ResearchArcade
Collection
23 items
•
Updated
Dataset Viewer
id
int64 1
24.5M
| paragraph_id
int64 0
159k
| content
stringlengths 1
571k
⌀ | paper_arxiv_id
stringlengths 10
13
| paper_section
stringlengths 1
262
| section_id
float64 4
360k
⌀ | paragraph_in_paper_id
int64 1
29.8k
|
|---|---|---|---|---|---|---|
22,512,201
| 2
|
Instead of manually designing prompts, CoOp \cite{zhou2022learning} proposes a prompt tuning approach that learns task-specific prompts. In this method, learnable vectors in the word embedding space are trained using image-text pairs from target tasks and serve as prompts. This approach achieves significant performance improvements over using hand-crafted prompts, emphasizing the importance of prompt designing. However, learned prompts often fail to generalize to unseen classes that were not targeted during prompt tuning, resulting in substantial performance degradation. Extensive research has focused on learning generalizable prompts by exploring both architectural improvements \cite{zhou2022conditional, khattak2023maple, lee2023read, zhang2024dept} and regularization methods \cite{zhu2023prompt, yao2023kgcoop, cho2023distribution, ding2024lobg}. A widely adopted regularization method minimizes the distance between the text embeddings generated by learnable prompts and those generated by hand-crafted prompts during prompt tuning \cite{yao2023kgcoop, khattak2023promptsrc, roy2024coprompt, yao2024tcp, ding2024lobg}. This approach improves the generalization ability of learnable prompts by preserving the general knowledge captured by hand-crafted prompts. However, a significant limitation of this method lies in its focus on individual classes in isolation, without considering the relationships among classes. As a result, inter-class consistency is not explicitly accounted for during prompt tuning, which indicates that there is room for further improvement in generalization.
|
2502.11969
|
Introduction
| 228,291
| 3
|
22,512,202
| 3
|
In this work, we first investigate how prompts learned from base classes fail to generalize to unseen classes in terms of semantic disruptions. We observe that when the text embeddings generated by learned prompts form incorrect semantic relationships with other classes, they act as low-quality classifiers, leading to poor generalization. In contrast, hand-crafted prompts capture meaningful semantic relationships among classes due to their strong generalization ability. Motivated by these findings, we propose \textbf{S}imilarity \textbf{A}lignment \textbf{R}egularization (\textbf{SAR}), a method for regularizing learnable prompts to preserve the semantic relationships captured by hand-crafted prompts. Specifically, we first obtain novel classes semantically aligned with the base classes by utilizing ChatGPT-4o. For both base and novel classes, SAR aligns the similarity relationships among text embeddings generated by learned prompts with those relationships from hand-crafted prompts. By preserving the meaningful semantic relationships among classes, SAR enables learnable prompts to effectively capture the semantics of unseen classes. Additionally, at each parameter update step, regularization is applied to the similarities computed on randomly sampled classes, instead of using all classes. This approach mitigates overfitting by introducing noise during prompt tuning and further improves the generalization of learned prompts to unseen classes.
|
2502.11969
|
Introduction
| 228,291
| 4
|
22,512,203
| 4
|
As a regularizer, SAR can be applied to existing prompt tuning models without any architectural modifications. We validate the effectiveness of SAR on various baselines, including textual prompt tuning models: CoOp \cite{zhou2022learning}, KgCoOp \cite{yao2023kgcoop}, TCP \cite{yao2024tcp} and multi-modal prompt tuning models: MaPLe \cite{khattak2023maple}, CoPrompt \cite{roy2024coprompt}. Notably, in base-to-new generalization experiments, SAR consistently improves overall accuracy on new classes across 11 datasets for five baselines, while either preserving or enhancing the base accuracy. Our contributions are summarized as follows:
\begin{itemize}[itemsep=0.5em, topsep=0.5em]
\item We experimentally analyze how learned prompts fail to generalize to unseen classes, focusing on semantic disruptions.
\item We propose SAR, a method to guide learnable prompts to capture meaningful semantic relationships among classes. SAR utilizes novel classes generated by ChatGPT-4o as potential unseen classes during prompt tuning.
\item We incorporate a random embedding sampling strategy into SAR, to mitigate overfitting.
\item We demonstrate the effectiveness of SAR in improving generalization to unseen classes through extensive experiments across 11 datasets and five baselines.
\end{itemize}\begin{itemize}[itemsep=0.5em, topsep=0.5em]
\item We experimentally analyze how learned prompts fail to generalize to unseen classes, focusing on semantic disruptions.
\item We propose SAR, a method to guide learnable prompts to capture meaningful semantic relationships among classes. SAR utilizes novel classes generated by ChatGPT-4o as potential unseen classes during prompt tuning.
\item We incorporate a random embedding sampling strategy into SAR, to mitigate overfitting.
\item We demonstrate the effectiveness of SAR in improving generalization to unseen classes through extensive experiments across 11 datasets and five baselines.
\end{itemize}
\item We experimentally analyze how learned prompts fail to generalize to unseen classes, focusing on semantic disruptions.
\item We propose SAR, a method to guide learnable prompts to capture meaningful semantic relationships among classes. SAR utilizes novel classes generated by ChatGPT-4o as potential unseen classes during prompt tuning.
\item We incorporate a random embedding sampling strategy into SAR, to mitigate overfitting.
\item We demonstrate the effectiveness of SAR in improving generalization to unseen classes through extensive experiments across 11 datasets and five baselines.
|
2502.11969
|
Introduction
| 228,291
| 5
|
22,512,207
| 3
|
\keypoint{Regularization for Prompt Tuning.}Regularization for Prompt Tuning. Various regularization techniques have been explored to learn generalizable prompts. ProGrad \cite{zhu2023prompt} ensures that the gradient for prompt tuning does not conflict with the gradient used to preserve the general knowledge of CLIP. Specifically, if the angle between the two gradient vectors is obtuse, the gradient for prompt tuning is projected to be orthogonal to the other gradient, and is used for the update. KgCoOp \cite{yao2023kgcoop} minimizes the distance between text embeddings generated by learnable prompts and those generated by hand-crafted prompts, preserving the general knowledge captured by hand-crafted prompts. LASP \cite{bulat2023lasp} introduces a text-to-text cross-entropy loss to ensure that text embeddings generated by learnable prompts are correctly classified as those generated by hand-crafted prompts for the same class. DAPT \cite{cho2023distribution} enforces a uniform distribution of text embeddings on a hypersphere to minimize overlap while encouraging image embeddings of the same class to be positioned closer together, to achieve better alignment. TPR \cite{chentpr} maximize Pearson correlation coefficient computed between the pairwise cosine similarity matrices of the original CLIP text embeddings and the learned text embeddings, to preserve the class topology. In this process, they utilize the textual descriptions of both base and new classes in the dataset. LOBG \cite{ding2024lobg} preserves the structural topology of image embeddings captured by hand-crafted prompts, which is achieved by maintaining local angular relationships among image embeddings. Unlike these methods, our method SAR utilize the similarity distribution among text embeddings, to effectively capture the relational similarities among classes.
