Provides advanced visual tooling and workflows to explore, label, improve and scale NLU training data and models.
For a lot of teams, it replaces tools like Excel 🙃
Much of the effort involved in developing models based on text is in the initial analysis phase. Identifying patterns and signals within the text (intents), and validating the presence of sufficiently similar data to train an NLU model, is very difficult to do without tools optimized for this purpose.
HumanFirst provides tools to explore both historical and real-time data at scale, with machine-learning assistance to find and organize data based on semantic similarity.
This analysis and exploration is not only useful at the beginning of a project, but throughout its lifecycle, as new data coming in will serve to discover new signals that can extend and improve the NLU's performance.
Labeling data is a necessary step in building datasets used to train NLU models. It's traditionally been a costly activity both in time and resources, with many teams resorting to inefficient tools like Excel to scale this.
HumanFirst provides machine-learning assisted workflows that not only accelerate this process (by more than 10x!), but also help discover the labels that exist within the data, and structure them in ways that allow the labeing process to scale to thousands of labels.
Training data contain errors, no matter how it was generated or labeled. The quality issues lead to poor NLU performance.
HumanFirst helps reduce ambiguity and improve accuracy, by allowing to identify problematic and ambiguous training examples, split & re-assign training data across intents, and merge intents.
The breadth and depth of an NLU model depends on the amount of individual intents - once you have more than a few dozen intents, it becomes very hard to manage and scale without introducing ambiguity and overlapping intents.
HumanFirst allows easy organization of intents in flexible hierarchies using simple drag & drop tooling, that make it easy to maintain and add new intents at the right level of abstraction, even at long-tail scale.