PaperlayerView exact revision on arXiv

Source-linked AI summary

Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition

Yongliang Shen, Xinyin Ma, Zeqi Tan, Shuai Zhang, Wen Wang, Weiming Lu

arXiv:2105.06804v2cs.CL

TL;DR

Nested NER requires handling entities that overlap, while existing span-based methods incur computational and boundary-related weaknesses. The paper proposes a two-stage identifier that filters and regresses span proposals before classifying them, and reports state-of-the-art performance on several nested NER datasets.

  • Problem

    Existing span-based nested NER methods have high computational cost, underuse boundary information and partially matched spans, and struggle with long entities.

  • Method

    A two-stage identifier filters seed spans, regresses proposal boundaries to locate entities, and classifies the adjusted proposals.

  • Results

    The model achieves state-of-the-art performance consistently on KBP17, ACE04, and ACE05, improving F1-score by +3.08%, +0.71%, and +1.27%, respectively.

  • Takeaways & Limitations

    The approach uses entity-boundary information, partially matched spans, and proposal filtering to address nested NER recognition and inference-cost challenges.

Abstract

from arXiv · show

Named entity recognition (NER) is a well-studied task in natural language processing. Traditional NER research only deals with flat entities and ignores nested entities. The span-based methods treat entity recognition as a span classification task. Although these methods have the innate ability to handle nested NER, they suffer from high computational cost, ignorance of boundary information, under-utilization of the spans that partially match with entities, and difficulties in long entity recognition. To tackle these issues, we propose a two-stage entity identifier. First we generate span proposals by filtering and boundary regression on the seed spans to locate the entities, and then label the boundary-adjusted span proposals with the corresponding categories. Our method effectively utilizes the boundary information of entities and partially matched spans during training. Through boundary regression, entities of any length can be covered theoretically, which improves the ability to recognize long entities. In addition, many low-quality seed spans are filtered out in the first stage, which reduces the time complexity of inference. Experiments on nested NER datasets demonstrate that our proposed method outperforms previous state-of-the-art models.

1 Introduction

The paper frames nested NER as a span-based localization and classification problem, then proposes a two-stage identifier that improves boundary use, partial-span training, efficiency, and recognition across nested NER benchmarks.

  • Span-based methods can recognize nested entities but face high computational cost, incomplete boundary use, under-utilized partial matches, and difficulty recognizing long entities.
  • The proposed identifier first locates entities through span filtering and boundary regression, then labels the adjusted proposals with entity categories.The model treats NER jointly as boundary regression and span classification.
  • The method uses boundary-level SmoothL1 and span-level overlap losses to improve entity-boundary localization.
  • Soft examples based on intersection-over-union use partially matched spans during training and alleviate positive-negative imbalance.
  • The model achieves state-of-the-art performance consistently on KBP17, ACE04, and ACE05, improving F1-score by +3.08%, +0.71%, and +1.27%, respectively.

2 Model

The model represents seed spans, filters low-quality candidates, regresses their boundaries, and classifies the adjusted spans. Training uses overlap-based supervision to exploit partially matched spans and jointly optimize the modules.

  • Seed Span Generation: Seed spans enumerate possible start and end positions under a maximum length constraint L.Each seed span is represented as b_i = (st_i, ed_i).
  • Span Proposal Module: The span proposal module filters contextual spans and retains higher-overlap proposal spans before boundary adjustment.The filter uses span representations built from pooled and boundary word representations.
  • Span Proposal Module: The boundary regressor adjusts the left and right boundaries of proposal spans to locate entities more precisely.It uses both span information and outer boundary words, then produces adjusted start and end positions.
  • Entity Classifier Module: The entity classifier assigns categories to adjusted proposals, using a NONE label when their IoU with the paired entity does not exceed α2.Training labels are reassigned after boundary adjustment.
  • Training Objective: Partially matched spans receive IoU-based soft weights, allowing their contribution to vary smoothly instead of treating them uniformly as negatives.The focusing parameter η controls the down-weighting rate; η = 0 recovers hard weighting.

3 Experiment Settings

The experiments evaluate the model on four nested NER datasets using strict entity-level correctness and precision, recall, and F1-score. Dataset domains, category counts, and split protocols vary across benchmarks.

  • Datasets: ACE04 and ACE05 each contain 7 entity categories and use an 8:1:1 train/dev/test split.The split follows prior work.
  • Datasets: KBP17 contains 5 entity categories and uses 866/20/167 documents for train, development, and test sets.The categories are GPE, ORG, PER, LOC, and FAC.
  • Datasets: GENIA contains five biomedical entity types and uses a 90%/10% train/test split.Its types are DNA, RNA, protein, cell line, and cell type.
  • Evaluation: Strict evaluation requires both entity boundaries and labels to be correct simultaneously, with precision, recall, and F1-score reported.This criterion evaluates exact entity identification and categorization together.
  • Implementation: The encoder generally uses GloVe and BERT, while GENIA uses BioWordvec and BioBERT.Training runs for 35 epochs with Adam, linear warmup-decay, and dropout rate 0.5.

4 Results and Comparisons

The proposed model consistently outperforms state-of-the-art systems on three nested NER datasets and performs comparably on GENIA. It also recognizes entities beyond enumerated training lengths, while filtering and boundary regression improve efficiency and long-entity identification.

