TIGER: A Generating-Then-Ranking Framework for Practical Python Type Inference
Python’s dynamic typing system offers flexibility and expressiveness but can lead to type-related errors, prompting the need for automated type inference despite efforts like Python Enhancement Proposals (PEPs) to enhance type hinting. While existing learning-based approaches show promising inference accuracy, they struggle with practical challenges in comprehensively handling various types, including complex generics and (unseen) user/library-defined types. To address these challenges, we introduce TIGER, employing a two-stage generating-then-ranking (GTR) framework. By fine-tuning pre-trained code models, TIGER trains a generation model with a generative span masking objective and a similarity model with a contrastive training objective. This enables TIGER to execute the GTR inference, generating diverse candidates and then ranking them alongside user/library-defined types. Evaluation on the ManyTypes4Py dataset demonstrates TIGER’s effectiveness across different type categories, particularly excelling in (unseen) user-defined types (with improvements of 11.2% and 20.1% in Top-5 Exact Match). The evaluation results also confirm the robustness and efficiency of TIGER, highlighting the contributions of the employed two stages.
Wed 30 AprDisplayed time zone: Eastern Time (US & Canada) change
11:00 - 12:30 | |||
11:00 15mTalk | A Multiple Representation Transformer with Optimized Abstract Syntax Tree for Efficient Code Clone Detection Research Track TianChen Yu School of Software Engineering, South China University of Technology, Li Yuan School of Software Engineering, South China University of Technology, Guangzhou, China, Liannan Lin School of Software Engineering, South China University of Technology, Hongkui He School of Software Engineering, South China University of Technology | ||
11:15 15mTalk | Can an LLM find its way around a Spreadsheet? Research Track Cho-Ting Lee Virginia Tech, Andrew Neeser Virginia Tech, Shengzhe Xu Virginia Tech, Jay Katyan Virginia Tech, Patrick Cross Virginia Tech, Sharanya Pathakota Virginia Tech, Marigold Norman World Forest ID, John C. Simeone Simeone Consulting, LLC, Jaganmohan Chandrasekaran Virginia Tech, Naren Ramakrishnan Virginia Tech | ||
11:30 15mTalk | QEDCartographer: Automating Formal Verification Using Reward-Free Reinforcement Learning Research Track Alex Sanchez-Stern University of Massachusetts at Amherst, Abhishek Varghese University of Massachusetts, Zhanna Kaufman University of Massachusetts, Shizhuo Zhang University of Illinois Urbana-Champaign, Talia Lily Ringer University of Illinois Urbana-Champaign, Yuriy Brun University of Massachusetts Link to publication Pre-print | ||
11:45 15mTalk | TIGER: A Generating-Then-Ranking Framework for Practical Python Type Inference Research Track Chong Wang Nanyang Technological University, Jian Zhang Nanyang Technological University, Yiling Lou Fudan University, Mingwei Liu Fudan University, Weisong Sun Nanyang Technological University, Yang Liu Nanyang Technological University, Xin Peng Fudan University | ||
12:00 15mTalk | ROCODE: Integrating Backtracking Mechanism and Program Analysis in Large Language Models for Code Generation Research Track Xue Jiang , Yihong Dong Peking University, Yongding Tao University of Electronic Science and Technology of China, Huanyu Liu Xidian University, Zhi Jin Peking University, Ge Li Peking University | ||
12:15 15mTalk | Rango: Adaptive Retrieval-Augmented Proving for Automated Software Verification Research Track Kyle Thompson University of California, San Diego, Nuno Saavedra INESC-ID and IST, University of Lisbon, Pedro Carrott Imperial College London, Kevin Fisher University of California San Diego, Alex Sanchez-Stern University of Massachusetts, Yuriy Brun University of Massachusetts, João F. Ferreira INESC-ID and IST, University of Lisbon, Sorin Lerner University of California at San Diego, Emily First University of California, San Diego Link to publication Pre-print File Attached | ||