Structuring: Knowledge Structuring to Help Retrieval and Augmented Generation

Pengcheng Jiang, Siru Ouyang, Yizhu Jiao,

Ming Zhong, Runchu Tian, Jiawei Han

Department of Computer Science

University of Illinois at Urbana-Champaign

August 3, 2025

1

Outline

• Text Classification
• Entity Typing and Entity Structure Mining
• Relation Extraction and Knowledge Graph Construction

2

Outline

• Text Classification
• Entity Typing and Entity Structure Mining
• Relation Extraction and Knowledge Graph Construction

3

Outline

• What is weakly-supervised text classification and why do we care?
• Weakly-supervised flat text classification
• ConWea [ACL’20], LOTClass [EMNLP’20], ClassKG [EMNLP’21], X-Class [NAACL’21], MEGClass [EMNLP’23], PIEClass [EMNLP’23], CARP [EMNLP’23]
• Weakly-supervised hierarchical text classification
• WeSHClass [AAAI’19], TaxoClass [NAACL’21], TELEClass [WWW’25]

4

Text Classification

• Given a set of text units (e.g., documents, sentences) and a set of categories, the task is to assign relevant category/categories to each text unit
• Text Classification has a lot of downstream applications

Sentiment Analysis

Location Prediction

News Topic Classification

Paper Topic Classification

Email Intent Identification

Hate Speech Detection

5

Different Text Classification Settings: �Single-Label vs. Multi-Label

• Single-label: Each document belongs to one category.
• E.g., Spam Detection

​

​

​

​

​

• Multi-label: Each document has multiple relevant labels.
• E.g., Paper Topic Classification

https://academic.microsoft.com/paper/2963341956/

6

Different Text Classification Settings: �Flat vs. Hierarchical

• Flat: All labels are at the same granularity level
• E.g., Sentiment Analysis of E-Commerce Reviews (1-5 stars)

​

​

​

​

• Hierarchical: Labels are organized into a hierarchy representing their parent-child relationship
• E.g., Paper Topic Classification (the arXiv category taxonomy)

https://www.amazon.com/gp/product/B089YFHYYS/

https://arxiv.org/abs/1810.04805

7

Weakly-Supervised Text Classification: Motivation

• Supervised text classification models (especially recent deep neural models) rely on a significant number of manually labeled training documents to achieve good performance.
• Collecting such training data is usually expensive and time-consuming. In some domains (e.g., scientific papers), annotations must be acquired from domain experts, which incurs additional cost.

​

• While users cannot afford to label sufficient documents for training a deep neural classifier, they can provide a small amount of seed information:
• Category names or category-related keywords
• A small number of labeled documents

​

8

Weakly-Supervised Text Classification: Definition

?

• Text classification without massive human-annotated training data
• Keyword-level weak supervision: category names or a few relevant keywords
• Document-level weak supervision: a small set of labeled docs

9

General Ideas to Perform �Weakly-Supervised Text Classification

• Joint representation learning
• Put words, labels, and documents into the same latent space using embedding learning or pre-trained language models

​

• Pseudo training data generation
• Retrieve some unlabeled documents or synthesize some artificial documents using text embeddings or contextualized representations
• Give them pseudo labels to train a text classifier

​

• Transfer the knowledge of pre-trained language models to classification tasks

10

11

LOTClass: Label-Name-Only Text Classification

11

LOTClass: Meaning of Word Is Context-Dependent

12

LOTClass: Contextualized Word-Level Topic Prediction

13

LOTClass: Experiment Results

14

14

PIEClass: Prompt-based Fine-tuning for �Text Classification

• Head token fine-tuning randomly initializes a linear classification head and directly predicts class distribution using the [CLS] token, which needs a substantial amount of training data.
• Prompt-based fine-tuning for MLM-based PLM converts the document into the masked token prediction problem by reusing the pre-trained MLM head.
• Prompt-based fine-tuning for ELECTRA-style PLM converts documents into the replaced token detection problem by reusing the pre-trained discriminative head.

