receiving responses from the domain-specific modules within the group. Based on the received responses, the automated agent may determine whether to provide, to the user, one of the domain-specific responses or a null response, in the event that none of the domain-specific responses is aligned with the intent of the message.

Abstract: A disambiguation dialog may be generated by determining candidate responses based on an intent of a user's message. A utility value and a relevance probability may be determined for each of the candidate responses. An intermediate ranking may be computed for each of the candidate responses based on the utility value and the relevance probability. Candidate dialogs may be formed with the top candidate response, the top two candidate responses, and so on. Additional candidate dialogs may be generated by varying a presentation format of the candidate responses. Discoverability probabilities may be associated with each of the candidate responses within a candidate dialog. A joint metric for each candidate dialog may be computed as a function of the utility value, relevance probability and discoverability probability associated with each of the candidate responses included in the candidate dialog. The highest ranked candidate dialog may be selected as the disambiguation dialog.

Abstract: A computer-implemented method for training a chatbot is provided. The method includes receiving a training input through a platform associated with the chatbot. The training input indicates user intent for interacting with the chatbot. The method includes calculating a confidence score associated with a prediction of the user intent identified by the chatbot. The method further includes providing a training score to the user providing the training input based on the confidence score.

Abstract: Embodiments provide methods and apparatus for improving responses of automated conversational agents. The method includes generating a vector representation of a conversational input provided by a user. The vector representation is used to determine an intent of the conversational input. Further, annotators generate bait sentences that cover multiple aspects of the intent. Then, sentences in a data pool are accessed. The bait sentences and the data pool sentences are converted into a first and a second set of vector representations, respectively. The first and the second set of vector representations are compared to retrieve a list of similar sentences. The list of similar sentences includes one or more sentences of the data pool that are semantically similar to the bait sentences. The list of similar sentences is analyzed for updating the intent data and thereby improve the responses.