TsinghuaNLP @TsinghuaNLP Twitter profile

Last Seen Profiles

@RainbowBirdy7

@YSoderlu

@vuufll

@BigDaddyDanger2

@stargzie

@the_zada

@mmdmmd76

@wjindemans

@Itsmeurlookin

@IELTSbuddy

@EduardoWeffort1

@jyothi6535

@gaybashingYE24

@Double_TroubleF

@JoeButlerPhD

@Naji_Lzah

@ev_miharuno

@miolly_0

@yowaimomo_

@reme_roman

@CERCA_IITD

@KhasisLieb

@EdgardoNovick

@BelgiuminVienna

@barubikinlol

@k_kinrade

@AlsmanyM58639

@mountainbunny1

@bokeplokalmalam

@glow_mistress07

@gyuned

@anandrewew

@KoronaMatata

@dene_schunck

@DDESGUIDANCE

@ShineFootPics

TsinghuaNLP

@TsinghuaNLP

4 years

Pre-trained language models (PLMs) have brought NLP into a new era. Here is the paper list () tracking the latest progress on PLM including modeling, compression, analysis, and finetuning. Welcome to follow and contribute! #PLM #NLProc #AI #TsinghuaNLP

2

173

540

TsinghuaNLP

@TsinghuaNLP

4 years

OpenMatch (): An open source library for neural information retrieval (IR) research, containing a variety of traditional and neural IR modules, standard training, testing and few-shot training framework. Give it a shot! #IR #NLProc #AI #TsinghuaNLP

1

62

245

TsinghuaNLP

@TsinghuaNLP

4 years

Graph neural networks (GNN) are emerging as an important ML tool, and we build a paper list tracking the most recent advances of GNNs and their applications in NLP/CV/... welcome to follow and contribute. #GNN #NLProc #ML #AI #TsinghuaNLP

GitHub - thunlp/GNNPapers: Must-read papers on graph neural networks (GNN)

Must-read papers on graph neural networks (GNN). Contribute to thunlp/GNNPapers development by creating an account on GitHub.

github.com

0

31

92

TsinghuaNLP

@TsinghuaNLP

3 years

🎉Thrilled to introduce our latest work, Delta Tuning: A Comprehensive Study of Parameter Efficient Methods for Pre-trained Language Models. We perform a comprehensive theoretical analysis and experimental validation of the parametrically efficient paradigm for #PLMs .

1

23

80

TsinghuaNLP

@TsinghuaNLP

3 years

We released a major update of #OpenAttack , an open-source textual adversarial attack toolkit developed by our lab. The related paper is published in the demo track of #ACL2021 . #NLProc #AI #TsinghuaNLP (1/2)

1

6

60

TsinghuaNLP

@TsinghuaNLP

4 years

Pre-train-then-fine-tuning paradigm can hardly help the model capture domain and task related knowledge. We propose a 3-stage framework instead, achieving satisfying performance with less than 50% of computation cost. #NLProc #AI #TsinghuaNLP #EMNLP #PLM

0

9

57

TsinghuaNLP

@TsinghuaNLP

4 years

Welcome to the @TsinghuaNLP Twitter feed, where we'll share new researches and information from TsinghuaNLP Group. Looking forward to interacting with you here! #NLProc

1

11

43

TsinghuaNLP

@TsinghuaNLP

3 years

Our wonderful @TsinghuaNLP team got 13 papers accepted at ACL 2022 Main conference and 5 papers accepted at Findings of ACL 2022 @aclmeeting ! #NLProc #AI #ACL2022 Congratulations to everyone! More details:

0

5

44

TsinghuaNLP

@TsinghuaNLP

3 years

Our work "Few-Shot Conversational Dense Retrieval" w/ @dgdsxyushi • @ZhenghaoLiu4 • @XiongChenyan • @Op_enmind • @zibuyu9 has been accepted at #SIGIR2021 as a full paper. We propose a few-shot strategy that effectively trains a conversational dense retriever. #NLProc #AI #TsinghuaNLP

0

8

40

TsinghuaNLP

@TsinghuaNLP

3 years

BERT-KPE toolkit provides several standard keyphrase extraction models that can be easily adapted to any custom data in any domain🤗. #NLProc #AI (1/3)

1

4

39

TsinghuaNLP

@TsinghuaNLP

2 years

Welcome to follow our #NAACL -HLT 2022 paper. We empirically investigate the transferability of soft prompts across tasks/PLMs and introduce a new transferability indicator. #NLProc #AI #TsinghuaNLP (1/4) Paper： Data/Code:

