Based on state-of-the-art Machine Learning Algorithms, the humanly.ai engine learns straight from Product Brochures, User Manuals etc. No training time. Higher accuracy than contemporary models.
I am still learning. So my predicted Accuracy is 67.64% and F1 Score is 77.28%. For Yes/No questions the accuracy predicted is 58.3%. But go ahead and test me. I do however, only understand English as of now.
Paste the paragraph you want to analyze. It could be from an article, an email, a brochure etc.
The humanly.ai solution, as the byline suggests is to allow chatbot agents to answer questions - as humans do! The current version of the product is centered around Machine Comprehension. The Machine comprehension model, in order to respond to questions does not need any kind of labelled data, but instead refers to artefacts - similar to what a human would read or refer to in order to answer any specific questions on the topic.
Enterprises usually do not have data ready in a (labelled) form that chatbots can be trained on. This solution circumvents that problem.
Also, once it has been trained on the relevant domain, the model does not require to be trained on the specific artefact.
For the Machine comprehension solution, we are using the BiDAF Model combined with Textual Entailment ( for Yes/No questions). The accuracy & the F1 Score can be improved once the model trains on relevant domain data.
We continue to evolve the solution and will expand further to a complete Dialog system. As a next product increment, we plan on using Recurrent Neural Network models, to improve the quality of the response and provide full sentence responses.
Please watch the space and to know more about our solution drop us a line at hello@humanly.ai