NER articles
Continuously updating...
Overview/Summary
-
[ Selection Technology ] NER introduces external information
- Mainly starting from the perspective of introducing external information to BERT-NER, investigating specific practices in industry and academia
-
- Alibaba DAMO Academy has compiled an introduction to common Chinese, English, multi-language, and multi-modal NER data sets.
-
Exploration and practice of NER technology in Meituan search
- This paper introduces the characteristics and technology selection of NER tasks in O2O search scenarios, and details the exploration and practice of entity dictionary matching and model construction.
-
- Using BERT as a time node, we will introduce in detail some methods used in the history of NER, as well as some methods after the emergence of BERT.
-
A brief discussion on Nested Named Entity Recognition (Nested NER)
- Traditional NER solutions currently in common use;
- Problems with traditional NER when solving nested NER tasks;
- How to deconstruct NER tasks and solve problems from different perspectives so that the model can identify nested NER;
- Introducing the representative solutions in the field of nested NER in recent years.
-
Summary of papers on named entity recognition at the 2021 ACL Conference (zhihu.com)
-
Summary of entity relationship extraction methods in nlp - Zhihu (zhihu.com) !!!
-
12 golden rules for solving NER problems in industry
- Xi Xiaoyao’s cute house
-
Natural language processing NLP text classification top conference paper reading notes (1)
-
Natural Language Processing NLP Text Classification Top Conference Paper Reading Notes (2)_
Paper/Interpretation
-
Enhanced Language Representation with Label Knowledge for Span Extraction
- Summary of Chinese NER task experiments: Re-optimization of BERT-MRC
- https://github.com/qiufengyuyi/lear_ner_extraction
- The proposed LEAR** (Label knowledge Enhanced Representation)** model architecture attempts to optimize some of the shortcomings of BERT-MRC
-
PCBERT: Parent and Child BERT for Chinese Few-shot NER
- COLING2022 | PCBERT: BERT model for Chinese small sample NER tasks
- A Chinese few-shot NER based on prompt-based P-BERT and C-BERT is proposed. This paper trains an annotation model on high-resource data sets, then discovers more implicit labels on low-resource data sets, and further designs a label expansion strategy to achieve label transmission of high-resource data sets.
-
A Boundary-aware Model for Nested Named Entity Recognition
- Code for EMNLP 2019 paper “A Boundary-aware Neural Model for Nested Named Entity Recognition” (github.com)
- This paper proposes a boundary-aware nested NER neural network model that utilizes the boundaries of entities to predict entity category labels.
-
A Neural Layered Model for Nested Named Entity Recognition
- NAACL2018
- meizhiju/layered-bilstm-crf A Neural Layered Model for Nested Named Entity Recognition(github.com)
- Dynamically overlay flat NER layers to identify nested entities . Each of the flat NER layers is based on the current state-of-the-art flat NER model, which uses a bidirectional long short-term memory (LSTM) network to capture sequence context representation and provide it to the cascaded CRF layer.
-
Deep Exhaustive Model for Nested Named Entity Recognition
- Exhaustive neural network model that exhaustively considers all possible regions of nested NER
-
Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition
- Nested tasks are converted into span predictions. Specifically, it adopts two stages: the first step is Locate, which is to locate the boundary of the entity; the second step is Label, which is to judge the entity type of the recognized span.
- tricktreat/locate-and-label: Code for Two-stage Identifier: “Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition”, accepted at ACL 2021. (github.com)
- A new paper on nested NER interprets ACL2021 - Zhihu (zhihu.com)
-
A Neural Transition-based Joint Model for Disease Named Entity Recognition and Normalization
- In the medical field, there are two ways to model the task of Disease Named Entity Recognition and Normalization . One uses a staged pipeline method, and the other uses joint learning to become multi -task learning. learning ) way.
- Experiments show that the latter learning model performs better, but there is still the problem of boundary inconsistency caused by using different decoders. In addition, the rich text information of candidate words is not considered in the normalization.