1

11

39

TsinghuaNLP

@TsinghuaNLP

2 years

We are excited to share our latest work, we present ProQA, a unified QA paradigm that solves various tasks through a single model, which got accepted by #NAACL -HLT 2022 as a long paper. #NLProc #AI #TsinghuaNLP (1/2)

1

11

38

TsinghuaNLP

@TsinghuaNLP

4 years

Relation extraction (RE) aims to extract relational facts of entities from plain text, which may help with building knowledge graphs. Check out our review paper summarizing the past and future directions of RE models (). #NLProc #AI #TsinghuaNLP (1/2)

More Data, More Relations, More Context and More Openness: A...

Relational facts are an important component of human knowledge, which are hidden in vast amounts of text. In order to extract these facts from text, people have been working on relation extraction...

arxiv.org

1

5

35

TsinghuaNLP

@TsinghuaNLP

2 years

We are excited to share our latest work " Knowledgeable Prompt-tuning: Incorporating Knowledge into Prompt Verbalizer for Text Classification", which got accepted by #ACL2022 as a long paper in the main conference. #NER #RE #NLProc #AI (1/2)

1

4

35

TsinghuaNLP

@TsinghuaNLP

2 years

Welcome to follow our #ACL2022 paper. We present a method of using PLMs to perform machine translation. #NLProc #AI #TsinghuaNLP (1/2)

1

4

35

TsinghuaNLP

@TsinghuaNLP

2 years

🎉We are excited to announce that our #ACL2022 paper “OpenPrompt: An Open-source Framework for Prompt-learning” has received the Best Demo Paper Award. Congratulations to all the co-authors! #acl2022nlp #NLProc

0

3

33

TsinghuaNLP

@TsinghuaNLP

2 years

Thrilled to introduce our latest work, we vectorize discrete lexical constraints so that constraints can be directly integrated into NMT models. Experimental results show that our method performs well interms of both BLEU and CSR. #ACL2022 #TsinghuaNLP . (1/2)

1

3

33

TsinghuaNLP

@TsinghuaNLP

3 years

Our wonderful @TsinghuaNLP team got 8 papers accepted at ACL-IJCNLP 2021 and 6 papers accepted at Findings of ACL 2021 @aclmeeting ! #NLProc #AI #ACL2021 Congratulations to everyone! More details:

0

4

32

TsinghuaNLP

@TsinghuaNLP

3 years

🎉Excited to present OpenDelta, a flexible parameter efficient tuning (we also call it delta tuning) toolkit for pre-trained language models! Github: OpenDelta has the following features:

GitHub - thunlp/OpenDelta: A plug-and-play library for parameter-efficient-tuning (Delta Tuning)

A plug-and-play library for parameter-efficient-tuning (Delta Tuning) - thunlp/OpenDelta

github.com

1

5

32

TsinghuaNLP

@TsinghuaNLP

3 years

We investigate prompt tuning in few-shot learning and find that soft prompts are hard to optimize which leads to their bad performance when the training data are insufficient. #ACL2022 #TsinghuaCoAI #TsinghuaNLP

1

7

29

TsinghuaNLP

@TsinghuaNLP

3 years

It is important to incorporate rich knowledge in KGs into pre-trained models. ERNIE () is one of the first works to pre-training with both large corpora and KGs for improving knowledge-intensive tasks. #NLProc #AI #TsinghuaNLP

0

5

28

TsinghuaNLP

@TsinghuaNLP

8 months

We're proud to introduce #𝐌𝐢𝐧𝐢𝐂𝐏𝐌🏆, our 2.4B parameter model that achieves state-of-the-art performance with minimal size🤏 Outperforming models 10x its size, MiniCPM sets a new bar for efficiency💡 and deployment flexibility. 📱🚀 GitHub：

1

10

28

TsinghuaNLP

@TsinghuaNLP

3 years

Our ACL 2021 paper presents one of the largest manually labeled, fine-grained NER datasets. Few-NERD can be used for supervised/few-shot/continual NER or entity typing. #NLProc #AI paper: code: website:

0

6

27

TsinghuaNLP

@TsinghuaNLP

3 years

CokeBERT can dynamically select and embed contextual knowledge according to textual context for PLMs, which outperforms conventional knowledge-enhanced PLMs in NLU tasks. #PLM #NLProc #AI #TsinghuaNLP paper: code:

