Text generation refinement

A computer-implemented method comprising:

• receiving a request to predict text;
• and
• generating, by a ranking service implemented as ranking service code executed by one or more processors, a predicted text output, wherein the predicted text output is selected from a plurality of predicted texts generated in response to the request, and wherein the generating comprises:
• predicting a plurality of predicted texts using at least one machine learning model,
• scoring, by the ranking service, each of the plurality of predicted texts by computing a first score using intra-sample scoring, computing a second score using entailment scoring in which a Natural Language Inference (NLI) model is used to quantify an entailment relation of a first one of the predicted texts to a second one of the predicted texts, and computing a third score using inter-sample scoring,
• computing a composite score for each of the plurality of predicted texts from the first score, the second score, and the third score,
• ranking the plurality of predicted texts based at least in part the composite scores, and
• selecting at least a highest ranked predicted text of the plurality of predicted texts as the predicted text output.

A system comprising:

• a first one or more computing devices to implement a text generation service in a multi-tenant provider network;
• and
• a second one or more computing devices to implement a ranking service in the multi-tenant provider network, the ranking service implemented as ranking service code that upon execution by one or more processors causes the ranking service to:
• generate a predicted text output from output of the text generation service, wherein the predicted text output is selected from a plurality of predicted texts, wherein to generate comprises to:
• predict a plurality of predicted texts using at least one machine learning model,
• score, by the ranking service, each of the plurality of predicted texts by computing a first score using intra-sample scoring, computing a second score using entailment scoring in which a Natural Language Inference (NLI) model is used to quantify an entailment relation of a first one of the predicted texts to a second one of the predicted texts, and computing a third score using inter-sample scoring,
• compute a composite score for each of the plurality of predicted texts from the first score, the second score, and the third score,
• rank the plurality of predicted texts based at least in part the composite scores, and
• select at least a highest ranked predicted text of the plurality of predicted texts as the predicted text output.