Academia.eduAcademia.edu

Figure 4 – uploaded by KOTHAMASU SAI RAHUL

See full PDFdownloadDownload figure
Figure 4: Confusion matrix for test data: 1-D CNN without any augmentation technique using FastText embeddings.

Figure 4 Confusion matrix for test data: 1-D CNN without any augmentation technique using FastText embeddings.

Related Figures (7)

CNN with Word2Vec embeddings: Word embeddings from Word2Vec are feed into 1-D CNN; the results were evaluated using weighted Fl-score. Various hyperparameters and word embeddings dimensions were experimented with. 1-D CNN with up-sampled data using Word2Vec embeddings (CNN_W2V_UP) obtained a weighted Fl-score of 0.73 on developmen data set. 1-D CNN without any augmentation technique (CNN_W2V_NORM) obtained a weighted Fl-score of 0.62 on development data set. 1-D CNN with data augmentation using back translation (CNN_W2V_AUG) obtained a weighted Fl-score of 0.69 on development data set. See Table 5 and Figures: 2,3,4 Table 5: Fil-score for development and _ test data.CNN_W2V_UP: 1-D CNN with up-sampled data using Word2Vec embeddings, CNN_W2V_NORM: I- D CNN without any augmentation technique using Word2Vec embeddings, CNN_W2V_AUG: 1-D CNN with data augmentation using back translation with Word2Vec embeddings.
CNN with Word2Vec embeddings: Word embeddings from Word2Vec are feed into 1-D CNN; the results were evaluated using weighted Fl-score. Various hyperparameters and word embeddings dimensions were experimented with. 1-D CNN with up-sampled data using Word2Vec embeddings (CNN_W2V_UP) obtained a weighted Fl-score of 0.73 on developmen data set. 1-D CNN without any augmentation technique (CNN_W2V_NORM) obtained a weighted Fl-score of 0.62 on development data set. 1-D CNN with data augmentation using back translation (CNN_W2V_AUG) obtained a weighted Fl-score of 0.69 on development data set. See Table 5 and Figures: 2,3,4 Table 5: Fil-score for development and _ test data.CNN_W2V_UP: 1-D CNN with up-sampled data using Word2Vec embeddings, CNN_W2V_NORM: I- D CNN without any augmentation technique using Word2Vec embeddings, CNN_W2V_AUG: 1-D CNN with data augmentation using back translation with Word2Vec embeddings.
Figure 3: Confusion matrix for test data: 1-D CNN with up-sampled data using Word2Vec embeddings Figure 2: Confusion matrix for test data: 1-D CNN with data augmentation using back translation with Word2Vec embeddings.
Figure 3: Confusion matrix for test data: 1-D CNN with up-sampled data using Word2Vec embeddings Figure 2: Confusion matrix for test data: 1-D CNN with data augmentation using back translation with Word2Vec embeddings.
Figure 1: Confusion matrix for test data: 1-D CNN without any augmentation technique using Word2Vec embeddings.
Figure 1: Confusion matrix for test data: 1-D CNN without any augmentation technique using Word2Vec embeddings.
Figure 5: Confusion matrix for test data: 1-D CNN with data augmentation using back translation with FastText embeddings.
Figure 5: Confusion matrix for test data: 1-D CNN with data augmentation using back translation with FastText embeddings.
Figure 6: Confusion matrix for test data: 1-D CNN with up-sampled data using FastText embeddings. Table 6: Fil-score for development and _ test data.CNN_ft_UP: 1-D CNN with up-sampled data using FastText embeddings, CNN_ft_NORM: 1-D CNN without any augmentation technique using FastText embeddings, CNN_ft_AUG: 1-D CNN with data augmentation using back translation with FastText embeddings. References development data set. We achieved an Fl-score
Figure 6: Confusion matrix for test data: 1-D CNN with up-sampled data using FastText embeddings. Table 6: Fil-score for development and _ test data.CNN_ft_UP: 1-D CNN with up-sampled data using FastText embeddings, CNN_ft_NORM: 1-D CNN without any augmentation technique using FastText embeddings, CNN_ft_AUG: 1-D CNN with data augmentation using back translation with FastText embeddings. References development data set. We achieved an Fl-score
Figure 7: Confusion matrix for test data: Bi-LSTM with up-sampled data using FastText embeddings. 5 Conclusion We evaluated the performance of our models on the test data using sklearn Fl-score °, we found that 1-D CNN on up-sampled data with both Word2Vec and FastText embeddings performed better than the Bi-LSTM model which we submitted for the task. However, Bi-LSTM performed relatively better on the development data set (see Table 7). We also observed that random up-sampling yielded better results compared to data augmentation with back translation method.
Figure 7: Confusion matrix for test data: Bi-LSTM with up-sampled data using FastText embeddings. 5 Conclusion We evaluated the performance of our models on the test data using sklearn Fl-score °, we found that 1-D CNN on up-sampled data with both Word2Vec and FastText embeddings performed better than the Bi-LSTM model which we submitted for the task. However, Bi-LSTM performed relatively better on the development data set (see Table 7). We also observed that random up-sampling yielded better results compared to data augmentation with back translation method.
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved