Microsoft Technology Licensing, LLC

Alexandre Rochette

Ruhi Sarikaya

Young-Bum Kim

Omar Zia Khan

Jean-Philippe Robichaud

Paul Anthony Crook

Vipul Agarwal

Marius Alexandru Marin

Nikhil Holenarsipur Ramesh

Asli Celikyilmaz

Puyang Xu

Tasos Anastasakos

Daniel Boies

Vasiliy Radostev

Zhaleh Feizollahi

Timothy J Hazen

Gabriel Arien Desgarennes

Soyoung Peraud

Niel Chah

Abhishek Kumar

A classifier may be trained with less than all datasets manually annotated with labels. A small subset of verbatims may be manually labeled with topic labels as seeds. Data augmentations can be used to acquire seed verbatim sets for known topics and to assign temporary pseudo labels to the rest of the verbatims based on their vector space proximity to the labeled seed verbatims. The training may involve classification epochs during which embeddings are updated with the assumption that the pseudo labels are ground-truth labels. The training may also involve labeling epochs during which the updated embeddings are used to update the vectors corresponding to the verbatims, and pseudo labels are updated based on updated vector coordinates in the vector space. As the training process progresses through the epochs, the embeddings will converge.

Classifier assistance using domain-trained embedding

Conversational understanding systems allow users to conversationally interface with a computing device. In examples, a query may be received that includes a request for execution of a task. A data exchange task definition may be accessed. The data exchange task definition assists a conversational understanding system in managing task state tracking for information needed for task execution. Using the data exchange task definition, a per-turn policy for interacting with the user computing device is generated based on the state of a dialogue with a computing device and an evaluation of a process flow chart provided by a task owner resource. The task owner resource may be independent from the conversational understanding system. A response to the query may be generated and output based on the per-turn policy. In examples, the per-turn policy is used to generate one or more responses during a dialogue with a user via a computing device.

Policy authoring for task state tracking during dialogue

Aspects herein provide third party application authors with a user interface authoring platform that automates and simplifies a task definition process while also providing the ability to leverage pre-existing language understanding models and canonicalization and resolution modules that are provided by the operating system on which the CU system resides or as provided by other third parties. In particular, the present disclosure provides a method and system for authoring a task using a user interface authoring platform.

Natural language task completion platform authoring for third party experiences

Multi-turn cross-domain natural language understanding (NLU) systems and platforms for building the multi-turn cross-domain NLU system are provided. Further, methods for using and building the multi-turn cross-domain NLU system are provided. More specifically, the multi-turn cross-domain NLU system supports multi-turn bot/agent/application scenarios for new domains without having to select a task definition and/or define a new schema during the building of the NLU system. Accordingly, the platform for building the multi-turn cross-domain NLU system that does not require the builder to select a task and/or build a schema for a selected task provides an easy to use, cost effective, and efficient service for building a NLU system. Further, the multi-turn cross-domain NLU system provides a more versatile NLU system than previously utilized NLU systems that were trained for and limited to a selected task and/or domain.

Multi-turn cross-domain natural language understanding systems, building platforms, and methods

A computer-implemented technique is described for facilitating the creation of a language understanding (LU) component for use with an application. The technique allows a developer to select a subset of parameters from a larger set of parameters. The subset of parameters pertains to a LU scenario to be handled by the application. The larger set of parameters pertains to a plurality of LU scenarios handled by an already-existing generic LU model. The technique creates a constrained LU component that is based on the subset of parameters in conjunction with the generic LU model. At runtime, the constrained LU component interprets input language items using the generic LU model in a manner that is constrained by the subset of parameters that have been selected, to provide an output result. The technique also allows the developer to create new rules and/or supplemental models.

Tool for facilitating the development of new language understanding scenarios

Methods and systems are provided for contextual language understanding. A natural language expression may be received at a single-turn model and a multi-turn model for determining an intent of a user. For example, the single-turn model may determine a first prediction of at least one of a domain classification, intent classification, and slot type of the natural language expression. The multi-turn model may determine a second prediction of at least one of a domain classification, intent classification, and slot type of the natural language expression. The first prediction and the second prediction may be combined to produce a final prediction relative to the intent of the natural language expression. An action may be performed based on the final prediction of the natural language expression.

Contextual language understanding for multi-turn language tasks

Systems and methods for building a dialog-state specific multi-turn contextual language understanding system are provided. More specifically, the systems and methods infer or are configured to infer a state-specific schema and/or state-specific rules from a formed single-shot language understanding model and/or a single-shot rule set. As such, the systems and methods only require the information necessary to form a single-shot language understanding model and/or a single-shot rule set from a builder to form or build the dialog-state specific multi-turn contextual language understanding system. Accordingly, the systems and methods for building a dialog-state specific multi-turn contextual language understanding system reduce the expertise, time, and resources necessary to build a dialog-state specific multi-turn contextual language understanding system for an application when compared to systems and methods that require further input from the builder than necessary to build a single-shot language understanding system.

Systems and methods for building state specific multi-turn contextual language understanding systems

User input expressed as text may be analyzed for determining a type of response, such as an application response, and/or determining a type of task that is requested by the user input. Entity representations may be identified, classified and/or or tagged based on a type of response, type of task and/or a set of entity types. A surface form of an entity, ambiguous entity representation and/or other type of expression within the user input may be resolved, normalized and/or mapped to a normalized value. Normalizing entities and/or entity attributes may involve using a set of normalization rules, a lookup table, one or more machined learned methods, and/or an entity normalization index that associates entities with alternate surface forms derived from web corpora. The normalized value may be used to construct a request to a structured knowledge source and/or an application.

Entity and attribute resolution in conversational applications

Embodiments provide for tracking a partial dialog state as part of managing a dialog state space, but the embodiments are not so limited. A method of an embodiment jointly models partial state update and named entity recognition using a sequence-based classification or other model, wherein recognition of named entities and a partial state update can be performed in a single processing stage at runtime to generate a distribution over partial dialog states. A system of an embodiment is configured to generate a distribution over partial dialog states at runtime in part using a sequence classification decoding or other algorithm to generate one or more partial dialog state hypothesis and/or a confidence score or measure associated with each hypothesis. Other embodiments are included.

Growing community of inventors

Montreal, Canada

Alexandre Rochette