Project Challenge

Mobilis faced significant challenges in providing efficient customer support to its extensive user base. With millions of subscribers across Algeria, the customer service team was overwhelmed with repetitive queries, often leading to long wait times and customer dissatisfaction.

Arabic NLP Challenges

Creating a chatbot for Arabic presented unique natural language processing challenges:

• Dialectal Variations: Accounting for the numerous Algerian Arabic dialects and expressions
• Text Directionality: Handling right-to-left text processing and mixed language inputs
• Morphological Complexity: Addressing Arabic’s rich morphological system with numerous word forms
• Lack of Resources: Overcoming the scarcity of Arabic NLP datasets and pre-trained models

Technical Implementation with Rasa

I leveraged the Rasa framework to build a robust chatbot solution:

Intent Classification & Entity Recognition

The chatbot’s intelligence centered around accurate intent classification and entity recognition:

• Custom Intent Classification: Trained models to identify 45+ distinct customer intents
• Entity Extraction: Built named entity recognition for phone numbers, plan names, and service types
• Contextual Understanding: Implemented context maintenance across conversation turns
• Sentiment Analysis: Added sentiment detection to escalate frustrated customers to human agents

# Example Rasa NLU configuration for Arabic
language: ar
pipeline:
  - name: WhitespaceTokenizer
  - name: RegexFeaturizer
  - name: LexicalSyntacticFeaturizer
  - name: CountVectorsFeaturizer
  - name: CountVectorsFeaturizer
    analyzer: char_wb
    min_ngram: 1
    max_ngram: 4
  - name: DIETClassifier
    epochs: 100
    constrain_similarities: true
  - name: EntitySynonymMapper
  - name: ResponseSelector
    epochs: 100
    constrain_similarities: true
  - name: FallbackClassifier
    threshold: 0.7

Training Data Development

A critical aspect of the project was developing comprehensive Arabic training data:

• Data Collection: Gathered real customer service conversations (anonymized)
• Augmentation: Used techniques to expand limited Arabic data through augmentation
• Annotation: Manually annotated thousands of Arabic customer queries
• Validation: Implemented rigorous testing to validate understanding across dialects

Integration with MyMobilis App

The chatbot was fully integrated into the MyMobilis mobile application:

• API Development: Created RESTful endpoints for the mobile app to communicate with the chatbot
• Authentication: Implemented secure user authentication to personalize responses
• Response Formatting: Designed message formatting for optimal display within the app interface
• Fallback Mechanisms: Created graceful degradation paths when the AI couldn’t confidently answer

Question-Answering Capabilities

Beyond simple intent classification, the chatbot features a sophisticated question-answering system:

• Knowledge Base: Developed a comprehensive knowledge base of Mobilis services and policies
• Information Extraction: Implemented algorithms to extract relevant information from knowledge base
• Response Generation: Created templates with dynamic content insertion for natural responses
• Multi-turn QA: Built capability to handle follow-up questions and maintain context

Results and Impact

The Arabic Chatbot project delivered significant results for ATM Mobilis:

• Customer Satisfaction: Improved customer satisfaction ratings by 27% for digital support
• Response Time: Reduced average response time from 15 minutes to under 10 seconds
• Service Adoption: Increased MyMobilis app usage by 34% following chatbot integration
• Cost Efficiency: Reduced customer support operational costs by approximately 22%

Continuous Learning System

A key innovation of the project was implementing continuous learning capabilities:

• Feedback Loop: Designed a system to collect user feedback on response quality
• Response Improvement: Created a pipeline to identify and improve low-confidence responses
• Pattern Recognition: Implemented analytics to identify new common questions
• Automated Training: Set up semi-automated retraining procedures for the NLU model

Conclusion

The ATM Mobilis Arabic Chatbot project represents a significant advancement in Arabic-language AI applications in the telecommunications sector. By successfully addressing the unique challenges of Arabic NLP while delivering a practical customer service solution, the chatbot now serves as a critical component of Mobilis’ customer service strategy and has been fully deployed in the MyMobilis application.

This project is part of my story: read Chapter 2: Building for Real People for the context around it.