DaLC: Difficulty-aware Label Completion for Crowdsourcing Truth Inference (Under Review)

Junjie Jiang, Jiachuan Wang*, Yongqi Zhang, Jiake Ge, Shuangyin Li, Lei Chen

• Proposes DaLC, a Difficulty-aware Label Completion framework for crowdsourcing truth inference
• Introduces a Neighbor Consistency Score (NCS) to quantify object inference difficulty from neighborhood consistency, splitting objects into easy, ambiguous, and hard subsets
• Estimates worker reliability in a difficulty-conditioned manner and completes the sparse matrix via a heterogeneous worker-object graph supervised by difficulty-aware soft targets under a joint training objective
• Experiments on 6 real-world and 34 simulated datasets show DaLC generalizes across aggregators and achieves the best downstream truth-inference performance

Code

Keywords: Crowdsourcing, Truth Inference, Label Completion, Worker Reliability, Graph Neural Network

ArchiDocGen: Multi-Agent Framework for Expository Document Generation in the Architectural Industry

Junjie Jiang, Haodong Wu, Yongqi Zhang*, Songyue Guo, Bingcen Liu, Caleb Chen Cao, Ruizhe Shao, Chao Guan, Peng Xu, Lei Chen

• A multi-agent framework that automates method-statement (expository document) generation in the architectural industry
• Decomposes the task into outline generation, section-based content generation, and polishing, each handled by a specialized agent
• A section-based retriever and a section-based chain-of-thought (SeCoT) scheme inject domain knowledge and iterative reasoning
• Achieves a 4.38 ContentScore and is deployed as a web application with industry partners

Keywords: Multi-Agent, Retrieval-Augmented Generation, Document Generation, Large Language Model, Architectural Industry

Retrieval augmented generation-driven information retrieval and question answering in construction management

Chengke Wu, Wenjun Ding, Qisen Jin, Junjie Jiang, Rui Jiang, Qinge Xiao, Longhui Liao, Xiao Li

• A retrieval-augmented generation (RAG) framework for information retrieval and question answering over construction-management documents
• Targets the mismatch between how stakeholders phrase queries and how information is stored in project documents
• Retrieves relevant evidence from heterogeneous project documents to support accurate, prompt responses
• Improves communication efficiency and project continuity in communication-intensive construction management

Keywords: Retrieval-Augmented Generation, Information Retrieval, Question Answering, Construction Management, Large Language Model

A compatible detector based on improved YOLOv5 for hydropower device detection in AR inspection system

Junjie Jiang, Zhile Yang, Chengke Wu, Yuanjun Guo, Meng Yang, Wei Feng*

• A Compatible detector is proposed for device detection in AR inspection system
• Computation cost is improved by a lightweight backbone network
• A coordinate-based attention module is designed for effective feature fusion
• Optimizing hyperparameters of the new architecture to make the detector robust

Code

Keywords: Object Detection, Augmented Reality, Genetic Algorithm, Power Grid

A teacher–student deep learning strategy for extreme low resolution unsafe action recognition in construction projects

Meng Yang, Chengke Wu, Yuanjun Guo, Yong He, Rui Jiang, Junjie Jiang, Zhile Yang*

• A teacher-student deep learning framework is proposed for identifying unsafe behaviours
• The useful feature is captured from extremely low resolution video
• Knowledge learning model and similarity loss function are designed for better training of student network

Keywords: Action Recognition, Knowledge Distillation, Low Resolution, Construction Safety