Granted Patent B2

US 12,518,229 · App. 17/677,257 · Granted Jan 6, 2026

Inventors: Monika Gupta (Gurugram, IN); Sampath Dechu (Bangalore, IN); Hussain Jagirdar (Ujjain, IN); Akansha Khanna (Bangalore, IN); Naveen Eravimangalath Purushothaman (Thrissur, IN)

Assignee: International Business Machines Corporation

G06Q10/0633G06N20/00

Abstract

Claims (74)

1 . A computer-implemented method for improving automated process discovery by generating an executable process variant from unstructured textual data, the method comprising:

obtaining, by at least one computing device, workflow data associated with one or more dynamic processes, wherein the workflow data comprise unstructured textual data corresponding to at least one of textual descriptions of the one or more dynamic processes and textual descriptions of one or more ad hoc tasks corresponding to the one or more dynamic processes;

creating, by the at least one computing device, event logs based at least in part on the textual descriptions;

applying, by the at least one computing device, a graph extraction process to derive process flow graphs from the created event logs, wherein the process flow graphs comprise nodes corresponding to the one or more ad hoc tasks and edges indicating at least one of an order and a duration of the one or more ad hoc tasks;

executing, by the at least one computing device, a first machine learning embedding process, wherein the first machine learning embedding process generates a first set of vector embeddings by processing the process flow graphs to encode a plurality of structural features of the process flow graphs, the plurality of structural features comprising at least a number of nodes and a number of edges;

executing, by the at least one computing device, a second machine learning embedding process, wherein the second machine learning embedding process generates a second set of vector embeddings by processing the textual descriptions, wherein the second set of vector embeddings encode one or more semantic features of the textual descriptions;

generating, by the at least one computing device, a unified set of attributed process graph vector embeddings by combining, for each process flow graph, the corresponding vector embedding encoding the plurality of structural features with the corresponding vector embedding encoding the one or more semantic features;

applying, by the at least one computing device, at least one unsupervised machine learning process to the unified set of attributed process graph vector embeddings to cluster the process flow graphs into one or more clusters;

computing, by the at least one computing device, automation scores for the process flow graphs based at least in part on the unified set of attributed process graph vector embeddings, wherein the automation scores indicate a complexity of automating the process flow graphs;

identifying, by the at least one computing device, at least one of the process flow graphs to be automated based at least in part on the computed automation scores;

generating, by the at least one computing device, at least one executable process variant of the identified at least one process flow graph, the executable process variant comprising a new process flow graph comprising a modified sequence of computer-executable tasks, wherein generating the at least one executable process variant comprises: (i) identifying a portion of the identified at least one process flow graph corresponding to an activity based on a comparison of an average time to perform the activity relative to other activities within a given one of the one or more clusters comprising the identified at least one process flow graph and a frequency of occurrence of the activity across multiple process flow graphs within the cluster, and (ii) modifying the identified portion based on the plurality of structural features and the one or more semantic features encoded in the attributed process graph vector embeddings corresponding to at least one other process flow graph within the cluster;

outputting, by the at least one computing device, the generated at least one executable process variant to a robotic process automation tool; and

in response to the outputting, using, by the at least one computing device, the robotic process automation tool to automatically execute the modified sequence of computer-executable tasks of the at least one executable process variant.

2 . The computer-implemented method of claim 1 , further comprising:

generating, by the at least one computing device, a respective keyphrase to represent each of the one or more clusters based on the textual descriptions associated with the process flow graphs within each cluster.

3 . The computer-implemented method of claim 1 , wherein the first set of vector embeddings of the process flow graphs generated by the first machine learning embedding process further encode one or more of: at least one temporal feature, and at least one frequency distribution feature.

4 . The computer-implemented method of claim 1 , wherein the plurality of structural features further comprises one or more of:

a maximum node fan-out count;

a maximum node fan-in count;

a number of variants; and

a number of loops.

