A 2MB Chinese NLP stack that runs on a budget phone

What it does LAC is a joint lexical analysis toolkit for Chinese. It takes raw text and performs word segmentation, part-of-speech tagging, named entity recognition, and word-importance scoring in one pass. The underlying model is a deep Bi-GRU-CRF network trained by Baidu’s NLP team, exposed through Python, Java, C++, and Android interfaces.

The interesting bit Rather than treating segmentation, tagging, and NER as separate pipeline stages, LAC trains them together as multi-task objectives. That shared representation is what lets it claim an overall F1 above 0.91 while still pushing 800 QPS on a single CPU thread—or 200 QPS on a budget phone using a 2 MB model.

Key highlights

• Jointly segments text, tags named entities (PER, LOC, ORG, TIME), and scores word importance on a 0–3 scale in a single inference pass.
• Ships with a user-dictionary mechanism that overrides model output with custom long-phrase matches for precise domain adaptation.
• Supports incremental training on your own UTF-8 data, provided you stick to the built-in tag schema.
• The mobile variant is a 2 MB model that reportedly hits 200 QPS on low-end Android hardware.
• Bindings for Python, Java, C++, and Android ship in the same repository.

Caveats

• Incremental training for the full lexical-analysis mode requires your data to use the exact tag schema shipped with LAC; custom schemas are not yet supported.
• The dictionary customization feature does not yet support wildcard entries, though the README says that is planned.
• The codebase still advertises Python 2 compatibility, which may signal aging maintenance.

Verdict If you need Chinese word segmentation, POS tagging, or named entity recognition in a single model with minimal assembly, LAC is worth a look. If you work primarily outside Chinese or need a fully custom tag schema today, look elsewhere.