paper_id
string
claim_id
string
claim
string
label
string
caption
string
evi_type
string
evi_path
string
context
string
domain
string
use_context
string
operation
string
paper_path
string
detail_others
string
license_name
string
license_url
string
claim_id_pair
string
evi_path_original
string
2206.00823
val_fig_0063
Using the same experiments, we show that bio-plausible learning rules tend to approach high-curvature regions in synaptic weight space as measured by the loss’ Hessian eigenspectrum (Figure 2 D-F).
Supported
Figure 2: Bio-plausible temporal credit assignment rules show worse and more variable generalization gap, which can be informed by loss landscape curvature . A-C) Generalization gap distributions computed at the end of training across different random weight initializations for several well-known neuroscience and machi...
figure
figures/dev/val_fig_0063.png
ml
no
Supported_claim_only
papers/dev/ml_2206.00823.json
CC BY-NC-SA 4.0
http://creativecommons.org/licenses/by-nc-sa/4.0/
no pair
null
2205.15827
val_fig_0064
We see that LUI is the only method capable of converging to optimal policies both before and after the change in environment.
Supported
Figure 5 : Environment change on the Chain Problem at different points {\dagger}\in\{10^{2},\dots,10^{5}\} using randomization parameter \xi=0.8 in the exploration.
figure
figures/dev/val_fig_0064.png
Finally, we investigate the behaviour of the learning methods when after a fixed number of trajectories the probabilities of the true MDP change, as introduced in Section 5.4 . Figure 5 shows the performance of the robust policy for each learning method on the Chain environment, results for the same experiment on the B...
ml
no
Legend Swap
papers/dev/ml_2205.15827.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0056
null
2205.15827
val_fig_0065
We see that LUI is the only method capable of converging to optimal policies both before and after the change in environment.
Refuted
Figure 5 : Environment change on the Chain Problem at different points {\dagger}\in\{10^{2},\dots,10^{5}\} using randomization parameter \xi=0.8 in the exploration.
figure
figures/dev/val_fig_0065.png
Finally, we investigate the behaviour of the learning methods when after a fixed number of trajectories the probabilities of the true MDP change, as introduced in Section 5.4 . Figure 5 shows the performance of the robust policy for each learning method on the Chain environment, results for the same experiment on the B...
ml
no
Legend Swap
papers/dev/ml_2205.15827.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0056
null
2206.00050
val_fig_0066
For instance, on the Cifar-10 benchmark it achieves diversity scores >6.9\% and up to 9.2\% (depending on ensemble size), against 6.8\% for a naive ensemble, see Fig. 1 .
Supported
Figure 1: Diversity analysis: CIFAR-10/VGG-11 experiment. With increasing number of members, FiLM-Ensemble achieves more diverse representations. See Section 3.1 .
figure
figures/dev/val_fig_0066.png
Inspired by that line of work, we propose a new, efficient ensemble method, FiLM-Ensemble. Our method adapts feature-wise linear modulation as an alternative way to construct an ensemble for (epistemic) uncertainty estimation. FiLM-Ensemble greatly reduces the computational overhead compared to the naïve ensemble appro...
ml
yes
Legend Swap
papers/dev/ml_2206.00050.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0057
null
2206.00050
val_fig_0067
For instance, on the Cifar-10 benchmark it achieves diversity scores >6.9\% and up to 9.2\% (depending on ensemble size), against 6.8\% for a naive ensemble, see Fig. 1 .
Refuted
Figure 1: Diversity analysis: CIFAR-10/VGG-11 experiment. With increasing number of members, FiLM-Ensemble achieves more diverse representations. See Section 3.1 .
figure
figures/dev/val_fig_0067.png
Inspired by that line of work, we propose a new, efficient ensemble method, FiLM-Ensemble. Our method adapts feature-wise linear modulation as an alternative way to construct an ensemble for (epistemic) uncertainty estimation. FiLM-Ensemble greatly reduces the computational overhead compared to the naïve ensemble appro...
ml
yes
Legend Swap
papers/dev/ml_2206.00050.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0057
null
2206.00050
val_fig_0069
In both metrics, FiLM-Ensemble achieves higher scores than the naïve, explicit deep ensemble.
Supported
Figure 1: Diversity analysis: CIFAR-10/VGG-11 experiment. With increasing number of members, FiLM-Ensemble achieves more diverse representations. See Section 3.1 .
figure
figures/dev/val_fig_0069.png
For both metrics, higher numbers correspond to higher diversity. See Fig. 1 . We observe that the average diversity increases with the number of ensemble members.
ml
no
Supported_claim_only
papers/dev/ml_2206.00050.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
no pair
null
2206.00050
val_fig_0070
Also, note that with the increasing number of members, improvements in diversity metrics for the naïve explicit deep ensemble are negligible.
Supported
Figure 1: Diversity analysis: CIFAR-10/VGG-11 experiment. With increasing number of members, FiLM-Ensemble achieves more diverse representations. See Section 3.1 .
figure
figures/dev/val_fig_0070.png
For both metrics, higher numbers correspond to higher diversity. See Fig. 1 . We observe that the average diversity increases with the number of ensemble members. In both metrics, FiLM-Ensemble achieves higher scores than the naïve, explicit deep ensemble. Meaning that the predictions of FiLM-Ensemble are less correlat...
ml
no
Legend Swap
papers/dev/ml_2206.00050.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0058
null
2206.00050
val_fig_0071
Also, note that with the increasing number of members, improvements in diversity metrics for the naïve explicit deep ensemble are negligible.
Refuted
Figure 1: Diversity analysis: CIFAR-10/VGG-11 experiment. With increasing number of members, FiLM-Ensemble achieves more diverse representations. See Section 3.1 .
figure
figures/dev/val_fig_0071.png
For both metrics, higher numbers correspond to higher diversity. See Fig. 1 . We observe that the average diversity increases with the number of ensemble members. In both metrics, FiLM-Ensemble achieves higher scores than the naïve, explicit deep ensemble. Meaning that the predictions of FiLM-Ensemble are less correlat...
ml
no
Legend Swap
papers/dev/ml_2206.00050.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0058
null
2206.00050
val_fig_0073
With regard to ECE (right subfigure), we see all methods improving with growing ensemble size M . FiLM-Ensemble achieves better calibration then the widely used deep ensemble and MC-dropout methods, with significant margins at large ensemble sizes.
Supported
Figure 2: Accuracy and ECE for CIFAR-10, with varying ensemble sizes, using VGG-11 as backbone.
figure
figures/dev/val_fig_0073.png
We perform several experiments on widely used benchmarks in computer vision, CIFAR-10 and CIFAR-100. As performance metrics, we plot the test set accuracy and the expected calibration error (ECE) against the ensemble size M in Fig. 2 , for all compared methods. In terms of accuracy (left subfigure), FiLM-Ensemble is ou...
ml
no
Legend Swap
papers/dev/ml_2206.00050.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0059
null
2206.00050
val_fig_0074
With regard to ECE (right subfigure), we see all methods improving with growing ensemble size M . FiLM-Ensemble achieves better calibration then the widely used deep ensemble and MC-dropout methods, with significant margins at large ensemble sizes.
Refuted
Figure 2: Accuracy and ECE for CIFAR-10, with varying ensemble sizes, using VGG-11 as backbone.
figure
figures/dev/val_fig_0074.png
We perform several experiments on widely used benchmarks in computer vision, CIFAR-10 and CIFAR-100. As performance metrics, we plot the test set accuracy and the expected calibration error (ECE) against the ensemble size M in Fig. 2 , for all compared methods. In terms of accuracy (left subfigure), FiLM-Ensemble is ou...
ml
no
Legend Swap
papers/dev/ml_2206.00050.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0059
null
2205.14612
val_fig_0075
This is illustrated in Figure 1 where we show that a deep ResNet can easily break the topology of the input space, which is impossible for a Neural ODE.
Supported
Figure 1: Trajectory of ResNets with 300 layers. Left: we learn x\to\frac{x^{2}}{2} , trajectories are smooth and do not intersect. Right: we learn x\to\frac{-x^{2}}{2} , trajectories are not smooth and intersect.
figure
figures/dev/val_fig_0075.png
Neural ODEs also provide a theoretical framework to study deep learning models from the continuous viewpoint, using the arsenal of ODE theory (Teh et al., 2019 ; Li et al., 2019 ; Teshima et al., 2020 ) . Importantly, they can also be seen as the continuous analog of ResNets. Indeed, consider for N an integer, the Eule...
ml
yes
Graph Swap
papers/dev/ml_2205.14612.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0060
null
2205.14612
val_fig_0076
This is illustrated in Figure 1 where we show that a deep ResNet can easily break the topology of the input space, which is impossible for a Neural ODE.