​

Zhang, Y., Jiang, M., Meng, Y., Zhang, Y., & Han, J. “PIEClass: Weakly-Supervised Text Classification

with Prompting and Noise-Robust Iterative Ensemble Training”, EMNLP’23.

15

PIEClass: Integrating Head Token �& Prompt-based Fine-tuning

• Why do we need prompts to get pseudo training data?
• Simple keyword matching may induce errors. We use prompts to contextualize the documents.
• E.g., “die” is a negative word, but a food review “It is to die for!” implies a strong positive sentiment.
• Noise-Robust Training with Iterative Ensemble
• Uses most confident predictions to improve and expand pseudo labels iteratively
• Difference with semi-supervised self-training: potential noise in the initial pseudo labels
• Noise-robustness: use two PLM fine-tuning methods as two views of data with model ensemble

​

​

16

PIEClass: Experiment Results

• PIEClass is on par with the fully supervised text classifier on sentiment analysis datasets (i.e., Yelp and IMDB).

​

​

​

​

​

​

​

​

• Why PIEClass can achieve similar performance to the fully supervised method?
• Annotation errors can affect the fully supervised model if used as training data

17

TELEClass: Taxonomy Enrichment and LLM-Enhanced�Hierarchical Text Classification with Minimal Supervision

• Task: Classifying documents into 10² to 10³ classes, without human annotation?
• Automatically enrich the label taxonomy with class-indicative topical terms mined from the corpus to facilitate classifier training
• Use LLMs for both data annotation and creation tailored for the hierarchical label space

18

TELEClass: Performance Study and Cost for Text Classification

Yunyi Zhang, et al., “TELEClass: Taxonomy Enrichment and LLM-Enhanced Hierarchical Text Classification with Minimal Supervision”, WWW’25

19

Outline

• Text Classification
• Entity Typing and Entity Structure Mining
• Relation Extraction and Knowledge Graph Construction

20

Entity Structure Mining

• Zero-shot entity typing: Assigns fine-grained semantic types to entities without any annotations
• Ex. Sammy Sosa [Person/Player] got a standing ovation at Wrigley Field [Location/Building/Stadium]
• Challenges of weak supervision based on masked language model (MLM) prompting
• A prompt generates a set of tokens, some likely vague or inaccurate, leading to erroneous typing
• Not incorporate the rich structural information in a given, fine-grained type ontology

21

21

OntoType : Three Steps

• Candidate type generation
• Candidate type generation with multiple MLM prompting
• Ensembled candidate type prediction
• Ex. Stadium, venue, location, games, things, teams
• High-level type alignment by entailment (local context + NLI)
• Progressively refine type resolution, from coarse to fine, following the type ontology
• Type ontology used at every step

22

OntoType: Step 1 - Candidate Type Generation

• Head Word Parsing
• Mention’s head word in the input text is often the cue that explicitly matches a mention to its type
• Ex. “Governor Arnold Schwarzenegger gives a speech …”
• Use the Stanford Dependency Parser to extract head word
• Leverage the head words of the input entity to select an initial context-sensitive coarse-grained type
• Ensembled MLM Prompting
• Leverage a BERT MLM and Hearst patterns to generate candidate types for the target mentions
• Ensemble 𝑛 patterns to generate the best candidate types
• Consolidated candidates are generated by a majority of Hearst patterns
• Ex. For 𝑒₁, "Stadiums, Venues, Locations, Games" retain, but the noisy types "Things" and "Teams" are removed

Four Hearst Patterns give the highest quality hypernyms with simple type mapping on the OntoNotes dataset

23

OntoType: Steps 2 & 3- High-Level Type Resolution & Progressive Type Refinement

• High-level type alignment by entailment
• Align generated candidate types to several high-level types in the type ontology by Word2Vec+ cosine similarity
• Then select the most accurate high-level types with a pre-trained entailment language model (NLI)

