Optimizing Efficiency Through Natural Language Understanding

Through advances in natural language processing and the ubiquity of messaging platforms, chatbots have become viable as a way for businesses to effectively scale responses to customer queries and engagements. At Trabeya, we believe that simulating human conversation has powerful potential beyond automating routine aspects of customer service.

Our conversational assistant, Taby, seeks to realize this potential by allowing users to interactively query it on articles, reports, communications, and other documents. Whilst the folk assertion that over 80% of business information originates in unstructured form has persisted over decades [1], beyond the rise of non-relational databases as document stores, there is no standard way to extract usable insights from this data deluge. Taby utilizes state-of-the-art models from question answering research to help users quickly get at answers that the rigid semantics of search engines is prone to miss. Taby can also train on a document corpus of interest, extracting summaries, relationships, named entities, keywords, and sentiments on the go, improving its skills as it encounters different usage contexts and parses more data.

Figure 1: Language models that learn in a self-supervised way, here illustrated with a model called BERT which masks certain tokens in sentences and learns to predict them, have opened the field to transfer learning, wherein representational knowledge in a model trained for one problem can be adapted to other problems (image source [3])

Taby is aimed at scalability, such that once deployed, it continues to learn the nuances of your queries as well as your data, across queries from multiple users. This allows the assistant to efficiently adapt to complex, multi-turn dialogue with layered intent, and in turn for users to engage with a natural flow of queries that much quicker. With Gartner expecting over 50% of medium to large enterprises to deploy chatbots by 2020 [2], these engagements themselves will open a new sluice gate contributing to the ever-increasing data deluge a business faces, albeit one that is ripe for extracting insights. Moreover, the direct effect of customer-facing chatbot deployments is bound to be worse than ever an imbalance in operational efficiency between customer engagement and the already laborious process of gathering timely and actionable business intelligence. Taming unstructured data enables a business to elevate its decision-making process from a heuristic-driven one to an objective, data-driven process where the impact of knowns and unknowns captured in the data is immediately quantifiable.

Topics: AI Machine Learning