Training machine learning models for multi-modal entity matching in electronic records

A non-transitory computer-readable storage medium storing computer program instructions executable to perform operations, the operations comprising:

• obtaining a set of training instances, a training instance including a series of tokens with a first series of tokens for a first input text obtained from a first electronic record of a cloud platform storing different types of electronic records related to a plurality of entities and a second series of tokens for a second input text obtained from a second electronic record of the cloud platform, an entity matching label indicating whether the first input text and the second input text refer to a same entity distinct from other ones of the plurality of entities, and a set of named entity recognition (NER) labels indicating a type of entity for a respective token of the series of tokens among a first set of entity categories, wherein the first electronic record is different from the second electronic record;
• for a training instance, generating a set of output embeddings by applying a set of encoders to the series of tokens, the set of output embeddings including an output embedding for a classification token and a first series of output embeddings for the first series of tokens and a second series of output embeddings for the second series of tokens;
• generating a set of NER predictions of the type of entity among the first set of entity categories for respective tokens of the series of tokens by applying a NER prediction layer to the first series of output embeddings and the second series of output embeddings;
• generating an entity matching prediction indicating whether the first input text and the second input text refer to the same entity by applying a match layer to the output embeddings for the classification token;
• determining a loss function including a NER loss and an entity matching loss, the NER loss indicating a first difference between the set of NER predictions and the corresponding NER labels for the training instance, and the entity matching loss indicating a second difference between the entity matching prediction and the entity matching label for the training instance;
• and
• backpropagating error terms obtained from the loss function to update parameters of the match layer, the NER prediction layer, and the set of encoders.

A system comprising:

• a processor;
• and
• a computer-readable medium comprising instructions executable on the processor for:
• obtaining a set of training instances, a training instance including a series of tokens with a first series of tokens for a first input text obtained from a first electronic record of a cloud platform storing different types of electronic records related to a plurality of entities and a second series of tokens for a second input text obtained from a second electronic record of the cloud platform, an entity matching label indicating whether the first input text and the second input text refer to a same entity distinct from other ones of the plurality of entities, and a set of named entity recognition (NER) labels indicating a type of entity for a respective token of the series of tokens among a first set of entity categories, wherein the first electronic record is different from the second electronic record;
• for a training instance, generating a set of output embeddings by applying a set of encoders to the series of tokens, the set of output embeddings including an output embedding for a classification token and a first series of output embeddings for the first series of tokens and a second series of output embeddings for the second series of tokens;
• generating a set of NER predictions of the type of entity among the first set of entity categories for respective tokens of the series of tokens by applying a NER prediction layer to the first series of output embeddings and the second series of output embeddings;
• generating an entity matching prediction indicating whether the first input text and the second input text refer to the same entity by applying a match layer to the output embeddings for the classification token;
• determining a loss function including a NER loss and an entity matching loss, the NER loss indicating a first difference between the set of NER predictions and the corresponding NER labels for the training instance, and the entity matching loss indicating a second difference between the entity matching prediction and the entity matching label for the training instance;
• and
• backpropagating error terms obtained from the loss function to update parameters of the match layer, the NER prediction layer, and the set of encoders.