- The author proposes a state transfer -based model to transform the end-to-end disease identification and normalization task into an action sequence prediction task, which not only has a model that shares input representations, but also searches the output in the same search space through state transitions.
-
A Span-Based Model for Joint Overlapped and Discontinuous Named Entity Recognition
- Discontinuous entity recognition
- This article also focuses on the learning of overlapping recognition, which means solving both nested and discontinuous entity recognition ( Overlapped and Discontinuous Named Entity Recognition )
- A model based on range recognition is proposed , which includes two steps: first, identifying entity fragments by traversing all possible text ranges to identify overlapping entities. Second, relationship classification is performed to determine whether a given pair of entity fragments is overlapping or continuous.
- ACL2021 | A clever way to solve the problem of NER coverage and discontinuity - Tencent Cloud Developer Community - Tencent Cloud (tencent.com)
-
- Entity identification and standardization in the medical field
- It also solves the problem of error transmission in the pipeline method, and proposes an end-to-end progressive multi-task learning framework ( End-to-End Progressive Multi-Task Learning ).
-
SPANNER: Named Entity Re-/Recognition as Span Prediction
- In recent years, the research paradigm of named entity recognition (NER) has shifted from sequence annotation to range prediction (spanNER)
- Various ways of treating named entity recognition as span-level prediction are mentioned
- question answering
- span classification
- dependency parsing tasks
- Despite showing preliminary effectiveness of spanNER, the model's architectural biases are not yet fully understood. This article first discusses the advantages and disadvantages of the range prediction model and the sequence labeling framework on NER tasks, and how to further improve the model, thereby promoting the complementary advantages of systems based on different paradigms.
-
Span-based Joint Entity and Relation Extraction with Transformer Pre-training (arxiv.org)
- Different from token pairs, which use head and tail tokens to represent this span, span-based uses each token of the span to do pooling to represent this span.
- https://github.com/markus-eberts/spert
- Related work was mentioned
-
Global Pointer: Novel Efficient Span-based Approach for Named Entity Recognition (arxiv.org)
- Su Jianlin's personal blog GlobalPointer: Handle nested and non-nested NER in a unified way - Scientific Spaces (kexue.fm)
- Improved version of Efficient GlobalPointer: fewer parameters, more effects - Scientific Spaces | Scientific Spaces
- Introducing a design called GlobalPointer, which uses the idea of global normalization to perform named entity recognition (NER), which can identify nested entities and non-nested entities without distinction . In the case of non-nested (Flat NER) It can achieve results comparable to CRF, and it also has good results in nested NER situations.
-
Lexicon Enhanced Chinese Sequence Labelling Using BERT Adapter
-
A model called LEBERT ( Lexicon Enhanced BERT ) is proposed to solve the Chinese sequence labeling task. Compared with FLAT, Lattice LSTM and other methods, LEBERT integrates lexical information into the underlying encoding process of BERT. Compared with Lex-BERT, LEBERT does not need a dictionary containing vocabulary type information, only ordinary word vectors.
-
An Embarrassingly Easy but Strong Baseline for Nested Named Entity Recognition
- CNN-NER - an extremely simple and effective nested named entity recognition method
- Fudan University NLP Group
-
A Unified Generative Framework for Various NER Subtasks【已阅】
-
It is proposed to describe the NER subtask as an entity cross-sequence generation task, which can be solved by a unified sequence-to-sequence (Seq2Seq) framework:
Flat NER, nested NER and discontinuous NER subtasks
-
Compare three types of entity representation methods, span, word, BPE
-
Used BART
-
Compare the impact of entity appearance position on recall rate
-
[2204.12031] Boundary Smoothing for Named Entity Recognition (arxiv.org)
- The entity boundary smoothing strategy smoothes the boundary position of the NER label to improve the generalization of the model. Boundary smoothing can prevent the model from being overconfident in predicting entities, thereby obtaining better calibration results.