Refuted
Figure 1: Trajectory of ResNets with 300 layers. Left: we learn x\to\frac{x^{2}}{2} , trajectories are smooth and do not intersect. Right: we learn x\to\frac{-x^{2}}{2} , trajectories are not smooth and intersect.
figure
figures/dev/val_fig_0076.png
Neural ODEs also provide a theoretical framework to study deep learning models from the continuous viewpoint, using the arsenal of ODE theory (Teh et al., 2019 ; Li et al., 2019 ; Teshima et al., 2020 ) . Importantly, they can also be seen as the continuous analog of ResNets. Indeed, consider for N an integer, the Eule...
ml
yes
Graph Swap
papers/dev/ml_2205.14612.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0060
null
2205.14612
val_fig_0077
{N} makes the network behave badly at large depth, while it still works well with our scaling \frac{1}{N} , as shown in Figure 3 (a).
Supported
Figure 3: (a) Test accuracy on CIFAR-10 as a function of the number of blocks in each layer of the ResNet. Within each layer, weights are tied ( 3 runs ). (b) Failure of the adjoint method with a ResNet-101 on ImageNet (the approximated gradients are only used in the third layer of the network, that contains 23 blocks)...
figure
figures/dev/val_fig_0077.png
We first train a ResNet-101 (He et al., 2016a ) on CIFAR-10 and ImageNet using the same hyper-parameters. Experimental details are in appendix B and results are summarized in table 1 , showing that the explicit addition of the step size \frac{1} {N} does not affect accuracy. In strike contrast, the classical ResNet rul...
ml
no
Legend Swap
papers/dev/ml_2205.14612.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0061
null
2205.14612
val_fig_0078
{N} makes the network behave badly at large depth, while it still works well with our scaling \frac{1}{N} , as shown in Figure 3 (a).
Refuted
Figure 3: (a) Test accuracy on CIFAR-10 as a function of the number of blocks in each layer of the ResNet. Within each layer, weights are tied ( 3 runs ). (b) Failure of the adjoint method with a ResNet-101 on ImageNet (the approximated gradients are only used in the third layer of the network, that contains 23 blocks)...
figure
figures/dev/val_fig_0078.png
We first train a ResNet-101 (He et al., 2016a ) on CIFAR-10 and ImageNet using the same hyper-parameters. Experimental details are in appendix B and results are summarized in table 1 , showing that the explicit addition of the step size \frac{1} {N} does not affect accuracy. In strike contrast, the classical ResNet rul...
ml
no
Legend Swap
papers/dev/ml_2205.14612.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0061
null
2205.14612
val_fig_0079
First, this results shows that the architecture needs to be deep enough, because it scales in \frac{1}{N} : for instance, we fail to train a ResNet-101 (He et al., 2016a ) on the ImageNet dataset using the adjoint method on its third layer (depth 23 ), as shown in Figure 3 (b).
Supported
Figure 3: (a) Test accuracy on CIFAR-10 as a function of the number of blocks in each layer of the ResNet. Within each layer, weights are tied ( 3 runs ). (b) Failure of the adjoint method with a ResNet-101 on ImageNet (the approximated gradients are only used in the third layer of the network, that contains 23 blocks)...
figure
figures/dev/val_fig_0079.png
In Prop. 3 we showed under assumption 2 , that is if the residuals are bounded and Lipschitz continuous with constant independent of the depth N , then the error for computing the activations backward would scale in \frac{1}{N} as well as the error for the gradients (Prop. 4 ).
ml
no
Legend Swap
papers/dev/ml_2205.14612.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0062
null
2205.14612
val_fig_0080
First, this results shows that the architecture needs to be deep enough, because it scales in \frac{1}{N} : for instance, we fail to train a ResNet-101 (He et al., 2016a ) on the ImageNet dataset using the adjoint method on its third layer (depth 23 ), as shown in Figure 3 (b).
Refuted
Figure 3: (a) Test accuracy on CIFAR-10 as a function of the number of blocks in each layer of the ResNet. Within each layer, weights are tied ( 3 runs ). (b) Failure of the adjoint method with a ResNet-101 on ImageNet (the approximated gradients are only used in the third layer of the network, that contains 23 blocks)...
figure
figures/dev/val_fig_0080.png
In Prop. 3 we showed under assumption 2 , that is if the residuals are bounded and Lipschitz continuous with constant independent of the depth N , then the error for computing the activations backward would scale in \frac{1}{N} as well as the error for the gradients (Prop. 4 ).
ml
no
Legend Swap
papers/dev/ml_2205.14612.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0062
null
2205.14612
val_fig_0081
The true backpropagation gives the same curves for the ResNet and the HeunNet.
Supported
Figure 4: Comparison of the best test errors as a function of depth when using Euler or Heun’s discretization method with or without the adjoint method.
figure
figures/dev/val_fig_0081.png
We then apply a batch norm, a ReLU and iterate relation ( 1 ) where f is a pre-activation basic block (He et al., 2016b ) . We consider the zero residual initialisation: the last batch norm of each basic block is initialized to zero. We consider different values for the depth N and notice that in this setup, the deeper...
ml
no
Legend Swap
papers/dev/ml_2205.14612.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0063
null
2202.10720
val_fig_0093
Despite that they outperform GCN, SGC and GAT, they still perform worse than implicit models with infinitely deep layers (i.e., EIGNN and IGNN) since finite layers for aggregations cannot effectively capture underlying dependencies along extremely long chains (e.g., with length > 30).
Supported
(a) Results of binary classification; (b) Results of multiclass classification; Averaged accuracies with respect to the length of chains.
figure
figures/dev/val_fig_0093.png
We first consider binary classification ( c=2 ) with 20 chains of each class as in Gu et al. [ 10 ] . We conduct experiments on graphs with chains of different lengths (10 to 200 with the interval 10). The averaged accuracies with respect to the length of chains are illustrated in Figure 1(a) . In general, EIGNN and IG...
ml
yes
Legend Swap
papers/dev/ml_2202.10720.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0064
null
2202.10720
val_fig_0094
Despite that they outperform GCN, SGC and GAT, they still perform worse than implicit models with infinitely deep layers (i.e., EIGNN and IGNN) since finite layers for aggregations cannot effectively capture underlying dependencies along extremely long chains (e.g., with length > 30).
Refuted
(a) Results of binary classification; (b) Results of multiclass classification; Averaged accuracies with respect to the length of chains.
figure
figures/dev/val_fig_0094.png
We first consider binary classification ( c=2 ) with 20 chains of each class as in Gu et al. [ 10 ] . We conduct experiments on graphs with chains of different lengths (10 to 200 with the interval 10). The averaged accuracies with respect to the length of chains are illustrated in Figure 1(a) . In general, EIGNN and IG...
ml
yes
Legend Swap
papers/dev/ml_2202.10720.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0064
null
2202.10720
val_fig_0096
Figure 3 shows that across different noise levels, EIGNN consistently outperforms IGNN.
Supported
Figure 3 : Accuracy on datasets with different feature noise.
figure
figures/dev/val_fig_0096.png
On synthetic chain datasets, we add random perturbations to node features as in Wu et al. [ 31 ] . Specifically, we add uniform noise \epsilon\sim\mathcal{U}(-\alpha,\alpha) to each node’s features for constructing noisy datasets.
ml
no
Legend Swap
papers/dev/ml_2202.10720.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0065
null
2201.11817
val_fig_0097
We find that the BIC value for RR-RL 2 is substantially lower compared to that of the hybrid model ( 5562.63 against 6158.91 ) when aggregated across all participants; all other models provide a less adequate fit to human choices.
Supported
Figure 2: Model comparison results on the two-armed bandit data from Gershman ( 2018 ). (a) Bayesian information criterion (BIC) values for the aggregated data of all participants. Lower values correspond to a better fit to human behavior. (b) Posterior probabilities for each model and participant. Higher values corres...
figure
figures/dev/val_fig_0097.png
Having established that different styles of exploration emerge in RR-RL 2 depending on its description length, we next set out to test how well it explains human choices. In order to do so, we conducted a Bayesian model comparison (Bishop, 2006 ) . A detailed summary of our comparison procedure is provided in Appendix ...
ml
no
Legend Swap
papers/dev/ml_2201.11817.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0066
null
2201.11817
val_fig_0098
We find that the BIC value for RR-RL 2 is substantially lower compared to that of the hybrid model ( 5562.63 against 6158.91 ) when aggregated across all participants; all other models provide a less adequate fit to human choices.
Refuted
Figure 2: Model comparison results on the two-armed bandit data from Gershman ( 2018 ). (a) Bayesian information criterion (BIC) values for the aggregated data of all participants. Lower values correspond to a better fit to human behavior. (b) Posterior probabilities for each model and participant. Higher values corres...
figure
figures/dev/val_fig_0098.png
Having established that different styles of exploration emerge in RR-RL 2 depending on its description length, we next set out to test how well it explains human choices. In order to do so, we conducted a Bayesian model comparison (Bishop, 2006 ) . A detailed summary of our comparison procedure is provided in Appendix ...
ml
no
Legend Swap
papers/dev/ml_2201.11817.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0066
null
2201.11817
val_fig_0099
Matching the main result of the experimental study, we find that strategic directed exploration increases with description length, while strategic random exploration remains unaffected.
