Introduction of QBox — January 18 update

Volume is a Gartner-recognised developer of conversational platforms with a number of live instances in market. The company’s early work in this field quickly revealed that the main NLP providers’ natural-language data models were essentially black boxes, and that tools to help assess performance and the impact of scaling simply weren’t available.

The key NLP providers give access to simple tooling to create data models but none enable developers to measure the performance and quality of the model. When trying to scale and train their models, developers are faced with manual, high-capital-cost, error-prone and slow process. With customer experience the key measurement of success, many brands don’t yet have the confidence or capability to deploy chatbots more widely.

Our own experience of working with natural-language data models led us to develop QBOX: a middleware technology-agnostic corpus-management tool that allows developers to quantify the quality of natural-language models and the impact of training and scaling.

QBOX ADDRESSES KEY CHALLENGES WHEN
WORKING WITH NL DATA MODELS.

MEASURABILITY: Benchmark your data model using automatically generated KPIs to validate performance.

VISUALISATION: Easily interpret and analyse your test results with different views to facilitate interpreting and analysing your test results.

IDENTIFICATION: Improve the accuracy of your natural-language data model by detecting poor-performing intents.

SCALABILITY: Rapidly scale your data model and maintain performance when new intents are added.

AFFORDABILITY: Significantly reduce your costs by simplifying maintenance processes.

WHAT THE QBOX TEAM HAS BEEN UP TO:

• We continue to improve QBOX based on Gartner analysts’ and external users’ feedback.
• The QBOX go-to-market strategy is to offer a vendor-agnostic solution. Currently, QBOX supports IBM Watson and Microsoft LUIS services.
• Dialogflow is currently in alpha, transitioning to beta by the end of Q1 2018.
• WIT.ai alpha is planned for mid-January 2018.

ADDITIONAL FUNCTIONALITIES:

• Three-dimensional scoring: The new scoring method evaluates the quality of the model and intents from three independent viewpoints: correctness, confidence and prediction stability. This equates to how correctly we can predict the right intent, how confident we are about this prediction, and finally, how stable over time this prediction will be.

• Concept recognition and mapping: This feature extracts concepts from the training data of an intent, to visualise whether a particular concept requires reinforcement.
• Payment: Payment functionality and tiered access are now available.