Natural language summarization using chain of soft-thought prompting

1 . A system comprising:

computer-readable memory storing a soft-prompt generator and a summary generator, wherein the soft-prompt generator includes a first language model, and the summary generator includes a second language model; and

one or more processors configured by executable instructions to:

receive, via user interaction with a user interface, a content data structure comprising content to be summarized and a user request to generate a summary of the content, wherein the user request specifies an intent to generate the summary based on a categorical description, and wherein the user request is in a natural language format;

generate, via the soft-prompt generator, a first soft prompt based on the user request and one or more tokens selected to configure the soft-prompt generator, wherein the first soft prompt is in a feature vector or matrix format configured for processing by the summary generator to cause the summary generator to generate an output based on the intent, and wherein the one or more tokens comprise word embeddings;

transmit the first soft prompt to the summary generator, wherein the first soft prompt is prepended to the content;

generate, via the summary generator, a categorical description output based on the first soft prompt and the content, wherein the categorical description output includes one or more categorical descriptions that are each in a natural language format;

transmit the categorical description output to the soft-prompt generator;

generate, via the soft-prompt generator, a second soft prompt based on the categorical description output and the one or more tokens, wherein the second soft prompt is in a feature vector or matrix format;

transmit the second soft prompt to the summary generator, wherein the second soft prompt is prepended to the content;

generate, via the summary generator, the summary based on the second soft prompt and the content; and

transmit the summary.

2 . The system of claim 1 , wherein the content comprises at least one of image information, textual information, audio information, video information, or other information.

3 . The system of claim 1 , wherein the one or more categorical descriptions comprise at least one of topics, issues, outcomes, action items, or other subject matter.

4 . The system of claim 1 , wherein the second language model is substantially larger than the first language model.

5 . The system of claim 1 , wherein the first language model is fine-tuned to generate soft prompts by a training set that includes at least one of a plurality of content items, a plurality of categorical descriptions, and a plurality of content summaries that correspond to the plurality of content items and the plurality of categorical descriptions.

6 . A computer-implemented method comprising:

as implemented by a hardware processor configured to execute computer-executable instructions:

receiving, via user interaction with a user interface, content and a user request to generate a summary of the content, wherein the user request specifies an intent to generate the summary based on a categorical description, and wherein the user request is in a natural language format;

generating, via a first language model, a first encoded output based on the user request and a first set of tokens selected to configure the first language model, wherein the first encoded output is in a feature vector or matrix configured for processing by a second language model to cause the second language model to generate an output based on the intent, and wherein the first set of tokens comprise word embeddings generated based on the user request;

generating, via the second language model, a categorical description output based on the first encoded output and the content;

generating, via the first language model, a second encoded output based on the categorical description output and a second set of tokens selected to configure the first language model;

generating, via the second language model, the summary based on the second encoded output and the content; and

transmitting the summary.

7 . The computer-implemented method of claim 6 , wherein the content includes one or more portions that are in a natural language format.

8 . The computer-implemented method of claim 6 , wherein the first set of tokens used to generate the first encoded output is different from the second set of tokens used to generate the second encoded output.

9 . The computer-implemented method of claim 6 , wherein at least one of first language model and the second language model is a large language model (LLM).

10 . The computer-implemented method of claim 6 , wherein the first encoded output is a soft prompt that is generated by fine-tuning the user request based, at least in part, on the first set of tokens, to allow the second language model to accurately generate the categorical description output.

11 . The computer-implemented method of claim 6 , wherein the second encoded output is a soft prompt that is generated by fine-tuning the categorical description output based, at least in part, on the second set of tokens, to allow the second language model to accurately generate the summary.

12 . The computer-implemented method of claim 6 , wherein at least one of the first encoded output and the second encoded output is in a feature vector or matrix format.

13 . The computer-implemented method of claim 6 , wherein at least one of the first encoded output and the second encoded output is in a format that is not processable by a human.

14 . A system comprising computer-readable memory and one or more processors, wherein the system is configured to:

receive, via user interaction with a user interface, content and a user request to generate a summary of the content, wherein the user request specifies an intent to generate the summary based on a categorical description, and wherein the user request is in a natural language format;

generate, via a first language model, a first encoded output based on the user request and a first set of tokens selected to configure the first language model, wherein the first encoded output is in a feature vector or matrix configured for processing by a second language model to cause the second language model to generate an output based on the intent, and wherein the first set of tokens comprise word embeddings generated based on the user request;

generate, via the second language model, a categorical description output based on the first encoded output and the content;

generate, via the first language model, a second encoded output based on the categorical description output and a second set of tokens selected to configure the first language model; and

generate, via the second language model, the summary based on the second encoded output and the content.

15 . The system of claim 14 , wherein the categorical description output includes multiple categorical descriptions identified in the content.

16 . The system of claim 15 , after the second language model generates the categorical description output, a user is permitted to select one or more of the multiple categorical descriptions, in which the second language model is configured to generate the summary based on one or more categorical descriptions selected by the user.

17 . The system of claim 14 , wherein a length of the summary is less than a length of the content.

18 . The system of claim 14 , wherein a number of words in the summary is limited to less than or equal to three hundred words.

19 . The system of claim 14 , wherein a format of the content is different from a format of the summary.

20 . The system of claim 14 , wherein a narrative in the content is in first person, and a narrative in the summary is in third person.