Supported
Figure 4: Illustration of strategic directed and random exploration in the horizon task. (a) Human data from Somerville et al. ( 2017 ). During adolescence, people start to engage more in strategic directed exploration, whereas strategic random exploration remains constant over time. (b) Data simulated from RR-RL 2 wit...
figure
figures/dev/val_fig_0099.png
Results: We trained RR-RL 2 with a targeted description length of \{100,200,\ldots,10000\} nats on the same distribution used in the original experimental study. Figure 4 (b) visualizes how strategic directed and random exploration change as the description length of RR-RL 2 increases.
ml
no
Graph Flip
papers/dev/ml_2201.11817.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0067
null
2202.02363
val_fig_0100
Ant-dir : We trained models for performing the task on an episode of 200 timesteps and found that MetODS can adapt in a few time-steps, similar to memory-based models such as RL^{2} , (Figure 3 -f) thanks to its continual adaptation mechanism.
Supported
Figure 3 : a-b) Schemas of the Harlow and Mujoco Ant-directional locomotion task. c-d) Evolution of accumulated reward over training. In the Harlow task, we conduct an ablation study by either reducing the number of recursive iterations (S=1) or removing the trainable plasticity weights \bm{\alpha} resulting in sub-opt...
figure
figures/dev/val_fig_0100.png
\diamond MuJoCo
ml
no
Legend Swap
papers/dev/ml_2202.02363.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0068
null
2501.09163
val_fig_0106
We can observe that classification errors rise with increasing noise levels and region sizes.
Supported
Figure 3 : TTA classification errors under different levels of shift severity levels and scopes.
figure
figures/dev/val_fig_0106.png
To investigate the trade-off between the shift scope (dense vs. sparse) and severity, we simulate different levels of corruption severity and corrupted region sizes and evaluate a classical TTA method TENT [ 15 ] on these configurations. Following [ 45 ] , we inject impulse noise to the CIFAR10 dataset, with noise leve...
ml
no
Graph Flip
papers/dev/ml_2501.09163.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0069
null
2211.16499
val_fig_0109
We plot the proportion of conserved correct predictions for the panoramic camera rotation in Fig. 7 . First, we observe that both networks struggle to a higher extent with images taken on opposite to the position, with the light source behind the camera.
Supported
Figure 7: Counterfactual study of all sizes of ConvNext and Swin networks for panoramic 360 ∘ camera rotation.
figure
figures/dev/val_fig_0109.png
ml
other sources
Graph Flip
papers/dev/ml_2211.16499.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0070
null
2410.17599
val_fig_0112
As shown in Figure 4(a) , we found that the performance was optimal when the \alpha value was 1.0, and increasing or decreasing \alpha resulted in decreased performance.
Supported
(a) Impact on instruction tuning; (b) Impact on unlearning; Impact of strength coefficient \alpha on performance
figure
figures/dev/val_fig_0112.png
For instruction tuning, we tested values within the range of [0.5, 2] and evaluated them on the first 50 data points of AlpacaEval.
ml
other sources
Supported_claim_only
papers/dev/ml_2410.17599.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
no pair
null
2410.17599
val_fig_0113
As shown in Figure 4(b) , increasing the value of \alpha can prevent the model from outputting more sensitive information, but it may also lead to the loss of necessary information.
Supported
(a) Impact on instruction tuning; (b) Impact on unlearning; Impact of strength coefficient \alpha on performance
figure
figures/dev/val_fig_0113.png
For unlearning, we found that adjusting the value of \alpha can serve as a balance between forgetting and retaining .
ml
other sources
Graph Flip
papers/dev/ml_2410.17599.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0071
null
2410.17599
val_fig_0114
As shown in Figure 4(b) , increasing the value of \alpha can prevent the model from outputting more sensitive information, but it may also lead to the loss of necessary information.
Refuted
(a) Impact on instruction tuning; (b) Impact on unlearning; Impact of strength coefficient \alpha on performance
figure
figures/dev/val_fig_0114.png
For unlearning, we found that adjusting the value of \alpha can serve as a balance between forgetting and retaining .
ml
other sources
Graph Flip
papers/dev/ml_2410.17599.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0071
null
2201.10328
val_fig_0116
With the increase of data size, the performance increases.
Supported
Figure 6: The Cumulative Reward of Anonymous in Validation Dataset. The performance of DAgger models is related to the number of iteration rounds.
figure
figures/dev/val_fig_0116.png
We show the performance of models with different settings in the validation set in Figure 6 . With the same data size, the model using DAgger performs better.
ml
other sources
Graph Flip
papers/dev/ml_2201.10328.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0072
null
2024.eacl-short.2
val_tab_0087
We also see that the representations we learn lead to competitive results with respect to BERT and RoBERTa despite Sentence-BERT not being suited to classification.
Supported
Table 4: Classification results in terms of F1 score on the eight data sets.
table
tables_png/dev/val_tab_0087.png
Similar to the clustering setting, we compare results from a Logistic Regression (LR) applied on the original sentence embeddings with and without the filtering operation we introduced. We also use fine-tuned BERT and RoBERTa (2 epochs) as baselines; we use the base versions due to computational restrictions. We report...
nlp
yes
Swap rows or columns
papers/dev/nlp_2024.eacl-short.2.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0073
tables/dev/val_tab_0087.tex
2024.eacl-short.2
val_tab_0088
We also see that the representations we learn lead to competitive results with respect to BERT and RoBERTa despite Sentence-BERT not being suited to classification.
Refuted
Table 4: Classification results in terms of F1 score on the eight data sets.
table
tables_png/dev/val_tab_0088.png
Similar to the clustering setting, we compare results from a Logistic Regression (LR) applied on the original sentence embeddings with and without the filtering operation we introduced. We also use fine-tuned BERT and RoBERTa (2 epochs) as baselines; we use the base versions due to computational restrictions. We report...
nlp
yes
Swap rows or columns
papers/dev/nlp_2024.eacl-short.2.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0073
tables/dev/val_tab_0088.tex
2024.eacl-short.4
val_tab_0089
Across 57 translation directions, chrF2 improves by 1.3 (M2M-100) and 1.1 (SMaLL-100) points with source-contrastive decoding.
Supported
Table 1: Automatic evaluation results. Averages over different sets of translation directions.
table
tables_png/dev/val_tab_0089.png
We report results using source-contrastive decoding ( C_{src} ), and combining source-contrastive and language-contrastive decoding ( C_{src+lang} ) in Table 1 . 10 10 10 See Appendix A for full results.
nlp
other sources
Change the cell values
papers/dev/nlp_2024.eacl-short.4.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0074
tables/dev/val_tab_0089.tex
2024.eacl-short.4
val_tab_0090
Across 57 translation directions, chrF2 improves by 1.3 (M2M-100) and 1.1 (SMaLL-100) points with source-contrastive decoding.
Refuted
Table 1: Automatic evaluation results. Averages over different sets of translation directions.
table
tables_png/dev/val_tab_0090.png
We report results using source-contrastive decoding ( C_{src} ), and combining source-contrastive and language-contrastive decoding ( C_{src+lang} ) in Table 1 . 10 10 10 See Appendix A for full results.
nlp
other sources
Change the cell values
papers/dev/nlp_2024.eacl-short.4.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0074
tables/dev/val_tab_0090.tex
2024.eacl-short.4
val_tab_0091
Language-contrastive decoding brings additional gains of 0.4 (M2M-100) and 0.2 (SMaLL-100) points.
Supported
Table 1: Automatic evaluation results. Averages over different sets of translation directions.
table
tables_png/dev/val_tab_0091.png
We report results using source-contrastive decoding ( C_{src} ), and combining source-contrastive and language-contrastive decoding ( C_{src+lang} ) in Table 1 . 10 10 10 See Appendix A for full results. Across 57 translation directions, chrF2 improves by 1.3 (M2M-100) and 1.1 (SMaLL-100) points with source-contrastive...
nlp
yes
Change the cell values
papers/dev/nlp_2024.eacl-short.4.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0075
tables/dev/val_tab_0091.tex
2024.eacl-short.4
val_tab_0092
Language-contrastive decoding brings additional gains of 0.4 (M2M-100) and 0.2 (SMaLL-100) points.
Refuted
Table 1: Automatic evaluation results. Averages over different sets of translation directions.
table
tables_png/dev/val_tab_0092.png
We report results using source-contrastive decoding ( C_{src} ), and combining source-contrastive and language-contrastive decoding ( C_{src+lang} ) in Table 1 . 10 10 10 See Appendix A for full results. Across 57 translation directions, chrF2 improves by 1.3 (M2M-100) and 1.1 (SMaLL-100) points with source-contrastive...
nlp
yes
Change the cell values
papers/dev/nlp_2024.eacl-short.4.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0075
tables/dev/val_tab_0092.tex
2023.starsem-1.10
val_tab_0093
The remaining prompts types ( Discourse, Mask ) are comparable to Corpus .
