Method of generating questions with blanks

One or more non-transitory computer-readable storage media storing one or more sequences of instructions which, when executed using one or more processors, cause the one or more processors to execute:

• using an application server computer that is communicatively coupled to a database via a telecommunication network, the database comprising a digitally stored relational table schema storing a plurality of term sets, each of the term sets comprising a plurality of terms, the application server computer hosting an application program programmed with a fill-in-the-blank (FITB) question generator service, receiving input specifying a particular term that does not include an FITB portion, and in response to the input, using the application server computer, executing an inference stage of a trained machine learning model over the particular term as input to generate output predictions comprising a token position and number of tokens of a target span of an FITB portion;
• using the application server computer, post-processing the output predictions by computing a penalty score, selecting a final position, and determining a final span to blank;
• using the application server computer, digitally storing a modified term comprising the FITB portion at the final position;
• using an application server computer that is communicatively coupled to a database via a telecommunication network, the database comprising a digitally stored relational table schema of a plurality of relational tables storing a plurality of term sets, each of the term sets comprising a plurality of terms, the application server computer hosting an application program programmed with a fill-in-the-blank (FITB) question generator service, configuring an untrained machine learning model as a bidirectional long short-term memory model via programmatic invocations of one or more machine learning code libraries or APIs that specify parameter values for embedding, dropout, and dense parameters of the machine learning model:
• using the application server computer, accessing a training dataset of digitally stored, user-generated FITB terms, executing one or more operations of preparation of the training dataset, and training the untrained machine learning model to create and store a trained machine learning model;
• using the application server computer, receiving input specifying a particular term that does not include an FITB portion, and in response to the input, using the application server computer, executing an inference stage of the trained machine learning model over the particular term as input to generate output predictions comprising a token position and number of tokens of a target span of an FITB portion, the target span representing a target location of the FITB portion for a question side of the particular term;
• using the application server computer, post-processing the output predictions by computing a penalty score, selecting a final position, and determining a final span to blank by calling a program library to recover an original character span corresponding to the number of tokens;
• using the application server computer, digitally storing a modified term comprising the FITB portion at the final position.