|
2502.11969
|
Related Work
| 228,292
| 9
|
22,512,208
| 0
|
In this section, we present the implementation of SAR and its integration within the prompt tuning framework.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 10
|
22,512,209
| 1
|
\subsection{Preliminaries}
\keypoint{Contrastive Language-Image Pre-training (CLIP) \cite{radford2021clip}.}Contrastive Language-Image Pre-training (CLIP) \cite{radford2021clip}. CLIP is a vision-language model pre-trained on 400 million image-text pairs, designed to align semantic relationships between image and text modalities. It consists of an image encoder $\theta(\cdot)$\theta(\cdot) and a text encoder $\phi(\cdot)$\phi(\cdot) that map their respective inputs into a shared embedding space. CLIP is trained using contrastive learning, where embeddings of matched image-text pairs are pulled closer together, while those of unmatched pairs are pushed apart. We can perform zero-shot image classification using CLIP by comparing the matching score, which is cosine similarity between the image embedding and the text embedding. Using appropriate prompts can significantly enhance CLIP’s classification performance. Finding task-optimized prompts is referred to as prompt engineering, which often relies on a trial-and-error process.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 11
|
22,512,210
| 2
|
\keypoint{Context Optimization.}Context Optimization. CoOp \cite{zhou2022learning} proposes a framework that trains prompts via a classification task using image-text pairs of target tasks. In CoOp, $P$P trainable vectors $[\boldsymbol{v}_1, \boldsymbol{v}_2, \dots, \boldsymbol{v}_P]$[\boldsymbol{v}v_1, \boldsymbol{v}v_2, \dots, \boldsymbol{v}v_P] serve as prompts. These vectors are added to $\boldsymbol{c}_i$\boldsymbol{c}c_i, the word embedding(s) of the $i$i-th class name $\boldsymbol{w}_i$\boldsymbol{w}w_i, and passed to the text encoder to compute the text embedding $\boldsymbol{g}_i = \phi([\boldsymbol{v}_1, \boldsymbol{v}_2, \dots, \boldsymbol{v}_P, \boldsymbol{c}_i])$\boldsymbol{g}g_i = \phi([\boldsymbol{v}v_1, \boldsymbol{v}v_2, \dots, \boldsymbol{v}v_P, \boldsymbol{c}c_i]). Similarly, the image embedding $\boldsymbol{f}$\boldsymbol{f}f is computed by passing each image $x$x through the image encoder $\theta(\cdot)$\theta(\cdot). The probability that $\boldsymbol{x}$\boldsymbol{x}x is predicted as $\boldsymbol{w}_i$\boldsymbol{w}w_i is computed as:
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 12
|
22,512,211
| 3
|
\begin{equation}
p(\boldsymbol{w}_i|\boldsymbol{x}) = \frac{\exp(\cos(\boldsymbol{g}_i,\boldsymbol{f}) / \tau)}{\sum_{j=1}^C \exp(\cos(\boldsymbol{g}_j,\boldsymbol{f}) / \tau)},
\end{equation}\begin{equation}
p(\boldsymbol{w}_i|\boldsymbol{x}) = \frac{\exp(\cos(\boldsymbol{g}_i,\boldsymbol{f}) / \tau)}{\sum_{j=1}^C \exp(\cos(\boldsymbol{g}_j,\boldsymbol{f}) / \tau)},
\end{equation}
p(\boldsymbol{w}w_i|\boldsymbol{x}x) = \frac{\exp(\cos(\boldsymbol{g}_i,\boldsymbol{f}) / \tau)}{\sum_{j=1}^C \exp(\cos(\boldsymbol{g}_j,\boldsymbol{f}) / \tau)},
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 13
|
22,512,212
| 4
|
where $\cos(\cdot, \cdot)$\cos(\cdot, \cdot) denotes cosine similarity, $C$C is the number of classes, and $\tau$\tau is the learned temperature parameter of CLIP. To optimize prompts for target tasks, trainable vectors are updated by cross-entropy loss:
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 14
|
22,512,213
| 5
|
\begin{equation}
\mathcal{L}_{\text{ce}}=-\sum_{i} \boldsymbol{y}_i \log p(\boldsymbol{w}_i|\boldsymbol{x}),
\end{equation}\begin{equation}
\mathcal{L}_{\text{ce}}=-\sum_{i} \boldsymbol{y}_i \log p(\boldsymbol{w}_i|\boldsymbol{x}),
\end{equation}
\mathcal{L}_{\text{ce}}\text{ce}=-\sum_{i}i \boldsymbol{y}y_i \log p(\boldsymbol{w}w_i|\boldsymbol{x}x),
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 15
|
22,512,216
| 8
|
\keypoint{Computing Semantic Similarities among Classes.}Computing Semantic Similarities among Classes. We begin by describing how to compute a similarity distribution matrix, which shows the semantic relationships among text embeddings. Let $\{\boldsymbol{g}_j\}_{j=1}^{N}$\{\boldsymbol{g}g_j\}_{j=1}j=1^{N}N denote a set of text embeddings corresponding to $N$N arbitrary classes, where $\boldsymbol{g}_j \in \mathbb{R}^{d}\ \forall j$\boldsymbol{g}g_j \in \mathbb{R}^{d}d\ \forall j. Since matched image-text pairs are located closer in CLIP’s embedding space, text pairs with similar semantics also tend to be closer, allowing us to measure their semantic similarities via cosine similarity. We can compute the cosine similarity matrix $\mathbf{C} \in \mathbb{R}^{N \times N}$\mathbf{C} \in \mathbb{R}^{N \times N}N \times N with entries $\mathbf{C}_{ij}=\cos(\boldsymbol{g}_i, \boldsymbol{g}_j)$\mathbf{C}_{ij}ij=\cos(\boldsymbol{g}g_i, \boldsymbol{g}g_j). To highlight relational similarities among classes, we compute the similarity distribution matrix $\mathbf{P}$\mathbf{P}, by applying a softmax function to each row of $\mathbf{C}$\mathbf{C}, excluding its diagonal elements (which represent self-similarity) from the computation:
\begin{equation}
\mathbf{P}_{ij} =
\begin{cases}
0, & \text{if } i = j, \\
\frac{\exp\left( \cos(\boldsymbol{g}_i, \boldsymbol{g}_j) / \tau \right)}{\sum_{k \neq i} \exp\left( \cos(\boldsymbol{g}_i, \boldsymbol{g}_k) / \tau \right)}, & \text{if } i \neq j.
\end{cases}
\end{equation}\begin{equation}
\mathbf{P}_{ij} =
\begin{cases}
0, & \text{if } i = j, \\
\frac{\exp\left( \cos(\boldsymbol{g}_i, \boldsymbol{g}_j) / \tau \right)}{\sum_{k \neq i} \exp\left( \cos(\boldsymbol{g}_i, \boldsymbol{g}_k) / \tau \right)}, & \text{if } i \neq j.
