Weakly Supervised Neural Symbolic Learning for Cognitive Tasks
Authors: Jidong Tian, Yitian Li, Wenqing Chen, Liqiang Xiao, Hao He, Yaohui Jin5888-5896
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate WS-Ne Sy L, we have conducted experiments on three cognitive datasets, including temporal reasoning, handwritten formula recognition, and relational reasoning datasets. Experimental results show that WS-Ne Sy L not only outperforms the end-to-end neural model but also beats the state-of-the-art neural symbolic learning models. |
| Researcher Affiliation | Academia | Jidong Tian1, 2, Yitian Li1, 2, Wenqing Chen1, 2, Liqiang Xiao1, 2, Hao He *1, 2, Yaohui Jin1, 2 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 State Key Lab of Advanced Optical Communication System and Network, School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University |
| Pseudocode | Yes | Algorithm 1: n-Step Metropolis-Hastings Sampler. Algorithm 2: 1-Step Sampler. |
| Open Source Code | No | The paper does not include an explicit statement or link to the source code for the methodology. |
| Open Datasets | Yes | We experiment on three datasets: multi-hop Time Bank-Dense (MHTBD), multi-hop HWF (Sinha et al. 2019) (MHHWF), and CLUTRR (Sinha et al. 2019). |
| Dataset Splits | Yes | Generalization tasks are to benchmark the reasoning ability by forcing the model to be trained on low-hop instances (2-3/2-4 hops) and tested on high-hop inference (up to 10-hop reasoning). Results on MHHWF are shown in Table 2. WS-Ne Sy L beats all baselines with a maximum accuracy of 96.70%, which performs well on both in-domain and out-of-domain evaluations. |
| Hardware Specification | No | The paper does not specify any hardware details (e.g., CPU, GPU models, or cloud resources) used for running the experiments. |
| Software Dependencies | No | For a fair comparison, we select the Bi-LSTM encoder and two LSTM attentional decoders on MHTBD and CLUTRR (Sinha et al. 2019), and use the CNN encoder and Multilayer Perceptrons decoders on MHHWF (Sinha et al. 2019) for baselines. |
| Experiment Setup | Yes | As a result, the final objective function (L) is shown in Eq. 10, where λ (0 λ 1) is a hyper-parameter to adjust the proportion of the cross-entropy loss and PLR. Table 1: Performance on MHTBD. λ is the hyper-parameter to reflect the proportion of PLR. ... WS-Ne Sy L λ = 0 67.5 0.0 λ = 0.2 68.3 0.9 λ = 0.5 71.1 0.1 λ = 0.8 72.9 3.1 |