Supported
Table 3: Mean perplexity (PPL) calculated using the GPT-J-6B model. Only the strings enclosed in square brackets are considered during calculation in order to provide a fair comparison with similar token length. For Corpus, PPL is calculated using the provided gold completion.
table
tables_png/dev/val_tab_0093.png
To further analyze the outputs, we calculate the perplexity (PPL) of the generated predictions to determine their plausibility (Wilcox et al., 2020 ) . Here, we choose the model with the best \textit{WHR}_{5} score on the MKR-NQ benchmark, and calculate the mean perplexity over all queries for each prompt type (5 compl...
nlp
yes
Swap rows or columns
papers/dev/nlp_2023.starsem-1.10.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0076
tables/dev/val_tab_0093.tex
2023.ijcnlp-main.36
val_tab_0094
It also has an equivalent amount or more for almost any ethnicity, except for Asians.
Supported
Table 7: Percentage of questions with changed answers between the biomedical and generic model as compared to a question with no demographic information about the patient. M =male; F =female; W =White; B =Black; A-A =African-American; H =Hispanic; As =Asian; SOr =sexual orientation.
table
tables_png/dev/val_tab_0094.png
Similar to our analysis between QAGNN and BioLinkBert above, our analysis between the biomedical and generic models can be split into the amount of answers and accuracy that changes when the dimensions change. From Table 7 it is visible that the generic transformer has more than double the amount of answers change for ...
nlp
yes
Change the cell values
papers/dev/nlp_2023.ijcnlp-main.36.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0077
tables/dev/val_tab_0094.tex
2023.ijcnlp-main.34
val_tab_0095
The reason is that the model struggles in distinguishing between refutes and NEI claims in these datasets, as reflected by Table 5 .
Supported
Table 5: Class-wise F1 of 3-class fact verification for the zero-shot generalization setup (left) and the in-domain training setup (right). S: supports; R: refutes; N: NEI.
table
tables_png/dev/val_tab_0095.png
Many works Jiang et al. ( 2020 ); Saakyan et al. ( 2021 ) do not consider NEI claims due to their ambiguity. To explore whether our previous observations also hold for the task of binary fact verification , we evaluate the generalization results for all 11 datasets using only the supports and refutes claims for trainin...
nlp
yes
Swap rows or columns
papers/dev/nlp_2023.ijcnlp-main.34.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0078
tables/dev/val_tab_0095.tex
2025.naacl-short.10
val_tab_0096
These are all assessed using the accuracy metric.
Supported
Table 1: This table presents the tasks implemented in this paper. The first column specifies the different tasks. The second details the metrics used (ROUGE includes ROUGE1, ROUGE2 and ROUGEL, and Perplexity includes Bits per Byte, Byte Perplexity, and Word Perplexity). The third column outlines the benchmarks used for...
table
tables_png/dev/val_tab_0096.png
This study considers four different close-ended healthcare tasks, which include nine different datasets ( e.g. , MedQA).
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.10.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0079
tables/dev/val_tab_0096.tex
2025.naacl-short.10
val_tab_0097
These are all assessed using the accuracy metric.
Refuted
Table 1: This table presents the tasks implemented in this paper. The first column specifies the different tasks. The second details the metrics used (ROUGE includes ROUGE1, ROUGE2 and ROUGEL, and Perplexity includes Bits per Byte, Byte Perplexity, and Word Perplexity). The third column outlines the benchmarks used for...
table
tables_png/dev/val_tab_0097.png
This study considers four different close-ended healthcare tasks, which include nine different datasets ( e.g. , MedQA).
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.10.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0079
tables/dev/val_tab_0097.tex
2025.naacl-short.10
val_tab_0098
At the same time, six open-ended tasks are studied, based on nine distinct datasets ( e.g. , MedText).
Supported
Table 1: This table presents the tasks implemented in this paper. The first column specifies the different tasks. The second details the metrics used (ROUGE includes ROUGE1, ROUGE2 and ROUGEL, and Perplexity includes Bits per Byte, Byte Perplexity, and Word Perplexity). The third column outlines the benchmarks used for...
table
tables_png/dev/val_tab_0098.png
This study considers four different close-ended healthcare tasks, which include nine different datasets ( e.g. , MedQA). These are all assessed using the accuracy metric.
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.10.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0080
tables/dev/val_tab_0098.tex
2025.naacl-short.10
val_tab_0099
At the same time, six open-ended tasks are studied, based on nine distinct datasets ( e.g. , MedText).
Refuted
Table 1: This table presents the tasks implemented in this paper. The first column specifies the different tasks. The second details the metrics used (ROUGE includes ROUGE1, ROUGE2 and ROUGEL, and Perplexity includes Bits per Byte, Byte Perplexity, and Word Perplexity). The third column outlines the benchmarks used for...
table
tables_png/dev/val_tab_0099.png
This study considers four different close-ended healthcare tasks, which include nine different datasets ( e.g. , MedQA). These are all assessed using the accuracy metric.
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.10.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0080
tables/dev/val_tab_0099.tex
2025.naacl-short.10
val_tab_0100
In this case, eleven different metrics are extracted.
Supported
Table 1: This table presents the tasks implemented in this paper. The first column specifies the different tasks. The second details the metrics used (ROUGE includes ROUGE1, ROUGE2 and ROUGEL, and Perplexity includes Bits per Byte, Byte Perplexity, and Word Perplexity). The third column outlines the benchmarks used for...
table
tables_png/dev/val_tab_0100.png
This study considers four different close-ended healthcare tasks, which include nine different datasets ( e.g. , MedQA). These are all assessed using the accuracy metric. At the same time, six open-ended tasks are studied, based on nine distinct datasets ( e.g. , MedText).
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.10.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0081
tables/dev/val_tab_0100.tex
2025.naacl-short.11
val_tab_0101
Our findings indicate that adding reinforcement learning (STRUX+RL) leads to stronger performance compared to using the SFT method alone.
Supported
Table 1: Our STRUX system outperforms strong benchmarks in making stock investment decisions. We present macro-averaged precision, recall, F-scores, accuracy for the test set. LLMs evaluated are: Llama3-8b-Instruct and gpt-4o-mini-2024-07-18 .
table
tables_png/dev/val_tab_0101.png
System Comparisons. Table 1 shows the macro-averaged precision, recall, F-scores, and accuracy for the test set. STRUX outperforms strong baselines in accuracy and F-scores for stock investment decisions.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-short.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0082
tables/dev/val_tab_0101.tex
2025.naacl-short.11
val_tab_0102
Our findings indicate that adding reinforcement learning (STRUX+RL) leads to stronger performance compared to using the SFT method alone.
Refuted
Table 1: Our STRUX system outperforms strong benchmarks in making stock investment decisions. We present macro-averaged precision, recall, F-scores, accuracy for the test set. LLMs evaluated are: Llama3-8b-Instruct and gpt-4o-mini-2024-07-18 .
table
tables_png/dev/val_tab_0102.png
System Comparisons. Table 1 shows the macro-averaged precision, recall, F-scores, and accuracy for the test set. STRUX outperforms strong baselines in accuracy and F-scores for stock investment decisions.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-short.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0082
tables/dev/val_tab_0102.tex
2025.naacl-short.11
val_tab_0103
The selection is predominantly positive, with 8 positive and 1 negative fact; about half of the negative fact has an impact strength of 2–3.
Supported
Table 3: Statistics of supporting facts.
table
tables_png/dev/val_tab_0103.png
Supporting Facts. We analyzed the supporting facts identified by the model in cases of correct decisions after reflections. Statistics are presented in Table 3 . Each transcript is distilled into a table of about 40 facts, from which the model selects 9.
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0083
tables/dev/val_tab_0103.tex
2025.naacl-short.11
val_tab_0104
The selection is predominantly positive, with 8 positive and 1 negative fact; about half of the negative fact has an impact strength of 2–3.
Refuted
Table 3: Statistics of supporting facts.
table
tables_png/dev/val_tab_0104.png
Supporting Facts. We analyzed the supporting facts identified by the model in cases of correct decisions after reflections. Statistics are presented in Table 3 . Each transcript is distilled into a table of about 40 facts, from which the model selects 9.
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0083
tables/dev/val_tab_0104.tex
2025.naacl-short.11
val_tab_0105
Moreover, reflection does not always yield perfect outcomes; the model can repeat decisions from previous cycles despite being instructed not to.
Supported
Table 2: The most common decision paths during reflection and their percentages in the training data.
table
tables_png/dev/val_tab_0105.png
Table 2 shows common decision paths during reflection. Interestingly, reflection can lead to abrupt decision changes, such as a direct jump from Buy to Strong Sell, instead of gradual shifts (e.g., Buy \rightarrow Hold \rightarrow Sell).
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0084
tables/dev/val_tab_0105.tex
2025.naacl-short.11
val_tab_0106
Moreover, reflection does not always yield perfect outcomes; the model can repeat decisions from previous cycles despite being instructed not to.