\end{cases}
\end{equation}
\mathbf{P}_{ij}ij =
0, & \text{if } i = j, \\
\frac{\exp\left( \cos(\boldsymbol{g}_i, \boldsymbol{g}_j) / \tau \right)}{\sum_{k \neq i} \exp\left( \cos(\boldsymbol{g}_i, \boldsymbol{g}_k) / \tau \right)}, & \text{if } i \neq j.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 18
|
22,512,217
| 9
|
Note that the matrix $\mathbf{P}$\mathbf{P} is not necessarily symmetric, as the normalization can differ for each row.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 19
|
22,512,219
| 10
|
\keypoint{Prompt Generalization Evaluation.}Prompt Generalization Evaluation. To evaluate the generalization of learned prompts to unseen classes, we perform base-to-new generalization experiments. Specifically, we first train prompts using image-text pairs from only half of dataset's classes (base classes). Then, we conduct zero-shot evaluation on the test data of remaining half classes (new classes) using the learned prompts. During this process, we compare the two similarity distribution matrices computed over all (base+new) classes: 1) $\mathbf{P}_{\mathtt{CoOp}}$\mathbf{P}_{\mathtt{CoOp}}\mathtt{CoOp}, produced by learned prompts, and 2) $\mathbf{P}_{\mathtt{hand}}$\mathbf{P}_{\mathtt{hand}}\mathtt{hand}, produced by hand-crafted prompts. Additionally, we examine the logit patterns for the test images. Specifically, we transform the logit vector for each image into a 2-dimensional vector using t-SNE \cite{van2008visualizing} and visualize the results.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 20
|
22,512,221
| 11
|
The experimental results on EuroSAT \cite{helber2019eurosat} are presented in \Cref{fig:motivation}. The leftmost heatmap, corresponding to $\mathbf{P}_{\mathtt{CoOp}}$\mathbf{P}_{\mathtt{CoOp}}\mathtt{CoOp}, reveals that the learned prompts generate semantically misaligned text embeddings for certain new classes. For example, in $\mathbf{P}_{\mathtt{CoOp}}$\mathbf{P}_{\mathtt{CoOp}}\mathtt{CoOp}, the class $River$River shows a higher similarity to $Forest$Forest than to $Sea\ or\ Lake$Sea\ or\ Lake, despite the latter being semantically closer. Similarly, the similarity between $Sea\ or\ Lake$Sea\ or\ Lake and $Highway\ or\ Road$Highway\ or\ Road is abnormally high, even though they are semantically unrelated. Moreover, the learned prompts fail to preserve subtle semantic distinctions among classes, forming overly confident associations. In contrast, $\mathbf{P}_{\mathtt{hand}}$\mathbf{P}_{\mathtt{hand}}\mathtt{hand} captures the more meaningful semantic relationships and better reflects subtle distinctions among classes. The logits scatterplots at the bottom further illustrate that semantically misaligned text embeddings lead to poor generalization in the classification task. In the scatterplot for CoOp, the logits for $Sea\ or\ Lake$Sea\ or\ Lake and $River$River form poorly defined clusters that overlap significantly, showing that the text embeddings struggle to capture the visual concepts of the two classes. In contrast, in the scatterplot for the hand-crafted prompt, the clusters for the two classes are clearly separated, highlighting its superior capability in generalization.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 21
|
22,512,223
| 12
|
\subsection{Similarity Alignment Regularization}
To prevent semantic disruptions caused by learned prompts and improve generalization, we propose Similarity Alignment Regularization (SAR). \Cref{fig:framework} illustrates the operation of SAR in prompt learning.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 22
|
22,536,844
| 1
|
The scalability of generative AI systems when handling big datasets presents another difficulty. The constraints imposed by API query limits and computational resources frequently cause traditional generative AI models to falter \cite{vaswani2017attention}. This limits their capacity to effectively manage large amounts of data, which lowers the caliber and comprehensiveness of the produced outputs. Longer processing times, higher computational costs, and the inability to handle real-time or nearly real-time data processing—all of which are crucial in many industrial applications—are all consequences of scalability problems.
|
2502.13164
|
Introduction
| 229,212
| 2
|
22,512,224
| 13
|
\keypoint{Prompt Learning with Novel Classes.}Prompt Learning with Novel Classes. We use ChatGPT-4o to generate novel classes semantically aligned with the base classes, utilizing them as potential unseen classes during prompt tuning. Specifically, we provide ChatGPT-4o with a list of the base classes and instruct it to generate semantically aligned novel classes. The prompt given to ChatGPT-4o is provided in \underline{\textbf{Supp. Material. A}}. To obtain more robust and representative embeddings for each class, we use ensembled text embeddings from multiple hand-crafted prompts. Let $T$T be the number of hand-crafted prompts used for prompt ensembling. For $j$j-th class, the ensembled text embedding is computed as $\bar{\boldsymbol{g}}_j=\frac{1}{T}\sum_{i=1}^T \boldsymbol{g}_j^i$\bar{\boldsymbol{g}}_j=\frac{1}{T}\sum_{i=1}i=1^T \boldsymbol{g}g_j^i, where $\boldsymbol{g}_j^i$\boldsymbol{g}g_j^i is the text embedding of $j$j-th class generated by the $i$i-th prompt. The hand-crafted prompts used for ensembling are listed in \underline{\textbf{Supp. Material. B}}. Let $\mathcal{G}_{\mathtt{hand}}=\{\bar{\boldsymbol{g}}_j\}_{j=1}^{M}$\mathcal{G}_{\mathtt{hand}}\mathtt{hand}=\{\bar{\boldsymbol{g}}_j\}_{j=1}j=1^{M}M denote the set of text embeddings for both base and novel classes, generated by the hand-crafted prompts. In contrast, $\mathcal{G}_{\mathtt{CoOp}}=\{\hat{\boldsymbol{g}}_j\}_{j=1}^{M}$\mathcal{G}_{\mathtt{CoOp}}\mathtt{CoOp}=\{\hat{\boldsymbol{g}}_j\}_{j=1}j=1^{M}M represents the set of text embeddings generated by the learnable prompts.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 23
|
22,512,225
| 14
|
For learnable prompts to preserve meaningful semantic relationships among classes, SAR aim to align the semantic relationships in $\mathcal{G}_{\mathtt{CoOp}}$\mathcal{G}_{\mathtt{CoOp}}\mathtt{CoOp} with those in $\mathcal{G}_{\mathtt{hand}}$\mathcal{G}_{\mathtt{hand}}\mathtt{hand} during prompt tuning.