-
[2112.10070] Unified Named Entity Recognition as Word-Word Relation Classification (arxiv.org)
- W2NERgithub.com
- Convert the NER task to word-word prediction, which can uniformly handle three NER tasks of flat entities, overlapping entities and discontinuous entities, that is, one-size-fits-all
-
[2005.07150] Named Entity Recognition as Dependency Parsing (arxiv.org)
- This article solves Nested NER by using the biaffine model. And this method has good performance in both "flat ner" and "nested ner" tasks.
- In the syntactic parsing (dependency parsing) task, the biaffine model predicts a "head" for each token and then specifies relationships for "head-child pairs".
-
A Local Detection Approach for Named Entity Recognition and Mention Detection - ACL Anthology
- ACL 2017
- Span-based nested named entity recognition method
- This paper proposes a local detection method to identify named entities through fixed-length sentence fragments and their context.
-
[Boundary Enhanced Neural Span Classification for Nested Named Entity Recognition](Boundary Enhanced Neural Span Classification for Nested Named Entity Recognition | Proceedings of the AAAI Conference on Artificial Intelligence)
- AAAI 2020
- Boundary Enhanced Neural Span Classification for Nested Named Entity Recognition | 闲记calculation method (lonepatient.top)
- This paper proposes a boundary-enhanced neural span classification model. In addition to classifying spans, an additional boundary detection task is added to predict words that serve as entity boundaries. And jointly trained under a multi-task learning framework, the span representation is enhanced with additional boundary supervision. In addition, the boundary detection model is able to generate high-quality candidate spans, which greatly reduces the time complexity of the inference process.
-
1908.06926v1Neural Architectures for Nested NER through Linearization
- Two neural network architectures are proposed for nested named entity recognition (NER), in a setting where named entities can overlap and be tagged by multiple labels.
- In the first approach, nested labels are modeled as multi-labels , corresponding to the Cartesian product of nested labels in the standard LSTM-CRF structure.
- In the second approach, nested named entity recognition is viewed as a Seq2Seq problem, in which the input sequence consists of Tokens and the output sequence consists of labels, and Hard Attention is used on the words for which the labels are being predicted. )
-
[2203.10545] Parallel Instance Query Network for Named Entity Recognition (arxiv.org)
- tricktreat/piqn: Code for “Parallel Instance Query Network for Named Entity Recognition”, accepted at ACL 2022. (github.com)
- There are three problems with treating NER as MRC:
- First, a specific type of query can only extract one type of entity per inference, which is inefficient.
- Secondly, the extraction of different types of entities is isolated, ignoring the dependencies between them.
- Third, query construction relies on external knowledge and is difficult to apply to real-life scenarios with hundreds of entity types.
- Parallel Instance Query Network (PIQN) is proposed, which sets global and learnable instance queries to extract entities from sentences in a parallel manner. Each instance query predicts one entity, and by serving all instance queries simultaneously, all entities can be queried in parallel. And instance queries are not built from external knowledge, but different query semantics can be learned during training.
-
- We introduce three kinds of prior knowledge at query sequences, including Wikipedia, annotation scheme, entity dictionary.
- Then, our model adopts a multi-task learning strategy to joint training the main task BioNER and the auxiliary task MRC.
- Finally, experimental results on three benchmark datasets validate the superiority of our BioNER model compared with various state-of-the-art baselines.
-
- This paper proposes a method to exploit the potential multi-granularity information in the dataset to alleviate the lack of training samples. Specifically, the proposed model is based on a multi-task approach, leveraging different training objectives by introducing auxiliary tasks (i.e., binary classification, multi-class classification, and multi-token classification).
-
-
This paper proposes two models based on entity definition information, MRC and SOne, for biomedical NER.
-
These models require no handcrafted features and can achieve micro-average state-of-the-art performance.
-
-
- This paper proposes a BERT-based multi-question MRC (NER-MQMRC) architecture for NER tasks. It can consider multiple entities simultaneously in a single run, thereby improving runtimes for training and inference.