Learning quality estimation device, method, and program

A system for evaluating aspects of discrete series of texts, the system comprises:

• a processor;
• and
• a memory storing computer-executable instructions that when executed by the processor cause the system to:
• receive a set of pairs of discrete series of texts for learning, wherein each pair comprises a first discrete series of input data and a second discrete series of output data, and wherein the first discrete series of input data and the second discrete series of output data are based on a correct relationship without noise;
• generate, based on training, a forward learning model as training of the forward learning model using the set of pairs of discrete series of tests for learning as first training data, wherein the forward learning model, after being trained, converts the first discrete model of input data into the second discrete series of output data according to the correct relationship;
• receive an erroneous pair of discrete series of texts of a plurality of erroneous pairs of discrete series of texts for learning, wherein the erroneous pair of discrete series of texts comprises a third discrete series of input data and a fourth discrete series of output data, and wherein the third discrete series of input data and the fourth discrete series of output data have a likelihood of being based on an erroneous relationship with noise;
• determine, using the generated forward learning model, a quality-score for each erroneous pair of discrete series of texts of the plurality of erroneous pairs of discrete series of texts according to the likelihood of being based an erroneous relationship with noise;
• rank said each erroneous pair of discrete series of texts according to the quality-score;
• generate, based on the plurality of erroneous pairs of discrete series of texts, second training data by removing at least the ranked said each erroneous pair of discrete series of texts for learning according to the quality-score;
• and
• update the forward learning model as iterative training of the forward learning model using second training data.

A computer-readable non-transitory recording medium storing computer-executable instructions that when executed by a processor cause a computer system to:

• receive a set of pairs of discrete series of texts for learning, wherein each pair comprises a first discrete series of input data and a second discrete series of output data, and wherein the first discrete series of input data and the second discrete series of output data are based on a correct relationship without noise;
• generate, based on training, a forward learning model as training of the forward learning model using the set of pairs of discrete series of tests for learning as first training data, wherein the forward learning model, after being trained, converts the first discrete model of input data into the second discrete series of output data according to the correct relationship;
• receive an erroneous pair of discrete series of texts of a plurality of erroneous pairs of discrete series of texts for learning, wherein the erroneous pair of discrete series of texts comprises a third discrete series of input data and a fourth discrete series of output data, and wherein the third discrete series of input data and the fourth discrete series of output data have a likelihood of being based on an erroneous relationship with noise;
• determine, using the generated forward learning model, a quality-score for each erroneous pair of discrete series of texts of the plurality of erroneous pairs of discrete series of texts according to the likelihood of being based an erroneous relationship with noise;
• rank said each erroneous pair of discrete series of texts according to the quality-score;
• generate, based on the plurality of erroneous pairs of discrete series of texts, second training data by removing at least the ranked said each erroneous pair of discrete series of texts for learning according to the quality-score;
• and
• update the forward learning model as iterative training of the forward learning model using second training data.