How NER Dominated NLPCC 2022: Techniques Behind the Winning Model

Introduction

At the 11th CCF International Conference on Natural Language Processing and Chinese Computing (NLPCC 2022), NetEase Cloud Commerce’s AI team won the open evaluation task for Named Entity Recognition (NER), demonstrating leading capabilities in NLP and deep learning. This article introduces NER technology, its development, and the optimizations made for the competition.

What Is a Named Entity?

Named entities are phrases in text that carry specific meaning or referential strength, such as person names, locations, and organization names. In e‑commerce scenarios, product names, models, colors, and sizes can also be defined as named entities.

Named Entity Recognition

NER aims to identify named entities in a document collection, solving two sub‑problems: locating entity phrases in the text and assigning a type to each phrase.

Find entity phrases in the text.

Label each phrase with its entity type.

Evolution of NER Techniques

Early NER relied on rule‑based dictionaries and templates. Later, feature‑based statistical machine‑learning methods emerged. With the rise of deep neural networks in 2014, feature engineering was replaced by deep models. Since 2018, large pre‑trained language models (e.g., BERT) have unified NLP tasks, bringing NER to a new performance level.

NER Modeling Paradigms

Five major approaches are commonly used:

Sequence Labeling : Tag each token and decode entities from the tag sequence.

Span Classification : Enumerate candidate spans and classify each span’s entity type.

Machine Reading Comprehension : Pose a question for each entity type and predict start/end positions.

Sequence Generation : Provide a generation prompt and let the model output entity phrases.

Relation Classification : Model the task as classifying relations between token pairs, treating entities as word‑word relations.

NLPCC 2022 Competition

The conference hosts seven tasks; NetEase Cloud Commerce participated in Task 5 (Multi‑label Classification, NER, Content Extraction for Scientific Literature) and achieved first place.

Winning Method: Relation Classification

The team adopted a relation‑classification framework, converting NER into a token‑pair relation task with three relation types: None Relation (no relation), Next‑Neighboring‑Word Relation (consecutive tokens within the same entity), and Tail‑Head‑[category] Relation (entity boundary with a specific category).

Model Architecture

The model consists of three components:

Document‑Level Token Feature Extractor : Uses a Transformer‑based pre‑trained language model to encode the sentence together with its document‑level context, enriching token representations.

Word Relation Feature Extractor : Applies Conditional Layer Normalization to transform token representations into token‑pair relations, adds positional and region embeddings, and extracts features with dilated convolutions.

Model Output : An MLP scores word‑relation representations, a biaffine predictor scores document‑level token representations, and the two scores are combined for final relation classification. A confidence filter removes low‑scoring entities.

Data Augmentation

Entity replacement was used to expand the training set fivefold. Entities were collected into a pool; for each sentence, entities were randomly replaced while preserving entity type, word count (difference ≤ 2), and character count (difference ≤ 6). This synthetic data helps the model learn structural patterns.

Result Ensemble

Predictions from multiple models (over 100 runs with different seeds) were combined using entity‑level voting to balance precision and recall.

Experiments

Standard NER baselines were evaluated; the relation‑classification model W2NER achieved the best baseline. Replacing its pre‑trained encoder with DeBERTa gave a slight edge. Adding context enhancement and data augmentation raised F1‑Score to 0.4430 (+1.5%). After ensembling 100+ models, the final score reached 0.4725.

Business Applications and Outlook

The winning NER system is already deployed in NetEase Cloud Commerce’s multi‑turn dialogue bots and outbound call robots, improving question‑matching and resolution rates. In multi‑turn QA, accurate entity extraction (e.g., “washing machine”, “WN54”, “Beijing”) enables faster automated responses, reducing human hand‑over.

Future work will focus on further optimizing the NER algorithm, lowering integration barriers, and providing lower‑latency APIs so that more small‑ and medium‑size enterprises can benefit from intelligent AI products.