A straightforward way to achieve this is to construct similarity distribution matrices for $\mathcal{G}_{\mathtt{CoOp}}$\mathcal{G}_{\mathtt{CoOp}}\mathtt{CoOp} and $\mathcal{G}_{\mathtt{hand}}$\mathcal{G}_{\mathtt{hand}}\mathtt{hand}, respectively, and minimize the KL divergence \cite{kullback1951information} between them during prompt learning. This method simultaneously takes into account all pairwise semantic similarities between the text embeddings. To mitigate overfitting, however, we introduce randomness into the regularization process through random embedding sampling.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 24
|
22,512,226
| 15
|
\keypoint{Computing Similarities for Sampled Classes.}Computing Similarities for Sampled Classes. For each $\hat{\boldsymbol{g}}_j$\hat{\boldsymbol{g}}_j, we compute similarities with $K$K other embeddings that are randomly sampled, rather than considering all other embeddings. Let $\mathcal{I}_j$\mathcal{I}_j denote the set of $K$K indices of the sampled embeddings, ensuring that $j \notin \mathcal{I}_j$j \notin \mathcal{I}_j to exclude self-similarity. Now the similarity disbribution matrix $\mathbf{P}_{\mathtt{CoOp}}^{\mathcal{I}}$\mathbf{P}_{\mathtt{CoOp}}\mathtt{CoOp}^{\mathcal{I}}\mathcal{I} is computed as:
\renewcommand{\arraystretch}{1.5}
\begin{equation}
\mathbf{P}_{\mathtt{CoOp}}^{\mathcal{I}} =
\begin{bmatrix}
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_1, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_1\big]\Big) \\
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_2, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_2\big]\Big) \\
\vdots \\
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_M, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_M\big]\Big)
\end{bmatrix}
\in \mathbb{R}^{M \times K},
\end{equation}\begin{equation}
\mathbf{P}_{\mathtt{CoOp}}^{\mathcal{I}} =
\begin{bmatrix}
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_1, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_1\big]\Big) \\
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_2, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_2\big]\Big) \\
\vdots \\
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_M, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_M\big]\Big)
\end{bmatrix}
\in \mathbb{R}^{M \times K},
\end{equation}
\mathbf{P}_{\mathtt{CoOp}}\mathtt{CoOp}^{\mathcal{I}}\mathcal{I} =
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 25
|
22,512,227
| 16
|
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_1, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_1\big]\Big) \\
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_2, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_2\big]\Big) \\
\vdots \\
\sigma\Big(\big[\cos(\hat{\boldsymbol{g}}_M, \hat{\boldsymbol{g}}_k) \,|\, k \in \mathcal{I}_M\big]\Big)
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 26
|
22,512,229
| 18
|
where $\sigma(\cdot)$\sigma(\cdot) denotes the softmax function with the CLIP-learned temperature $\tau$\tau. Using $\mathcal{G}_{\mathtt{hand}}$\mathcal{G}_{\mathtt{hand}}\mathtt{hand}, we also compute $\mathbf{P}_{\mathtt{hand}}^{\mathcal{I}} \in \mathbb{R}^{M \times K}$\mathbf{P}_{\mathtt{hand}}\mathtt{hand}^{\mathcal{I}}\mathcal{I} \in \mathbb{R}^{M \times K}M \times K with the same family $\mathcal{I}$\mathcal{I} of index sets, as the target of $\mathbf{P}_{\mathtt{CoOp}}^{\mathcal{I}}$\mathbf{P}_{\mathtt{CoOp}}\mathtt{CoOp}^{\mathcal{I}}\mathcal{I}. This approach allows learnable prompts to capture the relative semantic similarities among classes more effectively.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 28
|
22,512,230
| 19
|
\keypoint{Randomized Similarity Alignment.}Randomized Similarity Alignment. SAR minimizes the KL divergence between the corresponding rows of $\mathbf{P}^{\mathcal{I}}_{\mathtt{CoOp}}$\mathbf{P}^{\mathcal{I}}\mathcal{I}_{\mathtt{CoOp}}\mathtt{CoOp} and $\mathbf{P}^{\mathcal{I}}_{\mathtt{hand}}$\mathbf{P}^{\mathcal{I}}\mathcal{I}_{\mathtt{hand}}\mathtt{hand}:
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 29
|
22,512,231
| 20
|
\begin{equation}
\mathcal{L}_{\text{SAR}} = \frac{1}{M} \sum_{i=1}^M \mathcal{D}_{\mathcal{KL}}(\mathbf{P}^{\mathcal{I}}_{\mathtt{CoOp},i} \parallel \mathbf{P}^{\mathcal{I}}_{\mathtt{hand},i}).
\end{equation}\begin{equation}
\mathcal{L}_{\text{SAR}} = \frac{1}{M} \sum_{i=1}^M \mathcal{D}_{\mathcal{KL}}(\mathbf{P}^{\mathcal{I}}_{\mathtt{CoOp},i} \parallel \mathbf{P}^{\mathcal{I}}_{\mathtt{hand},i}).
\end{equation}
\mathcal{L}_{\text{SAR}}\text{SAR} = \frac{1}{M} \sum_{i=1}i=1^M \mathcal{D}_{\mathcal{KL}}\mathcal{KL}(\mathbf{P}^{\mathcal{I}}\mathcal{I}_{\mathtt{CoOp},i}\mathtt{CoOp},i \parallel \mathbf{P}^{\mathcal{I}}\mathcal{I}_{\mathtt{hand},i}\mathtt{hand},i).
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 30
|
22,512,232
| 21
|
where $\mathcal{I}$\mathcal{I} is newly sampled at every single parameter update step during prompt tuning.
Finally, the training objective for SAR-applied prompt tuning is:
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 31
|
22,512,233
| 22
|
\begin{equation}
\mathcal{L}=\mathcal{L}_{\text{ce}}+\lambda \mathcal{L}_{\text{SAR}}
\end{equation}\begin{equation}
\mathcal{L}=\mathcal{L}_{\text{ce}}+\lambda \mathcal{L}_{\text{SAR}}
\end{equation}
\mathcal{L}=\mathcal{L}_{\text{ce}}\text{ce}+\lambda \mathcal{L}_{\text{SAR}}\text{SAR}
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 32
|
22,512,234
| 23
|
where $\lambda$\lambda is the hyperparameter that controls the strength of the regularization.
|
2502.11969
|
Prompt Tuning with SAR
| 228,293
| 33
|
22,512,236
| 1
|
\keypoint{Datasets.}Datasets.
For base-to-new generalization, we evaluate our method on 11 datasets. ImageNet \cite{deng2009imagenet} and Caltech-101 \cite{fei2004learning} are used for generic object classification, while fine-grained classification is conducted on OxfordPets \cite{parkhi2012cats}, StanfordCars \cite{krause20133d}, Flowers102 \cite{nilsback2008automated}, Food-101 \cite{bossard2014food}, and FGVCAircraft \cite{maji2013fine}. For domain-specific classification tasks, we use EuroSAT \cite{helber2019eurosat} for satellite image classification, UCF101 \cite{soomro2012ucf101} for action recognition, DTD \cite{cimpoi2014describing} for texture classification, and SUN397 \cite{xiao2010sun} for scene recognition. For domain generalization, ImageNet serves as the source dataset, while ImageNet-V2 \cite{recht2019imagenet}, ImageNet-Sketch \cite{wang2019learning}, ImageNet-A \cite{gao2022generating}, and ImageNet-R \cite{hendrycks2021many} act as target domains.
|
2502.11969
|
Experiments
| 228,294
| 35
|
22,512,237
| 2
|
\keypoint{Implementation details.}Implementation details.