5 . The computer-implemented method of claim 1 , wherein the identifying comprises:

ranking the process flow graphs based on the computed automation scores.

6 . The computer-implemented method of claim 1 , wherein generating the at least one executable process variant comprises at least one of: adding, deleting, or changing one or more repetitive tasks associated with the identified process flow graph.

7 . A computer program product for improving automated process discovery by generating an executable process variant from unstructured textual data, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computing device to cause the computing device to:

obtain workflow data associated with one or more dynamic processes, wherein the workflow data comprise unstructured data corresponding to at least one of textual descriptions of the one or more dynamic processes and textual descriptions of one or more ad hoc tasks corresponding to the one or more dynamic processes;

create event logs based at least in part on the textual descriptions;

apply a graph extraction process to derive process flow graphs from the created event logs, wherein the process flow graphs comprise nodes corresponding to the one or more ad hoc tasks and edges indicating at least one of an order and a duration of the one or more ad hoc tasks;

execute a first machine learning embedding process, wherein the first machine learning embedding process generates a first set of vector embeddings by processing the process flow graphs to encode a plurality of structural features of the process flow graphs, the plurality of structural features comprising at least a number of nodes and a number of edges;

execute a second machine learning embedding process, wherein the second machine learning embedding process generates a second set of vector embeddings by processing the textual descriptions, wherein the second set of vector embeddings encode one or more semantic features of the textual descriptions;

generate a unified set of attributed process graph vector embeddings by combining, for each process flow graph, the corresponding vector embedding encoding the plurality of structural features with the corresponding vector embedding encoding the one or more semantic features;

apply at least one unsupervised machine learning process to the unified set of attributed process graph vector embeddings to cluster the process flow graphs into one or more clusters;

compute automation scores for the process flow graphs based at least in part on the unified set of attributed process graph vector embeddings, wherein the automation scores indicate a complexity of automating the process flow graphs;

identify at least one of the process flow graphs to be automated based at least in part on the computed automation scores;

generate at least one executable process variant of the identified at least one process flow graph, the executable process variant comprising a new process flow graph comprising a modified sequence of computer-executable tasks, wherein generating the at least one executable process variant comprises: (i) identifying a portion of the identified at least one process flow graph corresponding to an activity based on a comparison of an average time to perform the activity relative to other activities within a given one of the one or more clusters comprising the identified at least one process flow graph and a frequency of occurrence of the activity across multiple process flow graphs within the cluster, and (ii) modifying the identified portion based on the plurality of structural features and the one or more semantic features encoded in the attributed process graph vector embeddings corresponding to at least one other process flow graph within the cluster;

output the generated at least one executable process variant to a robotic process automation tool; and

in response to the outputting, use the robotic process automation tool to automatically execute the modified sequence of computer-executable tasks of the at least one executable process variant.

8 . The computer-implemented method of claim 1 , wherein executing the second machine learning embedding process comprises:

updating pre-trained sentence embeddings using domain-specific training data derived from the textual descriptions of the one or more ad hoc tasks.

9 . The computer-implemented method of claim 8 , wherein updating the pre-trained sentence embeddings comprises tuning the pre-trained sentence embeddings using a task sequence of a process instance associated with the textual descriptions as a context.

10 . The computer program product of claim 7 , wherein the computing device is further caused to:

generate a respective keyphrase to represent each of the one or more clusters based on the textual descriptions associated with the process flow graphs within each cluster.

11 . The computer program product of claim 7 , wherein the first set of vector embeddings of the process flow graphs generated by the first machine learning embedding process further encode one or more of: at least one temporal feature, and at least one frequency distribution feature.

12 . The computer program product of claim 7 , wherein the identifying comprises:

ranking the process flow graphs based on the computed automation scores.

13 . The computer program product of claim 7 , wherein executing the second machine learning embedding process comprises:

updating pre-trained sentence embeddings using domain-specific training data derived from the textual descriptions of the one or more ad hoc tasks, wherein updating the pre-trained sentence embeddings comprises tuning the pre-trained sentence embeddings using a task sequence of a process instance associated with the textual descriptions as a context.