  • The model outperforms state-of-the-art systems on KBP17, ACE04, and ACE05, with F1-score gains of +3.08%, +0.71%, and +1.27%, respectively.On GENIA, it achieves comparable performance.
  • Entities of length 6 remain comparable despite not being enumerated during training, and entities longer than the training maximum of 15 are still well recognized.The authors attribute this ability to boundary regression.
  • Removing the boundary regressor decreases long-entity F1-score by 36.73% on ACE04 and 30.54% on KBP17 for entities with L ≥10.This ablation identifies boundary regression as particularly effective for long-entity identification.
  • The full model improves F1-score over the w/o filter setting by 0.52% on ACE04 and 0.75% on KBP17.These ablations evaluate the filter alongside the boundary regressor on the two selected datasets.
  • Previous span-based methods classify almost all N(N+1)/2 spans with O(cN2) complexity, whereas the proposed filter retains span proposals containing entity words.The proposed approach has O(N2) filtering complexity and, in practice, generates approximately three times as many span proposals as entities.
  • With k entities and k << N2, the model’s total time complexity is O(N2 + ck).Although the worst case retains all seed spans, the observed proposal count is much smaller than the total possible span count.

5 Case Study

The case study shows accurate recognition of multi-level nested and long entities, while also exposing difficulty with ambiguous entity references.

  • Nested Entities: The model correctly recognizes three nested entities from inner to outer in the “united nations secretary general kofi annan” example.The entities are “united nations secretary general kofi annan,” “united nations secretary general,” and “united nations.”
  • Ambiguous References: The model incorrectly classifies the ambiguous reference “both sides,” which refers to ORG, as PER.This illustrates difficulty resolving ambiguous entity references.

6 Related Work

Nested NER research includes tagging-based, hypergraph-based, span-based, and reading-comprehension approaches, while object detection provides the paper’s two-stage inspiration.

  • Nested Named Entity Recognition: Traditional sequence labeling assigns one label per token and therefore cannot handle nested structures.Nested NER methods address this limitation with alternative tagging, hypergraph, span-based, or reading-comprehension formulations.
  • Tagging-Based Approaches: Tagging-based approaches encode nested entities with specialized sequential schemas, including layered and pyramid structures.Layered-CRF stacks flat NER layers, whereas Pyramid identifies entities bottom-up with CNNs.
  • Hypergraph-Based Approaches: Hypergraph-based models represent overlapping or nested mentions as hypergraphs and decode their entity structures.Mention Hypergraphs address overlapping mentions, while later work learns nested structures greedily with LSTM networks.
  • Span-Based Approaches: Span-based models extract subsequences and classify them, with Exhaustive Model enumerating all possible spans.Other span-based work uses sequence labeling to identify entity boundaries.
  • Object Detection: Object detection locates regions and assigns categories; two-stage detectors first obtain sparse proposals, then adjust positions and predict categories.This structure motivates locating entities before labeling them in the proposed NER model.

7 Conclusion

The paper proposes a two-stage identifier that filters and regresses span proposals before classification, using boundary and partial-match information to improve nested NER efficiency and performance.

  • Conclusion: The model generates span proposals with a filter and regressor, then classifies them into corresponding entity categories.NER is treated jointly as boundary regression and span classification.
  • Conclusion: The approach uses entity boundary information and soft samples to exploit spans that partially match entities during training.The model also reduces computational cost by filtering seed spans before classification.
  • Conclusion: Experiments demonstrate state-of-the-art performance on several nested NER datasets.The paper proposes future work combining NER and object detection for multimodal joint identification.

A.2 Baseline Methods

The paper compares its model with several nested NER baselines, including structured prediction, pyramid, graph, sequence-to-nuggets, sequence-to-sequence, and prior best-result approaches. It excludes BERT-MRC because that baseline uses external resources to construct category-informed questions.

  • Baseline Methods: The baselines represent nested NER through structured prediction, interconnected span layers, graph modules, head-tail interactions, sequence-to-nuggets, or sequence-to-sequence modeling.Biaffine uses dependency parsing; Pyramid predicts complete entity mentions by span length; BiFlaG separates outermost and inner entities; HIT uses head-tail pairs and token interaction; ARN models head-driven phrase structures; Seq2seq formulates nested NER as sequence generation.
  • Baseline Methods: KBP17-Best reports previous best results for the nested NER entity-discovery task.
  • Baseline Methods: BERT-MRC is omitted because its question construction uses additional external resources that introduce descriptive category information.
  • Baseline Methods: The experiments use detailed model parameter settings reported in Table 6.

A.4 Analysis of Boundary Offset Regression

The paper reports boundary-offset statistics on ACE04 and describes additional flat-NER datasets and dataset-statistics tables used in its experiments. Most predicted offsets are small, indicating that seed spans generally require only slight boundary adjustments.

  • A.4 Analysis of Boundary Offset Regression: Most ACE04 boundary offsets are 1, while offsets of 0 are also frequent because many short entities are already covered by seed spans.The reported counts for offsets 0, 1, 2, 3, and ≥4 are 2162, 2440, 888, 368, and 202, respectively.
  • A.4 Analysis of Boundary Offset Regression: Figure 3 reports ACE04 boundary-offset counts for offsets 0, 1, 2, 3, and ≥4.
  • Flat NER Evaluation: The flat-NER evaluation includes CoNLL03 English and Weibo Chinese, each containing four entity types.CoNLL03 uses Location, Organization, Person, and Miscellaneous; Weibo Chinese uses Person, Organization, Location, and Geo-political.
  • Parameter Settings: Table 6 contains the detailed parameter settings used for the model.

B.3 Results

The model is evaluated on flat NER datasets and outperforms the baselines on Weibo Chinese while remaining comparable on CoNLL03 English. The reported results are summarized in Table 7.

  • Results for flat NER tasks: 0.61% F1-score improvement over baseline models is reported on Weibo Chinese.
  • Results for flat NER tasks: Less than 1% performance drop compared with Yu et al. (2020) is reported on CoNLL03, where the model achieves comparable results.
  • Results for flat NER tasks: Table 7 reports the results for the flat NER tasks.
Loading 2105.06804v2…