Abstract: Systems and methods are directed to email threading based on machine learning determined categories and features. A network system accesses a plurality of emails addressed to a user. The network system then classifies, using a machine learning model, each email into at least one of a plurality of categories. For a category of the plurality of categories, one or more feature values are extracted from each email in the category. Based on the category and the extracted feature values, the network system groups messages having a same feature value in the same category together into a single email thread. Information related to the single email thread is then presented at a client device of the user.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for resolving temporal ambiguities are presented. A natural language input may be received. A temporal component of the input may be identified. A determination may be made that the temporal component includes a conjunction that separates temporal meeting block alternatives. A temporal ambiguity may be identified in one of the meeting block alternatives. A plurality of syntax tree permutations may be generated for the meeting block alternative where the ambiguity was identified. A machine learning model that has been trained to identify a most relevant permutation for a given natural language input may be applied to each of the plurality of permutations. A temporal meeting block alternative corresponding to the most relevant permutation may be surfaced.

Abstract: Systems and methods are directed to email threading based on machine learning determined categories and features. A network system accesses a plurality of emails addressed to a user. The network system then classifies, using a machine learning model, each email into at least one of a plurality of categories. For a category of the plurality of categories, one or more feature values are extracted from each email in the category. Based on the category and the extracted feature values, the network system groups messages having a same feature value in the same category together into a single email thread. Information related to the single email thread is then presented at a client device of the user.

Abstract: Various embodiments are directed to automatically determining when meetings are related to each other. The relationship between meetings may be stored in a meeting-oriented knowledge graph that can be analyzed to provide meeting analytics. Various technologies can leverage the meeting relationship information to provide improved meeting services to users. For example, meeting suggestions may be presented to a user with suggested meeting parameters (e.g., suggested attendees, suggested location, suggested topic) that are accurate because a relationship between meetings is used to predict the parameters. The information in the meeting-oriented knowledge graph can be used to generate various analytics and visualizations that help users plan or prepare for meetings.

Abstract: The disclosure relates to systems and methods for identifying entities related to a task in a natural language input. An entity detection model is provided which receives a natural language input. The entity detection model processes the natural language input using an entity encoder and an input encoder. The entity encoder identifies and encodes relevant entities while the input encoder generates a contextual encoding which represents contextual information associated with a relevant entity. The encoded entity and contextual encodings may then be combined and processed to generate a probability score for an identified entity. A negation constraint model is also disclosed. The negation constraint model receives the natural language input and the identified entities. The natural language input is analyzed to identify negation cues and determine if the negation cue is associated with an identified entity.

Abstract: A system and method for allocating a recurring resource is described. The system receives a request to allocate the recurring resource of an application to computing devices associated with one or more users. The system identifies a cadence of the recurring resource and a range based on the cadence. A period is determines based on the cadence and the range. The system accesses user resource data of the application for each period and iteratively computes a resource availability score for the period based on the corresponding user resource data. The system determines that the resource availability score of an instance of the period is less than a threshold value and allocates the recurring resource to instances of the period where the corresponding resource availability score exceeds the threshold value.

Abstract: A system and method providing an accessibility tool that enhances a graphical user interface of an online meeting application is described. In one aspect, a computer-implemented method performed by an accessibility tool (128), the method includes accessing (802), in real-time, audio data of a session of an online meeting application (120), identifying (804) a target user, a speaking user, and a task based on the audio data, the speaking user indicating the task assigned to the target user in the audio data, generating (806) a message (318) that identifies the speaking user, the target user, and the task, the message (318) including textual content, and displaying (808) the message (318) in a chat pane (906) of a graphical user interface (902) of the online meeting application (120) during the session.

Abstract: A software agent, that is used to assist in providing a service, receives communications from a set of users that are attempting to use the software agent. The communications include communications that are interacting with the software agent, and communications that are not interacting with the software agent. The software agent performs natural language processing on all communications to identify such things as user sentiment, user concerns or other items in the content of the messages, and also to identify actions taken by the users in order to obtain a measure of user satisfaction with the software agent. One or more action signals are then generated based upon the identified user satisfaction with the software agent.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for training machine learning models are presented. An automated task framework comprising a plurality of machine learning models for executing a task may be maintained. A natural language input may be processed by two or more of the machine learning models. An action corresponding to a task intent identified from the natural language input may be executed. User feedback related to the execution may be received. The feedback may be processed by a user sentiment engine. A determination may be made by the user sentiment engine that a machine learning model generated an incorrect output. The machine learning model that generated the incorrect output may be identified. The machine learning model that generated the incorrect output may be automatically penalized via training. Any machine learning models that a user expressed neutral or positive sentiment toward may be rewarded.

