Abstract
Conversation interfaces (CIs), or chatbots, are a popular form of intelligent agents that engage humans in task-oriented or informal conversation. In this position paper and demonstration, we argue that chatbots working in dynamic environments, like with sensor data, can not only serve as a promising platform to research issues at the intersection of learning, reasoning, representation and execution for goal-directed autonomy; but also handle non-trivial business applications. We explore the underlying issues in the context of Water Advisor, a preliminary multi-modal conversation system that can access and explain water quality data.
Introduction
Most contemporary chatbots are "static," relying on fixed databases such as product catalogs or user manuals. While these systems can process natural language effectively, they often struggle in dynamic environments where data changes constantly—such as real-time sensor feeds—and where users possess complex, multi-step goals. This paper argues for an evolution in chatbot architecture toward goal-directed autonomy, where agents can navigate dynamic environments to provide actionable advice.
The Challenge of Dynamic Environments
Determining the safety of a resource like water is a non-trivial task for the general public. It requires balancing specific activities (e.g., drinking vs. washing) against complex regulatory standards that vary by geography and jurisdiction. Furthermore, real-time sensor data may be missing, inconsistent, or provided at varying granularities. There is a significant gap between raw scientific data and the actionable advice required by a non-expert. Bridging this gap requires a system capable of sophisticated reasoning and representation.
The Water Advisor System
To address these challenges, we developed the "Water Advisor" (WA) system, a prototype designed to implement goal-directed autonomy. WA is built as a multi-modal interface that synchronizes a chat interface with interactive maps and document views, ensuring the user receives consistent information across all modalities.
Technical Framework
The WA system integrates four key technical pillars:
- Learning: The system utilizes commercial Natural Language Classifiers (NLC) to interpret user intent and employs regression models to identify trends within water quality sensor data.
- Representation: To manage the complexity of overlapping jurisdictions, water usage purposes and regulatory limits are represented as geographically-scoped attribute-value pairs in JSON. This allows the bot to reconcile federal, state, and local regulations dynamically.
- Reasoning: The agent uses rule-based logic to minimize the user's "cognitive cost." It autonomously decides whether to request clarification, fetch updated data, or present a specific visual response based on its confidence levels.
- Execution: The system autonomously selects actions, such as identifying the most reliable data for a specific time interval or generating advice based on filtered pollutant levels.
Data Integration and Results
The prototype currently utilizes water quality data from Flint, Michigan, with ongoing work to integrate real-time streams from the US Geological Survey (USGS). We have also consolidated a dataset of water standards from national (US, India) and state (Michigan, New York) agencies, which is available for public use.
The functional prototype demonstrates that a chatbot can serve as an integration hub, acting as the "glue" between raw sensor data and human decision-making. By moving from simple queries (e.g., "What is the price?") to contextual, goal-oriented questions (e.g., "Is the water safe for washing now?"), the system proves that chatbots can manage multi-modal actions to support user goals in real-time.
Future Directions
While the current prototype successfully demonstrates the feasibility of autonomous chatbots in dynamic domains, several areas for future research remain. We aim to better incorporate "sensed signals"—such as user frustration or confusion—to proactively improve the interaction. Additionally, as we scale to handle high-frequency data streams from thousands of USGS locations, we must address the ethical implications of data prioritization and the potential for algorithmic bias in safety-critical advice.
Summary updated