All experiments are conducted with the CLIP ViT-B/16 model. For each dataset, 200 novel classes generated by ChatGPT-4o are used for SAR. The embedding sampling size $K$K is fixed at 64 across all experiments. In base-to-new generalization, for CoOp and KgCoOp, the number of training epochs is set to 100 by default to reduce training time. Except for this, we follow the same training settings, such as learning rate, epochs, and training schedules, as the baseline models. The hyperparameter $\lambda$\lambda for SAR is adjusted based on the difficulty of optimizing its loss and the characteristics of the dataset. For TCP, $\lambda$\lambda is set to 0.1 across all datasets. On EuroSAT, $\lambda$\lambda is set to 0.5 in CoOp and 0.75 in both MaPLe and CoPrompt. On DTD and UCF101, $\lambda$\lambda is set to 0.5 in all three models. On FGVCAircraft, $\lambda$\lambda is set to 0.5 in CoOp, 0.75 in MaPLe, and 0.8 in CoPrompt. For all remaining datasets, $\lambda$\lambda is set to 0.1 in all three models. The final performance is reported as the average over seeds 1, 2, and 3 to ensure a fair evaluation. We reproduce all baseline performances using the original code provided by the authors.
|
2502.11969
|
Experiments
| 228,294
| 36
|
22,512,238
| 3
|
\subsection{Ablation Study}
\keypoint{Stepwise Analysis of SAR's Effectiveness.}Stepwise Analysis of SAR's Effectiveness.
To analyze the contribution of each component in SAR, we conduct a base-to-new generalization experiment in the 16-shot setting, using CoOp as the baseline. The results, averaged across 11 datasets, are presented in \Cref{tab:component_ab}. First, CoOp exhibits a significant gap between base and new accuracies, revealing its limited ability to generalize to unseen classes. By introducing $\mathcal{L}_{\text{SAR}}$\mathcal{L}_{\text{SAR}}\text{SAR}, this limitation is effectively addressed, leading to a substantial improvement in new accuracy (+5.3\%) while also slightly enhancing base class accuracy (+0.11\%). Building on this, the integration of prompt ensembling further enhances both base and new accuracies by guiding the learnable prompts with more robust semantic relationships among classes. Lastly, incorporating random embedding sampling further boosts new accuracy with a negligible decrease in base accuracy. Overall, SAR improves CoOp's base accuracy from 82.35\% to 82.79\% (+0.44\%), new accuracy from 66.61\% to 73.75\% (+7.14\%), and the harmonic mean (H) from 73.65\% to 78.01\% (+4.36\%), demonstrating its effectiveness by improving all three metrics.
|
2502.11969
|
Experiments
| 228,294
| 37
|
22,512,239
| 4
|
\keypoint{Impact of Number of Novel Classes on SAR.}Impact of Number of Novel Classes on SAR.
We evaluate the impact of the number of novel classes on the effectiveness of SAR. To this end, we train prompts by applying SAR under two different numbers of novel classes: one using 50 novel classes and the other using 200 novel classes. In the 50-class setup, the classes are randomly selected from the set of 200 novel classes, and the embedding sampling size is fixed at 16. As the results in \Cref{fig:novelnum_ab} show, using a larger number of novel classes leads to greater performance improvements. With more novel classes, richer similarity information among semantically diverse classes can be obtained, enabling a more refined and accurate capture of the semantics of unseen classes. Notably, even when using only 50 novel classes for SAR, significant performance improvements over the baseline are observed.
|
2502.11969
|
Experiments
| 228,294
| 38
|
22,512,240
| 5
|
\keypoint{Impact of Regularization Weight $\lambda$ on SAR.}Impact of Regularization Weight $\lambda$\lambda on SAR.
We evaluate the impact of the regularization weight $\lambda$\lambda on performance. The experimental results under 4-shot setting are presented in \Cref{fig:weight_ab}. The left plot in the figure shows the variations in base accuracy, new accuracy, and their harmonic mean (H) as $\lambda$\lambda changes. Increasing $\lambda$\lambda up to 1.0 leads to a significant improvement in new accuracy, showing a clear association between SAR's mechanism and the enhancement of generalization performance to new classes. However, when $\lambda$\lambda becomes excessively large (\eg 1.5), all accuracies slightly decline due to the effects of overly strong regularization. The right plot in \Cref{fig:weight_ab} depicts the average SAR loss at the final epoch of the model, averaged across all datasets. As $\lambda$\lambda increases, SAR loss decreases, which corresponds to the improvement in new accuracy observed in the left plot.
|
2502.11969
|
Experiments
| 228,294
| 39
|
22,512,241
| 6
|
\keypoint{Resource Costs for Introducing SAR.}Resource Costs for Introducing SAR.
We report the additional resource requirements incurred by SAR on ImageNet in \Cref{tab:incurred_resource}. Since SAR does not modify the model architecture, it does not introduce any extra trainable parameters. Moreover, as SAR operates during training, it has no impact on inference time. However, implementing SAR requires computing text embeddings not only for the base classes but also for the novel classes. This additional computation in the text encoder, along with the computation of similarity among classes, increases memory usage and training time. Furthermore, additional training time increases with the number of update steps, as text embeddings are computed at every update. Among the 11 datasets, ImageNet has the largest training set, resulting in the highest number of update steps. Consequently, the additional training time is notably smaller for the other 10 datasets. The number of novel classes for SAR can be chosen by considering the tradeoff between resource cost and performance, as the additional memory usage and training time incurred by SAR grow with the number of novel classes.
|
2502.11969
|
Experiments
| 228,294
| 40
|
22,536,845
| 2
|
Furthermore, a useful mechanism for creating and validating intricate analytical workflows is frequently absent from current systems. While code snippets and simple visualizations can be produced by generative AI, significant human intervention is usually necessary to ensure these outputs’ accuracy, effectiveness, and utility \cite{he2016deep, hochreiter1997long}. This restricts the use of generative AI in fields where accuracy and in-depth research are essential. The time-consuming and error-prone nature of the manual review and correction process further reduces the effectiveness of generative AI applications.
