Virtual Assistant Design for Water Systems Operation

Abstract

Water management systems such as wastewater treatment plants and water distributions systems are big systems which include a multitude of variables and performance indicators that drive the decision making process for controlling the plant. To help water operators make the right decisions, we provide them with a platform to get quick answers about the different components of the system that they are controlling in natural language.

In our research, we explore the architecture for building a virtual assistant in the domain of water systems. Our design focused on developing better semantic inference across the different stages of the process. We developed a named entity recognizer that is able to infer the semantics in the water field by leveraging state-of-the art methods for word embeddings. Our model achieved significant improvements over the baseline Term Frequency - Inverse Document Frequency (TF-IDF) cosine similarity model.

Additionally, we explore the design of intent classifiers, which involves more challenges than a traditional classifier due to the small ratio of text length compared to the number of classes. In our design, we incorporate the results of entity recognition, produced from previous layers of the Chatbot pipeline to boost the intent classification performance. Our baseline bidirectional Long Short Term Memory Network (LSTM) model showed significant improvements, amounting to 7-10\% accuracy boost on augmented input data and we contrasted its performance with a modified bidirectional LSTM architecture which embeds information about recognized entities.

In each stage of our architecture, we explored state-of-the-art solutions and how we can customize them to our problem domain in order to build a production level application. We additionally leveraged Chatbot frameworks architecture to provide a context aware virtual assistance experience which is able to infer implicit references from the conversation flow.