A collection of resources on multimodal named entity recognition.
Sorry, this repo will be not updated, since I already changed my research topic.
🐌 Markdown Format:
🌱: Novel idea
📌: The first...
- 🚀: State-of-the-Art
- 👑: Novel dataset/model
- 📚:Downstream Tasks
- 👑 Dataset
-
2020
-
2021
- (AAAI 2021) Multi-modal Graph Fusion for Named Entity Recognition with Targeted Visual Guidance [paper]
- (AAAI 2021) RpBERT: A Text-image Relation Propagation-based BERT Model for Multimodal NER [paper] [code]
- (EMNLP 2021) Can images help recognize entities? A study of the role of images for Multimodal NER [paper] [code]
-
2022
-
(COLING 2022) Flat Multi-modal Interaction Transformer for Named Entity Recognition [paper]
- 📌 1st interpolating FLAT with MNER
- 🚀 SOTA on Twitter15 with Bert_base_uncased but code is unavailable
-
(NAACL Findings 2022) Good Visual Guidance Makes A Better Extractor: Hierarchical Visual Prefix for Multimodal Entity and Relation Extraction [paper] [code]
- 📌 code using refined Twitter15 dataset
-
(WSDM 2022) MAF: A General Matching and Alignment Framework for Multimodal Named Entity Recognition [paper] [code]
-
(SIGIR 2022) Hybrid Transformer with Multi-level Fusion for Multimodal Knowledge Graph Completion [paper] [paper]
- 📌 1st fully Transformer structure
- 🚀 SOTA on Twitter17 using Bert_base_uncased but only implement on Twitter17
-
(NAACL 2022) ITA: Image-Text Alignments for Multi-Modal Named Entity Recognition [paper] [code]
- 📌 Roberta_large as backbone provides powerful improvements
- 🌱 Using OCR ect without directly using images
-
(MM 2022) Query Prior Matters: A MRC Framework for Multimodal Named Entity Recognition [paper]
- 🌱 1st MRC based framework for MNER
-
(SIGIR 2022) Learning from Different text-image Pairs: A Relation-enhanced Graph Convolutional Network for Multimodal NER [paper]
- 📌 Trustworthy performance by reimplementation
-
(ICME 2022) CAT-MNER: Multimodal Named Entity Recognition with Knowledge-Refined Cross-Modal Attention [paper]
- 🚀 SOTA on Twitter15 and Twitter17 with Roberta_large
- 📌 Require 8 V100 GPU
-
(DSAA 2022) PromptMNER: Prompt-Based Entity-Related Visual Clue Extraction and Integration for Multimodal Named Entity Recognition [paper]
- 🚀 SOTA on Twitter15 and Twitter17 with Roberta_large
- 📌 Require 8 V100 GPU
- 🌱 Prompt-based
-
(arxiv 2022) Multi-Granularity Cross-Modality Representation Learning for Named Entity Recognition on Social Media [paper] [code]
-
(arxiv 2021) Multi-Granularity Contrastive Knowledge Distillation for Multimodal Named Entity Recognition
- submitted to ACL2021 but not accepted
-
(arxiv 2022) MNER-QG: An End-to-End MRC framework for Multimodal Named Entity Recognition with Query Grounding [paper]
-
-
2023
- (EMNLP 2023) Named Entity and Relation Extraction with Multi-Modal Retrieval [paper] [code in construction]
- (MM 2023) Learning Implicit Entity-object Relations by Bidirectional Generative Alignment for Multimodal NER [paper]
- 🚀 generative alignment
- 🌱 free from image during inference
-
MNER is a hard task since it needs multimodal understanding in social media domain. However, existing methods simplify it to extacting helpful viusal clue to assist NER, with a simple showcase. In twitter datasets, the image-text pair always has no or vague relationship, which needs extra information or supervision for model to understand. Therefore, I believe that is why MNER-QG, MoRe, R-GCN and PromptMNER work. However, existing works are still nowhere near logical understanding, since they all introduce out-sample knowledge. Now I am trying to introduce knowledge graph in MNER to provide in-sample context.
-
Tricky task: When I developed my work (SOTA in two datasets but in submission), I found
1 only tuning task head of BERT (freeze Bert and ViT) can achieve comparable results (0.2-0.5% drop). So I believe we can directly introduce prompt eniggering for text.
2 large language model matters more than fancy innovation
3 using a empty image to replace all valid images during test only drop 2%
4 simple loss like contractive loss brings 2% improvement. So I think the model heavily focuses on text.
5 Applying the same code in different environments even in different GPUs has results with large variance...
6 It has two mainstreams depending on image+text or caption+text as input. DAMO-series works mainly focus on the latter one, which is proved to be more extandable and SOTA than others.
- Feel free to contact me by email ([email protected])