|
2502.13164
|
Introduction
| 229,212
| 3
|
22,512,242
| 7
|
\subsection{Base-to-New Generalization}
In the Base-to-New generalization task, we evaluate the performance of models with and without SAR in 16-shot setting across 11 datasets. The classes in each dataset are divided into two groups, where prompts trained on one group (base classes) are used to evaluate performance on the other group (new classes). As summarized in \Cref{tab:b2n}, SAR consistently improves the new class accuracy and the harmonic mean (H) across all baselines. For CoOp and MaPLe, SAR also delivers significant improvements in base class accuracy, while showing no critical base-new accuracy trade-offs for other baselines. Notably, SAR yields the largest improvements for CoOp, while also bringing meaningful improvements for the baselines equipped with regularization techniques designed to improve generalization. These results suggest that SAR effectively addresses gaps that cannot be resolved by existing regularization methods of the baselines. SAR brings marginal improvements over TCP, which can be attributed to the fact that TCP, by generating class-aware prompts, already captures the semantics of new classes effectively (as shown in \Cref{fig:motivation}), thereby reducing the necessity for SAR's contribution. Similarly, the limited improvements by SAR on the Caltech101 and OxfordPets datasets are likely because all baselines already generalize well on these datasets. This claim is supported by the observation that all baselines exhibit fairly high new accuracy on these datasets, as shown in \Cref{tab:b2n}, leaving little room for SAR to provide further improvement.
|
2502.11969
|
Experiments
| 228,294
| 41
|
22,512,243
| 8
|
\subsection{Effectiveness with Alternative Word Sources}
We investigate whether SAR remains effective when using novel classes extracted from sources other than LLMs. To this end, we train SAR with novel classes from (1) 200 randomly sampled nouns from WordNet \cite{miller1995wordnet}, a large lexical database, and (2) new class names from the dataset. The base-to-new generalization results for CoOp and CoPrompt are presented in \Cref{fig:wordsource_ab}. A key observation is that new accuracy improves significantly even when randomly sampled words are used as novel classes. This suggests that SAR is robust to the choice of novel classes and consistently enhances generalization to unseen classes. Notably, using words generated by ChatGPT-4o as novel classes leads to even greater performance improvements. This can be attributed to the fact that these words are more semantically aligned with the new classes in the dataset than randomly sampled words, making them more effective at capturing the semantics of new classes (Examples of generated words are listed in \underline{\textbf{Supp. Material C}}). However, using the actual new class names from the dataset as novel classes (SAR w/ `Oracle') does not yield the highest performance improvements. This is primarily due to the small number of them (on average, less than 100), which reduces the semantic diversity available for SAR and ultimately makes it less effective to capture the fine-grained semantics of them.
|
2502.11969
|
Experiments
| 228,294
| 42
|
22,512,244
| 9
|
\subsection{Effectiveness in Few Shot Setting}
To validate the effectiveness of SAR in a few-shot setting, we conduct base-to-new generalization experiments under 4-shot setting for CoOp, TCP, and CoPrompt. In the case of CoOp, the models show a stronger tendency to increase SAR loss during prompt tuning compared to the 16-shot setting. This may be due to the model attempting to capture spurious correlations in the limited training data, leading to overfitting. To address this issue, we set the regularization weight to 1.0 for CoOp across all datasets. For TCP and CoPrompt, the same regularization weights as those used in the 16-shot setting are applied to each dataset. The performance averaged across 11 datasets is summarized in \Cref{tab:4shot_b2n}, demonstrating that SAR remains consistently effective even in the few-shot setting.
|
2502.11969
|
Experiments
| 228,294
| 43
|
22,512,245
| 10
|
\subsection{Domain Generalization}
Domain generalization evaluates how well a model can generalize to domains with data distributions that differ from its training domain. Following convention, we use ImageNet as the source domain and evaluate the performance of the trained model on ImageNet-Sketch, ImageNet-A, ImageNet-R, and ImageNet-V2. The results, summarized in \Cref{tab:domain_gen}, show that SAR consistently improves the average target accuracy across the baselines.
|
2502.11969
|
Experiments
| 228,294
| 44
|
22,512,247
| 1
|
{
\small
\bibliographystyle{ieeenat_fullname}
\bibliography{main}
}
\small
\bibliographystyle{ieeenat_fullname}ieeenat_fullname
\bibliography{main}
|
2502.11969
|
Conclusion and Limitation
| 228,295
| 46
|
22,512,248
| 2
|
\clearpage
\setcounter{page}{1}
\maketitlesupplementary
|
2502.11969
|
Conclusion and Limitation
| 228,295
| 47
|
22,512,249
| 3
|
\definecolor{codegreen}{rgb}{0,0.6,0}
\definecolor{codegray}{rgb}{0.5,0.5,0.5}
\definecolor{codepurple}{rgb}{0.58,0,0.82}
\definecolor{backcolour}{rgb}{0.95,0.95,0.92}
|
2502.11969
|
Conclusion and Limitation
| 228,295
| 48
|
22,512,250
| 0
|
\begin{lstlisting}[
language={},
label=lst:gpt,
backgroundcolor=\color{backcolour},
commentstyle=\color{codegreen},
keywordstyle=\color{magenta},
numberstyle=\tiny\color{codegray},
stringstyle=\color{codepurple},
basicstyle=\ttfamily\footnotesize,
breakatwhitespace=false,
breaklines=true,
keepspaces=true,
numbers=none,
numbersep=5pt,
showspaces=false,
showstringspaces=false,
showtabs=false,
tabsize=2,
frame=single,
rulecolor=\color{black}
]
You are a helpful assistant that identifies the common, specific concept among a given set of words and provides additional words that fit within that concept. For example, given ["rose," "daisy," "tulip"], you might infer the concept of "plants," with a more specific focus on "flowers." Once you've identified the concept, generate exactly 200 words that belong to it.
|
2502.11969
|
Prompt for ChatGPT-4o
| 228,296
| 49
|
22,512,252
| 2
|
Here are the my words: [`Annual Crop Land', 'Forest', 'Herbaceous Vegetation Land', 'Highway or Road', 'Industrial Buildings']
\end{lstlisting}\begin{lstlisting}[
language={},
label=lst:gpt,
backgroundcolor=\color{backcolour},
commentstyle=\color{codegreen},
keywordstyle=\color{magenta},
numberstyle=\tiny\color{codegray},
stringstyle=\color{codepurple},
basicstyle=\ttfamily\footnotesize,
breakatwhitespace=false,
breaklines=true,
keepspaces=true,
numbers=none,
numbersep=5pt,
showspaces=false,
showstringspaces=false,
showtabs=false,
tabsize=2,
frame=single,
rulecolor=\color{black}
]
You are a helpful assistant that identifies the common, specific concept among a given set of words and provides additional words that fit within that concept. For example, given ["rose," "daisy," "tulip"], you might infer the concept of "plants," with a more specific focus on "flowers." Once you've identified the concept, generate exactly 200 words that belong to it.
