Abstract: Implementations relate to using deep reinforcement learning to train a model that can be utilized, at each of a plurality of time steps, to determine a corresponding robotic action for completing a robotic task. Implementations additionally or alternatively relate to utilization of such a model in controlling a robot. The robotic action determined at a given time step utilizing such a model can be based on: current sensor data associated with the robot for the given time step, and free-form natural language input provided by a user. The free-form natural language input can direct the robot to accomplish a particular task, optionally with reference to one or more intermediary steps for accomplishing the particular task. For example, the free-form natural language input can direct the robot to navigate to a particular landmark, with reference to one or more intermediary landmarks to be encountered in navigating to the particular landmark.

Abstract: Implementations relate to using deep reinforcement learning to train a model that can be utilized, at each of a plurality of time steps, to determine a corresponding robotic action for completing a robotic task. Implementations additionally or alternatively relate to utilization of such a model in controlling a robot. The robotic action determined at a given time step utilizing such a model can be based on: current sensor data associated with the robot for the given time step, and free-form natural language input provided by a user. The free-form natural language input can direct the robot to accomplish a particular task, optionally with reference to one or more intermediary steps for accomplishing the particular task. For example, the free-form natural language input can direct the robot to navigate to a particular landmark, with reference to one or more intermediary landmarks to be encountered in navigating to the particular landmark.

Abstract: Determining a dialog state of an electronic dialog that includes an automated assistant and at least one user, and performing action(s) based on the determined dialog state. The dialog state can be represented as one or more slots and, for each of the slots, one or more candidate values for the slot and a corresponding score (e.g., a probability) for each of the candidate values. Candidate values for a slot can be determined based on language processing of user utterance(s) and/or system utterance(s) during the dialog. In generating scores for candidate value(s) of a given slot at a given turn of an electronic dialog, various features are determined based on processing of the user utterance and the system utterance using a memory network. The various generated features can be processed using a scoring model to generate scores for candidate value(s) of the given slot at the given turn.

Abstract: The present disclosure is generally directed to methods, apparatus, and computer-readable media (transitory and non-transitory) for learning to automatically navigate interactive web documents and/or websites. More particularly, various approaches are presented for training various deep Q network (DQN) agents to perform various tasks associated with reinforcement learning, including hierarchical reinforcement learning, in challenging web navigation environments with sparse rewards and large state and action spaces. These agents include a web navigation agent that can use learned value function(s) to automatically navigate through interactive web documents, as well as a training agent, referred to herein as a “meta-trainer,” that can be trained to generate synthetic training examples. Some approaches described herein may be implemented when expert demonstrations are available. Other approaches described herein may be implemented when expert demonstrations are not available.

Abstract: Determining a dialog state of an electronic dialog that includes an automated assistant and at least one user, and performing action(s) based on the determined dialog state. The dialog state can be represented as one or more slots and, for each of the slots, one or more candidate values for the slot and a corresponding score (e.g., a probability) for each of the candidate values. Candidate values for a slot can be determined based on language processing of user utterance(s) and/or system utterance(s) during the dialog. In generating scores for candidate value(s) of a given slot at a given turn of an electronic dialog, various features are determined based on processing of the user utterance and the system utterance using a memory network. The various generated features can be processed using a scoring model to generate scores for candidate value(s) of the given slot at the given turn.

Abstract: The present disclosure is generally directed to methods, apparatus, and computer-readable media (transitory and non-transitory) for learning to automatically navigate interactive web documents and/or websites. More particularly, various approaches are presented for training various deep Q network (DQN) agents to perform various tasks associated with reinforcement learning, including hierarchical reinforcement learning, in challenging web navigation environments with sparse rewards and large state and action spaces. These agents include a web navigation agent that can use learned value function(s) to automatically navigate through interactive web documents, as well as a training agent, referred to herein as a “meta-trainer,” that can be trained to generate synthetic training examples. Some approaches described herein may be implemented when expert demonstrations are available. Other approaches described herein may be implemented when expert demonstrations are not available.

Abstract: Described herein is a system for improving sentiment detection and/or recognition using multiple inputs. For example, an autonomously motile device is configured to generate audio data and/or image data and perform sentiment detection processing. The device may process the audio data and the image data using a multimodal temporal attention model to generate sentiment data that estimates a sentiment score and/or a sentiment category. In some examples, the device may also process language data (e.g., lexical information) using the multimodal temporal attention model. The device can adjust its operations based on the sentiment data. For example, the device may improve an interaction with the user by estimating the user's current emotional state, or can change a position of the device and/or sensor(s) of the device relative to the user to improve an accuracy of the sentiment data.

Abstract: A dialog generator receives data corresponding to desired dialog, such as application programming interface (API) information and sample dialog. A first model corresponding to an agent simulator and a second model corresponding to a user simulator take turns creating a plurality of dialog outlines of the desired dialog. The dialog generator may determine that one or more additional APIs are relevant to the dialog and may create further dialog outlines related thereto. The dialog outlines are converted to natural dialog to generate the dialog.

Abstract: A dialog manager receives text data corresponding to a dialog with a user. Entities represented in the text data are identified. Context data relating to the dialog is maintained, which may include prior dialog, prior API calls, user profile information, or other data. Using the text data and the context data, an N-best list of one or more dialog models is selected to process the text data. After processing the text data, the outputs of the N-best models are ranked and a top-scoring output is selected. The top-scoring output may be an API call and/or an audio prompt.

Abstract: The present disclosure is generally directed to methods, apparatus, and computer-readable media (transitory and non-transitory) for learning to automatically navigate interactive web documents and/or websites. More particularly, various approaches are presented for training various deep Q network (DQN) agents to perform various tasks associated with reinforcement learning, including hierarchical reinforcement learning, in challenging web navigation environments with sparse rewards and large state and action spaces. These agents include a web navigation agent that can use learned value function(s) to automatically navigate through interactive web documents, as well as a training agent, referred to herein as a “meta-trainer,” that can be trained to generate synthetic training examples. Some approaches described herein may be implemented when expert demonstrations are available. Other approaches described herein may be implemented when expert demonstrations are not available.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media for dialogue systems. A transcription of a user utterance is obtained. The transcription of the utterance is tokenized to identify multiple tokens for the utterance. Token-level utterance encodings corresponding to different tokens of the transcription are generated. A system action encoding from data indicating system actions previously performed by the dialogue system are generated. A dialogue context vector based on the utterance encoding and the system action encoding are generated. The token-level utterance encodings, the system action encoding, and the dialogue context vector are processed using a slot tagger to produce token-level output vectors. A limited set of candidate token classifications for the tokens of the user utterance are determined based on the token-level utterance encodings. A response for output is provided in response to the user utterance.

Abstract: Implementations relate to using deep reinforcement learning to train a model that can be utilized, at each of a plurality of time steps, to determine a corresponding robotic action for completing a robotic task. Implementations additionally or alternatively relate to utilization of such a model in controlling a robot. The robotic action determined at a given time step utilizing such a model can be based on: current sensor data associated with the robot for the given time step, and free-form natural language input provided by a user. The free-form natural language input can direct the robot to accomplish a particular task, optionally with reference to one or more intermediary steps for accomplishing the particular task. For example, the free-form natural language input can direct the robot to navigate to a particular landmark, with reference to one or more intermediary landmarks to be encountered in navigating to the particular landmark.

Abstract: Determining a dialog state of an electronic dialog that includes an automated assistant and at least one user, and performing action(s) based on the determined dialog state. The dialog state can be represented as one or more slots and, for each of the slots, one or more candidate values for the slot and a corresponding score (e.g., a probability) for each of the candidate values. Candidate values for a slot can be determined based on language processing of user utterance(s) and/or system utterance(s) during the dialog. In generating scores for candidate value(s) of a given slot at a given turn of an electronic dialog, various features are determined based on processing of the user utterance and the system utterance using a memory network. The various generated features can be processed using a scoring model to generate scores for candidate value(s) of the given slot at the given turn.

Abstract: Determining a dialog state of an electronic dialog that includes an automated assistant and at least one user, and performing action(s) based on the determined dialog state. The dialog state can be represented as one or more slots and, for each of the slots, one or more candidate values for the slot and a corresponding score (e.g., a probability) for each of the candidate values. Candidate values for a slot can be determined based on language processing of user utterance(s) and/or system utterance(s) during the dialog. In generating scores for candidate value(s) of a given slot at a given turn of an electronic dialog, various features are determined based on processing of the user utterance and the system utterance using a memory network. The various generated features can be processed using a scoring model to generate scores for candidate value(s) of the given slot at the given turn.

