Design and Implementation of an Intelligent Chatbot System: Intent Recognition Algorithms and Architecture

This article introduces the design and implementation of an intelligent dialogue system tailored for urban management applications, addressing challenges such as high user thresholds, cumbersome task processing, and unidirectional information flow.

The system supports three primary scenarios: information query robots for system knowledge and data retrieval, task service robots for attendance and case handling, and intelligent customer service robots for core dialogue management.

The overall architecture consists of a terminal layer (Android, iOS, WeChat), a dialogue service layer that processes requests and integrates with MIS/GIS services, and a natural language processing layer responsible for intent recognition. The dialogue service also analyzes interaction data to optimize future performance.

Intent recognition is the core component, utilizing multiple logical adapters. The simplest method employs regular expression matching for direct keyword detection. To overcome the rigidity of regex, the system implements fuzzy search based on Levenshtein edit distance, calculating the minimum single-character edits required to transform one string into another:

1. sitten(k->s)
2. sittin(e->i)
3. sitting(_->g)

For more advanced feature extraction, the system applies Chinese word segmentation followed by Principal Component Analysis (PCA) to reduce dimensionality and eliminate multicollinearity. A Naive Bayes classifier then computes the probability of each intent category based on the extracted features, selecting the highest probability as the final classification.

To ensure continuous improvement, the system incorporates a dynamic feedback mechanism where users can rate responses. These labeled interactions are continuously fed back into the training pipeline, allowing the classification model to adapt and improve its accuracy over time.

The frontend display layer communicates with the backend via a flexible JSON structure, supporting various message types such as text, links, markdown, and action cards. An example markdown payload is structured as follows:

{
"data": {
"msgType": "markdown",
"markdown": {
"title": "Wuhan Weather",
"content": "#### Wuhan Weather @150XXXXXXXX \n> 9 degrees, NW wind level 1, air quality 89, humidity 73%\n> ![screenshot](https://img.alicdn.com/tfs/TB1NwmBEL9TBuNjy1zbXXXpepXa-2400-1218.png)\n> ###### Released at 10:20 [Weather](https://www.egova.com) \n"
},
"msgID":1234
}
}

In conclusion, while the Naive Bayes classifier offers fast performance, its independence assumption may not always hold true in practice. Future iterations will explore alternative algorithms like Support Vector Machines and focus on parameter tuning to enhance classification accuracy, demonstrating the system's broad applicability beyond urban management.