• Progressively refine type resolution, from coarse to fine, following the type ontology
• Ex. At the 2^nd level of ontology, it generates the hypotheses and ranks all child types of "location“
• This consolidates and selects "building" as the highest ranked label
• At a deeper level, it selects the final type "stadium”
• Type ontology is used at every step

24

OnEFET: Ontology Enrichment for FET

Usually organized as a structure — ontology

• Directly prompting LLMs (GPT) cannot handle Zero-shot FET
• Nuanced semantic relations as ontology goes deeper and types become more fine-grained
• Contextualized information

Siru Ouyang, et al., “Ontology Enrichment for Effective Fine-grained Entity Typing”, KDD’24

• Task: Zero-shot fine-grained entity typing (Zero-shot FET)
• Input: a sentence with a given entity mention
• Output: the entity type label from a predefined set of types

25

Overall Framework of OnEFET : Three Steps

26

OnEFET: Step 1 – Ontology Enrichment

[1] Zhang, Yu, et al. "Seed-guided topic discovery with out-of-vocabulary seeds." NAACL 2022.

[2] Jiao, Yizhu, et al. "Open-vocabulary argument role prediction for event extraction." EMNLP 2022 Findings.

[3] Zhang, Yu, et al. "Entity Set Co-Expansion in StackOverflow." Big Data 2022.

• Enrichment for topic information
• Providing entity type T, first select 20 related documents in Wikipedia using Elastic search.
• Filtering out noisy documents and reduce memory usage
• SeeTopic [1] for out-of-vocabulary topic words/phrases mining

​

• Enrichment for instance information
• LM-based instance seeds curation in question-answering style [2]
• Providing entity type T, first retrieve related sentences in Wikipedia.
• QA template: [CLS] What is the instance of <T> in this sentence? [SEP] <Sentence> [SEP]
• SECoExpan [3] for instance expansion

27

OnEFET: Steps 2 & 3 – Coarse-to-fine Typing

• Instance information → contextualized training samples for each fine-grained types
• Leverage language model to generate a sentence that contain instance e of a specific type
• Rewarding and paneling mechanism in LM decoding for diversity

28

OnEFET: Performance Study

• Use 3 benchmark FET datasets: BBN, Ontonotes, and FIGER:

• Ablation Study

• Transferability test on UFET

• OnEFET significantly outperform previous baseline models; on par with supervised settings.
• OnEFET could be smoothly transferred to unseen settings, even with more entity types

29

Open-Vocabulary Relation Type Discovery

• Related Work:
• Most of existing studies rely on hand-crafted ontologies (costly, cannot generalize)
• A few studies try to automatically induce argument roles (limited pre-defined glossary)
• New Task: Infer a set of relation type names for a given entity type to describe the crucial relations between the entity type and its related entities

Yizhu Jiao, Sha Li, Yiqing Xie, Ming Zhong, Heng Ji and Jiawei Han “Open-Vocabulary Argument Role Prediction for Event Extraction”, EMNLP’22

30

Framework for RolePred

Entity Type

Candidate Relation Type

Relation Types

Candidate Relation Types

31

Experiment: Relation Type Discovery

Relation Type Prediction

Relation Extraction w/o Golden Roles

Example of the generated relation types

Extracted Results by RolePred and baselines

32

Chemical Reaction Extraction with Weak Supervision

• Task: Chemical Reaction Extraction
• Goal: Extract chemical reactions from a scientific paper
• Input: A scientific paper
• Output: Multiple structured chemical reactions

​

Ming Zhong, Siru Ouyang, Minhao Jiang, Vivian Hu, Yizhu Jiao, Xuan Wang, Jiawei Han, “ReactIE: Enhancing Chemical Reaction Extraction with Weak Supervision”, ACL’23 Findings