1

5

25

TsinghuaNLP

@TsinghuaNLP

2 years

🎉🎉Our group's latest paper about in the frontier of pre-trained language models, "Parameter-efficient Fine-tuning of Large-scale Pre-trained Language Models", was published in Nature Machine Intelligence! #NLProc #AI #TsinghuaNLP

2

3

25

TsinghuaNLP

@TsinghuaNLP

2 years

We are excited to share our latest work "Exploring the Universal Vulnerability of Prompt-based Learning Paradigm", which got accepted by #NAACL -HLT 2022 as a findings paper. #NLProc #AI #TsinghuaNLP (1/2)

1

6

25

TsinghuaNLP

@TsinghuaNLP

4 years

Adversarial attack and defense has become a very important research topic for deep learning. We build a reading list that includes almost all published papers about textual adversarial attack and defense (). (1/2)

GitHub - thunlp/TAADpapers: Must-read Papers on Textual Adversarial Attack and Defense

Must-read Papers on Textual Adversarial Attack and Defense - thunlp/TAADpapers

github.com

1

2

23

TsinghuaNLP

@TsinghuaNLP

2 years

Welcome to follow our #ACL2022 paper. We build a large-scale quotation recommendation dataset named QuoteR and a quotation recommendation model. #NLProc #AI #TsinghuaNLP (1/2)

2

4

23

TsinghuaNLP

@TsinghuaNLP

3 years

Check out our recent paper in ACL 2021! We propose a contrastive learning framework ERICA to improve the understanding of the entities and their relations for pre-trained language models. See paper [] #NLProc #AI #TsinghuaNLP #ACL2021

0

9

23

TsinghuaNLP

@TsinghuaNLP

4 years

Pre-trained models show effectiveness in knowledge transfer, potentially alleviating data sparsity problem in recommender systems. We published a review and prospect of recommender systems with pre-training on Front. Big Data (). #NLProc #AI #TsinghuaNLP

0

6

21

TsinghuaNLP

@TsinghuaNLP

2 years

Welcome to follow our #ACL2022 Findings paper. We propose DPT, the first prompt tuning framework for discriminative pre-trained language models, which reformulates NLP tasks into a discriminative language modeling problem. #NLProc #AI #TsinghuaNLP (1/2)

1

8

23

TsinghuaNLP

@TsinghuaNLP

3 years

Our lab released a unified prompt-learning toolkit #OpenPrompt , which aims to make it easy to deploy prompt-learning framework to solve various NLP problems based on pre-trained models. Welcome to use and contribute! #NLProc #AI #TsinghuaNLP Github:

0

6

23

TsinghuaNLP

@TsinghuaNLP

3 years

How to design a powerful graph neural network (GNN) model? In our updated GNN survey, we break down the design process into four steps: find graph structure, specify graph type and scale, design loss function and build the model using computational modules. #GNN #NLProc #AI (1/2)

1

6

21

TsinghuaNLP

@TsinghuaNLP

4 years

Few-shot relation classification aims to alleviate the heavy reliance on labeled data. Our recent work () leverages meta-information to guide meta-learning for few-shot relation classification. #NLProc #AI #TsinghuaNLP

Meta-Information Guided Meta-Learning for Few-Shot Relation Classification

Bowen Dong, Yuan Yao, Ruobing Xie, Tianyu Gao, Xu Han, Zhiyuan Liu, Fen Lin, Leyu Lin, Maosong Sun. Proceedings of the 28th International Conference on Computational Linguistics. 2020.

aclanthology.org

0

4

21

TsinghuaNLP

@TsinghuaNLP

3 years

MAVEN is a massive general domain event detection dataset, which contains 4,480 Wikipedia documents, 118,732 event mentions, and 168 event types. Check out our EMNLP 2020 paper (). #NLProc #AI #TsinghuaNLP #EMNLP

0

4

21

TsinghuaNLP

@TsinghuaNLP

3 years

On Transferability of Prompt Tuning for Natural Language Understanding We empirically investigate the transferability of soft prompts across different tasks and models to demonstrate some promising directions worth exploring in prompt. ArXiv: #nlp #prompt

1

5

21

TsinghuaNLP

@TsinghuaNLP

2 years

Recently, we propose a cross-lingual contrastive learning framework to learn fine-grained entity typing (FGET) models for low-resource languages. This paper has been accepted to appear at the #ACL2022 main conference. #NLProc #AI #TsinghuaNLP (1/2)

