Techniques for automatic subject line generation

An apparatus for data processing, comprising:

• at least one processor;
• at least one memory coupled with the at least one processor;
• and
• instructions stored in the at least one memory and executable by the at least one processor to cause the apparatus to:
• receive a single exemplary subject line from a cloud client of a content generation service, wherein the single exemplary subject line is a reference subject line for generation of candidate subject lines and comprises a first entity name;
• generate a prompt for a hybrid machine learning model by interleaving soft tokens and hard tokens into an interleaved prompt;
• generate, by inputting the interleaved prompt into the hybrid machine learning model of the content generation service and in response to receiving the single exemplary subject line from the cloud client, a plurality of candidate subject lines similar to the single exemplary subject line, wherein the hybrid machine learning model is trained using a dataset of annotated subject lines, and wherein generating the plurality of candidate subject lines comprises:
• determining that an annotated subject line in the dataset includes a second entity name based at least in part on a set of markers in the annotated subject line that surround the second entity name;
• adding at least a portion of the annotated subject line comprising the second entity name to a candidate subject line of the plurality of candidate subject lines;
• and
• replacing the second entity name in the candidate subject line with the first entity name extracted from the single exemplary subject line;
• derive respective weights for a first semantic similarity score, a first surface form dissimilarity score, and a first length consistency score that correspond to an annotated A/B dataset, wherein the respective weights relatively maximize a rank correlation between the annotated A/B dataset and an aggregation of the first semantic similarity score, the first surface form dissimilarity score, and the first length consistency score according to the respective weights;
• calculate a respective semantic similarity score, a respective surface form dissimilarity score, and a respective length consistency score between the single exemplary subject line and each candidate subject line of the plurality of candidate subject lines;
• select, by the content generation service, a quantity of candidate subject lines from the plurality of candidate subject lines based at least in part on filtering the plurality of candidate subject lines according to an aggregation of the respective semantic similarity score, the respective surface form dissimilarity score, and the respective length consistency score that are associated with each candidate subject line and are weighted by the respective weights;
• cause the quantity of candidate subject lines to be displayed at the cloud client;
• and
• receive, from the cloud client, feedback associated with the quantity of candidate subject lines and a selection of at least one candidate subject line displayed at the cloud client.

A non-transitory computer-readable medium storing code for data processing, the code comprising instructions executable by at least one processor to:

• receive a single exemplary subject line from a cloud client of a content generation service, wherein the single exemplary subject line is a reference subject line for generation of candidate subject lines and comprises a first entity name;
• generate a prompt for a hybrid machine learning model by interleaving soft tokens and hard tokens into an interleaved prompt;
• generate, by inputting the interleaved prompt into the hybrid machine learning model of the content generation service and in response to receiving the single exemplary subject line from the cloud client, a plurality of candidate subject lines similar to the single exemplary subject line, wherein the hybrid machine learning model is trained using a dataset of annotated subject lines, and wherein generating the plurality of candidate subject lines comprises:
• determining that an annotated subject line in the dataset includes a second entity name based at least in part on a set of markers in the annotated subject line that surround the second entity name;
• adding at least a portion of the annotated subject line comprising the second entity name to a candidate subject line of the plurality of candidate subject lines;
• and
• replacing the second entity name in the candidate subject line with the first entity name extracted from the single exemplary subject line;
• derive respective weights for a first semantic similarity score, a first surface form dissimilarity score, and a first length consistency score that correspond to an annotated A/B dataset, wherein the respective weights relatively maximize a rank correlation between the annotated A/B dataset and an aggregation of the first semantic similarity score, the first surface form dissimilarity score, and the first length consistency score according to the respective weights;
• calculate a respective semantic similarity score, a respective surface form dissimilarity score, and a respective length consistency score between the single exemplary subject line and each candidate subject line of the plurality of candidate subject lines;
• select, by the content generation service, a quantity of candidate subject lines from the plurality of candidate subject lines based at least in part on filtering the plurality of candidate subject lines according an aggregation of the respective semantic similarity score, the respective surface form dissimilarity score, and the respective length consistency score that are associated with each candidate subject line and are weighted by the respective weights;
• cause the quantity of candidate subject lines to be displayed at the cloud client;
• and
• receive, from the cloud client, feedback associated with the quantity of candidate subject lines and a selection of at least one candidate subject line displayed at the cloud client.