|
2502.11969
|
Prompt for ChatGPT-4o
| 228,296
| 51
|
22,512,254
| 4
|
Here are the my words: [`Annual Crop Land', 'Forest', 'Herbaceous Vegetation Land', 'Highway or Road', 'Industrial Buildings']
\end{lstlisting}[
language={},
label=lst:gpt,
backgroundcolor=\color{backcolour},
commentstyle=\color{codegreen},
keywordstyle=\color{magenta},
numberstyle=\tiny\color{codegray},
stringstyle=\color{codepurple},
basicstyle=\ttfamily\footnotesize,
breakatwhitespace=false,
breaklines=true,
keepspaces=true,
numbers=none,
numbersep=5pt,
showspaces=false,
showstringspaces=false,
showtabs=false,
tabsize=2,
frame=single,
rulecolor=\color{black}
]
You are a helpful assistant that identifies the common, specific concept among a given set of words and provides additional words that fit within that concept. For example, given ["rose," "daisy," "tulip"], you might infer the concept of "plants," with a more specific focus on "flowers." Once you've identified the concept, generate exactly 200 words that belong to it.
|
2502.11969
|
Prompt for ChatGPT-4o
| 228,296
| 53
|
22,512,255
| 5
|
The new words should be as semantically distinct from each other as possible while staying relevant to the shared concept. Aim for diversity within the category to showcase a broad range of examples. Please return the words in a Python list format.
|
2502.11969
|
Prompt for ChatGPT-4o
| 228,296
| 54
|
22,512,257
| 7
|
We provide the prompt given to ChatGPT-4o for generating novel classes for EuroSAT. The prompt is designed based on the templates from \cite{anperceptionclip}. For other datasets, only the list of base classes is modified.
|
2502.11969
|
Prompt for ChatGPT-4o
| 228,296
| 56
|
22,512,258
| 0
|
We use the prompt templates provided in the implementation of CoOp \cite{zhou2022learning} for prompt ensembling. For each dataset, a total of eight prompts are ensembled, comprising seven general prompts and one dataset-specific custom prompt.
|
2502.11969
|
Prompts for Ensembling
| 228,297
| 57
|
22,512,259
| 0
|
We compare the words generated by ChatGPT-4o with the actual new classes in the dataset. This qualitative comparison suggests that the words generated by ChatGPT-4o are semantically aligned with the new classes, making them suitable candidates for unseen classes.
|
2502.11969
|
Word Examples
| 228,298
| 58
|
22,512,270
| 5
|
Abstract Meaning Representation \citep[AMR;][]{banarescu-etal-2013-abstract}
on which UMR is based, has seen success and adoption by the broader NLP community, and this is in large part due to the substantial efforts made towards high-quality text-to-AMR parsing and AMR-to-text generation models \citep{sadeddine-etal-2024-survey}.
\footnote{The breadth of parsing and generation work for AMR is shown in the \href{https://nert-nlp.github.io/AMR-Bibliography/}{AMR Bibliography}.}
This progress on parsing and generation is integral to being able to incorporate AMR graphs into downstream applications and studies of language \citep{wein-opitz-2024-survey}. Thus, in order to see similar success for UMR, efforts towards UMR-to-text generation and text-to-UMR parsing are critical.
|
2502.11973
|
Introduction
| 228,299
| 6
|
22,512,272
| 6
|
In this work, we leverage the recent release of human-annotated UMR data (detailed in \Cref{ssec:data})
to tackle UMR-to-text generation and introduce the first UMR-to-text generation models. We investigate generation from the six languages included in the UMR v1.0 dataset: English, Chinese, Sanapaná, Arápaho, Kukama, and Navajo.
Our approaches and contributions include:
\begin{itemize}
\addtolength\itemsep{-3mm}
\item A baseline analysis as to how well six AMR-to-text generation models generate text from UMR out-of-the-box (\Cref{sec:baseline}).
% \ri{A baseline analysis as to how well six AMR-to-text generation models generate **text** from UMR.}
\item A novel pipelined approach to UMR-to-text generation, which converts UMR graphs into AMRs, then uses them as input to AMR-to-text generation models (\Cref{sec:conversion}).
\item
% \sw{split into two bullet points via new section 7 fine-tuned llms}
Seven fine-tuned UMR-to-text generation models, three using pretrained large language models and four pretrained AMR-to-text generation models as bases, demonstrating the potential benefits of leveraging AMR graphs for UMR tools (\Cref{sec:fine-tuned}).
% \item Thorough human evaluations for English and Chinese with respect to generated text fluency and adequacy, as well as automatic metric evaluations for all six languages, resulting in a model ranking and assertion of our best-performing model (\Cref{ssec:evaluation}).
\end{itemize}\begin{itemize}
\addtolength\itemsep{-3mm}
\item A baseline analysis as to how well six AMR-to-text generation models generate text from UMR out-of-the-box (\Cref{sec:baseline}).
% \ri{A baseline analysis as to how well six AMR-to-text generation models generate **text** from UMR.}
\item A novel pipelined approach to UMR-to-text generation, which converts UMR graphs into AMRs, then uses them as input to AMR-to-text generation models (\Cref{sec:conversion}).
\item
% \sw{split into two bullet points via new section 7 fine-tuned llms}
Seven fine-tuned UMR-to-text generation models, three using pretrained large language models and four pretrained AMR-to-text generation models as bases, demonstrating the potential benefits of leveraging AMR graphs for UMR tools (\Cref{sec:fine-tuned}).
% \item Thorough human evaluations for English and Chinese with respect to generated text fluency and adequacy, as well as automatic metric evaluations for all six languages, resulting in a model ranking and assertion of our best-performing model (\Cref{ssec:evaluation}).
\end{itemize}
\addtolength\itemsep{-3mm}-3mm
\item A baseline analysis as to how well six AMR-to-text generation models generate text from UMR out-of-the-box (\Cref{sec:baseline}).
\item A novel pipelined approach to UMR-to-text generation, which converts UMR graphs into AMRs, then uses them as input to AMR-to-text generation models (\Cref{sec:conversion}).
\item
Seven fine-tuned UMR-to-text generation models, three using pretrained large language models and four pretrained AMR-to-text generation models as bases, demonstrating the potential benefits of leveraging AMR graphs for UMR tools (\Cref{sec:fine-tuned}).
|
2502.11973
|
Introduction
| 228,299
| 7
|
22,512,274
| 0
|
UMR is based on Abstract Meaning Representation (AMR), which was designed for English
\citep{banarescu-etal-2013-abstract} but has since seen various cross-lingual extensions and applications \citep{10.1162/coli_a_00503}.
|
2502.11973
|
Uniform Meaning Representation
| 228,300
| 8
|
22,512,275
| 1
|
Similarly to AMR, UMR annotations are rooted, directed graphs that capture the meaning of a sentence \citep{van2021designing}.
UMR incorporates aspect and modality at the sentence-level and additionally includes document-level graphs, enabling annotation of coreferential relations.
Alignments between the coreferential elements are also provided.
|
2502.11973
|
Uniform Meaning Representation
| 228,300
| 9
|
22,512,276
| 2
|
In order to ensure that UMR could reflect meaning for many languages, the UMR schema is flexible in its annotation while ensuring consistency across languages.
At the sentence-level, UMR accounts for linguistic diversity across languages through the use of a lattice-like annotation schema \citep{van-gysel-etal-2019-cross}. This enables annotators to choose a more coarse-grained versus a more fine-grained annotation for numerous annotation categories (including discourse relations, modality, number, spatial relations, aspect, and temporality) based on the features of the language being annotated.