1

4

20

TsinghuaNLP

@TsinghuaNLP

4 years

Multi-hop reasoning aims to complete knowledge graphs in an interpretable way. Our recent work DacKGR () proposes two strategies (dynamic anticipation and completion) to improve the performance on sparse knowledge graphs. #KG #NLProc #AI #TsinghuaNLP

0

5

18

TsinghuaNLP

@TsinghuaNLP

3 years

🔖Wanna continually pre-train PLMs for emerging data? Check our recent work ELLE, which is accepted by the findings of ACL 2022! 🥳 ELLE could recycle an existing PLM’s parameters and efficiently acquire new knowledge from emerging corpora. 🧐Paper:

1

6

19

TsinghuaNLP

@TsinghuaNLP

3 years

We propose a novel text style transfer model, StyIns, which combines the attention mechanism and a generative flow model to achieve a better balance between content preservation and style transfer accuracy. #IJCAI #NLProc #TsinghuaNLP

0

8

19

TsinghuaNLP

@TsinghuaNLP

3 years

Our recent work on Chinese word segmentation got accepted in #EMNLP2021 ! Here's a research blog for a brief introduction:

Segment, Mask, and Predict: Augmenting Chinese Word Segmentation with Self-Supervision

We propose a self-supervised CWS approach with a straightforward and effective architecture, which outperforms previous methods on 9 different CWS datasets.

thumtblog.github.io

0

1

19

TsinghuaNLP

@TsinghuaNLP

2 years

We are excited to share our latest work "Packed Levitated Marker for Entity and Relation Extraction ", which got accepted by #ACL2022 as a long paper in the main conference. #NER #RE #NLProc #AI (1/2)

1

4

19

TsinghuaNLP

@TsinghuaNLP

3 years

KEPLER () is a unified model for knowledge embedding (KE) and pre-trained language model (PLM) representation. It encodes textual entity descriptions with a PLM and then jointly optimize KE and language modeling objectives. #NLProc #AI #TsinghuaNLP

1

8

19

TsinghuaNLP

@TsinghuaNLP

3 years

💥Proud to announce that OpenPrompt supports the "generation for all tasks" paradigm! By using a GenerationVerbalizer, anything in the InputExample could become the target words, and any tasks (classification, generation) could be carried out in a unified generation style.

1

3

19

TsinghuaNLP

@TsinghuaNLP

3 years

🎉 OpenPrompt hits 1,000 stars ! Thanks for all the contributors and supporters of this awesome project. #NLProc #AI URL:

0

3

19

TsinghuaNLP

@TsinghuaNLP

4 years

OpenKE () is an open-source framework for knowledge graph embedding (KGE) including many popular KGE methods such as TransE, TransR, RotatE. A new version supporting distributed training of large KGs is coming soon, stay tuned~ #KGE #NLProc #AI #TsinghuaNLP

GitHub - thunlp/OpenKE: An Open-Source Package for Knowledge Embedding (KE)

An Open-Source Package for Knowledge Embedding (KE) - thunlp/OpenKE

github.com

0

1

19

TsinghuaNLP

@TsinghuaNLP

2 years

We proposed a novel span representation approach to consider the interrelation between the spans by strategically packing the markers in the encoder. For more details, please refer to: . Welcome to leave your valuable comments on the post. #TsinghuaNLP (2/2)

0

2

18

TsinghuaNLP

@TsinghuaNLP

2 years

We construct a pluggable entity vocabulary on demand by aggregating the output representations of an entity in different sentences. The paper got accepted by #ACL2022 as a short paper in the main conference. #NLProc #AI #TsinghuaNLP (1/2)

1

6

18

TsinghuaNLP

@TsinghuaNLP

3 years

Grammatical error correction (GEC) aims to correct writing errors. We present VERNet for GEC quality estimation with multi-hypotheses. VERNet achieves SOTA and improves GEC performance. #NLProc #AI #TsinghuaNLP Paper: Code:

1

5

17

TsinghuaNLP

@TsinghuaNLP

3 years

We propose a latent model to generate stories from titles. By modeling an outline as a discrete latent variable to bridge the title and story, our model can predict explainable outlines to generate good stories. Paper: #NLProc #AI #TsinghuaNLP #TASLP

1

6

17

TsinghuaNLP

@TsinghuaNLP

3 years

Conversational search systems interact with users through dialogues. Our recent work () demonstrates the few-shot ability of GPT-2 for conversational query rewriting and proposes two methods to generate weak supervision data. #NLProc #AI #TsinghuaNLP

1

2

17

TsinghuaNLP

@TsinghuaNLP

3 years

On the occasion of the 37th Chinese Teacher's Day, THUNLP students prepared flowers and gifts for their professors to thank them for the guidance and dedication. Happy #TeachersDay to all the teachers!

