Seq2Parse: Neurosymbolic Parse Error Repair
summary
Combines a neural sequence classifier that predicts a small subset of grammar-derived error rules with a symbolic error-correcting L-parser to synthesize repairs; trains a 5M-parameter Transformer on 1,100,000 Python examples labeled by diffed fixes and abstracted parser states; reduces rule search from hundreds to a top-k set so the parser fixes programs quickly; parses 94% of test cases with median 2.1s and reproduces the historical user fix about one in three times, with human judgments nearly as helpful as originals.
AI-written summary · about
chapters (23)
- 0:00Intro
- 0:17Parse Errors are Easy to Fix
- 1:37Parse Errors are Common & Time-Consuming
- 2:44Automatic Parse Error Repair
- 5:04Parsing Bad Programs
- 6:25Extract Error Rules from a Grammar
- 7:21Non-Terminal Error Rules
- 9:29Use Error Production Rules with ECE-Parser
- 11:08Find a subset of relevant error rules
- 11:41Sequence Models
- 12:16Program Token Sequences
- 12:58Abstracted Token Sequences
- 13:28Abstract Programs into Token Sequences
- 14:12Sequence Encoder
- 15:00Training Sequence Classifiers
- 16:37Predict Error Rules & Repair Programs
- 18:05Evaluation
- 18:40How accurate are Seq2Parse's models?
- 20:40How precise are Seq2Parse's repairs?
- 22:24How efficient is Seq2Parse?
- 23:02How useful is Seq2Parse's feedback?
- 26:55Further Discussion in Paper
- 27:33I. The Goal Repair III-Parsed Code Efficiently & Automatically
related talks
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
01h 00mHuman and machine inductive biases for compositional linguistic generalization
26mTowards controllable language generation
26mMachine Translation into Learning
18mActive Learning for Neurosymbolic Program Synthesis
01h 31mNatural Language Understanding: Foundations and State-of-the-Art
23mEnhancing Neural Network Robustness via Synthesis of Repair Programs
24mParsing natural languages in Swift