Learning Sentiment-Specific Word Embedding via Global Sentiment Representation
Authors: Peng Fu, Zheng Lin, Fengcheng Yuan, Weiping Wang, Dan Meng
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on several benchmark datasets demonstrate that the proposed architecture outperforms the state-of-the-art methods for sentiment classification. |
| Researcher Affiliation | Academia | Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China1 School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China2 |
| Pseudocode | No | No structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures) were found. |
| Open Source Code | No | No explicit statement about providing source code (e.g., a specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper was found. |
| Open Datasets | Yes | The movie reviews are extracted from the SAR14(Nguyen et al. 2014) dataset that contains 233,600 IMDB reviews along with their associated ratings on a 1-10 scale. We extract tweets from a collection of dataset1 that labeled positive or negative, and filter the tweets that less than 7 words. For document-level sentiment classification, we use the IMDB movie review dataset (Maas et al. 2011). For sentence-level sentiment classification, we use the Twitter dataset from the Sem Eval 2013 task 2 (Nakov et al. 2013). |
| Dataset Splits | Yes | We use the default train/test split, and randomly select 10% of the training data as the development set. Dataset Subset #Positive #Negative #Total IMDB Train 11,250 11,250 22,500 Dev 1,250 1,250 2,500 Test 12,500 12,500 25,000 Twitter Train 2978 1162 4,140 Dev 328 170 498 Test 1306 485 1,791 |
| Hardware Specification | No | No specific hardware details (e.g., exact GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running experiments were mentioned. |
| Software Dependencies | No | The paper mentions "Our models are implemented in tensorflow2." and "We use the LIBLINEAR (Fan et al. 2008) as the classifier." but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We use Ada Grad (John Duchi 2011) to update the parameters and the learning rate is 1.0. We empirically set the context window size as 3 and batch size as 128. We set the hyper-parameter α = 1e 4 and p = 0.9 for corruption. We evaluate the effect of the embedding size and choose 150 for both models. For the IMDB dataset, SWPredict and SWRank obtain best results when setting β as 0.5 on both global representation methods. While, for the Twitter dataset, both models get best results when β is 0.6. We change λ from 1 to 10, and the performance results are shown in Figure 5. For the IMDB dataset, both global representation methods obtain the best results when λ is 3, and drop steadily after that. For the Twitter dataset, the best results are achieved at λ = 2 |