Figure - uploaded by Pranav Nair
Content may be subject to copyright.
Results for variants of our first approach on the validation set.
Contexts in source publication
Context 1
... conduct ablation study to try out various configurations, which may or may not produce better results. We present these results in Table 3 and Table 4. For the part where we use external commonsense knowledge in the decoder, we also try a version where this knowledge is provided in the encoder self attention section (NBEmbed+encNBEmbed), the results are presented in Table 3. ...Context 2
... present these results in Table 3 and Table 4. For the part where we use external commonsense knowledge in the decoder, we also try a version where this knowledge is provided in the encoder self attention section (NBEmbed+encNBEmbed), the results are presented in Table 3. We see that the model performance significantly decreases, probably because providing a knowledge distribution different from that of the encoder for self-attention adds noise which reduces the informativeness of the encoder self-attention. ...Context 3
... see that the performance is similar to the baseline. The model is not able to utilize this knowledge since it had not seen it during training, these results are also presented in Table 3 for the validation data. ...Context 4
... we add adjectives in the document to our concept set as well, we call these models NBEmbed+Adjective and ConTra+Adjective. The results on the validation set are shown in Table 3 and Table 4 respectively. As we can see from the results there is no increase in the performance, since adjectives in isolation, when disconnected from the noun they modify, cannot truly express the noun's property. ...Context 5
... conduct ablation study to try out various configurations, which may or may not produce better results. We present these results in Table 3 and Table 4. For the part where we use external commonsense knowledge in the decoder, we also try a version where this knowledge is provided in the encoder self attention section (NBEmbed+encNBEmbed), the results are presented in Table 3. ...Context 6
... present these results in Table 3 and Table 4. For the part where we use external commonsense knowledge in the decoder, we also try a version where this knowledge is provided in the encoder self attention section (NBEmbed+encNBEmbed), the results are presented in Table 3. We see that the model performance significantly decreases, probably because providing a knowledge distribution different from that of the encoder for self-attention adds noise which reduces the informativeness of the encoder self-attention. ...Context 7
... see that the performance is similar to the baseline. The model is not able to utilize this knowledge since it had not seen it during training, these results are also presented in Table 3 for the validation data. ...Context 8
... we add adjectives in the document to our concept set as well, we call these models NBEmbed+Adjective and ConTra+Adjective. The results on the validation set are shown in Table 3 and Table 4 respectively. As we can see from the results there is no increase in the performance, since adjectives in isolation, when disconnected from the noun they modify, cannot truly express the noun's property. ...Similar publications
This chapter summarizes and concludes the contribution of this thesis in Section 6.1 and Section 6.2, respectively. Section 6.3 provides an overview of future work projects.