Abstract: Determining a dialog state of an electronic dialog that includes an automated assistant and at least one user, and performing action(s) based on the determined dialog state. The dialog state can be represented as one or more slots and, for each of the slots, one or more candidate values for the slot and a corresponding score (e.g., a probability) for each of the candidate values. Candidate values for a slot can be determined based on language processing of user utterance(s) and/or system utterance(s) during the dialog. In generating scores for candidate value(s) of a given slot at a given turn of an electronic dialog, various features are determined based on processing of the user utterance and the system utterance using a memory network. The various generated features can be processed using a scoring model to generate scores for candidate value(s) of the given slot at the given turn.

Abstract: Prosodic features are used for discriminating computer-directed speech from human-directed speech. Statistics and models describing energy/intensity patterns over time, speech/pause distributions, pitch patterns, vocal effort features, and speech segment duration patterns may be used for prosodic modeling. The prosodic features for at least a portion of an utterance are monitored over a period of time to determine a shape associated with the utterance. A score may be determined to assist in classifying the current utterance as human directed or computer directed without relying on knowledge of preceding utterances or utterances following the current utterance. Outside data may be used for training lexical addressee detection systems for the H-H-C scenario. H-C training data can be obtained from a single-user H-C collection and that H-H speech can be modeled using general conversational speech. H-C and H-H language models may also be adapted using interpolation with small amounts of matched H-H-C data.

Abstract: Semantic entity relation detection classifier training implementations are presented that are generally used to train a semantic entity relation detection classifier to identify relations expressed in a natural language query. In one general implementation, queries are found in a search query click log that exhibit relations and entity types found in a semantic knowledge graph. Explicit relations are inferred from the found queries and an explicit relations data set is generated that includes queries associated with the inferred explicit relations. In addition, implicit relations are inferred from the found queries and an implicit relations data set is generated that includes queries associated with the inferred implicit relations. A semantic entity relation detection classifier is then trained using the explicit and implicit data sets.

Abstract: Techniques are described related to turn-based reinforcement learning for dialog management. In various implementations, dialog states and corresponding responsive actions generated during a multi-turn human-to-computer dialog session may be obtained. A plurality of turn-level training instances may be generated, each including: a given dialog state of the plurality of dialog states at an outset of a given turn of the human-to-computer dialog session; and a given responsive action that was selected based on the given dialog state. One or more of the turn-level training instances may further include a turn-level feedback value that reflects on the given responsive action selected during the given turn. A reward value may be generated based on an outcome of the human-to-computer dialog session. The dialog management policy model may be trained based on turn-level feedback values of the turn-level training instance(s) and the reward value.

Abstract: Techniques are described related to turn-based reinforcement learning for dialog management. In various implementations, dialog states and corresponding responsive actions generated during a multi-turn human-to-computer dialog session may be obtained. A plurality of turn-level training instances may be generated, each including: a given dialog state of the plurality of dialog states at an outset of a given turn of the human-to-computer dialog session; and a given responsive action that was selected based on the given dialog state. One or more of the turn-level training instances may further include a turn-level feedback value that reflects on the given responsive action selected during the given turn. A reward value may be generated based on an outcome of the human-to-computer dialog session. The dialog management policy model may be trained based on turn-level feedback values of the turn-level training instance(s) and the reward value.

Abstract: Structured web pages are accessed and parsed to obtain implicit annotation for natural language understanding tasks. Search queries that hit these structured web pages are automatically mined for information that is used to semantically annotate the queries. The automatically annotated queries may be used for automatically building statistical unsupervised slot filling models without using a semantic annotation guideline. For example, tags that are located on a structured web page that are associated with the search query may be used to annotate the query. The mined search queries may be filtered to create a set of queries that is in a form of a natural language query and/or remove queries that are difficult to parse. A natural language model may be trained using the resulting mined queries. Some queries may be set aside for testing and the model may be adapted using in-domain sentences that are not annotated.