Machine Reading Comprehension Revolutionizes E‑commerce: Alibaba’s XiaoMi and AI Models

Research Background

Alibaba XiaoMi is an intelligent human‑machine interaction product focused on e‑commerce services, guiding users and assisting tasks. During last year’s Double‑11 shopping festival, XiaoMi handled 6.43 million smart service requests with a 95% intelligent resolution rate. Its QA technology has evolved from retrieval‑based knowledge‑base answering to deep semantic modeling.

Recently, XiaoMi is exploring machine reading comprehension (MRC) to give the system human‑like reading abilities, aiming to provide truly intelligent QA and improve service efficiency.

E‑commerce Application Scenarios

Typical QA in e‑commerce relies on manually crafted question‑answer pairs, which is labor‑intensive and cannot cover the long‑tail of user queries. MRC can directly read unstructured rule texts and product descriptions to answer questions.

Transaction rule interpretation : During events like Double‑11, users frequently ask about activity rules. Instead of manually extracting possible questions, MRC can read the rule documents and provide answers such as "When will my Double‑11 order be automatically confirmed?".

Q: How much revenue did Alibaba generate in fiscal year 2017? A: 1582.73 billion RMB

Pre‑sale product consultation : Store‑XiaoMi serves millions of merchants. Users often inquire about product details (e.g., camera features) that already exist in the product description page. MRC can read these pages and answer intelligently, reducing service cost and increasing conversion.

Related Work Survey

Knowledge‑base based Machine Reading

Traditional pipelines involve entity detection, entity linking, attribute filling, and knowledge retrieval. While controllable and interpretable, they struggle with domain‑specific variations and error propagation.

End‑to‑end Machine Reading

Public datasets have driven rapid progress:

Facebook bAbI reasoning dataset

Microsoft MCTest multiple‑choice dataset

DeepMind CNN/DailyMail cloze dataset

Facebook CBT cloze dataset

iFlytek & Harbin Institute of Technology Chinese cloze dataset (≈870 k articles)

Stanford SQuAD (over 100 k questions on Wikipedia)

These datasets differ in size, language, and answer format, providing benchmarks for various MRC tasks.

Models Based on SQuAD

Match‑LSTM with Answer Pointer : Combines match‑LSTM and pointer networks; the Boundary Model predicts only start and end positions, simplifying search.

Bidirectional Attention Flow (BiDAF) : Introduces bidirectional attention (Context‑to‑Query and Query‑to‑Context) to produce query‑aware passage representations.

FastQAExt : A lightweight architecture that adds two simple statistical features (question word overlap and weighted term frequency) to accelerate convergence.

r‑net : Uses dual interaction layers (question‑passage matching and passage self‑matching) and a pointer‑based answer layer, achieving top SQuAD leaderboard performance.

Challenges and Practices in Business Scenarios

Chinese dataset construction : High‑quality annotated data is costly; we supplement with translated public datasets and batch translation to quickly build Chinese corpora.

Model optimization for business : Incorporate document structure (headings, hierarchy) into encoders to better handle long e‑commerce rule documents.

Model simplification : Reduce complex bi‑LSTM components to meet online latency constraints while keeping performance degradation controllable.

Multi‑model fusion : Combine deep learning models with traditional rule‑based systems to balance intelligence and controllability.

Conclusion

Machine reading comprehension has become a hot research area in natural language processing, with rapid advances driven by datasets like SQuAD and a variety of neural models. While these models achieve strong results on Wikipedia‑style factual QA, they remain limited for complex, long‑form e‑commerce queries. Integrating academic breakthroughs with Alibaba’s real‑world scenarios—through data collection, model adaptation, simplification, and hybrid systems—promises significant value and more intelligent services for users.

References

Weston et al., 2015. Towards AI‑Complete Question Answering: A Set of Prerequisite Toy Tasks.

Richardson et al., 2013. MCTest: A Challenge Dataset for Open‑Domain Machine Comprehension.

Hermann et al., 2015. Teaching Machines to Read and Comprehend.

Hill et al., 2015. The Goldilocks Principle: Reading Children’s Books with Explicit Memory Representations.

iFlytek Chinese RC dataset.

Rajpurkar et al., 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text.

Wang et al., 2016. Machine Comprehension Using Match‑LSTM and Answer Pointer.

Seo et al., 2016. Bidirectional Attention Flow for Machine Comprehension.

Weissenborn et al., 2017. Making Neural QA as Simple as Possible but not Simpler.

Wang et al., 2017. Gated Self‑Matching Networks for Reading Comprehension and Question Answering.