14 . A system for improving automated process discovery by generating an executable process variant from unstructured textual data, the system comprising:

a memory configured to store program instructions;

a processor operatively coupled to the memory to execute the program instructions to:

obtain workflow data associated with one or more dynamic processes, wherein the workflow data comprise unstructured data corresponding to at least one of textual descriptions of the one or more dynamic processes and textual descriptions of one or more ad hoc tasks corresponding to the one or more dynamic processes;

create event logs based at least in part on the textual descriptions;

apply a graph extraction process to derive process flow graphs from the created event logs, wherein the process flow graphs comprise nodes corresponding to the one or more ad hoc tasks and edges indicating at least one of an order and a duration of the one or more ad hoc tasks; execute a first machine learning embedding process, wherein the first machine learning embedding process generates a first set of vector embeddings by processing the process flow graphs to encode a plurality of structural features of the process flow graphs, the plurality of structural features comprising at least a number of nodes and a number of edges;

execute a second machine learning embedding process, wherein the second machine learning embedding process generates a second set of vector embeddings by processing the textual descriptions, wherein the second set of vector embeddings encode one or more semantic features of the textual descriptions;

generate a unified set of attributed process graph vector embeddings by combining, for each process flow graph, the corresponding vector embedding encoding the plurality of structural features with the corresponding vector embedding encoding the one or more semantic features;

apply at least one unsupervised machine learning process to the unified set of attributed process graph vector embeddings to cluster the process flow graphs into one or more clusters;

compute automation scores for the process flow graphs based at least in part on the unified set of attributed process graph vector embeddings, wherein the automation scores indicate a complexity of automating the process flow graphs;

identify at least one of the process flow graphs to be automated based at least in part on the computed automation scores;

generate at least one executable process variant of the identified at least one process flow graph, the executable process variant comprising a new process flow graph comprising a modified sequence of computer-executable tasks, wherein generating the at least one executable process variant comprises: (i) identifying a portion of the identified at least one process flow graph corresponding to an activity based on a comparison of an average time to perform the activity relative to other activities within a given one of the one or more clusters comprising the identified at least one process flow graph and a frequency of occurrence of the activity across multiple process flow graphs within the cluster, and (ii) modifying the identified portion based on the plurality of structural features and the one or more semantic features encoded in the attributed process graph vector embeddings corresponding to at least one other process flow graph within the cluster;

output the generated at least one executable process variant to a robotic process automation tool; and

in response to the outputting, use the robotic process automation tool to automatically execute the modified sequence of computer-executable tasks of the at least one executable process variant.

15 . The system of claim 14 , wherein the processor further executes the program instructions to:

generate a respective keyphrase to represent each of the one or more clusters based on the textual descriptions associated with the process flow graphs within each cluster.

16 . The system of claim 14 , wherein the first set of vector embeddings of the process flow graphs generated by the first machine learning embedding process further encode one or more of: at least one temporal feature, and at least one frequency distribution feature.

17 . The system of claim 14 , wherein the plurality of structural features further comprises one or more of:

a maximum node fan-out count;

a maximum node fan-in count;

a number of variants; and

a number of loops.

18 . The system of claim 14 , wherein the identifying comprises:

ranking the process flow graphs based on the computed automation scores.

19 . The system of claim 14 , wherein generating the at least one executable process variant comprises at least one of: adding, deleting, or changing one or more repetitive tasks associated with the identified process flow graph.

20 . The system of claim 14 , wherein executing the second machine learning embedding process comprises:

updating pre-trained sentence embeddings using domain-specific training data derived from the textual descriptions of the one or more ad hoc tasks, wherein updating the pre-trained sentence embeddings comprises tuning the pre-trained sentence embeddings using a task sequence of a process instance associated with the textual descriptions as a context.

Assignments (1)

Continuity (1)

References Cited (53)