Apparatus for determining role fitness while eliminating unwanted bias

A method, the method comprising:

• a plurality of parallel multicore graphics processing devices, within a multicore system processing device, training a deep learning model based on training data that is stored within a storage device and that comprises positive examples and negative examples, the positive examples representing candidates who would be invited to an interview and the negative examples representing candidates without at least one of relevant skills and experience who would not be invited to an interview;
• and applying the deep learning model to score a level of fitness that a specific candidate has for a role, wherein a system interface bus operably couples the multicore system processing device, the plurality of parallel multicore graphics processing devices, the storage device, and a network interface device that provides remote connection to the multicore system processing device, the training data is generated by stripping a most recent role from a resume for a candidate who has three or more instances of work experience listed in a listing of professional history, a positive example is generated by concatenating the most recent role that was stripped from the resume with a remainder of preceding work experience, a negative example is generated by concatenating a random job description from a randomly selected and different resume with the remainder of preceding work experience based on an unlikelihood that the candidate will be a good match for the random job description, a set of indicative pronouns is excluded from a process of extracting relevant information-rich natural language, thereby limiting capacity associated with the deep learning model to develop explicit bias against a demographic group, and the deep learning model features three parallel deep learning natural language model paths where sigmoid output neurons are featured at an output stage of each portion of the deep learning model with rectified linear unit neurons used between internal layers within the deep learning model.

A non-transitory computer-readable storage medium storing computer-executable instructions that, when executed, cause the performance of a method comprising:

• a plurality of parallel multicore graphics processing devices, within a multicore system processing device, training a deep learning model based on training data that is stored within a storage device and that comprises positive examples and negative examples, the positive examples representing candidates who would be invited to an interview and the negative examples representing candidates without at least one of relevant skills and experience who would not be invited to an interview;
• and applying the deep learning model to score a level of fitness that a specific candidate has for a role, wherein a system interface bus operably couples the multicore system processing device, the plurality of parallel multicore graphics processing devices, the storage device, and a network interface device that provides remote connection to the multicore system processing device, the training data is generated by stripping a most recent role from a resume for a candidate who has three or more instances of work experience listed in a listing of professional history, a positive example is generated by concatenating the most recent role that was stripped from the resume with a remainder of preceding work experience, a negative example is generated by concatenating a random job description from a randomly selected and different resume with the remainder of preceding work experience based on an unlikelihood that the candidate will be a good match for the random job description, a set of indicative pronouns is excluded from a process of extracting relevant information-rich natural language, thereby limiting capacity associated with the deep learning model to develop explicit bias against a demographic group, and the deep learning model features three parallel deep learning natural language model paths where sigmoid output neurons are featured at an output stage of each portion of the deep learning model with rectified linear unit neurons used between internal layers within the deep learning model.