0

16

TsinghuaNLP

@TsinghuaNLP

4 years

THUMT () is an open-source toolkit for neural machine translation with implementations of PyTorch and TensorFlow. It is user-friendly and extendable for the latest advances in NMT research. Give it a try! #AI #NLProc #MT #TsinghuaNLP

GitHub - THUNLP-MT/THUMT: An open-source neural machine translation toolkit developed by Tsinghua...

An open-source neural machine translation toolkit developed by Tsinghua Natural Language Processing Group - THUNLP-MT/THUMT

github.com

0

3

17

TsinghuaNLP

@TsinghuaNLP

5 months

We're excited to introduce 💫𝗜𝘁𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗘𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲 𝗥𝗲𝗳𝗶𝗻𝗲𝗺𝗲𝗻𝘁 (IER), a method that lets AI Agents quickly adapt to new tasks by improving experiences across tasks. 📑 Stay tuned for future integration! 🌟

GitHub - OpenBMB/ChatDev: Create Customized Software using Natural Language Idea (through LLM-pow...

Create Customized Software using Natural Language Idea (through LLM-powered Multi-Agent Collaboration) - OpenBMB/ChatDev

github.com

0

4

17

TsinghuaNLP

@TsinghuaNLP

4 years

OpenNE is an open source toolkit for Network Embedding, providing a standard NE training and testing framework. Recently, pytorch version () has been released with more models, more datasets and a new framework. Come and try it! #NLProc #AI #TsinghuaNLP

0

5

14

TsinghuaNLP

@TsinghuaNLP

3 years

Back translation (BT) is a widely-used approach to leverage synthetic data in NMT. However, adding large-scale synthetic data often deteriorates translation performance. Our ACL Findings paper proposes alternated training on synthetic and authentic data to alleviate this problem.

1

5

15

TsinghuaNLP

@TsinghuaNLP

2 years

We propose a fully hyperbolic framework based on Lorentz transformations to build hyperbolic neural networks. The framework contains basic blocks such as linear layer, attention, etc. #ACL2022 #NLProc #AI (1/2)

1

5

14

TsinghuaNLP

@TsinghuaNLP

3 years

Our ACL Findings paper proposed a large-scale multi-task benchmark for knowledge abstraction, concretization and completion (KACC), which focused on two-view knowledge graphs and provided thorough experiments on existing methods. #NLProc #ACL2021 Paper:

1

14

TsinghuaNLP

@TsinghuaNLP

4 years

AI techniques may benefit legal practice with better fairness, transparency & efficiency, and thus Legal AI draws attention from both AI & Law community. We build legal AI paper list () for the most recent advances. #NLProc #legaltech #AI #TsinghuaNLP

GitHub - thunlp/LegalPapers: Must-read Papers on Legal Intelligence

Must-read Papers on Legal Intelligence. Contribute to thunlp/LegalPapers development by creating an account on GitHub.

github.com

0

6

13

TsinghuaNLP

@TsinghuaNLP

4 years

Topjudge builds a directed acyclic graph by topological order to pass information between legal judgment prediction tasks. Experiment results are promising when applying Topjudge with simple encoders. Paper: #legalAI #NLProc #AI #TsinghuaNLP

0

6

14

TsinghuaNLP

@TsinghuaNLP

3 years

We propose a novel model, MixPoet to achieve stylistic Chinese poetry generation. Based on a VAE framework, MixPoet mixes factors that influence the composition of human poets to create distinct styles. #Jiuge #NLProc #TsinghuaNLP

0

5

14

TsinghuaNLP

@TsinghuaNLP

4 months

The RLAIF-V-Dataset, the preference dataset for #MiniCPM -Llama3-V 2.5, has reached 𝟮𝗻𝗱 𝗽𝗹𝗮𝗰𝗲🏆 on Hugging Face Trending! Great data leads to great models. This large-scale, high-quality multimodal dataset contains 83,132 preference pairs from a diverse range of tasks.

