APPLICATION LABEL TAGGING USING LARGE LANGUAGE MODELS
A graph includes nodes representing applications and tags describing subjective qualities of the applications. The system generates tags for applications using a large language model (LLM). The system receives the name of an application and generates a prompt for an LLM based on the name. The prompt includes a request for one or more tags associated with the application. The system provides the prompt to the LLM for execution and receives, as output from the LLM, candidate tags. The system inputs the candidate tags into a classifier trained to classify candidate tags into known tags, tags that already exist in a graph. The system receives, as output from the classifier, known tags. The system updates the graph to include a node corresponding to the application, the node linked to the known tags with one or more edges.
1 . A method for generating tagged interests for an application, the method comprising:
receiving a name of an application;
generating a prompt for a large language model (LLM) based on the name of the application, wherein the prompt includes a request for one or more tags associated with the application, the one or more tags describing tagged interests associated with the application;
providing the prompt to the LLM for execution;
receiving, as output from the LLM, a plurality of candidate tags;
inputting the plurality of candidate tags into a classifier, the classifier trained to classify candidate tags into known tags, wherein known tags are tags that already exist in a graph;
receiving, as output from the classifier, a plurality of known tags; and
updating the graph to include a node corresponding to the application, the node linked to the plurality of known tags with one or more edges.
2 . The method of claim 1 , further comprising sanitizing the name of the application by removing non-text characters from the name and translating the name into English.
3 . The method of claim 1 , wherein the classifier is one of a supervised machine learning model or an LLM that performs zero-shot classification.
4 . The method of claim 1 , wherein receiving the name of an application comprises extracting the name of the application as a signal of the application.
5 . The method of claim 1 , wherein the prompt is based on a portion of the graph.
6 . The method of claim 1 , wherein generating the prompt for the LLM comprises selecting the prompt from a set of pre-generated prompts.
7 . The method of claim 1 , wherein receiving, as output from the classifier, the plurality of known tags further comprises:
receiving, as output from the classifier, an unknown tag, wherein an unknown tag is a tag that does not already exist in the graph; and
updating the graph to add the unknown tag.
8 . A non-transitory computer-readable medium comprising memory with instructions encoded thereon, the instructions, when executed by one or more processors, causing the one or more processors to perform operations, the instructions comprising instructions to:
receive a name of an application;
generate a prompt for a large language model (LLM) based on the name of the application, wherein the prompt includes a request for one or more tags associated with the application, the one or more tags describing tagged interests associated with the application;
provide the prompt to the LLM for execution;
receive, as output from the LLM, a plurality of candidate tags;
input the plurality of candidate tags into a classifier, the classifier trained to classify candidate tags into known tags, wherein known tags are tags that already exist in a graph;
receive, as output from the classifier, a plurality of known tags; and
update the graph to include a node corresponding to the application, the node linked to the plurality of known tags with one or more edges.
9 . The non-transitory computer-readable medium of claim 8 , wherein the instructions further comprise instructions to sanitize the name of the application by removing non-text characters from the name and translating the name into English.
10 . The non-transitory computer-readable medium of claim 8 , wherein the classifier is one of a supervised machine learning model or an LLM that performs zero-shot classification.
11 . The non-transitory computer-readable medium of claim 8 , wherein the instructions for receiving the name of an application comprise instructions to extract the name of the application as a signal of the application.
12 . The non-transitory computer-readable medium of claim 8 , wherein the prompt is based on a portion of the graph.
13 . The non-transitory computer-readable medium of claim 8 , wherein the instructions for generating the prompt for the LLM comprise instructions to select the prompt from a set of pre-generated prompts.
14 . The non-transitory computer-readable medium of claim 8 , wherein the instructions for receiving, as output from the classifier, the plurality of known tags further comprise instructions to:
receive, as output from the classifier, an unknown tag, wherein an unknown tag is a tag that does not already exist in the graph; and
update the graph to add the unknown tag.
15 . A system comprising:
memory with instructions encoded thereon; and
one or more processors that, when executing the instructions, are caused to perform operations comprising:
receiving a name of an application;
generating a prompt for a large language model (LLM) based on the name of the application, wherein the prompt includes a request for one or more tags associated with the application, the one or more tags describing tagged interests associated with the application;
providing the prompt to the LLM for execution;
receiving, as output from the LLM, a plurality of candidate tags;
inputting the plurality of candidate tags into a classifier, the classifier trained to classify candidate tags into known tags, wherein known tags are tags that already exist in a graph;
receiving, as output from the classifier, a plurality of known tags; and
updating the graph to include a node corresponding to the application, the node linked to the plurality of known tags with one or more edges.
16 . The system of claim 15 , wherein the operations further comprise sanitizing the name of the application by removing non-text characters from the name and translating the name into English.
17 . The system of claim 15 , wherein the classifier is one of a supervised machine learning model or an LLM that performs zero-shot classification.
18 . The system of claim 15 , wherein the operations for receiving the name of an application comprise extracting the name of the application as a signal of the application.
19 . The system of claim 15 , wherein the prompt is based on a portion of the graph.
20 . The system of claim 15 , wherein the operations for generating the prompt for the LLM comprise selecting the prompt from a set of pre-generated prompts.