How to Train a Chatbot

Every chatbot project is different but often clients come to us with a large knowledge base which they want a chatbot to support from its release but with very little training data.

The Challenge

Chatbots require extensive training data. Dialogflow requires at least 10 training phrases to each intent, with thousands of inquiries potentially needed depending on application domain. The training process of an AI-powered chatbot means that chatbots learn from each new inquiry, requiring months of optimisation to achieve 90-100% recognition rates.

Data Scarcity

Training corpora historically required manual creation, increasing project costs and timelines. Clients want to see quick results in a chatbot implementation, yet many lack sufficient chat transcripts or emails for initial training data development.

The Solution

The Bot Forge's synthetic training data service can automatically generate intent variation datasets that cover all of the different ways that users from different demographic groups might call the same intent. The service supports multiple NLP platforms (Rasa, LUIS, Wit.ai, Watson, Lex, Dialogflow) and provides training datasets in 100+ languages.

Testing Results

Three Dialogflow experiments using identical agent configurations with identical intents about angel investors were conducted:

• Test A: 2 hand-tagged phrases per intent -- 50% correct match rates with 0.64 confidence
• Test B: 10 synthetic phrases -- 91% accuracy with 0.76 confidence
• Test C: 25-60 phrases -- 100% accuracy with 0.86 confidence

Conclusion

Continuous monitoring and training remain essential for long-term chatbot success. Synthetic training data provides a powerful way to bootstrap chatbot performance from day one.