33

Chemical Reaction Extraction with Weak Supervision

• Method: ReactIE
• Linguistic cues
• Domain Knowledge

​

34

Chemical Reaction Extraction with Weak Supervision

• Result for Product Extraction

​

• Result for Role Extraction

​

• Case Study

35

ZOES: Zero-shot Open-Schema Entity Structure Discovery

• Entity structure extraction: Extract entities and their associated attribute-value structures from text
• Open-schema entity structure discovery: automatically identify entities and their corresponding structures, from an input document and a given set of entity types of interest, without relying on any pre-defined schemas
• Zero-Shot Open-schema Entity Structure Discovery (ZOES): Extract entity structure without using any schema or annotated samples
• Enrichment, refinement & unification, based that an entity and its associated structure are mutually reinforcing

Xueqiang Xu, Jinfeng Xiao, James Barry, Mohab Elkaref, Jiaru Zou, Pengcheng Jiang, Yunyi Zhang, Max Giammona, Geeth de Mel, and Jiawei Han. "Zero-shot open-schema entity structure discovery", arxiv-2506.04458, 2025.

Mutual Dependency Principle: For a triplet t = ⟨e, a, v⟩, appropriate granularity is achieved when any one component can be reliably inferred fr. the other two within the context d

36

ZOES: Methodology Overview

• Triplet Candidates Extraction: Expand the initial zero-shot EAV triplet set by leveraging generalized root attributes induced from initial extractions as guidance to uncover additional triplets
• Triplet Granularity Refinement: Apply the triplet mutual dependency principle to detect and revise under-specified or inconsistent triplets
• Entity Structure Construction: Assemble refined triplets into entity structures, which are filtered based on user-specified target entity types

Induce root attribute (e.g., discharge capacity): abstract over semantically similar attrs to further guide the triplet enrichment (i.e., embed all extracted attrs using a dense encoder and cluster them based on semantic similarity―This clustering step can group attributes that express the same underlying general attribute)

Value-Anchored Enrichment: Once root attributes are identified, use them to guide the discovery of additional value mentions. For each root attribute, prompt the LLM to revisit the document and list all corresponding values.

Some entities may lack explicitly stated attr.-value structures. But each semantically meaningful value (e.g., “80.6%”) should correspond to at least one valid triplet, and is treated as an anchor to elicit a missing triplet

37

ZOES: Performance Study

Example extracted results on a doc from the Economics domain for “Toyota,” using Granite-8B

Ablation results on the Finance domain using GPT-4o as the backbone

Dataset statistics across “Battery Science”, “Finance”, and “Politics” domains

Evaluation with user interested entity types across different backbone models & methods

38

Outline

• Text Classification
• Entity Typing and Entity Structure Mining
• Relation Extraction and Knowledge Graph Construction

39

TAGREAL: Pattern Mining for Prompt Generation

40

Textual Pattern Mining for Prompt Generation (I)

• Sub-corpora mining
• Given a KG with a relation set R = (r₁, r₂, ..., r_k), first extract tuples paired by head entities and tail entities for each relation r_i ∈ R from the KG
• For each tuple t_j, search sentences s_tj containing both head and tail from a large corpus (e.g., Wiki) and other reliable sources, add to compose the sub-corpus C_ri
• Phrase segmentation and frequent pattern mining: AutoPhrase + FPGrowth
• Prompt selection and optimization: Using MetaPAD (Jiang et al., 2017) and TruePIE (Li et al., 2018) for prompt selction
• Prompt optimization based on the formula:

where w_i,j is the weight of j^th prompt for i^th relation

All weights are learned from PLM

41

Textual Pattern Mining for Prompt Generation (II)

• Support information retrieval: Use BM25 to retrieve highly ranked support texts (with score greater than δ and length shorter than ϕ from the reliable corpus) and select one of them as the support information
• Training the model: Create negative triples by replacing the head and tail in each positive triple with the "incorrect" entity that achieves high probability by KGE (knowledge graph embedding) model, plus random negative examples; then, transform all training triples of each relation r into sentences with the prompt ensemble
• Inference: To predict (h, r, ?), replace [MASK] with each entity in the known entity set and rank nd ensemble their classification scores