Refuted
Table 2: The most common decision paths during reflection and their percentages in the training data.
table
tables_png/dev/val_tab_0106.png
Table 2 shows common decision paths during reflection. Interestingly, reflection can lead to abrupt decision changes, such as a direct jump from Buy to Strong Sell, instead of gradual shifts (e.g., Buy \rightarrow Hold \rightarrow Sell).
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-short.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0084
tables/dev/val_tab_0106.tex
2025.naacl-long.11
val_tab_0107
All in all, gender-specific projector gets the higher correlation score.
Supported
Table 2 : Gender diversity. Average Spearman correlations between proportion of female voices and diversity scores induced by speech representations.
table
tables_png/dev/val_tab_0107.png
In Table 2 we report correlation scores for the gender diversity. Again, we split in two groups: male-voice and female voice-dominant. From the results, we notice that the correlations showed by the non-specialized embeddings are extremely weak. However, both specialized models reach very high correlations (e.g., up to...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0085
tables/dev/val_tab_0107.tex
2025.naacl-long.11
val_tab_0108
All in all, gender-specific projector gets the higher correlation score.
Refuted
Table 2 : Gender diversity. Average Spearman correlations between proportion of female voices and diversity scores induced by speech representations.
table
tables_png/dev/val_tab_0108.png
In Table 2 we report correlation scores for the gender diversity. Again, we split in two groups: male-voice and female voice-dominant. From the results, we notice that the correlations showed by the non-specialized embeddings are extremely weak. However, both specialized models reach very high correlations (e.g., up to...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0085
tables/dev/val_tab_0108.tex
2025.naacl-long.11
val_tab_0109
From these results, we see that, generally, SpeechSim performs better than other general-purpose representations across all configurations.
Supported
Table 3 : Emotion diversity. Average Spearman correlations between the classes entropy in EmoV and Expresso and diversity scores induced by the speech representations. SpeechSim/Emotion-Expresso (resp. Speech/Emotion-EmoV) refers to SpeechSim emotion head trained on Expresso (resp. EmoV).
table
tables_png/dev/val_tab_0109.png
In Table 3 we report average Spearman correlations for the tested representation models, when tested on EmoV and Expresso separately.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0086
tables/dev/val_tab_0109.tex
2025.naacl-long.11
val_tab_0110
From these results, we see that, generally, SpeechSim performs better than other general-purpose representations across all configurations.
Refuted
Table 3 : Emotion diversity. Average Spearman correlations between the classes entropy in EmoV and Expresso and diversity scores induced by the speech representations. SpeechSim/Emotion-Expresso (resp. Speech/Emotion-EmoV) refers to SpeechSim emotion head trained on Expresso (resp. EmoV).
table
tables_png/dev/val_tab_0110.png
In Table 3 we report average Spearman correlations for the tested representation models, when tested on EmoV and Expresso separately.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0086
tables/dev/val_tab_0110.tex
2025.naacl-long.11
val_tab_0111
Equally, the specialized projection of SpeechSim for the emotion facet, SpeechSim/Emotion performs best across all representations.
Supported
Table 3 : Emotion diversity. Average Spearman correlations between the classes entropy in EmoV and Expresso and diversity scores induced by the speech representations. SpeechSim/Emotion-Expresso (resp. Speech/Emotion-EmoV) refers to SpeechSim emotion head trained on Expresso (resp. EmoV).
table
tables_png/dev/val_tab_0111.png
In Table 3 we report average Spearman correlations for the tested representation models, when tested on EmoV and Expresso separately. From these results, we see that, generally, SpeechSim performs better than other general-purpose representations across all configurations.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0087
tables/dev/val_tab_0111.tex
2025.naacl-long.11
val_tab_0112
Finally, the original SpeechSim embeddings have negative correlation with the gender diversity.
Supported
Table 6 : Changing voice and gender diversity in the opposite directions. Average Spearman correlation between the gender diversity and the diversity scores induced by speech representations.
table
tables_png/dev/val_tab_0112.png
We report results in Table 6 . From this results we see that SpeechSim/Gender is positively correlated with the gender diversity. In contrast, the SpeechSim/Voice metric tracks the speaker diversity and hence is negatively correlated with the gender diversity.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0088
tables/dev/val_tab_0112.tex
2025.naacl-long.11
val_tab_0113
Finally, the original SpeechSim embeddings have negative correlation with the gender diversity.
Refuted
Table 6 : Changing voice and gender diversity in the opposite directions. Average Spearman correlation between the gender diversity and the diversity scores induced by speech representations.
table
tables_png/dev/val_tab_0113.png
We report results in Table 6 . From this results we see that SpeechSim/Gender is positively correlated with the gender diversity. In contrast, the SpeechSim/Voice metric tracks the speaker diversity and hence is negatively correlated with the gender diversity.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0088
tables/dev/val_tab_0113.tex
2025.naacl-long.11
val_tab_0114
Moreover, its correlation coefficient is smaller than that of SpeechSim/Voice.
Supported
Table 6 : Changing voice and gender diversity in the opposite directions. Average Spearman correlation between the gender diversity and the diversity scores induced by speech representations.
table
tables_png/dev/val_tab_0114.png
We report results in Table 6 . From this results we see that SpeechSim/Gender is positively correlated with the gender diversity. In contrast, the SpeechSim/Voice metric tracks the speaker diversity and hence is negatively correlated with the gender diversity. Finally, the original SpeechSim embeddings have negative co...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0089
tables/dev/val_tab_0114.tex
2025.naacl-long.11
val_tab_0115
Moreover, its correlation coefficient is smaller than that of SpeechSim/Voice.
Refuted
Table 6 : Changing voice and gender diversity in the opposite directions. Average Spearman correlation between the gender diversity and the diversity scores induced by speech representations.
table
tables_png/dev/val_tab_0115.png
We report results in Table 6 . From this results we see that SpeechSim/Gender is positively correlated with the gender diversity. In contrast, the SpeechSim/Voice metric tracks the speaker diversity and hence is negatively correlated with the gender diversity. Finally, the original SpeechSim embeddings have negative co...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.11.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0089
tables/dev/val_tab_0115.tex
2025.naacl-long.10
val_tab_0116
All models can relatively accurately identify hateful posts with negative emotions and non-hateful posts with positive emotions.
Supported
Table 4: Classification accuracy of HateBERT, ToxDect-roberta, Perspective API, and Llama Guard 3 1/8B on GPT-HateCheck grouped by the hatefulness label (hate) and the polarity of the detected emotions (emo). We highlight the “positive” labels in green (“non-hateful” and positive emotions) and “negative” labels in red ...
table
tables_png/dev/val_tab_0116.png
We further group the fine-grained emotions into positive (1), negative (-1), and ambiguous (0), based on Demszky et al. ( 2020 ) ’s taxonomy and present the models’ classification accuracy in the presence of emotions with different polarities in Table 4 . The result is revelatory:
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.10.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0090
tables/dev/val_tab_0116.tex
2025.naacl-long.10
val_tab_0117
All models can relatively accurately identify hateful posts with negative emotions and non-hateful posts with positive emotions.
Refuted
Table 4: Classification accuracy of HateBERT, ToxDect-roberta, Perspective API, and Llama Guard 3 1/8B on GPT-HateCheck grouped by the hatefulness label (hate) and the polarity of the detected emotions (emo). We highlight the “positive” labels in green (“non-hateful” and positive emotions) and “negative” labels in red ...
table
tables_png/dev/val_tab_0117.png
We further group the fine-grained emotions into positive (1), negative (-1), and ambiguous (0), based on Demszky et al. ( 2020 ) ’s taxonomy and present the models’ classification accuracy in the presence of emotions with different polarities in Table 4 . The result is revelatory:
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.10.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0090
tables/dev/val_tab_0117.tex
2025.naacl-long.10
val_tab_0119
This result is alarming since it suggests that HS detectors are entangled with emotion polarity, and some safe posts with negative emotions, such as counter-speech expressing disapproval or sadness, are likely marked as hateful, potentially silencing the voices of vulnerable groups.
Supported
Table 4: Classification accuracy of HateBERT, ToxDect-roberta, Perspective API, and Llama Guard 3 1/8B on GPT-HateCheck grouped by the hatefulness label (hate) and the polarity of the detected emotions (emo). We highlight the “positive” labels in green (“non-hateful” and positive emotions) and “negative” labels in red ...
table
tables_png/dev/val_tab_0119.png
We further group the fine-grained emotions into positive (1), negative (-1), and ambiguous (0), based on Demszky et al. ( 2020 ) ’s taxonomy and present the models’ classification accuracy in the presence of emotions with different polarities in Table 4 . The result is revelatory: All models can relatively accurately i...
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-long.10.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0091
tables/dev/val_tab_0119.tex
2025.naacl-long.10
val_tab_0120
This result is alarming since it suggests that HS detectors are entangled with emotion polarity, and some safe posts with negative emotions, such as counter-speech expressing disapproval or sadness, are likely marked as hateful, potentially silencing the voices of vulnerable groups.
