US 12,124,491 B1
Identifying and categorizing adverse remarks from audit reports for knowledge base creation and generating recommendations
Aditi Anil Pawde, Pune (IN); Akshada Ananda Shinde, Pune (IN); Manoj Madhav Apte, Pune (IN); Sachin Sharad Pawar, Pune (IN); Sushodhan Sudhir Vaishampayan, Pune (IN); and Girish Keshav Palshikar, Pune (IN)
Assigned to TATA CONSULTANCY SERVICES LIMITED, Mumbai (IN)
Filed by Tata Consultancy Services Limited, Mumbai (IN)
Filed on Sep. 7, 2023, as Appl. No. 18/462,589.
Claims priority of application No. 202321023187 (IN), filed on Mar. 29, 2023.
Int. Cl. G06F 16/35 (2019.01); G06F 16/335 (2019.01)
CPC G06F 16/353 (2019.01) [G06F 16/335 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A processor implemented method, comprising:
receiving, via one or more hardware processors, a corpus of audit reports pertaining to one or more entities;
applying, a first set of rules via the one or more hardware processors, on a plurality of sentences comprised in the corpus of audit reports to obtain a set of relevant sentences;
labelling, by using a second set of rules via the one or more hardware processors, each sentence amongst the set of relevant sentences to obtain a set of labelled sentences, wherein each sentence in the set of labelled sentences is indicative of (i) an adverse remark-based sentence, (ii) a non-adverse remark-based sentence, and (iii) a sentence with no label;
identifying, by using a pre-trained sentence classifier having at least one attention layer via the one or more hardware processors, one or more adverse remarks with an associated explainability in one or more sentences having no label in the set of labelled sentences to obtain a set of adverse remark-based sentences, wherein the at least one attention layer is configured to generate the associated explainability for each of the one or more adverse remarks identified in the set of labelled sentences based on a high-attention weight assigned by the at least one attention layer;
applying, a windowing technique having a specific length via the one or more hardware processors, on each adverse remark-based sentence amongst the set of adverse remark-based sentences and performing a comparison of (i) an output of the windowing technique and (ii) a reference category having a pre-defined length to obtain at least one category tag for each adverse remark-based sentence amongst the set of adverse remark-based sentences;
generating, via the one or more hardware processors, a knowledge base based on the at least one category tag for each adverse remark-based sentence amongst the set of adverse remark-based sentences; and
automatically extracting the knowledge base of actionable knowledge elements from the corpus of past audit reports, wherein each knowledge element from the knowledge base of actionable knowledge elements consists of the tuple (S; v; Xv), where S is an adverse remark sentence, v is the text fragment in S which is a mention of a FV in S, Xv is a set of XBRL categories corresponding to v.