Scalable Neural Methods for Reasoning With a Symbolic Knowledge Base
Authors: William W. Cohen, Haitian Sun, R. Alex Hofer, Matthew Siegler
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 3 EXPERIMENTS |
| Researcher Affiliation | Industry | William W. Cohen & Haitian Sun & R. Alex Hofer & Matthew Siegler Google, Inc {wcohen,haitiansun,rofer,msiegler}@google.com |
| Pseudocode | No | The paper describes methods and equations in text, but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The relation-set following operation using reified KBs is implemented in an open-source package called NQL, for neural query language. To reproduce these experiments, first download and install the Google language package5. |
| Open Datasets | Yes | Meta QA (Zhang et al., 2018), Web Questions SP (Yih et al., 2016) contains 4737 natural language questions, all of which are answerable using Free Base (Bollacker et al., 2008), NELL-995 dataset (Xiong et al., 2017). |
| Dataset Splits | Yes | In the experiments we tune the hyperparameters T {1, . . . , 6} and N {1, 2, 3} on a dev set. and we trained on 360,000 sentences requiring between 1 and T hops and tested on an additional 12,000 sentences. |
| Hardware Specification | Yes | These experiments were conducted on a Titan Xp GPU with 12Gb of memory. and We split the KB across three 12-Gb GPUs, and used a fourth GPU for the rest of the model. and four P100 GPUs |
| Software Dependencies | No | Our implementations are based on Tensorflow (Abadi et al., 2016) and To reproduce these experiments, first download and install the Google language package5. No specific version numbers are provided for these software components. |
| Experiment Setup | Yes | In the experiments we tune the hyperparameters T {1, . . . , 6} and N {1, 2, 3} on a dev set. and Each function f t(q) is a different linear projection of a common encoding for q, specifically a mean-pooling of the tokens in q encoded with a pre-trained 128-dimensional word2vec model (Mikolov et al., 2013). |