Tianyu Yu

@YiranyYu47806

4 months

🚀Excited to introduce the RLAIF-V framework, which aligns MLLMs through open-source AI feedback for super GPT-4V trustworthiness. Github: Paper: Dataset:

0

1

7

0

2

14

TsinghuaNLP

@TsinghuaNLP

3 years

Distant supervision (DS) has achieved success in sentence-level relation extraction, while cannot be directly applied to document-level data due to the vast inherent noise. We propose 3 pre-training tasks to denoise the data. #NLProc #TsinghuaNLP #EMNLP

0

1

12

TsinghuaNLP

@TsinghuaNLP

2 years

Happy that our #TsinghuaNLP team got a paper accepted at # ECCV 2022 oral presentation. Congratulations to everyone! #NLProc #AI (1/2)

2

1

13

TsinghuaNLP

@TsinghuaNLP

3 years

We propose to incorporate sememes, the minimum semantic units of human languages, into recurrent neural networks (RNNs) to improve their sequence modeling ability, which is beneficial to all kinds of downstream tasks. #NLProc #AI #TsinghuaNLP #TASLP

0

5

13

TsinghuaNLP

@TsinghuaNLP

4 years

Recently, @paperdigestorg published the list of the most influential papers in various fields, and 9 papers of our group were selected into the list of the most influential papers of #AAAI , #ACL , #IJCAI and #EMNLP . The 9 selected papers are as follows: #NLProc #AI #TsinghuaNLP

2

4

13

TsinghuaNLP

@TsinghuaNLP

1 year

🎉 We present 1.5 million high-quality instructional conversation data, UltraChat, to facilitate the research of chat LLMs. And model trained on UltraChat shows extraordinary performance that surpasses all the open-source baselines!

Ning Ding

@stingning

1 year

Meet (one of) the strongest LLaMA LLM🦙. Lots of studies show that getting satisfactory instruction-following LLMs is easy and only needs a few thoughtfully selected samples. However, surpassing "satisfactory performance" to chase stronger models is extremely challenging.

1

6

16

0

4

13

TsinghuaNLP

@TsinghuaNLP

2 years

Our group released ProtoVerb on ACL 2022, which learns prototypes as verbalizer directly from training data by contrastive learning. In this way, prototypes are able to contain rich class-level semantics. Welcome to use and contribute! #NLProc #AI #TsinghuaNLP

0

6

13

TsinghuaNLP

@TsinghuaNLP

3 years

A sememe is defined as the minimum semantic unit of human languages. Some linguists believe that the meanings of all words can be expressed by a limited set of sememes. Many languages have no sememe knowledge bases (SKBs), and the construction of SKBs is very expensive. (1/3)

2

3

13

TsinghuaNLP

@TsinghuaNLP

3 years

Oracle bone script (OBS) is the earliest known ancient Chinese writing system. Our information system for OBS, IsOBS, applies advanced machine learning methods to promote OBS research. IsOBS now support char-level and doc-level retrieval, try it on !

1

2

13

TsinghuaNLP

@TsinghuaNLP

3 years

Congratulations! Our recent paper “A Deep-learning System Bridging Molecule Structure and Biomedical Text with Comprehension Comparable to Human Professionals” was selected as @NatureComms Editors’ Highlights webpage under the “Biotechnology and methods”. #NLProc #AI #KVPLM

1

2

13

TsinghuaNLP

@TsinghuaNLP

2 years

Thrilled to introduce our latest work, we explore the question how could existing PLMs benefit training large-scale PLMs in future. The paper got accepted by # NAACL-HLT 2022 as a long paper (oral) in the main conference. #NLProc #AI #TsinghuaNLP (1/2)

1

0

13

TsinghuaNLP

@TsinghuaNLP

3 years

Fact Verification requires to find subtle clues and identify misinformation. Kernel Graph Attention Network (KGAT) conducts kernel-based attention mechanism for fine-grained fact verification. #FactVerification #KernelGAT #NLProc #AI #TsinghuaNLP

Fine-grained Fact Verification with Kernel Graph Attention Network

Zhenghao Liu, Chenyan Xiong, Maosong Sun, Zhiyuan Liu. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.

aclanthology.org

0

6

13

TsinghuaNLP

@TsinghuaNLP

3 years

To learn incessantly emerging novel relations, we introduce a memory-based method for continual relation learning, as well as a prototype-based reconsolidation mechanism to make better use of memorized examples. #NLProc #AI #TsinghuaNLP #ACL

