Improving conversations with digital assistants through extracting, recommending, and verifying user inputs

Abstract

Digital assistants, including chat bots and voice assistants, suffer from discrepancies and uncertainty in human text and speech inputs. Human dialogue is often varied, ambiguous, and inconsistent, making data entry prone to error and difficult for digital assistants to process. Finding and extracting pertinent information from unstructured user inputs improves and expands the use of digital assistants on any platform. By confirming data entries and providing relevant recommendations when invalid information is provided, the digital assistant enables the use of natural language and introduces a higher degree of flow into the conversation.

This paper describes a series of input logic codifiers that form a corrective method to overcome errors and ambiguity typical of voice and text inputs. When users make a common mistake or forget data, the digital assistant can bridge the gap by recommending the most similar data that is available. The assistant measures the delta between the user’s utterance and valid entries using fuzzy logic to identify the closest and next closest data that relates to the unstructured text.

Furthermore, there are endless ways to denote dates, locations, etc., making it difficult for digital assistants to extract accurate and relevant data from the user’s natural language. However, the assistant may infer the desired data format or reference from the dialogue provided and validate this with the user as a follow-on question. The desired data format or type is inferred using fuzzy extraction methods, such as fuzzy date extraction, to isolate the desired data format from the unstructured text. This extracted information is then verified or confirmed by the user to maintain data accuracy and avoid downstream data quality issues.