Language model arbitration for natural language processing

1 . A computer-implemented method comprising:

receiving, by a first input device, a first natural language input;

generating, using an automatic speech recognition (ASR) component, first ASR output data comprising a first text transcription of the first natural language input;

determining first device identifier data that identifies the first input device;

determining first state data describing a state of at least the first input device;

inputting the first ASR output data, the first device identifier data, and the first state data into a machine learning classifier model, wherein the machine learning classifier model is trained to select from among a large language model (LLM) based processing flow and an intent-based processing flow for processing natural language inputs;

selecting, by the machine learning classifier model, the LLM-based processing flow for processing the first natural language input;

generating, by the LLM-based processing flow using the first ASR output data, first executable data; and

sending the first executable data to the first input device or another device, wherein the first input device or another device is configured to execute the first executable data to perform a first action.

2 . The computer-implemented method of claim 1 , further comprising:

determining first dialog session data indicating that a second natural language input of a current dialog session that includes the first natural language input was routed to the LLM-based processing flow for processing; and

generating second state data representing the first dialog session data, wherein the second state data is input into the machine learning classifier model, and the LLM-based processing flow is selected for processing the first natural language input based at least in part on the second state data.

3 . The computer-implemented method of claim 1 , further comprising:

determining first domain data associated with the LLM-based processing flow, the first domain data indicating at least one of a capability or natural language processing domain of the LLM-based processing flow;

determining a correspondence between the first domain data and a subject of the first natural language input; and

selecting the LLM-based processing flow for processing the first natural language input based on the correspondence.

4 . A method comprising:

receiving, from a first input device, a first natural language input;

determining first automatic speech recognition (ASR) output data representing the first natural language input;

selecting, by a first machine learning model using the first ASR output data, a first natural language processing flow, wherein the first machine learning model is trained to select between at least the first natural language processing flow and a second natural language processing flow, wherein the first natural language processing flow is a large language model (LLM)-based processing flow and the second natural language processing flow is an intent-based processing flow;

processing the first ASR output data using the first natural language processing flow; and

generating, by the first natural language processing flow, first executable data; and

sending the first executable data to the first input device or a second device, wherein the first input device or the second device executes the first executable data to perform a first action.

5 . The method of claim 4 , further comprising:

determining first state data indicating a current dialog session including the first natural language input, wherein the first state data further indicates that the first natural language processing flow was selected to process at least one previous natural language input of the current dialog session; and

inputting the first state data into the first machine learning model, wherein the first machine learning model selects the first natural language processing flow based at least in part on the first state data.

6 . The method of claim 4 , further comprising:

determining a subject of the first natural language input; and

determining a correspondence between the subject of the first natural language input and a first natural language processing domain associated with the first natural language processing flow, wherein the first machine learning model selects the first natural language processing flow based at least in part on the correspondence.

7 . The method of claim 4 , further comprising:

receiving, by the first machine learning model, a first training instance comprising a second natural language input, first data indicating that the first natural language processing flow was selected for processing the second natural language input, and first result data assessing performance of the first natural language processing flow in processing the second natural language input;

receiving, by the first machine learning model, a second training instance comprising a third natural language input, second data indicating that the second natural language processing flow was selected for processing the third natural language input, and second result data assessing performance of the second natural language processing flow in processing the third natural language input; and

updating parameters of the first machine learning model using the first training instance and the second training instance to optimize performance of selected natural language processing flows for natural language inputs.

8 . The method of claim 4 , further comprising:

generating first flag data indicating that the first natural language processing flow was used to process the first natural language input;

storing the first flag data in memory in association with a first time to live (TTL) value;

receiving, by the first input device, a second natural language input prior to an expiration of the first TTL value; and

selecting the first natural language processing flow for processing the second natural language input based at least in part on the first flag data.

9 . The method of claim 8 , further comprising extending the first TTL value based at least in part on receipt of the second natural language input prior to the expiration of the first TTL value.

10 . The method of claim 4 , further comprising determining first state data describing a current state of at least the first input device, wherein the first machine learning model selects the first natural language processing flow further based at least in part on the first state data.

11 . The method of claim 4 , further comprising determining first device type data indicating a type of the first input device, wherein the first machine learning model selects the first natural language processing flow further based at least in part on the first device type data.

12 . The method of claim 4 , wherein the first machine learning model is configured to select, from among a set of natural language processing flows comprising the first natural language processing flow and the second natural language processing flow, a subset of natural language processing flows for processing a second natural language input.

13 . A system comprising:

at least one processor; and

non-transitory computer-readable memory storing instructions that, when executed by the at least one processor, are effective to:

receive, from a first input device, a first natural language input;

determine first automatic speech recognition (ASR) output data representing the first natural language input;

select, by a first machine learning model using the first ASR output data, a first natural language processing flow, wherein the first machine learning model is trained to select between at least the first natural language processing flow and a second natural language processing flow, wherein the first natural language processing flow is a large language model (LLM)-based processing flow and the second natural language processing flow is an intent-LLM-based processing flow;

process the first ASR output data using the first natural language processing flow;

generate, by the first natural language processing flow, first executable data; and send the first executable data to the first input device or a second device, wherein the first input device or the second device executes the first executable data to perform a first action.

14 . The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine first state data indicating a current dialog session including the first natural language input, wherein the first state data further indicates that the first natural language processing flow was selected to process at least one previous natural language input of the current dialog session; and

input the first state data into the first machine learning model, wherein the first machine learning model selects the first natural language processing flow based at least in part on the first state data.

15 . The system of claim 14 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine a subject of the first natural language input; and

determine a correspondence between the subject of the first natural language input and a first natural language processing domain associated with the first natural language processing flow, wherein the first machine learning model selects the first natural language processing flow based at least in part on the correspondence.

16 . The system of claim 15 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

receive, by the first machine learning model, a first training instance comprising a second natural language input, first data indicating that the first natural language processing flow was selected for processing the second natural language input, and first result data assessing performance of the first natural language processing flow in processing the second natural language input;

receive, by the first machine learning model, a second training instance comprising a third natural language input, second data indicating that the second natural language processing flow was selected for processing the third natural language input, and second result data assessing performance of the second natural language processing flow in processing the third natural language input; and

update parameters of the first machine learning model using the first training instance and the second training instance to optimize performance of selected natural language processing flows for natural language inputs.

17 . The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

generate first flag data indicating that the first natural language processing flow was used to process the first natural language input;

store the first flag data in memory in association with a first time to live (TTL) value;

receive, by the first input device, a second natural language input prior to an expiration of the first TTL value; and

select the first natural language processing flow for processing the second natural language input based at least in part on the first flag data.

18 . The system of claim 17 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

extend the first TTL value based at least in part on receipt of the second natural language input prior to the expiration of the first TTL value.

19 . The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine first state data describing a current state of at least the first input device, wherein the first machine learning model selects the first natural language processing flow further based at least in part on the first state data.

20 . The system of claim 13 , the non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to:

determine first device type data indicating a type of the first input device, wherein the first machine learning model selects the first natural language processing flow further based at least in part on the first device type data.