Cognitive Service in AR platform Platform provides cognitive

Cognitive Service in AR platform. • Platform provides cognitive service that can be used by any application. • Uses Innovation Suite’s (Referred as IS) Cognitive service • Supports auto categorization today. • Given a text , classify/categorize into a category based on training provided • Key is training data!

High Level Architecture

Consuming Cognitive Capabilities - Steps. • Contact BMC to get IS service URL, username, and password. • Identify application form(s) that needs auto categorization capabilities • Identify Ifield(s) on the form that has description • Identify field(s) on the form that stores corresponding category • Identify existing data in this form that you would want to use for training the ML provider. • Generate training data & Train the service • Use the service to automatically categorize new data you would create in the form.

Configuration Setting Name Default Value Description *Cognitive-Service-Url "" URL of the IS cognitive service *Cognitive-Service-User-Name "" Username provided by BMC to access IS cognitive service *Cognitive-Service-User. Password "" Password provided by BMC to access IS cogntive service Cognitive-Service-Confidence. Threshold 90 AI providers typically return confidence when categorizing a given text. Server will not return suggested category if confidence level is below this threshold. Enable-Cognitive-Service F Enable/Disable Platform's Cognitive capabilities.

Data Set Descriptor form - AR System Administration: Cognitive Service Training Data Set Descriptor

Data Set Descriptor form • Entry in this form describes training data and how to generate the training data • Each entry has an ID – Data Set ID. • This is GUID – If not provided , server will generate one! • Training will be identified using this ID • Auto categorization of a text requires this ID – ML provider needs to know which training to use to categorize! • Why qualification and Training data generation algorithm? • You can’t send unlimited data for training –current limit is 15, 000 rows • Qualification will identify the data • Algorithm will pick subset from the identified data.

Training Jobs form - AR System Administration: Cognitive Service Training Data Set Pending

Training Jobs form • Create an entry into this form in order to generate training data and train the service. • Data Set ID will be used to find data set descriptor entry that describes how to generate the training data. • Allows one to provide training data. If training data is provided in the entry (attachment) , server WILL just use that data to train the service. • Creating entry generates entry into Application pending form and server responsible for generating training data gets signaled. (See next slide)

Generating training data • Generating training data is server group operation! • Keeping future need in mind - Generating training data can become time and resource consuming. • Provides failover capability – other ranked server will pick up job of generating the training data should current server gets booted. • New server group operation “Cognitive Service” is introduced for this purpose. • Training data will be sent over to IS service which will in turn send the training data to ML Provider – Watson currently. • Training ML provider is done asynchronously as it can take time to train. • Ranked server will poll the status and update the status of the job accordingly.

Generating training data • Generating training data is server group operation! • Keeping future need in mind - Generating training data can become time and resource consuming. • Provides failover capability – other ranked server will pick up job of generating the training data should current server gets booted. • New server group operation “Cognitive Service” is introduced for this purpose. • Training data will be sent over to IS service which will in turn send the training data to ML Provider – Watson currently. • Training ML provider is done asynchronously as it can take time to train. • Ranked server will poll the status and update the status of the job accordingly.

Using the service - AR System Administration: Cognitive Service

Using the service • Service entry call on display only form - AR System Administration: Cognitive Service • Input • Command – categorize • 50 fields for you to map text to classify • output • 10 fields for returning category (Hierarchical categories) • Confidence field that ML provide returns • In case confidence provided by ML provider < configured threshold, category fields will not be returned.

Few limits • Training is not cumulative – you can not incrementally train. • As of now, you can only create 8 trainings. • Creating 9 th training will give you entitlement error –you can always delete one or more of the 8 training and then create new once. • For training only 15, 000 rows can be sent. • Only 1024 characters will be used from input text column of CSV file.

Troubleshooting • Log file arextension. log at <Server – db –directory> • Check the configuration – logs will indicate if server was able to connect to IS service or not. • Make sure service is enabled – again logs will indicate if not enabled. • Error text will say “IS Service” in case error returned by IS service.

Code • New internal plugin is written • Internal plugin issues REST api calls to IS service • Nothing else was changed in server to deliver this feature except writing this new internal plugin. • More in code walkthrough.