Refuted
Table 4: Classification accuracy of HateBERT, ToxDect-roberta, Perspective API, and Llama Guard 3 1/8B on GPT-HateCheck grouped by the hatefulness label (hate) and the polarity of the detected emotions (emo). We highlight the “positive” labels in green (“non-hateful” and positive emotions) and “negative” labels in red ...
table
tables_png/dev/val_tab_0120.png
We further group the fine-grained emotions into positive (1), negative (-1), and ambiguous (0), based on Demszky et al. ( 2020 ) ’s taxonomy and present the models’ classification accuracy in the presence of emotions with different polarities in Table 4 . The result is revelatory: All models can relatively accurately i...
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-long.10.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0091
tables/dev/val_tab_0120.tex
2025.naacl-long.12
val_tab_0121
Additional synthetic datasets and high-quality retrieval datasets further improve the embedding model quality despite the tasks these datasets solve already being well represented in the basic datasets.
Supported
Table 2: Different data sources impact. Model performance is measured on ruMTEB. Avg. stands for the average score and is computed as the mean of the category scores. The best score is put in bold, the second best is underlined.
table
tables_png/dev/val_tab_0121.png
Results presented in Table 2 indicate that the embedding model gets better results when trained on data in Russian and English simultaneously.
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.12.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0092
tables/dev/val_tab_0121.tex
2025.naacl-long.12
val_tab_0122
Merging the encoder layers after language adaptation with the original model improves the model quality while using RetroMAE leads to decreased results.
Supported
Table 3: Results of the model, method, and data variation. † The reference results for the training objective and training examples sections is model based on ru-en-RoBERTa. Each experiment changes a single component (e.g., use AnglE similarity instead of cosine). Model performance is evaluated on ruMTEB. The best scor...
table
tables_png/dev/val_tab_0122.png
We perform contrastive fine-tuning for each model and then evaluate them on ruMTEB. Results (see Table 3 ) show that ru-en-RoBERTa outperforms both baselines by a significant margin. Additionally, the fact that XLM-R slightly outperforms ruRoBERTa may indicate that XLM-R copes better with knowledge transfer from basic ...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-long.12.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0093
tables/dev/val_tab_0122.tex
2025.naacl-short.13
val_tab_0124
The book of Genesis contains the highest number of half-verse chiasmi, while Numbers contains the most verse-level chiasmi.
Supported
Table 2: Summary of detected chiasmi. 2700+ chiasmi were detected at the verse and half-verse level. The highest number of chiasmi was found in the Book of Genesis and Book of Numbers. Both the precision and the inter-annotator agreement increase for the verse-level chiasmi.
table
tables_png/dev/val_tab_0124.png
Table 2 presents an overview of the system’s output for chiastic structures at the half-verse and verse levels. A total of 1,896 chiastic structures were identified at the half-verse level, with an average length of 5.93 textual units ( \pm 1.34) and an average score of 0.32 ( \pm 0.1). For verse-level groupings, 879 c...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-short.13.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0094
tables/dev/val_tab_0124.tex
2025.naacl-short.8
val_tab_0126
Meantime, Self-Taught-Llama-70B with extensive multilingual pre-training demonstrates the strongest cross-lingual transfer, achieving the best accuracies in all four non-English Multilingual RewardBench .
Supported
Table 2: Averaged Multilingual RewardBench results in two classifier RMs (top) and two generative RMs (bottom). Off-the-shelf RMs based on MLMs show strong cross-lingual transfer as in Table 1 .
table
tables_png/dev/val_tab_0126.png
Interestingly, we can observe strong cross-lingual transfer in the generative RMs in Table 2 , as in the classifier RMs. As discussed in Section 3.2 , fine-grained representation learning is a crucial component for having strong downstream generative abilities. While the extent of multilingual pre-training in GPT-4o is...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-short.8.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0095
tables/dev/val_tab_0126.tex
2025.naacl-short.8
val_tab_0127
Meantime, Self-Taught-Llama-70B with extensive multilingual pre-training demonstrates the strongest cross-lingual transfer, achieving the best accuracies in all four non-English Multilingual RewardBench .
Refuted
Table 2: Averaged Multilingual RewardBench results in two classifier RMs (top) and two generative RMs (bottom). Off-the-shelf RMs based on MLMs show strong cross-lingual transfer as in Table 1 .
table
tables_png/dev/val_tab_0127.png
Interestingly, we can observe strong cross-lingual transfer in the generative RMs in Table 2 , as in the classifier RMs. As discussed in Section 3.2 , fine-grained representation learning is a crucial component for having strong downstream generative abilities. While the extent of multilingual pre-training in GPT-4o is...
nlp
yes
Swap rows or columns
papers/dev/nlp_2025.naacl-short.8.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0095
tables/dev/val_tab_0127.tex
2023.emnlp-main.12
val_tab_0130
The results show that UTs are able to generalize better for the A and B splits, outperforming the LSTM and VT.
Supported
Table 5: Test accuracy of the models, trained on operation lengths of \leq 6 , with their out-of-distribution results shown here (lengths 7-12). LSTM baseline from Bowman et al. ( 2015 ) , and Transformer baseline from Shen et al. ( 2019 )
table
tables_png/dev/val_tab_0130.png
We train a 12 layer model with Attention (E=12,k=4,H=2,D=32,W=1) and FFD (E=12,K=4,D=128) and halting. Refer to Appendix A for further details. Training a 12-layer Vanilla Transformer achieves approximately the same results as in Shen et al. ( 2019 ) , so we report their results. Our results in Table 5 confirm the find...
nlp
yes
Swap rows or columns
papers/dev/nlp_2023.emnlp-main.12.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0096
tables/dev/val_tab_0130.tex
2023.emnlp-main.12
val_tab_0131
The results show that UTs are able to generalize better for the A and B splits, outperforming the LSTM and VT.
Refuted
Table 5: Test accuracy of the models, trained on operation lengths of \leq 6 , with their out-of-distribution results shown here (lengths 7-12). LSTM baseline from Bowman et al. ( 2015 ) , and Transformer baseline from Shen et al. ( 2019 )
table
tables_png/dev/val_tab_0131.png
We train a 12 layer model with Attention (E=12,k=4,H=2,D=32,W=1) and FFD (E=12,K=4,D=128) and halting. Refer to Appendix A for further details. Training a 12-layer Vanilla Transformer achieves approximately the same results as in Shen et al. ( 2019 ) , so we report their results. Our results in Table 5 confirm the find...
nlp
yes
Swap rows or columns
papers/dev/nlp_2023.emnlp-main.12.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0096
tables/dev/val_tab_0131.tex
2023.emnlp-main.1
val_tab_0132
As shown in Table 1 , for CSQA2.0, higher scores are reported for retrieval only than for induction only, while the result is contrary for StrategyQA.
Supported
Table 1: Performance on two ODQA tasks. The first two columns report scores on CSQA2.0 dev set and StreategyQA test set respectively. The last two columns compare IAG with ChatGPT on a randomly held-out subset containing 50 examples for each task.
table
tables_png/dev/val_tab_0132.png
Besides, the results on different setups of IAG-GPT suggest that, the relative contributions of the retrieval and the inductive knowledge can be different, depending on the tasks.
nlp
yes
Change the cell values
papers/dev/nlp_2023.emnlp-main.1.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0097
tables/dev/val_tab_0132.tex
2023.ijcnlp-main.9
val_tab_0133
The only exception is the case of style strength evaluation task in the direction of H_{1} \rightarrow H_{2} , using the GPT-NeoX model.
Supported
Table 6: Inter-annotator agreement scores for the three human evaluation tasks. Standard deviations over all data points are shown in brackets for the style strength and appropriateness evaluation tasks. The detailed procedure for calculating the agreement scores can be found in Appendix D .
table
tables_png/dev/val_tab_0133.png
The inter-annotator agreement in all of the tasks are shown in Table 6 . Note that, for calculating agreement in the semantic correctness evaluation task, all of the data points are aggregated to measure the agreement score as they represent categorical evaluation measures. On the other hand, that is not possible in ca...
nlp
yes
Change the cell values
papers/dev/nlp_2023.ijcnlp-main.9.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0098
tables/dev/val_tab_0133.tex
2025.naacl-long.1
val_tab_0136
Yet still only half of its explanations are considered adequate.
Supported
Table 6: Adequacy and Preference rates for generated explanations.
table
tables_png/dev/val_tab_0136.png
In Table 6 , we show adequacy and preference rates for explanations from the 3 systems, where an explanation is deemed adequate if both annotators agreed it is, and inadequate if both agreed it is not. The preference percentage is also taken among instances where the annotators agreed that the model’s explanation is pr...
nlp
yes
Change the cell values
papers/dev/nlp_2025.naacl-long.1.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0099
tables/dev/val_tab_0136.tex
2025.clpsych-1.4
val_tab_0138
It is also robust, with a standard deviation of 0.022 .