​

​

42

Example of Link Prediction with TAGREAL

Man denotes manual prompt. Optim denotes optimized prompt ensemble. Supp denotes support information. The ground truth tail entity is in red/yellow, helpful information in green and optimized prompts in blue

43

Performance of TagReal: Knowledge Graph Completion

• Performance Study: Results on FB60K-NYT10

​

Performance Variation of F1-score

Relation-wise Hits@10 on FB60K-NYT10

• TAGREAL has superb performance even with limited training data, outperforming existing embedding-based, graph-based, and PLM-based methods.

44

Evaluation Bottleneck of Generative KG Construction in the LLM Era

*GRE: Generative Relation Extraction (by LLMs)

Ideal metrics for GRE should be able to evaluate :

​

• How much content of the source text are covered by the relationships extracted
• How many unique relationships are extracted
• How factual the extracted triples are, referring to the source text
• How atomic the extracted triples are
• How many ground truth relations are predicted

45

GenRES: A Comprehensive Evaluation Framework for Generative RE

46

GenRES: A Comprehensive Evaluation Framework for Generative RE

Four workers died in a massive oil rig fire that raged for hours off the coast of Mexico Wednesday. Mexican state oil company Pemex said 45 workers were injured in the blaze, which began early Wednesday morning. Two of them are in serious condition, the company said. Authorities evacuated about 300 people from the Abkatun Permanente platform after the fire started, Pemex said. At least 10 boats worked to battle the blaze for hours. The fire had been extinguished by Wednesday night, Pemex said in a Twitter post. The company denied rumors that the platform had collapsed and said there was no oil spill as a result of the fire. The state oil company hasn't said what caused the fire on the platform, which is located in the Gulf of Mexico's Campeche Sound. The fire began in the platform's dehydration and pumping area, Pemex said. CNN's Mayra Cuevas contributed to this report.

[Four workers | were died in | oil rig fire],

[45 workers | were injured in | the blaze],

[Two workers | are in | serious condition],

[300 people | were evacuated from | the platform],

[The fire | had been extinguished by | Wednesday night],

[The fire | did not result in | oil spill].

Generative Relation Extraction

Triples

Text

Topical Similarity Score (TS)

“How much content of the source text are covered by the relationships extracted (by comparing triples* to the source text)”

47

GenRES: A Comprehensive Evaluation Framework for Generative RE

Evaluate the factualness of an extracted relationship (triplet) based on the given source text. Indicate whether the relationship accurately reflects the information in the source text by responding with "true" or "false".

You should only output "true" or "false" with no additional information.

​

Example 1:

Source Text: The Great Barrier Reef, located off the coast of Australia, is the world's largest coral reef system. It has been severely affected by climate change, leading to coral bleaching.

Relationship: ["Great Barrier Reef", "affected by", "climate change"]

Factualness: true

​

Example 2:

Source Text: The Eiffel Tower was constructed in 1889 and is located in Paris, France. It is one of the most recognizable structures in the world.

Relationship: ["Eiffel Tower", "located in", "London"]

Factualness: false

​

Example 3:

Source Text: The novel "Moby-Dick" by Herman Melville features a ship named Pequod. The narrative follows the ship and its crew in their pursuit of a giant white sperm whale.

