Determining dialog states for language models
Systems, methods, devices, and other techniques are described herein for determining dialog states that correspond to voice inputs and for biasing a language model based on the determined dialog states. In some implementations, a method includes receiving, at a computing system, audio data that indicates a voice input and determining a particular dialog state, from among a plurality of dialog states, which corresponds to the voice input. A set of n-grams can be identified that are associated with the particular dialog state that corresponds to the voice input. In response to identifying the set of n-grams that are associated with the particular dialog state that corresponds to the voice input, a language model can be biased by adjusting probability scores that the language model indicates for n-grams in the set of n-grams. The voice input can be transcribed using the adjusted language model.
1. A computer-implemented method, comprising:
receiving, at a computing system, audio data that indicates a first voice input that was provided to a computing device;
determining that the first voice input is part of a voice dialog that includes a plurality of pre-defined dialog states arranged to receive a series of voice inputs related to a particular task, wherein each dialog state is mapped to:
(i) a set of display data characterizing content that is designated for display when voice inputs for the dialog state are received, and
(ii) a set of n-grams;
receiving, at the computing system, first display data that characterizes content that was displayed on a screen of the computing device when the first voice input was provided to the computing device;
selecting, by the computing system, a particular dialog state of the plurality of pre-defined dialog states that corresponds to the first voice input, including determining a match between the first display data and the corresponding set of display data that is mapped to the particular dialog state;
biasing a language model by adjusting probability scores that the language model indicates for n-grams in the corresponding set of n-grams that are mapped to the particular dialog state; and
transcribing the voice input using the biased language model.
2. The computer-implemented method of claim 1 , further comprising:
receiving a second voice input at the computing system;
selecting a second particular dialog state, from among the plurality of pre-defined dialog states, that corresponds to the second voice input; and
biasing the language model by adjusting probability scores that the language model indicates for n-grams in the corresponding set of n-grams that are mapped to the second particular dialog state, wherein the set of n-grams that are mapped to the second particular dialog state are different than the set of n-grams that are mapped to the particular dialog state that corresponds to the first voice input.
3. The computer-implemented method of claim 1 , wherein selecting the particular dialog state that corresponds to the first voice input comprises:
identifying a second particular dialog state, from among the plurality of dialog states, that corresponds to a second voice input that was provided to the computing device before the first voice input for the same particular task as the first voice input; and
selecting the particular dialog state that corresponds to the first voice input based on (i) an indication that the second particular dialog state is a preceding dialog state in an order of dialog states for the plurality of pre-defined dialog states and (ii) the first display data that characterizes the content that was displayed on the screen of the computing device when the first voice input was provided to the computing device.
4. The computer-implemented method of claim 1 , wherein selecting the particular dialog state that corresponds to the first voice input further comprises:
generating a transcription of the first voice input; and
determining a match between one or more n-grams that occur in the transcription of the first voice input and one or more n-grams in the corresponding set of n-grams that are mapped to the particular dialog state.
5. The computer-implemented method of claim 4 , wherein determining the match comprises determining a semantic relationship between the one or more n-grams that occur in the transcription of the first voice input and the one or more n-grams in the corresponding set of n-grams that are mapped to the particular dialog state.
6. The computer-implemented method of claim 1 , further comprising:
receiving context information associated with the first voice input other than the first display data; and
selecting the particular dialog state that corresponds to the first voice input further based on the second context information associated with the voice input.
7. The computer-implemented method of claim 1 , further comprising receiving, at the computing system, an application identifier that indicates an application to which the first voice input was directed at the computing device,
wherein the plurality of dialog states pertain to an application-specific task for the application to which the first voice input was directed.
8. The computer-implemented method of claim 1 , wherein the computing device is separate from the computing system.
9. The computer-implemented method of claim 1 , wherein the first display data comprises a hash value that results from processing data that represents information about the content that was displayed on the screen of the computing device when the first voice input was provided to the computing device.
10. A computing system comprising:
one or more processors; and
one or more computer-readable media having instructions stored thereon that, when executed, cause performance of operations comprising:
receiving audio data that indicates a first voice input that was provided to a computing device;
determining that the first voice input is part of a voice dialog that includes a plurality of pre-defined dialog states arranged to receive a series of voice inputs related to a particular task, wherein each dialog state is mapped to:
(i) a set of display data characterizing content that is designated for display when voice inputs for the dialog state are received, and
(ii) a set of n-grams;
receiving first display data that characterizes content that was displayed on a screen of the computing device when the first voice input was provided to the computing device;
selecting a particular dialog state of the plurality of pre-defined dialog states; that corresponds to the first voice input, including determining a match between the first display data and the corresponding set of display data that is mapped to the particular dialog state;
biasing a language model by adjusting probability scores that the language model indicates for n-grams in the corresponding set of n-grams that are mapped to the particular dialog state; and
transcribing the voice input using the biased language model.
11. The computing system of claim 10 , wherein the operations further comprise:
receiving a second voice input;
selecting a second particular dialog state, from among the plurality of pre-defined dialog states, that corresponds to the second voice input; and
biasing the language model by adjusting probability scores that the language model indicates for n-grams in the corresponding set of n-grams that are mapped to the second particular dialog state, wherein the set of n-grams that are mapped to the second particular dialog state are different than the set of n-grams that are mapped to the particular dialog state that corresponds to the first voice input.
12. One or more non-transitory computer-readable media having instructions stored thereon that, when executed by one or more processors, cause the processors to perform operations comprising:
receiving audio data that indicates a first voice input that was provided to a computing device;
determining that the first voice input is part of a voice dialog that includes a plurality of pre-defined dialog states arranged to receive a series of voice inputs related to a particular task, wherein each dialog state is mapped to:
(i) a set of display data characterizing content that is designated for display when voice inputs for the dialog state are received, and
(ii) a set of n-grams;
receiving first display data that characterizes content that was displayed on a screen of the computing device when the first voice input was provided to the computing device;
selecting a particular dialog state of the plurality of pre-defined dialog states that corresponds to the first voice input, including determining a match between the first display data and the corresponding set of display data that is mapped to the particular dialog state;
biasing a language model by adjusting probability scores that the language model indicates for n-grams in the corresponding set of n-grams that are mapped to the particular dialog state; and
transcribing the voice input using the adjusted biased language model.