Cross-Lingual Propagation for Deep Sentiment Analysis

Authors: Xin Dong, Gerard de Melo

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now turn to our extensive empirical evaluation, which assesses the effectiveness of using cross-lingual projections of three different sources of sentiment word vectors. Table 4: Accuracy on 9 language datasets using 12 embedding alternatives
Researcher Affiliation Academia Xin Dong Rutgers University New Brunswick, NJ, USA xd48@rutgers.edu Gerard de Melo Rutgers University New Brunswick, NJ, USA gdm@demelo.org
Pseudocode No The paper describes the approach using textual descriptions and mathematical equations, but it does not provide any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper concludes with 'Please refer to http://gerard.demelo.org/sentiment/ to obtain a copy of our data.', which refers to data, not the source code for their methodology.
Open Datasets Yes For evaluation, we use real-world datasets for several different languages, taken from 5 different sources that cover a range of different domains. These are summarized in Table 1. The Stanford Sentiment Treebank (SST) dataset (Socher et al. 2013) ... The Sem Eval-2016 Task 5 (SE16-T5) dataset (Pontiki et al. 2016) ... The Amazon Fine Food Reviews AFF (Mc Auley and Leskovec 2013) dataset...
Dataset Splits Yes Given the lack of provided test splits for TA, AFF, and AC, we randomly partitioned each into training/validation/test splits with a 80%/10%/20% ratio. Additionally, 10% of the training sets from SE16-T5 were randomly extracted and reserved for validation, while SST provides its own validation set.
Hardware Specification No The paper does not specify any hardware details such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper mentions software components like GloVe, fastText, VADER, Adam optimization, and CNN-non-static architecture, but it does not provide specific version numbers for any of these.
Experiment Setup Yes For CNNs, we make use of the CNN-non-static architecture and hyperparameters proposed in previous work (Kim 2014). The learning rate used to train all languages for it is 0.0006. For our DC-CNN models, the configuration of the regular channel is the same as for CNNs and the remaining hyperparameter values were tuned on the validation sets. An overview of further network parameters resulting from this tuning is given in Table 2.