STREAM: A Data-Centric Framework for Mining High-Value Task-Oriented Dialogues from Streaming Media
AuthorsLiang Xue, Haoyu Liu, Cheng Wang, Pengyu Chen, Haozhuo Zheng, Yang Liu
STREAM turns live streams and short videos into a large task-oriented dialogue dataset, aiming to fill the gap in realistic domain conversations for training and evaluating LLMs.
Key results
total dialogue sessions in the released dataset
total dialogue turns in the released dataset
StreamDial Hybrid intrinsic score
public test set performance with StreamDial Hybrid
X-RiSAWOZ Automotive with StreamDial Hybrid
What the paper found
STREAM, from Harbin Institute of Technology and Byering Technology, tackles the scarcity of high-value task-oriented dialogue data by mining public streaming media and converting noisy interaction signals into structured supervision through four phases: Streaming Signal Ingestion, Adaptive Persona Synthesis, Conversational Blueprinting, and RAG-enhanced Interactive Dialogue Generation. The released StreamDial dataset spans 87,498 dialogue sessions and 1,497,320 turns across Automotive, Restaurant, and Hotel, with each session stored as a quadruplet <Pu, Pa, B, H> that encodes user persona, agent persona, a conversational blueprint, and dialogue history. In intrinsic evaluation, StreamDial raises overall quality from 6.32 to 8.98 under Qwen3-Max, from 5.89 to 7.81 under GPT-5.2, and from 6.91 to 8.82 under Gemini3-Pro, with the largest gains in informativeness, diversity, and flexibility. Under a controlled 2,000-dialogue DST budget, the hybrid setting improves Qwen3-8B on the public test set to 96.72 JGA and 99.41 Slot-value F1, and on cross-lingual X-RiSAWOZ it reaches 87.63/96.86 in English, 81.06/94.45 in French, and 85.61/96.03 in Korean. The construction pipeline also reports ASR word error rates of 3.5% to 10.5% and uses a 4.1M-entry automotive lexicon plus 43K address entries to normalize domain-specific entities.
Original abstract
Large language models for vertical domains are bottlenecked by the scarcity of complex, domain-specific task-oriented dialogues. Existing data acquisition pipelines face a persistent trilemma: expert annotation is expensive, real-world service conversations are constrained by privacy and commercial restrictions, and static corpora quickly become temporally stale. We propose Stream, a data-centric framework that leverages publicly available streaming media (live streams and short videos) to synthesize high-value service dialogues at scale. Stream mines authentic interaction signals from noisy streams and synthesizes conversations by integrating role-grounded persona construction with Conversational Blueprint construction; it further adopts retrieval-augmented generation (RAG) to support knowledge-aware responses. Based on Stream, we release StreamDial, a large-scale multi-domain dataset covering Automotive, Restaurant, and Hotel. StreamDial contains 87,498 dialogue sessions and 1,497,320 turns in total, with an average of 17.11 turns per session and a comparable scale across domains. Each session is organized as a structured quadruplet $\langle P_u, P_a, B, H \rangle$ that pairs dialogue history with explicit user/agent personas and a Conversational Blueprint, capturing realistic service behaviors such as requirement mining, constraint conflicts, negotiation, and recovery. Evaluations with automatic judges and downstream tasks show that StreamDial improves intrinsic dialogue quality over strong baselines, and models trained with StreamDial improve Dialogue State Tracking across backbones; we further report a completed human-evaluation set and encouraging multilingual transfer on Qwen3-8B under a controlled training budget. The data is released in https://github.com/hitxueliang/DialogDataSetBySTREAM.
Read the original paperMore in Natural Language Processing
Browse all 26 papers →AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research
Kailin Jiang, Lei Liu, Jian Xi, Yangqi Chen, Hui Xu, Hongwei Zhao, Bin Li, Yu Lu, Haibo Shi
AdaTutoRank teaches rerankers to assemble complementary evidence sets rather than merely picking individually relevant documents, improving RAG and deep-research retrieval with fewer calls.
Computation Over Geometry: Meaning Identity Is Computed, Not Shipped in the Embeddings
Jiaqi Deng
The paper argues that sentence meaning equivalence is not reliably stored in separate embeddings but is computed when models process both sentences together.
SlopShape: Identifying AI-Generated Commercial Web Content
Jochen Madler
SlopShape detects and identifies AI-written commercial content by recognizing its underlying organizational style, even after the text has been reworded.