Facilitating reduction of noise in non-standard printed circuit board assembly component descriptions using a zero-shot model to identify salient component class descriptions
Facilitating reduction of noise in non-standard printed circuit board assembly component descriptions using a zero-shot model to identify salient component class descriptions is presented herein. A system receives defined valid label designations(s) representing an accepted domain of component class descriptions; receives defined invalid label designations(s) representing a rejected domain of component class descriptions; and replaces non-alphanumeric characters of respective component descriptions with respective spaces to obtain revised component descriptions, removes, from the revised component descriptions, word(s) that include number(s) to obtain reduced component descriptions, expands, using a defined knowledge base comprising an online library of information, respective words of the reduced component descriptions to obtain respective expanded words representing natural-language expressions of the respective words, and based on the defined valid and invalid label designation(s) and the respective expanded words, selects, using a zero-shot model, words from the reduced component description for inclusion in a final reduced component description.
1. A system, comprising:
at least one processor; and
at least one memory that stores executable components that, when executed by the at least one processor, facilitate performance of operations by the system, the operations comprising:
receiving a group of defined valid label designations representing an accepted domain of component class descriptions;
receiving a group of defined invalid label designations representing a rejected domain of component class descriptions; and
for each component description of a group of component descriptions,
replacing non-alphanumeric characters of the component description with respective spaces to obtain a revised component description,
removing, from the revised component description, words that comprise numbers to obtain a reduced component description,
expanding, using a defined knowledge base comprising an online library of information, respective words of the reduced component description to obtain respective expanded words representing natural-language expressions of the respective words of the reduced component description,
based on the group of defined valid label designations, the group of defined invalid label designations, and the respective expanded words, selecting, using a zero-shot model comprising a pre-trained machine learning model, a group of words from the reduced component description to be included in a final reduced component description representing the accepted domain of component class descriptions, and
based on a defined error condition and using a group of machine learning models, selecting the final reduced component description as a candidate of a component failure.
2. The system of claim 1 , wherein the group of defined valid label designations represents the accepted domain comprising an electronic device manufacturing domain, a medical domain, an industrial domain, or a financial domain.
3. The system of claim 1 , wherein the component description comprises unstructured data.
4. The system of claim 1 , wherein the component description comprises a printed circuit board assembly component description.
5. The system of claim 4 , wherein the selecting comprises:
including the final reduced component description in a bill of materials (BOM); and
selecting, via the BOM, the final reduced component description as the candidate of the component failure.
6. The system of claim 1 , wherein the online library of information comprises a multilingual online encyclopedia maintained by volunteer input.
7. The system of claim 6 , wherein the expanding of the respective words further comprises:
determining, utilizing a python-based interface of a multilingual online encyclopedia maintained by volunteer input, the respective expanded words representing the natural-language expressions of the respective words.
8. The system of claim 7 , wherein the expanding of the respective words further comprises:
in response to determining that a word of the respective words cannot be represented by an expanded word of the respective expanded words, using the word as the expanded word.
9. The system of claim 1 , wherein selecting the group of words from the reduced component description comprises:
generating, for an expanded word of the respective expanded words corresponding to a word of the group of words, respective zero-shot similarity scores for the group of defined valid label designations and the group of defined invalid label designations; and
based on the respective zero-shot similarity scores, selecting the word to be included in the final reduced component description.
10. The system of claim 9 , wherein the generating of the respective zero-shot similarity scores further comprises:
in response to determining that the group of defined valid label designations comprises more than one defined valid label designation, determining an average valid label zero-shot similarity score of the respective zero-shot similarity scores corresponding to the group of defined valid label designations; and
based on the average valid label zero-shot similarity score, selecting the word to be included, or otherwise excluded, in the final reduced component description.
11. The system of claim 10 , wherein the operations further comprise:
in response to determining that the group of defined invalid label designations comprises more than one defined invalid label designation, determining an average invalid label zero-shot similarity score of the respective zero-shot similarity scores corresponding to the group of defined invalid label designations; and
based on the average invalid label zero-shot similarity score, selecting the word to be included, or otherwise excluded, in the final reduced component description.