Relationship: ["Moby-Dick", "is about", "a whale named Pequod"]

Factualness: false

​

​

Source Text: $TEXT$

Relationship: $TRIPLE$

Factualness:

“How factual the extracted triples are, referring to the source text (by factualness verification treating source text as the “knowledge base”)”

48

GenRES: A Comprehensive Evaluation Framework for Generative RE

Granularity-checking prompt:

“How atomic the extracted triples are (by asking LLM to split each triple)”

49

GenRES: Robustness and Human Preference Alignment Results

• Leading LLM’s consistent performance across different runs

​

• The robustness of GenRES as an evaluation framework across different metrics

• Consistent alignment with human evaluation across different metrics

50

PriORE: Open Relation Extraction with a Priori Seed Generation

• MOTIVATION
• Reduce randomness
• enhance the robustness of long-tail scenarios

​

• STEP1: Seed Relations Generation (A Priori)
• Context-agnostic
• Type-centric

​

• STEP2: Relation classification

51

Methods Outlined with An Example

52

PrioRE: Experiment Results

• PrioRE is more effective on narrow topics compared to others
• PrioRE has a large improvement on the informativeness

• Datasets
• General：DocRED
• Domain：FewRel Domain Adaption（Biomedical）
• Theme：Electric vehicle (EV) batteries & Redox reactions

53

Case study: Inheritance_type_of

• Lack of uniformity
• Overly general

• Informative
• Uniform

​

54

Embedding-Based KG Construction – KG-FIT

Up-to-Date Global Knowledge

Fast-Iterating LLMs

Global Knowledge

Local Knowledge

Small-scale PLMs

Structure-based Methods

Pros over PLMs/LLMs:

• Fast Training/Inference
• Low Resource
• Interpretable Embeddings
• Robustness to Sparse Data

Pros over Structure-based Methods:

• Abundant External Knowledge
• Handling Linguistic Ambiguity

Pros over PLMs:

• Up-to-Date Knowledge
• More comprehensive understanding of entities

55

Embedding-Based KG Construction – KG-FIT

56

KG-FIT: Experimental Results

Metrics:

Mean Rank (MR):

• Measures the average rank of true entities.

Mean Reciprocal Rank (MRR):

• Averages the reciprocal ranks of true entities.

Hits@N:

• Measures the proportion of true entities in the top N predictions.

Findings:

• KG-FIT consistently and significantly outperforms state-of-the-art PLM-based and structure-based methods across all datasets and metrics.
• With LLM-guided hierarchy refinement, KG-FIT achieves huge performance gains compared to the base models and KG-FIT with seed hierarchy.
• KG-FIT is more effective for smaller KGs, e.g., more performance gains on PrimeKG (~ 0.1 million triples) than YAGO3-10 (~1 million triples).

Datasets:

57

KG-FIT: Embedding Visualization

KG-FIT preserves both local and global semantics!

“CAA”

Hits@1022

“Exertional dyspnea”

Hits@981

“CAA”

Hits@45

“Exertional dyspnea”

Hits@26

“CAA”

Hits@12

“Exertional dyspnea”

Hits@5

58

References

• Xiang Chen, Ningyu Zhang, Xin Xie, Shumin Deng, Yunzhi Yao, Chuanqi Tan, Fei Huang, Luo Si, Huajun Chen, “KnowPrompt: Knowledge-aware Prompt-tuning with Synergistic Optimization for Relation Extraction”, WWW’22
• Yew Ken Chia, Lidong Bing, Soujanya Poria, Luo Si, “RelationPrompt: Leveraging Prompts to Generate Synthetic Data for Zero-Shot Relation Triplet Extraction”, ACL’22 Findings
• Linyi Ding, Sizhe Zhou, Jinfeng Xiao, Jiawei Han, “Automated Construction of Theme-specific Knowledge Graphs”, arXiv’24
• Xiaotao Gu , Zihan Wang , Zhenyu Bi , Yu Meng, Liyuan Liu, Jiawei Han, Jingbo Shang. “UCPhrase: Unsupervised Context-aware Quality Phrase Tagging.” (KDD’21)
• Xiaotao Gu, Yikang Shen, Jiaming Shen, Jingbo Shang, Jiawei Han, “Phrase-aware Unsupervised Constituency Parsing” (ACL’22)
• Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, Maosong Sun, “PTR: Prompt Tuning with Rules for Text Classification”, AI Open
• Pere-Lluís Huguet Cabot, Roberto Navigli, “REBEL: Relation Extraction By End-to-end Language generation”, ACL’21
• Pengcheng Jiang , Shivam Agarwal , Bowen Jin , Xuan Wang, Jimeng Sun, Jiawei Han, “Text-Augmented Open Knowledge Graph Completion via Pre-Trained Language Models”, ACL’23 Findings
• Yizhu Jiao, Sha Li, Yiqing Xie, Ming Zhong, Heng Ji, Jiawei Han. "Open-vocabulary argument role prediction for event extraction." EMNLP’22 Findings.
• Yizhu Jiao, Ming Zhong, Sha Li, Ruining Zhao, Siru Ouyang, Heng Ji, and Jiawei Han. “Instruct and Extract: Instruction Tuning for On-Demand Information Extraction.” EMNLP 2023.
• Yizhu Jiao, Sha Li, Sizhe Zhou, Heng Ji, and Jiawei Han. “ Text2DB: Integration-Aware Information Extraction with Large Language Model Agents .” ACL 2024 findings.
• Seoyeon Kim, Kwangwook Seo, Hyungjoo Chae, Jinyoung Yeo, Dongha Lee, “VerifiNER: Verification-augmented NER via Knowledge-grounded Reasoning with Large Language Models”, ACL’24

