Integrating Machine Learning and Symbolic Reasoning: Learning to Generate Symbolic Representations from Weak SupervisionPublic Deposited
Machine learning and symbolic reasoning have been two main approaches to build intelligent systems. Symbolic reasoning has been used in many applications by making use of expressive symbolic representations to encode prior knowledge, conduct complex reasoning and provide explanations. Recently, machine learning has enabled various successful applications by learning from a large amount of noisy data. In this thesis, I propose to integrate these two approaches to build more expressive, efficient and interpretable learning systems. The main idea is, instead of training a model to predict the output directly, training a model to generate symbolic representations and then predict the output based on the generated symbolic representations. Incorporating symbolic representations into machine learning helps the model conduct complex reasoning, leverage external knowledge sources and learn more efficiently with better inductive biases. The main challenge is that the symbolic representations are usually hard to collect because it requires expertise, so we propose to induce them from weak supervision, which is much easier to collect. We analyze the challenges when learning from weak supervision and propose several novel techniques in reinforcement learning and latent structured prediction to overcome the problems. The proposed approach is investigated in two settings. In the first setting, to estimate the similarity between two relational structures, we use the structural alignment between them as the symbolic representation, which is then fed into a classifier to estimate their similarity. Experiments have shown that, with the inductive bias from structural alignment, the learned model achieves results competitive to state-of-the-art on paraphrase identification and knowledge base completion benchmarks while being much simpler or using orders of magnitude less data. In the second setting, we use compositional programs as the symbolic representations, which can be executed against a knowledge base or database tables to answer open-domain questions. By generating programs, the model can leverage existing knowledge and operations to perform complex reasoning compositionally. To our knowledge, this is the first end-to-end model without feature engineering that significantly outperforms previous state-of-the-art results on two very competitive semantic parsing benchmarks. Besides, I will also show that the generated symbolic representations, e.g., the structural alignment and the programs, can be inspected and verified by the users, which makes the model more interpretable and easier to debug.
- Last modified
- Date created
- Resource type
- Rights statement