12. The system of claim 11 , wherein the selecting of the word further comprises:
in response to determining that a ratio of the average valid label zero-shot similarity score to the average invalid label zero-shot similarity score satisfies a defined condition representing that the word is to be included in the final reduced component description, including the word in the final reduced component description.
13. A method, comprising:
obtaining, by a system comprising at least one processor, a component description;
replacing, by the system, non-alphanumeric characters of the component description with respective spaces to obtain a revised component description;
removing, by the system, words that comprise numbers from the revised component description, resulting in a reduced component description;
expanding, by the system using a defined knowledge base comprising an online library of information, respective words of the reduced component description to obtain respective expanded words comprising natural-language words representing the respective words;
based on a first group of defined valid label designations, a second group of defined invalid label designations, and the respective expanded words, selecting, by the system using a zero-shot model, a group of the respective words to be included in a final reduced component description representing an accepted domain of component class descriptions; and
based on a defined error condition, selecting, by the system via a machine learning model, a candidate of a component failure from the final reduced component description.
14. The method of claim 13 , wherein the expanding of the respective words comprises:
determining, utilizing a python-based interface of the online library of information, the respective expanded words.
15. The method of claim 14 , further comprising:
in response to determining that a word of the respective words cannot be represented by an expanded word of the respective expanded words, using, by the system, the word as the expanded word.
16. The method of claim 13 , wherein the selecting of the group of the respective words to be included in the final reduced component description comprises:
generating, for the respective expanded words, respective zero-shot similarity scores corresponding to the group of defined valid label designations and the group of defined invalid label designations; and
based on the respective zero-shot similarity scores, selecting the group of the respective words to be included in the final reduced component description.
17. The method of claim 16 , wherein the generating of the respective zero-shot similarity scores further comprises:
determining an average valid label zero-shot similarity score of the respective zero-shot similarity scores corresponding to the group of defined valid label designations; and
determining an average invalid label zero-shot similarity score of the respective zero-shot similarity scores corresponding to the group of defined invalid label designations.
18. The method of claim 17 , wherein the selecting of the group of the respective words to be included in the final reduced component description further comprises:
based on a ratio of the average valid label zero-shot similarity score to the average invalid label zero-shot similarity score, selecting the group of the respective words to be included in the final reduced component description.
19. A non-transitory machine-readable medium comprising instructions that, in response to execution, cause a system comprising at least one processor to perform operations, the operations comprising:
obtaining a group of defined valid label designations representing an accepted domain of component class descriptions and a group of defined invalid label designations representing a rejected domain of component class descriptions;
replacing non-alphanumeric characters of a component description with respective spaces to obtain a revised component description;
removing, from the revised component description, words that comprise numbers to obtain a reduced component description;
expanding, using a python-based interface of an online encyclopedic repository of knowledge maintained using volunteer input, respective words of the reduced component description to obtain respective expanded words representing natural-language expressions of the respective words;
determining, via a pre-trained zero-shot machine learning model, respective zero-shot similarity scores corresponding to the group of defined valid label designations, the group of defined invalid label designations, and the respective expanded words;
based on the respective zero-shot similarity scores, removing at least one word of the respective words from the reduced component description to obtain a final reduced component description representing the accepted domain of component class descriptions; and
based on a defined error condition, selecting, using a machine learning model, a description of the final reduced component description as a candidate of a component failure corresponding to a bill of materials.
20. The non-transitory machine-readable medium of claim 19 , wherein the removing of the at least one word comprises:
determining an average valid label zero-shot similarity score of the respective zero-shot similarity scores corresponding to the group of defined valid label designations and an expanded word of the respective expanded words;
determining an average invalid label zero-shot similarity score of the respective zero-shot similarity scores corresponding to the group of defined invalid label designations and the expanded word; and
based on a ratio of the average valid label zero-shot similarity score to the average invalid label zero-shot similarity score, removing, from the reduced component description, a word of the respective words corresponding to the expanded word.