Linguistic schema mapping via semi-supervised learning

A method, comprising:

• receiving, by an input receiving module, a customer schema;
• identifying, by a processor, an industry-specific schema corresponding to the received customer schema;
• selecting, by the processor, a first attribute included within the received customer schema for labeling;
• generating, by the processor, a candidate pair based on the selected first attribute, the candidate pair including the first attribute and a second attribute;
• generating, by processor using a linguistic featurizer, an embedding featurizer, and a syntactic featurizer for the candidate pair, a featurized candidate based on an identified similarity between the first attribute and the second attribute, the featurized candidate being a 3-tuple numerical vector corresponding to the candidate pair based on the identified similarity including a linguistic similarity, an embedded similarity, and a syntactic similarity between the first attribute and the second attribute in the candidate pair, wherein the linguistic similarity, the embedded similarity, and the syntactic similarity in the 3-tuple numerical vector are identified using the linguistic featurizer, the embedding featurizer, and the syntactic featurizer respectively;
• generating, by processor using a semi-supervised machine learning (ML) model, a matching score for the candidate pair based on the featurized candidate, the matching score measuring how closely the second attribute matches the first attribute in the candidate pair;
• identifying, by the processor using the semi-supervised ML model, a predetermined quantity of matching scores having a highest value of the generated matching scores;
• identifying, by the processor using the semi-supervised ML model, the candidate pair with the predetermined quantity of matching scores having the highest value of the generated matching scores;
• generating, by the processor using the ML model, a suggested label for the first attribute in the candidate pair, the suggested label corresponding to the second attribute in the candidate pair, wherein the semi-supervised ML model is trained on labeled training data including the featurized candidate and used to generate pseudo-labels for unlabeled data points, and wherein the semi-supervised ML model is iteratively retrained based on received feedback on the generated pseudo-labels in a train-predict loop until all candidate pairs are labeled or a maximum number of iterations is reached;
• and
• applying, by the processor, the suggested label to the first attribute, wherein applying the suggested label comprises updating the linguistic featurizer, based on the applied suggested label, including adding a sentence representation of the candidate pair and the applied suggested label with a larger sample weight than samples generated using only the industry-specific schema.

A computer-storage memory device embodied with executable instructions that, when executed by a processor, cause the processor to:

• receive a customer schema;
• identify an industry-specific schema corresponding to the received customer schema;
• select a first attribute included within the received customer schema for labeling;
• generate a candidate pair based on the selected first attribute, the candidate pair including the first attribute and a second attribute;
• generate a featurized candidate based on an identified similarity between the first attribute and the second attribute, the featurized candidate being a 3-tuple numerical vector corresponding to the candidate pair based on the identified similarity including a linguistic similarity, an embedded similarity, and a syntactic similarity between the first attribute and the second attribute in the candidate pair, wherein the linguistic similarity, the embedded similarity, and the syntactic similarity in the 3-tuple numerical vector are identified using a linguistic featurizer, an embedding featurizer, and a syntactic featurizer respectively;
• generate, using a semi-supervised machine learning (ML) model, a matching score for the candidate pair based on the featurized candidate, the matching score measuring how closely the second attribute matches the first attribute in the candidate pair;
• identify, using the semi-supervised ML model, a predetermined quantity of matching scores having a highest value of the generated matching scores;
• identify, using the semi-supervised ML model, the candidate pair with the predetermined quantity of matching scores having the highest value of the generated matching scores;
• generate, using the semi-supervised ML model, a suggested label for the first attribute in the candidate pair, the suggested label corresponding to the second attribute in the candidate pair having the matching score with the highest value, wherein the semi-supervised ML model is trained on labeled training data including the featurized candidate and used to generate pseudo-labels for unlabeled data points, and wherein the semi-supervised ML model is iteratively retrained based on received feedback on the generated pseudo-labels in a train-predict loop until all candidate pairs are labeled or a maximum number of iterations is reached;
• and
• apply the suggested label to the first attribute, wherein applying the suggested label comprises updating the linguistic featurizer, based on the applied suggested label, including adding a sentence representation of the candidate pair and the applied suggested label with a larger sample weight than samples generated using only the industry-specific schema.