1

4

13

TsinghuaNLP

@TsinghuaNLP

4 years

Jiuge is an online Chinese classical poetry generation system developed by the THUNLP lab, which is built on several specially designed neural models and supports multi-modal inputs and various genres and styles. Welcome to use it! #NLProc #AI #TsinghuaNLP

1

2

11

TsinghuaNLP

@TsinghuaNLP

3 years

CorefBERT () leverages repeated mentions that appear multiple times in the passage to acquire abundant co-referring relations, and then explicitly model the coreference information during pre-training. #EMNLP #NLProc #AI #TsinghuaNLP

1

2

12

TsinghuaNLP

@TsinghuaNLP

2 years

Recently, our group opened up OpenBackdoor, which can be used to reproduce, evaluate and develop related algorithms for text backdoor attack and defense. Welcome to use and leave your valuable comments. Link:

0

2

12

TsinghuaNLP

@TsinghuaNLP

3 years

Our demo paper "OpenMatch: An Open-Source Library for Neu-IR Research" has been accepted by #sigir2021 ! 👏 Congrats to the authors @ZhenghaoLiu4 @An4Lm9vT7peHQyg @XiongChenyan @zibuyu9 🎉🎉

TsinghuaNLP

@TsinghuaNLP

4 years

OpenMatch (): An open source library for neural information retrieval (IR) research, containing a variety of traditional and neural IR modules, standard training, testing and few-shot training framework. Give it a shot! #IR #NLProc #AI #TsinghuaNLP

1

62

245

0

5

12

TsinghuaNLP

@TsinghuaNLP

3 years

Our #acl2021nlp Findings paper presents a simple solution that enhances adversarial data augmentation with mixup, which significantly pretrained language models’ robustness accuracy under adversarial attacks. paper: code:

1

0

12

TsinghuaNLP

@TsinghuaNLP

2 years

Congratulations to our five students on their graduation！Fanchao Qi, Xuancheng Huang and Xu Han received Doctorate degrees. Huimeng Zhang and Yizhou Xu received Masters degrees. We wish you all the very best for the future. #TsinghuaNLP

1

10

TsinghuaNLP

@TsinghuaNLP

4 years

Our group members also develop OpenNRE (), a package that contains several SOTA neural relation extraction models, including the trending pre-trained ones. Welcome to try it out! #NLProc #AI #TsinghuaNLP (2/2)

GitHub - thunlp/OpenNRE: An Open-Source Package for Neural Relation Extraction (NRE)

An Open-Source Package for Neural Relation Extraction (NRE) - thunlp/OpenNRE

github.com

0

4

10

TsinghuaNLP

@TsinghuaNLP

3 years

Happy that our @TsinghuaNLP team got 5 papers accepted at #NAACL2022 @naaclmeeting Congratulations to everyone! More details:

0

11

TsinghuaNLP

@TsinghuaNLP

4 years

The open-access book RL4NLP by our group members Zhiyuan Liu, Yankai Lin & Maosong Sun is now one of the Springer Nature 2020 highlights! 😁 Welcome download here: #NLProc #TsinghuaNLP

1

10

TsinghuaNLP

@TsinghuaNLP

4 years

How does NLP benefit legal systems? Legal AI focuses on applying AI techniques to legal tasks. In this work (), we summarize the history, the current state and future directions of researches in legal AI. #NLProc #AI #legalAI #TsinghuaNLP

0

4

11

TsinghuaNLP

@TsinghuaNLP

2 years

We released an open-source event extraction toolkit OmniEvent: . OmniEvent is a powerful open-source toolkit for event extraction, including event detection and event argument extraction.

GitHub - THU-KEG/OmniEvent: A comprehensive, unified and modular event extraction toolkit.

A comprehensive, unified and modular event extraction toolkit. - THU-KEG/OmniEvent

github.com

0

1

11

TsinghuaNLP

@TsinghuaNLP

3 years

We do an empirical study on neural relation extraction and propose an entity-masked contrastive pre-training framework for RE. Check out our EMNLP 2020 paper () for more details. #NLProc #AI #TsinghuaNLP

1

3

11

TsinghuaNLP

@TsinghuaNLP

3 years

We propose QAjudge, which simulates element trial by iteratively questioning and answering between a virtual judge and a criminal. Our model provides interpretability for judgment prediction. Paper: #legalAI #NLProc #AI #TsinghuaNLP