Abstract: A system and method providing an accessibility tool that enhances a graphical user interface of an online meeting application is described. In one aspect, a computer-implemented method performed by an accessibility tool (128), the method includes accessing (802), in real-time, audio data of a session of an online meeting application (120), identifying (804) a target user, a speaking user, and a task based on the audio data, the speaking user indicating the task assigned to the target user in the audio data, generating (806) a message (318) that identifies the speaking user, the target user, and the task, the message (318) including textual content, and displaying (808) the message (318) in a chat pane (906) of a graphical user interface (902) of the online meeting application (120) during the session.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for resolving temporal ambiguities are presented. A natural language input may be received. A temporal component of the input may be identified. A determination may be made that the temporal component includes a conjunction that separates temporal meeting block alternatives. A temporal ambiguity may be identified in one of the meeting block alternatives. A plurality of syntax tree permutations may be generated for the meeting block alternative where the ambiguity was identified. A machine learning model that has been trained to identify a most relevant permutation for a given natural language input may be applied to each of the plurality of permutations. A temporal meeting block alternative corresponding to the most relevant permutation may be surfaced.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for assigning conference rooms are presented. A meeting request may be received by an electronic meeting service. Meeting fit scores may be calculated for the meeting request and one or more conference rooms. The meeting fit scores may be based on location, capacity, and/or audio-visual capabilities. The meeting request may be assigned to a conference room with a highest meeting fit score. A meeting request may be re-assigned to a different conference room based on a conference room becoming available that has a higher meeting fit score. A meeting request may be re-assigned to a different conference room based on characteristics of the meeting request being modified (e.g., fewer invitees, more invitees, different location specified), and thus, the meeting fit scores for conference rooms changing based on those modifications.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for identifying relevant content in a natural language input are presented. An email may be received. A machine learning model may be applied to the email. The machine learning model may have been trained to rank sentences based on their relevance to a schedule meeting task. The machine learning model may comprise: an embedding layer for generating an embedding for each word in the email; a distinct sentence aggregation layer for aggregating the embeddings for each word in the email into a distinct embedding for each of the sentences in the email; a contextual aggregation layer for aggregating each distinct embedding for each of the sentences into a contextual embedding for each of the sentences; and a scoring layer for scoring and ranking each of the sentences based on their relevance to the schedule meeting task.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for resolving temporal ambiguities are presented. A natural language input may be received. A temporal component of the input may be identified. A determination may be made that the temporal component includes a conjunction that separates temporal meeting block alternatives. A temporal ambiguity may be identified in one of the meeting block alternatives. A plurality of syntax tree permutations may be generated for the meeting block alternative where the ambiguity was identified. A machine learning model that has been trained to identify a most relevant permutation for a given natural language input may be applied to each of the plurality of permutations. A temporal meeting block alternative corresponding to the most relevant permutation may be surfaced.

Abstract: The disclosure relates to systems and methods for identifying entities related to a task in a natural language input. An entity detection model is provided which receives a natural language input. The entity detection model processes the natural language input using an entity encoder and an input encoder. The entity encoder identifies and encodes relevant entities while the input encoder generates a contextual encoding which represents contextual information associated with a relevant entity. The encoded entity and contextual encodings may then be combined and processed to generate a probability score for an identified entity. A negation constraint model is also disclosed. The negation constraint model receives the natural language input and the identified entities. The natural language input is analyzed to identify negation cues and determine if the negation cue is associated with an identified entity.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for prioritizing calendar events with artificial intelligence are presented. A request to schedule a new calendar event for a specified period may be received. A conflicting calendar event during the specified time period may be identified. An event priority score for the new calendar event may be compared with an event priority score for the conflicting calendar event, wherein the event priority scores are generated via application of a statistical machine learning model to a plurality of factors associated with the new calendar event and the conflicting calendar event. A selectable option to replace the conflicting calendar event with the new calendar event may be presented if the event priority score for the new calendar event is higher than the event priority score for the conflicting calendar event.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for assisting with scheduling a meeting are presented. A message comprising a plurality of sentences may be received. A hierarchical attention model may be utilized to identify a subset of sentences of the plurality of sentences that are relevant to a scheduling of the meeting. A subset of words in the subset of sentences that are potentially relevant to scheduling of the meeting may be identified based on relating to at least one meeting parameter. The subset of words may be split into a first group comprising words from the subset of words that are above a meeting relevance threshold value, and a second group comprising words from the subset of words that are below a meeting relevance threshold value. An automated action associated with scheduling the meeting may be caused to be performed.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for assigning conference rooms are presented. A meeting request may be received by an electronic meeting service. Meeting fit scores may be calculated for the meeting request and one or more conference rooms. The meeting fit scores may be based on location, capacity, and/or audio-visual capabilities. The meeting request may be assigned to a conference room with a highest meeting fit score. A meeting request may be re-assigned to a different conference room based on a conference room becoming available that has a higher meeting fit score. A meeting request may be re-assigned to a different conference room based on characteristics of the meeting request being modified (e.g., fewer invitees, more invitees, different location specified), and thus, the meeting fit scores for conference rooms changing based on those modifications.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for routing meeting requests by a digital assistant service are presented. A request to schedule a meeting between an invitee and a principal may be received by a digital assistant service, wherein the request is sent by an agent of the principal. The digital assistant service may determine that the agent is a delegate of the principal with scheduling authority. The digital assistant service may further determine that follow-up information for the meeting is required, and the digital assistant service may route an electronic message requesting the follow-up information directly the agent-delegate. Other aspects describe mechanisms for routing meeting requests from third parties directly to delegates, rather than sending those communications directly to principals.

Abstract: In non-limiting examples of the present disclosure, systems, methods and devices for identifying relevant content in a natural language input are presented. An email may be received. A machine learning model may be applied to the email. The machine learning model may have been trained to rank sentences based on their relevance to a schedule meeting task. The machine learning model may comprise: an embedding layer for generating an embedding for each word in the email; a distinct sentence aggregation layer for aggregating the embeddings for each word in the email into a distinct embedding for each of the sentences in the email; a contextual aggregation layer for aggregating each distinct embedding for each of the sentences into a contextual embedding for each of the sentences; and a scoring layer for scoring and ranking each of the sentences based on their relevance to the schedule meeting task.