Granted Patent B2

US 12,405,986 · App. 18/627,097 · Granted Sep 2, 2025

Inventors: Yun Huang (Champaign, IL); Yiren Liu (Champaign, IL)

Assignee: The Board of Trustees of the University of Illinois

G06F16/337G06F16/3344G06N20/00

Abstract

Claims (50)

1. A method comprising:

receiving initial text from a user;

applying the initial text to a classifier to select, from an enumerated set of incident types, a session incident type, wherein the session incident type is associated with a set of session incident properties;

applying the initial text and an indication of the set of session incident properties to a machine learning model to generate a value for a session incident property of the set of session incident properties;

based on the session incident property, selecting, from a list of queries, a query about a further session incident property whose value has not yet been generated by the machine learning model in relation to the initial text;

providing the query to the user;

receiving additional text from the user in response to the query; and

applying the additional text and the indication of the set of session incident properties to the machine learning model to generate an additional value for an additional session incident property of the set of session incident properties.

2. The method of claim 1 , wherein applying the initial text to the classifier comprises:

performing keyword matching between words of the initial text and sets of one or more keywords for each incident type in the enumerated set of incident types to determining that the initial text does not match any of the sets of one or more keywords for any of the enumerated set of incident types; and

responsive to determining that the initial text does not match any of the sets of one or more keywords, applying the initial text to a machine learning text classification model to select, from an enumerated set of incident types, the session incident type.

3. The method of claim 1 , wherein enumerated set of incident types include suspicious activity, drugs or alcohol, harassment or abuse, theft or lost item, and mental health.

4. The method of claim 1 , further comprising:

subsequent to applying the additional text and the indication of the set of session incident properties to the machine learning model, determining that values have been generated by the machine learning model in relation to text from the user for at least a threshold number of the set of session incident properties and responsively terminating a chat session with the user.

5. The method of claim 1 , wherein the machine learning model has been trained using chat logs that have been annotated to indicate portions of the chat logs that indicate values of incident properties and identities of the incident properties.

6. The method of claim 1 , wherein receiving initial text from a user comprises:

indicating, to the user, a query about whether an incident is ongoing;

responsively receiving, from the user, a response indicating that the incident is not ongoing; and

responsive to receiving the response indicating that the incident is not ongoing, indicating, to the user, a prompt to describe the incident, wherein the initial text is received from the user in response to the prompt to describe the incident.

7. The method of claim 1 , wherein the machine learning model comprises a transformer.

8. A method comprising:

receiving initial text from a user;

applying the initial text to a classifier to select, from an enumerated set of incident types, a session incident type, wherein the session incident type is associated with a set of session incident properties;

applying the initial text and an indication of the set of session incident properties to a machine learning model to generate an output that indicates (i) a value for a session incident property of the set of session incident properties and (ii) a query about a further session incident property of the set of session incident properties whose value has not yet been generated by the machine learning model in relation to the initial text;

based on the output, updating a list indicating which of the set of session incident properties have values that have been generated by the machine learning model in relation to the initial text;

providing the query to the user;

receiving additional text from the user in response to the query; and

applying the additional text, the indication of the set of session incident properties, and an representation of the list to the machine learning model to generate an additional output that indicates an additional value for an additional session incident property of the set of session incident properties.

9. The method of claim 8 , wherein applying the initial text to the classifier comprises:

performing keyword matching between words of the initial text and sets of one or more keywords for each incident type in the enumerated set of incident types to determining that the initial text does not match any of the sets of one or more keywords for any of the enumerated set of incident types; and

responsive to determining that the initial text does not match any of the sets of one or more keywords, applying the initial text to a machine learning text classification model to select, from an enumerated set of incident types, the session incident type.

10. The method of claim 8 , wherein enumerated set of incident types include suspicious activity, drugs or alcohol, harassment or abuse, theft or lost item, and mental health.

11. The method of claim 8 , further comprising:

subsequent to applying the additional text, the indication of the set of session incident properties, and the representation of the list to the machine learning model, determining that values have been generated by the machine learning model in relation to text from the user for at least a threshold number of the set of session incident properties and responsively terminating a chat session with the user.

12. The method of claim 8 , wherein the machine learning model has been trained using chat logs that have been annotated to indicate portions of the chat logs that indicate values of incident properties and identities of the incident properties.

13. The method of claim 8 , wherein the machine learning model has been trained using non-annotated chat logs to predict dispatcher responses to user text based on sets of prior user text and dispatcher queries.

14. The method of claim 8 , wherein receiving initial text from a user comprises:

indicating, to the user, a query about whether an incident is ongoing;

responsively receiving, from the user, a response indicating that the incident is not ongoing; and

responsive to receiving the response indicating that the incident is not ongoing, indicating, to the user, a prompt to describe the incident, wherein the initial text is received from the user in response to the prompt to describe the incident.

15. The method of claim 8 , further comprising:

determining, based on the initial text and the additional text, a user emotional state, wherein applying the additional text, the indication of the set of session incident properties, and the representation of the list to the machine learning model to generate the additional output comprises applying the additional text, the indication of the set of session incident properties, the representation of the list to the machine learning model, and an indication of the user emotional state to generate the additional output.

16. The method of claim 15 , wherein determining the user emotional state based on the initial text and the additional text comprises applying the initial text and the additional text to a sentiment prediction model.

17. The method of claim 16 , wherein the machine learning model has been trained using chat logs that have been augmented by applying the sentiment prediction model thereto to annotate user text in the chat logs with user emotional state information predicted from the user text.

18. The method of claim 8 , wherein the machine learning model has been trained using chat logs that have been augmented by applying the chat logs to a generative machine learning language model to generate therefrom simulated chat logs that exhibit increased emotional support.

19. The method of claim 8 , wherein receiving the initial text, applying the initial text to a classifier, providing the query to the user, and receiving the additional text are performed by a first computational system, and wherein applying the initial text and the indication of the set of session incident properties to the machine learning model comprises:

transmitting, from the first computational system to a second computational system that is remote from the first computational system, an indication of the initial text and the set of session incident properties;

applying, by the second computational system, the initial text and the indication of the set of session incident properties to the machine learning model to generate the output; and

transmitting, from the second computational system to the first computational system, an indication of the output.

20. The method of claim 8 , wherein the machine learning model comprises a large language model.

Assignments (1)

Continuity (2)

References Cited (269)