Baidu Chinese Text Correction Technology Overview

The presentation focuses on the traditional NLP problem of text correction, introducing its background and mainstream techniques before detailing Baidu's major work and showcasing product experience upgrades through specific application scenarios.

Text correction is crucial for NLP tasks such as lexical and syntactic analysis because accurate input data is a prerequisite for reliable results.

Historically, rule‑based methods and dictionaries were used; later, statistical machine translation (SMT) and neural machine translation (NMT) approaches became dominant, with recent research combining both.

Baidu's Chinese correction aims to support multiple error types (word choice, grammatical, knowledge‑based) and multimodal inputs, while providing fast scenario migration and deep customization.

The system decomposes correction into three key steps: error detection, candidate recall, and correction ranking, supporting both SMT‑based and NMT‑based frameworks.

Error detection uses a Transformer/LSTM + CRF sequence model, leveraging linguistic priors, hard statistical features, and fusion of character‑ and word‑level representations.

Candidate recall combines large‑scale offline error‑alignment corpora with online pre‑ranking, using language models and confusion matrices to filter candidates.

Correction ranking employs a Deep&Wide hybrid model (deep contextual DNN and wide feature‑based layers) together with GBDT & LR to prioritize the correct output.

The core technologies revolve around language knowledge, context understanding, and knowledge computation, using restricted‑vocabulary language models, contextual DNN with AOA attention, and external knowledge retrieval to enhance correction.

Two system frameworks are offered: ECNet, a pipeline of specialized models, and Restricted‑V NEC, an end‑to‑end joint optimization model, each with its own trade‑offs.

Applications include open‑domain correction (writing assistance, content review) and scenario‑specific correction (map search, voice dialogue), both leveraging the platform’s high accuracy and stability.