AliMe Assist: An Intelligent Assistant for Creating an Innovative E-commerce Experience

Abstract: We present AliMe Assist, an intelligent assistant designed for creating an innovative online shopping experience in E-commerce. Based on question answering (QA), AliMe Assist offers assistance service, customer service, and chatting service. It is able to take voice and text input, incorporate context to QA, and support multi-round interaction. Currently, it serves millions of customer questions per day and is able to address 85% of them. In this paper, we demonstrate the system, present the underlying techniques, and share our experience in dealing with real-world QA in the E-commerce field.

• The paper presents AliMe Assist’s main contribution as an intelligent assistant leveraging CNN-based intent classification to enhance customer interactions.
• The paper employs a multi-layered architecture combining semantic parsing, slot-filling, and knowledge graph retrieval to boost answer precision.
• The paper demonstrates that AliMe Assist processes 85% of queries without human intervention, showcasing scalable efficiency in a real-world e-commerce setting.

AliMe Assist: An Intelligent Assistant for E-commerce

The paper presents AliMe Assist, an intelligent assistant tailored for the E-commerce sector, aiming to enhance customer interaction through advanced question answering (QA) mechanisms. AliMe Assist addresses various customer service scenarios via automation, thus reducing dependency on human agents and optimizing resource utilization. The system is architected to handle a substantial volume of customer inquiries while providing meaningful responses.

System Architecture and Processing

The architecture of AliMe Assist is structured to facilitate both voice and text input processing, utilizing a multi-layered approach to classify user intentions and retrieve relevant answers. Initially, the input parsing through semantic rules and machine learning models allows the identification of customer intent. The subsequent layers discriminate between tasks requiring structured assistance, information retrieval tasks leveraging a knowledge graph, and general chat interactions managed by hybrid approaches of information retrieval (IR) and sequence generation models.

System Features and Methodologies

Intention Identification

The system leverages Convolutional Neural Networks (CNN) to classify user intents based on input embeddings, coupled with semantic tags extracted via Trie-based parsers. The CNN model outperforms traditional models (e.g., SVM, MaxEnt) with a precision of approximately 89.91%, showing robustness and scalability suitable for industrial applications.

Task-Oriented Assistance

AliMe Assist utilizes a schema-based slot-filling technique to manage task-oriented inquiries such as flight bookings. This approach ensures that necessary user inputs are systematically collected and processed, allowing for task completion via third-party services.

Knowledge Graph-Based Service

For information-oriented queries, a knowledge graph framework is applied. Entities are extracted and structured hierarchically, with relations guiding answer retrieval. Semantic normalization is critical to map diverse customer utterances to standard representations within the graph, thereby enhancing retrieval accuracy by 10% compared to traditional IR models.

Chatting Service Integration

The paper introduces a hybrid mechanism combining attentive Seq2Seq models with IR systems to manage chat-oriented queries. The hybrid approach significantly enhances response quality over independent models, achieving a $\text{P}_{\text{top}_1}$ of 60.01% in offline testing and 60.36% in online trials, thus advocating for improved user interaction in casual conversations.

Real-World Application and Performance

Implemented in Alibaba's customer service ecosystem, AliMe Assist efficiently processes millions of queries daily and successfully addresses 85% of them without human intervention. Scenarios demonstrate its efficacy in triaging customer intent and providing structured assistance across distinct service domains.

Related Work and Innovations

The work builds upon traditional closed-domain QA methods and open-domain conversational models, advancing by integrating hybrid approaches to enhance response accuracy and conversational coherence. Moreover, the system capitalizes on CNNs for speed and context-awareness, balancing performance with industrial-scale computational demands.

Conclusion

AliMe Assist epitomizes the potential of QA systems in revolutionizing customer interactions within E-commerce platforms. Future enhancements could target extended context handling, reinforcement learning for real-time guidance, and integrating visual recognition capabilities, thereby broadening the system's applicability and efficiency. These advancements would solidify AI-Assisted E-commerce as a vital component in modern consumer engagement strategies.