Access comprehensive training tools and processes to improve the
learnability, usability, and performance of your bots.
Why businesses need this
Human discourse is complex and dynamic as no two individuals are the same. People tend to use different words, syntax, and style while typing compared to speaking, though their intention may be same. An intelligent, well-trained, conversational chatbot should be able to account for these nuances, while understanding the intent and context of the user’s request
Kore.ai’s Approach to Chatbot Training
To ensure your bot can understand what people are saying – and respond with the most relevant task – it’s important that you provide sufficient training data and test the bot with a variety of user inputs. Evaluating your bot with a large sample of expected user inputs provides insight into bot responses and gives the opportunity to further train the bot to interpret diverse human expressions.The Kore.ai platform uses three NL engines, each with the own training specialities.
Training with Machine Learning
Developers can train the ML engine by providing sample utterances for each intent (task) the bot needs to be able to identify. The engine will build a model that maps a user utterance to one of the bot intents. If the training utterances are tagged with entity details then a second model is created to assist with named entity recognition (NER).
Auto-Train
Auto Train allows a bot to do unsupervised learning. When enabling a bot will improve the intent detection by automatically adding the successfully executed utterances into its machine learning training set. Additions can always be reviewed by a bot developer.
Training with Fundamental Meaning
Languages are full of words than mean the same thing and that users are likely to use as alternatives. Defining synonyms for the bot’s task names quickly increases intent recognition with very little effort.
Patterns allows the bot developer to express precise control over the set and order of words that match an utterance. For example, idiomatic expressions only make sense when all of the words are present and in the same order, and often contain stop words that would otherwise be ignored.
Create patterns by using concepts, which are hierarchical collections of words and synonyms, to efficiently train intents.
Training with Knowledge Graph
An ontology, through domain terms and the relationships between them, provides a framework that limits false positives. Synonyms for each term easily extend the identification possibilities without the need to provide very large numbers of alternative training questions.
- Training with Machine Learning
-
Training with Machine Learning
Developers can train the ML engine by providing sample utterances for each intent (task) the bot needs to be able to identify. The engine will build a model that maps a user utterance to one of the bot intents. If the training utterances are tagged with entity details then a second model is created to assist with named entity recognition (NER).
Auto-Train
Auto Train allows a bot to do unsupervised learning. When enabling a bot will improve the intent detection by automatically adding the successfully executed utterances into its machine learning training set. Additions can always be reviewed by a bot developer.
- Training with Fundamental Meaning
-
Training with Fundamental Meaning
Languages are full of words than mean the same thing and that users are likely to use as alternatives. Defining synonyms for the bot’s task names quickly increases intent recognition with very little effort.
Patterns allows the bot developer to express precise control over the set and order of words that match an utterance. For example, idiomatic expressions only make sense when all of the words are present and in the same order, and often contain stop words that would otherwise be ignored.
Create patterns by using concepts, which are hierarchical collections of words and synonyms, to efficiently train intents. - Training with Knowledge Graph
-
Training with Knowledge Graph
An ontology, through domain terms and the relationships between them, provides a framework that limits false positives. Synonyms for each term easily extend the identification possibilities without the need to provide very large numbers of alternative training questions.
Training is a continuous process
Users may always express something unexpected or their usage will change. It’s imperative to track and analyze failed conversations through statistical analysis and take corrective actions.
Failure Analysis: Analyzing Bot’s Performance
Gain in-depth insight into your bot’s performance in identifying and executing the tasks
- View information about user utterances that matched and those that didn’t match with intents
- View chat transcripts to the point of the user utterance
- View NLP analysis for intent scoring and selection
- View exit points for failed tasks
- View response times and details for all scripts and service calls
- Ability to train the bot for the failed intent from the same dashboard
Automate Bot Testing
Batch testing helps you discern the ability of your bot to correctly recognize the expected intents for a given set of utterances and helps validate that a bot will continue to work as expected after making changes, such as to ML utterances, patterns, or synonyms. Our platform allows you to execute a series of tests to get a detailed statistical analysis and gauge the effectiveness of your bot’s training.
Run regular automated testing suites to get detailed statistical analysis, view NLP success rates and detailed metrics like precision, recall & F1 scores.
• Support test driven development
• Validates that bot continues to work as expected post changes
• Pre-built Test suites
• Ability to create custom test suites
• Record and run bot-user conversation suites