Supported
Table 2: Results of QuestMF ( ImbOLL ) framework over 3 different seed runs compared with ablation frameworks.
table
tables_png/dev/val_tab_0138.png
The results comparing our proposed QuestMF ( ImbOLL ) framework with its ablations and current state-of-the-art methods are presented in Table 1 . Since the prior works only show their best results on a single run, we also pick our best results on CCC for a fair comparison. As we can see, our proposed QuestMF ( ImbOLL ...
nlp
yes
Change the cell values
papers/dev/nlp_2025.clpsych-1.4.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0100
tables/dev/val_tab_0138.tex
2025.clpsych-1.4
val_tab_0139
It is also robust, with a standard deviation of 0.022 .
Refuted
Table 2: Results of QuestMF ( ImbOLL ) framework over 3 different seed runs compared with ablation frameworks.
table
tables_png/dev/val_tab_0139.png
The results comparing our proposed QuestMF ( ImbOLL ) framework with its ablations and current state-of-the-art methods are presented in Table 1 . Since the prior works only show their best results on a single run, we also pick our best results on CCC for a fair comparison. As we can see, our proposed QuestMF ( ImbOLL ...
nlp
yes
Change the cell values
papers/dev/nlp_2025.clpsych-1.4.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0100
tables/dev/val_tab_0139.tex
2024.eacl-short.15
val_tab_0141
However, by some simple maths, the readers can easily verify that even considering this, the accuracy of answers containing calculations is still much lower than that of answers that do not include calculations.
Supported
Table 1: The redundancy ( Red. ) and accuracy of LLMs’ responses. We report the average accuracy ( Avg. ) on all questions (second column), the accuracy for answers without calculation (Cal. ✗, third column) and with calculation (Cal. ✓, fourth column). \dagger : The accuracy of GPT-4 is 100% by construction since we u...
table
tables_png/dev/val_tab_0141.png
Next, for accuracy, even if LLMs are wrong for all the invalid questions, their accuracy should be around 85% if they get all the valid questions correct. However, this is clearly not the case for all LLMs except Claude-2. Next, for the rightmost column in Table 1 , if we assume that all the invalid samples happen to b...
nlp
yes
Swap rows or columns
papers/dev/nlp_2024.eacl-short.15.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0101
tables/dev/val_tab_0141.tex
2024.eacl-short.15
val_tab_0142
However, by some simple maths, the readers can easily verify that even considering this, the accuracy of answers containing calculations is still much lower than that of answers that do not include calculations.
Refuted
Table 1: The redundancy ( Red. ) and accuracy of LLMs’ responses. We report the average accuracy ( Avg. ) on all questions (second column), the accuracy for answers without calculation (Cal. ✗, third column) and with calculation (Cal. ✓, fourth column). \dagger : The accuracy of GPT-4 is 100% by construction since we u...
table
tables_png/dev/val_tab_0142.png
Next, for accuracy, even if LLMs are wrong for all the invalid questions, their accuracy should be around 85% if they get all the valid questions correct. However, this is clearly not the case for all LLMs except Claude-2. Next, for the rightmost column in Table 1 , if we assume that all the invalid samples happen to b...
nlp
yes
Swap rows or columns
papers/dev/nlp_2024.eacl-short.15.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0101
tables/dev/val_tab_0142.tex
2025.clpsych-1.1
val_tab_0143
At the same time, the RMSE obtained using the majority voting is also considerably lower than the results shown in the previous subsection from just one run (1.12 vs 1.45; columns GPT-4 {}_{\text{maj}} and GPT-4 in Table 2 ).
Supported
Table 2: RMSE results for research questions Q2, Q3, and Q4. GPT-4 stands for the GPT-4 annotator, Human for the human reader-annotator, avg : average of five GPT-4 completions/human guesses, maj : majority vote of five GPT-4 completions/human guesses, conf : majority vote of five GPT-4 completions/human guesses with c...
table
tables_png/dev/val_tab_0143.png
The application of the majority voting algorithm improved the RMSE on average by about 30% from 1.61 to 1.12 (columns GPT-4 {}_{\text{avg}} and GPT-4 {}_{\text{maj}} in Table 2 ). However, compared to the results shown in previous Section 4.2 (also shown in Table 2 ), generating five runs and taking the average increas...
nlp
yes
Change the cell values
papers/dev/nlp_2025.clpsych-1.1.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0102
tables/dev/val_tab_0143.tex
2025.clpsych-1.3
val_tab_0146
Also in this case, it can be observed that the choice of fine-tuning a generative model, leveraging its autoregressive characteristics for classification, leads to the best results.
Supported
Table 2: Emotional accuracy comparison between RACLETTE and other benchmarks (best results highlighted in green).
table
tables_png/dev/val_tab_0146.png
Table 2 shows how RACLETTE outperforms the benchmarks considered.
nlp
other sources
Swap rows or columns
papers/dev/nlp_2025.clpsych-1.3.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0103
tables/dev/val_tab_0146.tex
2025.clpsych-1.3
val_tab_0147
Also in this case, it can be observed that the choice of fine-tuning a generative model, leveraging its autoregressive characteristics for classification, leads to the best results.
Refuted
Table 2: Emotional accuracy comparison between RACLETTE and other benchmarks (best results highlighted in green).
table
tables_png/dev/val_tab_0147.png
Table 2 shows how RACLETTE outperforms the benchmarks considered.
nlp
other sources
Swap rows or columns
papers/dev/nlp_2025.clpsych-1.3.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0103
tables/dev/val_tab_0147.tex
2024.eacl-long.18
val_tab_0149
However, for NusaX and the African languages, a considerable number of them are not covered.
Supported
Table 1: Languages used in this paper. “✓” (green) and “✗” (red) mean that the language has and has not been seen by the models or language resources. CW indicates code-switching text. The language coverage for GPT-3.5 is derived from GPT-3 Brown et al. ( 2020 ) .
table
tables_png/dev/val_tab_0149.png
In Table 1 , we provide an overview of the languages and datasets used in this paper, categorized into: (1) high/medium-resource languages; (2) NusaX, covering local Indonesian languages (low resource); (3) African languages from SemEval 2023 (low resource); and (4) code-switching texts. The high/medium-resource langua...
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0104
tables/dev/val_tab_0149.tex
2024.eacl-long.18
val_tab_0150
However, for NusaX and the African languages, a considerable number of them are not covered.
Refuted
Table 1: Languages used in this paper. “✓” (green) and “✗” (red) mean that the language has and has not been seen by the models or language resources. CW indicates code-switching text. The language coverage for GPT-3.5 is derived from GPT-3 Brown et al. ( 2020 ) .
table
tables_png/dev/val_tab_0150.png
In Table 1 , we provide an overview of the languages and datasets used in this paper, categorized into: (1) high/medium-resource languages; (2) NusaX, covering local Indonesian languages (low resource); (3) African languages from SemEval 2023 (low resource); and (4) code-switching texts. The high/medium-resource langua...
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0104
tables/dev/val_tab_0150.tex
2024.eacl-long.18
val_tab_0151
Language coverage of the NRC-VAD multilingual lexicons remains limited in 109 languages.
Supported
Table 1: Languages used in this paper. “✓” (green) and “✗” (red) mean that the language has and has not been seen by the models or language resources. CW indicates code-switching text. The language coverage for GPT-3.5 is derived from GPT-3 Brown et al. ( 2020 ) .
table
tables_png/dev/val_tab_0151.png
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0105
tables/dev/val_tab_0151.tex
2024.eacl-long.18
val_tab_0152
Language coverage of the NRC-VAD multilingual lexicons remains limited in 109 languages.
Refuted
Table 1: Languages used in this paper. “✓” (green) and “✗” (red) mean that the language has and has not been seen by the models or language resources. CW indicates code-switching text. The language coverage for GPT-3.5 is derived from GPT-3 Brown et al. ( 2020 ) .
table
tables_png/dev/val_tab_0152.png
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0105
tables/dev/val_tab_0152.tex
2024.eacl-long.18
val_tab_0153
The improvements shown in Table 3 are particularly noticeable in few-shot training, with increases of +6.6 and +5.2 for binary and 3-way classification, respectively.
Supported
Table 3: Preliminary results, based on averaged macro-F1 scores across 34 languages. “EN Lex.” and “ML Lex.” indicate the English and multilingual NRC-VAD lexicons.
table
tables_png/dev/val_tab_0153.png
nlp
other sources
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0106
tables/dev/val_tab_0153.tex
2024.eacl-long.18
val_tab_0154
The improvements shown in Table 3 are particularly noticeable in few-shot training, with increases of +6.6 and +5.2 for binary and 3-way classification, respectively.
Refuted
Table 3: Preliminary results, based on averaged macro-F1 scores across 34 languages. “EN Lex.” and “ML Lex.” indicate the English and multilingual NRC-VAD lexicons.
table
tables_png/dev/val_tab_0154.png
nlp
other sources
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0106
tables/dev/val_tab_0154.tex
2024.eacl-long.18
val_tab_0155
In the full training scenario, the increments are smaller, at only +1.4 and +1.1 .
