Method for automatically labeling data items
A method for labeling data items includes receiving, from a client, user input generated by a user of the client. The method also includes generating query conditions based on the user input. Further, the method includes retrieving, from a custom knowledge system integration module and using the query conditions, a list of label candidates. In addition, the method includes determining, via the client, that the user approves the list of label candidates. Moreover, the method includes generating, based on the determination, a label candidate query using at least the list of label candidates. Further, the method includes submitting the label candidate query to a large language model (LLM) engine. Moreover, the method includes obtaining, in response to the label candidate query, a list of target labels from the LLM engine. Also, the method includes labeling a plurality of unlabeled data items using the list of target labels.
1 . A method for labeling data items, the method comprising:
receiving, from a client, user input generated by a user of the client;
generating query conditions based on the user input;
retrieving, from a custom knowledge system integration module and using the query conditions, a list of label candidates, wherein the list of label candidates is associated with a labeling domain;
determining, via the client, that the user approves the list of label candidates;
generating, based on the determination, a label candidate query using at least the list of label candidates;
submitting the label candidate query to a large language model (LLM) engine;
obtaining, in response to the label candidate query, a list of target labels from the LLM engine; and
labeling a plurality of unlabeled data items using the list of target labels, wherein the labeling comprises:
selecting an unlabeled data item from the plurality of unlabeled data items;
obtaining a data item embedding for the selected unlabeled data item;
performing matrix multiplication between the data item embedding with each textual embedding from a plurality of textual embeddings to obtain a set of similarity scores for the selected unlabeled data item,
wherein each of the plurality of textual embeddings is associated with a target label from the list of target labels, and
wherein each of the similarity scores in the set of similarity scores is associated with a target label from the list of target labels;
selecting a target label from the list of target labels with a highest similarity score in the set of similarity scores, wherein the target label reflects a most accurate label of the labeling domain for the unlabeled data item; and
associating the target label with the unlabeled data item to obtain a labeled data item.
2 . The method of claim 1 , the method further comprising:
prior to the labeling of the plurality of unlabeled data items:
obtaining, from a multimodal embedding generator, a plurality of data item embeddings for each of the plurality of unlabeled data items, wherein the plurality of data item embeddings are used in the labeling.
3 . The method of claim 2 , wherein a multimodal embedding generator utilizes contrastive language-image pre-training (CLIP) models to generate the plurality of data item embeddings.
4 . The method of claim 1 , the method further comprising:
generating, by a data labeling module, label sentences for each label in the list of target labels; and
generating, by the multimodal embedding generator, textual embeddings for each label sentence to obtain a plurality of textual embeddings, wherein the plurality of textual embeddings are used in the labeling.
5 . The method of claim 4 , wherein the multimodal embedding generator utilizes contrastive language-image pre-training (CLIP) models to generate the plurality of textual embeddings.
6 . The method of claim 1 , wherein the custom knowledge system integration module utilizes retrieval augmented generation (RAG) methodology to retrieve a list of label candidates.
7 . The method of claim 6 , wherein the custom knowledge system integration module comprises at least knowledge graphs, ontologies, and text based knowledge base articles.
8 . The method of claim 1 , wherein determining, via the client, that the user approves the list of label candidates comprises:
receiving additional user input from the user; and
wherein the label candidate query is generated, at least in part, using the additional user input.
9 . A non-transitory computer readable medium (CRM) comprising computer readable program code, which when executed by a computer processor, enables the computer to perform a method for labeling data items, the method comprising:
receiving, from a client, user input generated by a user of the client;
generating query conditions based on the user input;
retrieving, from a custom knowledge system integration module and using the query conditions, a list of label candidates, wherein the list of label candidates is associated with a labeling domain;
determining, via the client, that the user approves the list of label candidates;
generating, based on the determination, a label candidate query using at least the list of label candidates;
submitting the label candidate query to a large language model (LLM) engine;
obtaining, in response to the label candidate query, a list of target labels from the LLM engine; and
labeling a plurality of unlabeled data items using the list of target labels, wherein the labeling comprises:
selecting an unlabeled data item from the plurality of unlabeled data items;
obtaining a data item embedding for the selected unlabeled data item;
performing matrix multiplication between the data item embedding with each textual embedding from a plurality of textual embeddings to obtain a set of similarity scores for the selected unlabeled data item,
wherein each of the plurality of textual embeddings is associated with a target label from the list of target labels, and
wherein each of the similarity scores in the set of similarity scores is associated with a target label from the list of target labels;
selecting a target label from the list of target labels with a highest similarity score in the set of similarity scores, wherein the target label reflects a most accurate label of the labeling domain for the unlabeled data item; and
associating the target label with the unlabeled data item to obtain a labeled data item.
10 . The non-transitory CRM of claim 9 , the method further comprising:
prior to the labeling of the plurality of unlabeled data items:
obtaining, from a multimodal embedding generator, a plurality of data item embeddings for each of the plurality of unlabeled data items, wherein the plurality of data item embeddings are used in the labeling.
11 . The non-transitory CRM of claim 10 , wherein the multimodal embedding generator utilizes contrastive language-image pre-training (CLIP) models to generate the plurality of data item embeddings.
12 . The non-transitory CRM of claim 9 , the method further comprising:
generating, by a data labeling module, label sentences for each label in the list of target labels; and
generating, by a multimodal embedding generator, textual embeddings for each label sentence to obtain a plurality of textual embeddings, wherein the plurality of textual embeddings are used in the labeling.
13 . The non-transitory CRM of claim 12 , wherein the multimodal embedding generator utilizes contrastive language-image pre-training (CLIP) models to generate the plurality of textual embeddings.
14 . The non-transitory CRM of claim 9 , wherein the custom knowledge system integration module utilizes retrieval augmented generation (RAG) methodology to retrieve a list of label candidates.
15 . The non-transitory CRM of claim 14 , wherein the custom knowledge system integration module comprises at least knowledge graphs, ontologies, and text based knowledge base articles.
16 . The non-transitory CRM of claim 9 , wherein determining, via the client, that the user approves the list of label candidates comprises:
receiving additional user input from the user; and
wherein the label candidate query is generated, at least in part, using the additional user input.
17 . A method for labeling data items, the method comprising:
receiving, from a client, user input generated by a user of the client;
generating query conditions based on the user input;
retrieving, from a custom knowledge system integration module and using the query conditions, a list of label candidates, wherein the list of label candidates is associated with a labeling domain;
determining, via the client, that the user approves the list of label candidates;
generating, based on the determination, a label candidate query using at least the list of label candidates;
submitting the label candidate query to a large language model (LLM) engine;
obtaining, in response to the label candidate query, a list of target labels from the LLM engine;
labeling a plurality of unlabeled data items using the list of target labels, wherein the labeling comprises:
selecting an unlabeled data item from the plurality of unlabeled data items,
obtaining a data item embedding for the selected unlabeled data item;
performing matrix multiplication between the data item embedding with each textual embedding from a plurality of textual embeddings to obtain a set of similarity scores for the selected unlabeled data item,
selecting a target label from the list of target labels with a highest similarity score in the set of similarity scores, wherein the target label reflects a most accurate label of the labeling domain for the unlabeled data item; and
associating the target label with the unlabeled data item to obtain a labeled data item.
18 . The method of claim 17 , wherein determining, via the client, that the user approves the list of label candidates comprises:
receiving additional user input from the user; and
wherein the label candidate query is generated, at least in part, using the additional user input.