59

59

References

• Yubo Ma, Yixin Cao, YongChing Hong, Aixin Sun, “Large Language Model Is Not a Good Few-shot Information Extractor, but a Good Reranker for Hard Samples!”, EMNLP’23 Findings
• Siru Ouyang, Jiaxin Huang, Pranav Pillai, Yunyi Zhang, Yu Zhang, Jiawei Han, “Ontology Enrichment for Effective Fine-grained Entity Typing”, KDD’24
• Tanay Komarlu, Minhao Jiang, Xuan Wang, Jiawei Han, “OntoType: Ontology-Guided and Pre-Trained Language Model Assisted Fine-Grained Entity Typing”, KDD’24
• Yiqing Xie, Jiaming Shen, Sha Li, Yuning Mao, Jiawei Han, “EIDER: Evidence-enhanced Document-level Relation Extraction”, ACL’22 Findings
• Changlong Yu, Weiqi Wang, Xin Liu, Jiaxin Bai, Yangqiu Song, Zheng Li, Yifan Gao, Tianyu Cao, and Bing Yin. “FolkScope: Intention Knowledge Graph Construction for E-commerce Commonsense Discovery”, ACL’23 Findings
• Kai Zhang, Bernal Jiménez Gutiérrez, Yu Su, “Aligning Instruction Tasks Unlocks Large Language Models as Zero-Shot Relation Extractors”, ACL'23 Findings
• Yu Zhang, Yu Meng, Xuan Wang, Sheng Wang, Jiawei Han. "Seed-guided topic discovery with out-of-vocabulary seeds." NAACL’22.
• Yu Zhang, Yunyi Zhang, Yucheng Jiang, Martin Michalski, Yu Deng, Lucian Popa, ChengXiang Zhai, Jiawei Han. "Entity Set Co-Expansion in StackOverflow." IEEE Big Data 2022
• Ming Zhong, Siru Ouyang, Minhao Jiang, Vivian Hu, Yizhu Jiao, Xuan Wang, Jiawei Han, “ReactIE: Enhancing Chemical Reaction Extraction with Weak Supervision”, ACL’23 Findings
• Sizhe Zhou, Suyu Ge, Jiawei Han, “Corpus-Based Relation Extraction by Identifying and Refining Relation Patterns”, ECMLPKDD’23
• Sizhe Zhou, Yu Meng, Bowen Jin, Jiawei Han, “Grasping the Essentials: Tailoring Large Language Models for Zero-Shot Relation Extraction”, arXiv’24
• Wenxuan Zhou, Muhao Chen, “An Improved Baseline for Sentence-level Relation Extraction”, AACL’22

​

​

60

60