0

3

10

TsinghuaNLP

@TsinghuaNLP

3 years

We investigate the application of prompt-learning on fine-grained entity typing in fully supervised, few-shot and zero-shot scenarios, where entity types are pre-defined. We first develop a naive pipeline of prompt-learning for entity typing. #NLProc #AI #TsinghuaNLP (1/2)

1

10

TsinghuaNLP

@TsinghuaNLP

4 years

Calibration aims to make the model confidence equal to the true correctness of model predictions. In this work, we find that miscalibration still remains a challenge for NMT at the inference time due to the exposure bias problem. #NLProc #TsinghuaNLP #ACL

0

2

10

TsinghuaNLP

@TsinghuaNLP

1 year

🚀 Meet XAgent ( @XAgentTeam ) , the supermodel of AI agents, developed in collaboration with @OpenBMB ✨It's a game-changer, outperforming AutoGPT and GPT-4. 🌟Excited to see more updates on this game-changing work! #AI #XAgent

Yujia Qin

@TsingYoga

1 year

🚀 Introducing XAgent: The next evolution of AI agents designed for intricate task-solving. XAgent completely outperforms AutoGPT and GPT-4 on various tasks and benchmarks. 💡 XAgent's dual-loop mechanism bridges the gap between high-level planning and detailed task execution.

2

46

230

1

2

10

TsinghuaNLP

@TsinghuaNLP

3 years

We propose TRICE for transferring pretrained models to multi-source sequence generation tasks (e.g., automatic post-editing, multi-source translation, and multi-document summarization). Blog: #NLProc #AI #TsinghuaNLP #ACL2021

0

4

10

TsinghuaNLP

@TsinghuaNLP

4 years

To facilitate research on Information Retrieval (IR), we develop OpenMatch (), a package providing functional interfaces for IR modules. OpenMatch supplies a series of widely-used sparse/dense retrievers and neural rerankers. #NLProc #AI #TsinghuaNLP

GitHub - thunlp/OpenMatch: An Open-Source Package for Information Retrieval.

An Open-Source Package for Information Retrieval. Contribute to thunlp/OpenMatch development by creating an account on GitHub.

github.com

0

5

10

TsinghuaNLP

@TsinghuaNLP

2 years

We present structured sememe prediction to organize word meanings as a tree structure of sememes. The paper was accepted by #ACL2022 as a findings paper. #NLProc #AI #TsinghuaNLP (1/2)

1

3

10

TsinghuaNLP

@TsinghuaNLP

2 years

we focused on integrating external knowledge into the presenter, forming a knowledge-enhanced prompt-tuning method (KPT) to improve and stabilize the verbalizer. Extensive experiments on zero-shot and low-shot text classification tasks demonstrate the effectiveness of KPT. (2/2)

0

1

10

TsinghuaNLP

@TsinghuaNLP

3 years

HMEAE is an event argument extraction model. It provides effective inductive bias from the concept hierarchy of event argument roles by using a neural module network to imitate the hierarchical structure of concepts. #NLProc #AI #TsinghuaNLP #EMNLP Paper:

0

5

10

TsinghuaNLP

@TsinghuaNLP

3 years

Adversarial attacks reveal the vulnerability of deep neural networks. In our recent work (), we conduct word-level textual attacks for NLP models by a two-step combinatorial optimization method. #NLProc #AI #TsinghuaNLP (1/2)

1

4

9

TsinghuaNLP

@TsinghuaNLP

3 years

We propose an approach for system combination in machine translation which aims to combine erroneous but complementary hypotheses to obtain better translations by modeling voting among hypotheses. Paper: #NLProc #TsinghuaNLP #IJCAI2020

0

2

9

TsinghuaNLP

@TsinghuaNLP

4 years

Congratulations @TsinghuaNLP students and collaborators for their papers accepted to #NAACL -HLT 2021! Topics include natural language generation, pre-trained language models and relation extraction. Preprints coming. #NLProc #AI

0

8

TsinghuaNLP

@TsinghuaNLP

3 years

Our #ACL2021 paper presents LWS, a kind of textual backdoor attacks based on dynamic word substitution. Experimental results show its high attack performance and invisibility. #TsinghuaNLP Paper: Code:

1

9

TsinghuaNLP

@TsinghuaNLP

3 years

Our ACL 2021 paper proposed a contrastive pre-training framework CLEVE for event extraction task, utilizing AMR structures to make PLMs better understand semantic relations between triggers and arguments. Our paper: . #NLProc #ACL2021 #AI #TsinghuaNLP

0

3

9