How Alibaba’s NLP Team Dominated Global Entity Extraction and Chinese Grammar Competitions

Recently, Alibaba announced two major breakthroughs in natural language processing: winning the global KBP2017 English entity discovery competition and sweeping all three levels of the Chinese Grammatical Error Diagnosis (CGED) contest.

Dr. Si Luo, a world‑renowned machine intelligence scholar and former tenured professor at Purdue University, leads Alibaba’s iDST NLP team. The team supports Alibaba’s ecosystem—retail, finance, logistics, entertainment, travel—through the AliNLP platform, which handles up to 60 billion NLP requests daily.

The team spans China (Hangzhou, Beijing) and the United States (Silicon Valley, Seattle), with most members having over ten years of NLP experience and more than 30% holding PhDs from institutions such as CMU, Berkeley, Princeton, Tsinghua, and Peking University.

KBP2017 Victory

The KBP competition, organized by NIST and co‑hosted by the U.S. Department of Defense, tasks participants with extracting entities and their relationships from unstructured English text to build a knowledge base. Over 20 top teams, including IBM Research, Stanford, CMU, and Tencent, competed.

Alibaba’s algorithm demonstrated deep contextual understanding—recognizing that “Apple” followed by “Jobs” refers to Steve Jobs, not the fruit, and that “Microsoft” increases the likelihood that “Apple” denotes the company.

The team employed a modified deep neural network architecture with three key features: automatic reading of massive corpora (e.g., Wikipedia), intelligent selection of training data for accuracy, and post‑regularization to ensure consistent results.

According to Si Luo, the system’s transfer‑learning component extracts common patterns across domains, improving accuracy on domain‑specific data.

Alibaba is integrating this entity‑extraction technology into its AliIE platform, offering an open, extensible solution for developers and researchers.

CGED 2017 Championship

The Chinese Grammatical Error Diagnosis competition, co‑organized by IJCNLP, challenges participants to automatically detect grammatical and semantic errors in Chinese essays written by non‑native speakers. Errors are categorized as Redundant, Missing, Selection, and Word Order, and evaluated across three levels: detection, identification, and position.

Alibaba’s system achieved the highest accuracy across all three levels, thanks to a model built on a bilstm‑crf backbone enriched with word segmentation, POS tags, dependency parsing, and unsupervised language‑model embeddings. The approach can spot short‑range errors (e.g., “一头牛” vs. “一只牛”) and long‑range inconsistencies (e.g., mismatched conjunctions in complex sentences).

For each level, the team designed specialized snapshot ensemble methods to boost performance.

Si Luo emphasized that while AI for natural language understanding is still in its early stages, breakthroughs in entity extraction and grammar diagnosis are essential steps toward strong artificial intelligence, and that achieving true semantic comprehension may still require 5–10 more years of research.