Supported
Table 3: Preliminary results, based on averaged macro-F1 scores across 34 languages. “EN Lex.” and “ML Lex.” indicate the English and multilingual NRC-VAD lexicons.
table
tables_png/dev/val_tab_0155.png
The improvements shown in Table 3 are particularly noticeable in few-shot training, with increases of +6.6 and +5.2 for binary and 3-way classification, respectively.
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0107
tables/dev/val_tab_0155.tex
2024.eacl-long.18
val_tab_0156
In the full training scenario, the increments are smaller, at only +1.4 and +1.1 .
Refuted
Table 3: Preliminary results, based on averaged macro-F1 scores across 34 languages. “EN Lex.” and “ML Lex.” indicate the English and multilingual NRC-VAD lexicons.
table
tables_png/dev/val_tab_0156.png
The improvements shown in Table 3 are particularly noticeable in few-shot training, with increases of +6.6 and +5.2 for binary and 3-way classification, respectively.
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0107
tables/dev/val_tab_0156.tex
2024.eacl-long.18
val_tab_0157
Overall, we observe that multilingual models fine-tuned on SST tend to perform the best in high/medium-resource languages, with mBERT {}_{\text{Base}} being the exception.
Supported
Table 4: Full zero-shot results. The underlined score indicates the highest performance within the respective group, while scores in bold indicate the best global performance. “HM-R” = high/medium-resource languages, excluding English, “CS” = code-switched text, and “ML Lex.” indicates the multilingual NRC-VAD lexicon....
table
tables_png/dev/val_tab_0157.png
Table 4 presents the averaged zero-shot performance of all models categorized by four language groups. The reported results use regression and classification in lexicon-based pretraining for binary and 3-way classification, respectively. In the case of high/medium-resource languages (HM-R), English is excluded to ensur...
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0108
tables/dev/val_tab_0157.tex
2024.eacl-long.18
val_tab_0158
Overall, we observe that multilingual models fine-tuned on SST tend to perform the best in high/medium-resource languages, with mBERT {}_{\text{Base}} being the exception.
Refuted
Table 4: Full zero-shot results. The underlined score indicates the highest performance within the respective group, while scores in bold indicate the best global performance. “HM-R” = high/medium-resource languages, excluding English, “CS” = code-switched text, and “ML Lex.” indicates the multilingual NRC-VAD lexicon....
table
tables_png/dev/val_tab_0158.png
Table 4 presents the averaged zero-shot performance of all models categorized by four language groups. The reported results use regression and classification in lexicon-based pretraining for binary and 3-way classification, respectively. In the case of high/medium-resource languages (HM-R), English is excluded to ensur...
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0108
tables/dev/val_tab_0158.tex
2024.eacl-long.18
val_tab_0159
Our findings indicate that regression performs better for binary classification, while classification leads to better results for 3-way classification.
Supported
Table 5: Regression vs. classification in lexicon-based pretraining for zero-shot sentiment analysis.
table
tables_png/dev/val_tab_0159.png
In Table 5 we present the average performance across the four language groups to compare the effectiveness of lexicon-based pretraining in regression and classification tasks for both binary and 3-way classification.
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0109
tables/dev/val_tab_0159.tex
2024.eacl-long.18
val_tab_0160
Our findings indicate that regression performs better for binary classification, while classification leads to better results for 3-way classification.
Refuted
Table 5: Regression vs. classification in lexicon-based pretraining for zero-shot sentiment analysis.
table
tables_png/dev/val_tab_0160.png
In Table 5 we present the average performance across the four language groups to compare the effectiveness of lexicon-based pretraining in regression and classification tasks for both binary and 3-way classification.
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0109
tables/dev/val_tab_0160.tex
2024.eacl-long.18
val_tab_0161
Although this setting works reasonably well for XLM-R, it yields poor performance for mT5.
Supported
Table 5: Regression vs. classification in lexicon-based pretraining for zero-shot sentiment analysis.
table
tables_png/dev/val_tab_0161.png
In Table 5 we present the average performance across the four language groups to compare the effectiveness of lexicon-based pretraining in regression and classification tasks for both binary and 3-way classification. Our findings indicate that regression performs better for binary classification, while classification l...
nlp
yes
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0110
tables/dev/val_tab_0161.tex
2024.eacl-long.18
val_tab_0162
Table 6 shows the average F1 scores on each task, demonstrating that lexicon-based pretraining boosts the performance of vanilla mBERT {}_{\text{Base}} .
Supported
Table 6: Lexicon-based pretraining performance (macro-F1) over stance detection, hate speech detection, and emotion classification. The results are based on the limited training data scenario.
table
tables_png/dev/val_tab_0162.png
For each task, we take two datasets and perform experiments in few-shot training using mBERT {}_{\text{Base}} , following the setup described in Section 4.2 . Instead of using the multilingual lexicon, we use the English NRC-VAD lexicon since all the data is in English. For detailed information about the datasets and r...
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.18.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0111
tables/dev/val_tab_0162.tex
2024.eacl-long.16
val_tab_0164
As reported in Table 2 , in Q Drawer the value of the entropy threshold \theta has, by design, a direct impact on the number of questions asked: i.e., a model that only asks for clarification in the face of high uncertainty ( \theta=1.1 ) gives rise to fewer dialogues with questions and to a lower average number of que...
Supported
Table 2: The effect of different \theta values on the number of questions asked by Q Drawer. We report the percentage of dialogues with at least one clarification question and, within this subset, the average number of clarification questions per dialogue.
table
tables_png/dev/val_tab_0164.png
Asking clarification questions carries a cost, both from the perspective of the agent asking the question and the agent processing it (Clark, 1996 ; Purver, 2004 ) . We take this partially into account by adopting a simple approach: we evaluate Q Drawer by using different entropy threshold values \theta to control for ...
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.16.json
CC BY 4.0
http://creativecommons.org/licenses/by/4.0/
0112
tables/dev/val_tab_0164.tex
2024.eacl-long.20
val_tab_0165
Here, while finetuning achieves higher accuracy for gendered words, it also degrades translation quality by 0.9 COMET.
Supported
Table 3: Grammatical gender control results in supervised condition (cross-domain: controller trained on gender-annotated data from a different domain).
table
tables_png/dev/val_tab_0165.png
As motivated in § 4.2 , the gender control results in Table 3 allow us to assess the impact of domain mismatch between the controller training data and the test data, a very realistic scenario in practice.
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.20.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0113
tables/dev/val_tab_0165.tex
2024.eacl-long.20
val_tab_0166
Here, while finetuning achieves higher accuracy for gendered words, it also degrades translation quality by 0.9 COMET.
Refuted
Table 3: Grammatical gender control results in supervised condition (cross-domain: controller trained on gender-annotated data from a different domain).
table
tables_png/dev/val_tab_0166.png
As motivated in § 4.2 , the gender control results in Table 3 allow us to assess the impact of domain mismatch between the controller training data and the test data, a very realistic scenario in practice.
nlp
no
Change the cell values
papers/dev/nlp_2024.eacl-long.20.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0113
tables/dev/val_tab_0166.tex
2024.eacl-long.20
val_tab_0168
For formality control, on Korean, the most distant language, CG consistently achieves stronger control results than finetuning, indicating more robustness when transferring to unfamiliar settings.
Supported
Table 4: Zero-shot formality control results. Best and second best results under the same data condition are marked.
table
tables_png/dev/val_tab_0168.png
While finetuning was consistently leading in supervised conditions (§ 5 ), now under zero-shot conditions with unseen target languages, the gap shrinks.
nlp
other sources
Swap rows or columns
papers/dev/nlp_2024.eacl-long.20.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0114
tables/dev/val_tab_0168.tex
2024.eacl-long.20
val_tab_0169
For formality control, on Korean, the most distant language, CG consistently achieves stronger control results than finetuning, indicating more robustness when transferring to unfamiliar settings.
Refuted
Table 4: Zero-shot formality control results. Best and second best results under the same data condition are marked.
table
tables_png/dev/val_tab_0169.png
While finetuning was consistently leading in supervised conditions (§ 5 ), now under zero-shot conditions with unseen target languages, the gap shrinks.
nlp
other sources
Swap rows or columns
papers/dev/nlp_2024.eacl-long.20.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0114
tables/dev/val_tab_0169.tex
2024.eacl-long.20
val_tab_0173
In contrast, for the pretrained NLLB, there is no clear distinction between the multilingual systems and rest.
Supported
Table 4: Zero-shot formality control results. Best and second best results under the same data condition are marked.
table
tables_png/dev/val_tab_0173.png
In Table 4 , controllers trained on multiple translation directions ( multi ) are compared to those trained on single directions ( en \rightarrow es or de ). On Transformer-base, multi consistently outperforms its single-direction counterparts, regardless whether the controller is finetuning- or CG-based.
nlp
other sources
Change the cell values
papers/dev/nlp_2024.eacl-long.20.json
CC BY 4.0
https://creativecommons.org/licenses/by/4.0/
0115
tables/dev/val_tab_0173.tex