Integrating Deep Learning with Logic Fusion for Information Extraction
Authors: Wenya Wang, Sinno Jialin Pan9225-9232
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate the effectiveness of our proposed method, we conduct experiments on 5 datasets from 2 tasks: OTE: We use Restaurant and Laptop reviews from Sem Eval 2014 and 2016 (Pontiki et al. 2014; 2016). End-to-End RE: 1) TREC: entity and relation dataset introduced in (Roth and Yih 2004). 2) ACE05: annotated data with 7 coarse-grained entity types and 6 coarse-grained relation types between entities. We follow the same setting as (Li and Ji 2014). For the OTE task, we follow the setting in (Wang et al. 2016) by first pre-training the word embedding using word2vec (Mikolov et al. 2013) on Yelp Challenge dataset2 and electronic dataset in Amazon reviews3 for restaurant domain and laptop domain, respectively. For RE task, the word embedding is pre-trained on wikipedia corpus using Glove (Pennington, Socher, and Manning 2014). For all experiments, the dimension for word embedding and POS embedding is set to 300 and 50, respectively. The hidden layers has dimension 200. We set label embedding with dimension 25. Following (Vaswani et al. 2017), we also use positional encoding that is added to the input vectors. The multi-head self-attentions adopts 10 heads that leads to 10dim attention weight vectors. For RE task, we use scheduled sampling, similar to (Miwa and Bansal 2016). To train the model, adadelta is adopted with initial rate as 1.0 and with dropout rate 0.1. For evaluation, we use micro-F1 scores on non-negative classes. An entity is counted as correct based on exact match. A relation is correct if both of its entities are correct and the relation type matches the ground-truth label. |
| Researcher Affiliation | Academia | Wenya Wang, Sinno Jialin Pan Nanyang Technological University, Singapore {wangwy, sinnopan}@ntu.edu.sg |
| Pseudocode | Yes | Algorithm 1 Deep Logic |
| Open Source Code | No | The paper does not contain any explicit statements about the release of source code or provide a link to a code repository. |
| Open Datasets | Yes | We use Restaurant and Laptop reviews from Sem Eval 2014 and 2016 (Pontiki et al. 2014; 2016). End-to-End RE: 1) TREC: entity and relation dataset introduced in (Roth and Yih 2004). 2) ACE05: annotated data with 7 coarse-grained entity types and 6 coarse-grained relation types between entities. We follow the same setting as (Li and Ji 2014). For the OTE task, we follow the setting in (Wang et al. 2016) by first pre-training the word embedding using word2vec (Mikolov et al. 2013) on Yelp Challenge dataset2 and electronic dataset in Amazon reviews3 for restaurant domain and laptop domain, respectively. For RE task, the word embedding is pre-trained on wikipedia corpus using Glove (Pennington, Socher, and Manning 2014). |
| Dataset Splits | No | The paper mentions using training and test data, but does not explicitly describe a validation split or provide specific percentages or counts for how the datasets were divided into training, validation, and test sets. It mentions 'For all experiments...' and then describes model parameters but not data splits. |
| Hardware Specification | No | The paper does not specify the hardware used for running experiments, such as CPU or GPU models, or cloud computing instances. |
| Software Dependencies | No | The paper mentions 'adadelta' as an optimizer and 'word2vec' and 'Glove' for word embeddings but does not provide specific version numbers for these or other software libraries (e.g., 'PyTorch 1.9', 'Python 3.8'). |
| Experiment Setup | Yes | For all experiments, the dimension for word embedding and POS embedding is set to 300 and 50, respectively. The hidden layers has dimension 200. We set label embedding with dimension 25. Following (Vaswani et al. 2017), we also use positional encoding that is added to the input vectors. The multi-head self-attentions adopts 10 heads that leads to 10dim attention weight vectors. For RE task, we use scheduled sampling, similar to (Miwa and Bansal 2016). To train the model, adadelta is adopted with initial rate as 1.0 and with dropout rate 0.1. |