Source-linked AI summary
AliMe Assist: An Intelligent Assistant for Creating an Innovative E-commerce Experience
Feng-Lin Li, Minghui Qiu, Haiqing Chen, Xiongwei Wang, Xing Gao, Jun Huang, Juwei Ren, Zhongzhou Zhao, Weipeng Zhao, Lei Wang, Guwei Jin, Wei Chu
TL;DR
E-commerce customer service faces growing demand and limitations in availability, response times, and peak-period service. AliMe Assist provides assistance, customer service, and chat through contextual, multi-round QA with voice or text input; it serves millions of questions daily and addresses 85%.
Problem
E-commerce customer service faces growing demand while traditional service is limited by availability, waiting times, and peak-period delays.
Source: §1 · p. 1Method
AliMe Assist offers assistance, customer service, and chat, with voice and text input, contextual multi-round QA, and techniques for intention identification, knowledge-oriented QA, and chat.
Source: §1 · p. 1Results
AliMe Assist serves millions of customer questions per day and addresses 85% of them.
Source: §4 · p. 3Takeaways & Limitations
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
E-commerce customer service faces growing demand alongside limited availability, slow responses, and peak-period delays. AliMe Assist addresses this setting with three services and contextual, multi-round QA, and the paper presents its techniques and real-world experience.
1 INTRODUCTION
- Traditional E-commerce customer service faces staffing demand and limits in availability, response time, and peak-period capacity.The introduction cites the fast-growing market and notes that customers may wait minutes for service.Source: §1 · p. 1
- The paper asks whether an assistant can handle simple, repetitive questions so staff can focus on cases needing human participation.The stated potential benefits are improved productivity, user experience, and cost reduction.Source: §1 · p. 1
- AliMe Assist offers assistance, customer service, and chat, supports voice and text with context and multi-round interaction, and serves millions of questions daily.The system addresses 85% of those questions; the questions are mainly Chinese, with some English.Source: §1 · p. 1
- The paper contributes a context-aware CNN for intention identification, semantic normalization and knowledge-graph QA, and an attentive Seq2Seq approach combining retrieval and generation for chat.It also presents the industrial assistant and outlines the paper's remaining sections.Source: §1 · p. 1
2 SYSTEM OVERVIEW
AliMe Assist processes voice or text questions through routing, question-processing components, and answer sources. Its flow handles business-related requests through task, configured-answer, knowledge-graph, or retrieval paths, while routing chat and other questions separately.
2 SYSTEM OVERVIEW
- Figure 1's architecture has input, intention-routing, question-processing, and knowledge-source layers, with answers retrieved from QA pairs or a knowledge graph.The input layer supports voice and text from devices such as phones, pads, and PCs.Source: §2 · p. 1, Figure 1 · p. 2
- Figure 2's flow first checks business-rule patterns, routing task requests to slot filling and promotional questions to configured answers.Requests for online service staff prompt customers to describe their problem; unmatched questions go to intention classification.Source: §2 · pp. 1–2, Figure 2 · p. 2
- For knowledge-oriented questions, semantic tags identify entities used by the knowledge-graph engine to retrieve an answer.If retrieval fails, the system retries semantic parsing with the previous question as context, using that context at most once when available.Source: §2 · p. 2
- Business-irrelevant questions recognized as chat go to the chat engine; other non-chat questions receive a pre-configured answer.Source: §2 · p. 2
3 SYSTEM FEATURES
AliMe Assist routes customer questions across task assistance, knowledge-oriented customer service, and chat using intention identification and specialized QA methods. The section reports improved results for intention classification, knowledge retrieval, and hybrid chatting.
3.1 Intention Identification
- A business-rule parser built from mined patterns works with a CNN intention classifier to route customer questions.The one-layer convolution-pooling CNN was selected for context capture and industrial efficiency, supporting around 200 queries per second; deeper CNNs or RNNs can perform better but scale less well.Source: §3.1 · p. 2
- The reported models achieved 89%, above the SVM/MaxEnt ensemble baseline (82.71%).Source: §3.1 · p. 2
3.2 Task-Oriented Assistance Service
- Task-oriented assistance defines mandatory and optional slots, fills them from customer inputs, and requests third-party services to complete the task.The engine uses dictionaries and patterns to identify fifteen attribute types.Source: §3.2 · p. 2
3.3 Knowledge-Oriented Customer Service
- The knowledge graph approach increases accuracy by 10% over traditional IR, while IR complements it because the graph covers only frequently asked knowledge items.Source: §3.3 · p. 3
3.4 Chatting Service
- An attentive Seq2Seq model scores retrieved chat answers; the system returns a candidate above a threshold, otherwise it outputs a generated answer.The approach combines retrieval and generation models.Source: §3.4 · p. 3
- 60.01% Ptop1 for the hybrid approach versus 47.11% for IR and 52.02% for Seq2Seq; online, it scored 60.36% versus IR’s 40.86%.The test used 600 questions; the online A/B test used 2,136 questions.Source: §3.4 · p. 3
4 DEMONSTRATION
The demonstration shows AliMe Assist operating at industrial scale and illustrates its assistance, customer-service, and chat features through realistic scenarios.
4 DEMONSTRATION
- AliMe Assist serves millions of customer questions per day and addresses 85% of them.Source: §4 · p. 3
- Three scenarios demonstrate flight booking through stepwise information gathering, contextual customer-service QA using a knowledge graph, and a chat conversation.The customer-service example combines the previous question “I want to check” with “Taobao account.”Source: §4 · p. 3
5 RELATED WORK
The related work covers task-, knowledge-, and chat-oriented QA. AliMe Assist’s described approaches combine slot filling, knowledge graphs with IR, and attentive reranking of retrieval and generation results.
5 RELATED WORK
- Task-oriented QA is closed-domain and commonly uses rule- or template-based methods and dialog state tracking; AliMe Assist’s slot filling resembles the template-based approach.Source: §5 · p. 4
- For knowledge-oriented QA, AliMe Assist combines a knowledge graph designed for high-precision E-commerce answers with IR as a complement.Source: §5 · p. 4
- Chat QA commonly uses IR or generation; AliMe Assist instead uses attentive Seq2Seq reranking to optimize their joint results.The passage contrasts IR’s reliance on close matches with generation’s potential for inconsistent or meaningless answers.Source: §5 · p. 4
6 CONCLUSION
AliMe Assist is presented as a real-world E-commerce assistant offering three services and handling millions of customer questions daily. Future work includes strengthening multi-round context, shopping guidance, and image recognition.