Siru Ouyang

I am now a senior undergraduate student at Shanghai Jiao Tong University (SJTU), majoring in Computer Science at IEEE honor class. I'm expected to get my B.S. degree in June, 2022.

I have been working as a research intern at BCMI Lab since Sept. 2020, advised by Prof. Hai Zhao. I am fortunate to collaborate with Zhuosheng Zhang. Also, I work as an intern in SALT Lab at Georgia Institute of Technology since July 2021, advised by Prof. Diyi Yang.

I am actively looking for PhD position starting from 2022 Fall.

Email  /  CV  /  Semantic Scholar  /  Twitter  /  Github

profile photo

My primary research interests lie in the intersection of natural language processing (NLP) and machine learning (ML), especially in its application of question answering and dialogue systems. My long-term research goal is to understand human intelligence and build AI agents that possess a human-level reasoning ability in an effective and interpretable way.

blind-date Dialogue Graph Modeling for Conversational Machine Reading

Siru Ouyang*, Zhuosheng Zhang*, and Hai Zhao

ACL 2021, Findings [paper] [code] [video]

We propose to model discourse structure and relations which is not considered by existing approaches in CMR. We employed GCNs to explicitly bridge the gap between user scenario with rule documents by injecting it as a special global node. A decoupling graph is designed to decouple the complex rule document so that it can capture the local representation and global interaction. The proposed method DGM achieved the new state-of-the-art results on ShARC.

clean-usnob Smoothing Dialogue States for Open Conversational Machine Reading

Zhuosheng Zhang*, Siru Ouyang*, Hai Zhao, Masao Utiyama, and Eiichiro Sumita

EMNLP 2021, Proceedings [paper] [code] [video]

In the open-retrieval setting for CMR, we bridged decision making and question generation for the challenging CMR task, which is the first practice to our best knowledge. Designed an end-to-end framework where the dialogue states for decision making are employed for question generation, in contrast to the independent models or pipeline systems in previous studies. Besides, a variety of strategies are empirically studied for smoothing the two dialogue states in only one decoder.

clean-usnob Fact-driven Logical Reasoning

Siru Ouyang*, Zhuosheng Zhang*, and Hai Zhao

in submission for ICLR 2022 [paper] [code] [video]

We proposed to extract a kind of broad “Fact Unit” according to backbone constituents of a sentence to effectively cover such indispensable logic reasoning basis, filling the gap of local, non-commonsense, non-entity, or even non-knowledge clues in existing methods. The proposed model Focal Reasoner builds super-graphs on top of fact units to capture both global connections between facts and the local concepts or actions inside the fact. Focal Reasoner achieved the new state-of-the-art results on ReClor and LogiQA with single model.

clean-usnob Compositional Data Augmentation for Abstractive Conversation Summarization

Siru Ouyang, Jiaao Chen, and Diyi Yang

in submission for ACL 2022 [paper] [code] [video]

We presented a simple yet effective compositional data augmentation method, Compo, for generating diverse and high-quality pairs of conversations and summaries through first extracting conversation snippets and summary sentences based on conversation stages and then randomly composing them. We further utilize knowledge distillation to learn concise representation from a teacher model to avoid potential noise. Compo significantly outperforms prior state-of-the-art baselines in terms of both quantitative and qualitative evaluation, and exhibits a reasonable level of interpretability.

clean-usnob Two-Hop Relay Deployment Based on User Trajectory in Wireless Networks

Zhiyao Li, Siru Ouyang, Xiaofeng Gao, and Guihai Chen

the Computer Journal 2021 [paper] [code] [video]

We proposed the concept "Demand Nodes" representing the locations where users frequently pass or stay, and convert the relay deployment problem into a Demand Node Coverage (DNC) problem, which proved to be NP-complete. We also designed an approximation algorithm to solve DNC problem, and simulated the method on five real-world trajectory datasets in CRAWDAD, which proved to have higher coverage and thus leading to better user experience.

Awards and Scholarships

Rongchang Scientific and Technological Innovation Scholarship, 2021 (rank 1st in SEIEE)

SenseTime Scholarship, 2021 (31 in total nationwide)

Google Women Tech Makers Scholarship, 2020 (110 out of 2800 in APAC)

School Scholarship B Prize, 2019-2021 (top 20%)


I love playing the piano. I have been practicing the piano ever since I was 5. I got my Associated Diploma (popularly known as "performing level") of Central Conservatory of Music in my senior high years.

I am a crazy fan of Frédéric Chopin, and I am also fond of music pieces arranged by Animenz. Some of my repertoires include Ballade No.1 Op.23, Polonaise in A-flat major, Op.53, My Dearest, Unravel, and Prelude g-moll, op.23, No.5

Updated at Dec. 2021
Thanks Jon Barron for this amazing template