Particular care is also given to the annotation of low- or no-resource languages \citep{vigus-etal-2020-cross}, as ``Stage 0'' annotation is performed for languages that do not have existing rolesets available. Stage 0 annotation enables annotators to establish predicate-argument structures and develop rolesets while performing UMR annotation.
|
2502.11973
|
Uniform Meaning Representation
| 228,300
| 10
|
22,512,277
| 3
|
UMR successfully accommodates the multilingual issues addressed in individual language adaptations of AMR, showing its promise as a multilingual representation \citep{wein-bonn-2023-comparing}. Its effectiveness at capturing meaning across languages was also indicated in the pilot annotation of UMR in four indigenous languages \citep{van-gysel-etal-2021-theoretical}.
|
2502.11973
|
Uniform Meaning Representation
| 228,300
| 11
|
22,512,278
| 4
|
\Citet{chun-xue-2024-pipeline} released a text-to-UMR parser, which produces the document-level UMR graph based on the contents of the sentence, and the sentence-level UMR graph by running existing text-to-AMR parsers and then converting the AMR into a UMR.\footnote{This pipelined parsing approach mirrors our baseline approach, in the reverse direction.}
|
2502.11973
|
Uniform Meaning Representation
| 228,300
| 12
|
22,512,281
| 1
|
Our baseline approach uses six AMR-to-text generation models, passing the sentence-level UMR graphs as input.
Next, using the same six models used for our baseline, we developed a pipeline approach (\Cref{sec:conversion}) to UMR-to-text generation, which involved first converting the UMR graphs to AMR graphs, and then passing the converted AMR graphs as input to the models.
Then, based on the baseline performance of the AMR-to-text generation models, we fine-tuned the top performing models as well as three large language models (with no AMR input) on the UMR training set (\Cref{sec:fine-tuned}).
|
2502.11973
|
Methodology
| 228,301
| 15
|
22,512,282
| 2
|
\subsection{Data}
\label{ssec:data}
We used the first release of UMR data \citep{bonn-etal-2024-building-broad}, which contains annotations in six languages: English, Chinese, and four languages indigenous to the Americas: Arápaho, Navajo, Sanapaná, and Kukama.
The English data contains
LORELEI news text and a description of a silent film. The Chinese data consists of wikinews sentences.
Arápaho, Navajo, and Kukama annotations all represent narrative documents, while it is not clear what genre the Sanapaná annotations are \citep{bonn-etal-2024-building-broad}.
We first took out the English UMR data that contains equivalent AMRs in the AMR3.0 dataset \citep{knight_amr_3.0}, as to avoid leakage and unfair evaluations of models which contain the corresponding AMR data. This consisted of 66 English UMRs.
Then, we divided it into training and testing sets.
|
2502.11973
|
Methodology
| 228,301
| 16
|
22,512,283
| 3
|
Not all annotations contained sentence-level and document-level graphs.
This is because some annotations without document-level data contained alignment data or could be referenced in other document-level annotations.
Our final data split was 70\% for training, 10\% for dev, and 20\% for testing (the number of sentences for each language is shown in \Cref{tab:split}).
|
2502.11973
|
Methodology
| 228,301
| 17
|
22,512,285
| 5
|
We compare the generated text from each of our approaches against the references (the ground-truth sentences that were annotated) by using BERTscore \citep{zhang2019bertscore}, given its previously evidenced correlation with human judgments for AMR-to-English text generation \citep{manning-etal-2020-human}. Specifically, we compare sentence similarity using multilingual BERTscore. We also use METEOR \citep{banerjee-lavie-2005-meteor} and BLEU \citep{papineni2002bleu} to enable ease of comparison with future work.
|
2502.11973
|
Methodology
| 228,301
| 19
|
22,512,286
| 6
|
\subsection{Indigenous Language Evaluation}
During initial experimentation, we saw primarily positive quantitative indications as far as the ability of our models to produce text in the four indigenous languages,
for example, multilingual BERTscores of 0.799 for Navajo, 0.816 for Sanapaná, 0.780 for Arápaho, and 0.673 for Kukama, as generated by amrlib trained on all languages' sentence-level data.
|
2502.11973
|
Methodology
| 228,301
| 20
|
22,512,287
| 7
|
However, given the fact that these languages are extremely low-resource and are not likely to be well-evaluated by BERTscore, METEOR, and BLEU,
we consulted with speakers of Arápaho and Navajo to provide a human evaluation of the generated text.
We provided the output from the amrlib sentence-level model fine-tuned on all languages, as the quantitative indications of the text quality were positive. Our Arápaho and Navajo speakers indicated that, while
these fine-tuned models did do a fairly good job at imitating the script of the four indigenous languages, the output nonetheless was nonsensical and ungrammatical. As such, we opted against a full human evaluation for the four indigenous languages, with the understanding that even our top-performing models were outputting gibberish for these languages. This is likely because these four languages are all extremely low-resource and exhibit morphological complexity that would require additional resources for coherent text generation.
|
2502.11973
|
Methodology
| 228,301
| 21
|
22,512,288
| 8
|
As a result, we moved forward with evaluating our approaches and models exclusively in English and Chinese (both with regard to automatic evaluation and human evaluation). We still leveraged the indigenous language data in the UMR splits (as indicated in \Cref{tab:split}) for our model fine-tuning (\Cref{sec:fine-tuned}).
|
2502.11973
|
Methodology
| 228,301
| 22
|
22,512,289
| 9
|
\subsection{Human Evaluation}
In order to validate the quantitative results obtained by the automatic metrics, we perform a human evaluation. Six college students participated in the evaluation of the English and Chinese texts, who were native speakers of English and Chinese accordingly.
Each annotator judged fluency and adequacy on a scale from 1-4. Fluency was judged first, without exposure to the reference, and then adequacy was judged in relation to the reference.
|
2502.11973
|
Methodology
| 228,301
| 23
|
22,512,290
| 10
|
There were a total of four surveys: English fluency, English adequacy, Chinese fluency, and Chinese adequacy, each of which contained 25 questions. For English, we chose to exclude the 5 shortest sentences, which left us with a final set of 25 sentences. For Chinese, we randomly selected the sentences.
Each question in the English fluency and adequacy surveys displayed six sentences, with five being from each of our top-performing models from each approach (with two models included for some approaches), as well as the reference.
The instructions provided to the human raters for the English fluency and adequacy surveys can be seen in \Cref{fig:fluency_instructions,fig:adequacy_instructions}.\footnote{The Chinese instructions are identical to the English instructions but include the word ``Chinese'' instead of ``English.''}
|
2502.11973
|
Methodology
| 228,301
| 24
|
22,512,291
| 11
|
We evaluated the inter-annotator agreement using the Pearson correlation coefficient \citep{Pearson1895}, calculating pairwise agreement for each of the four surveys (Chinese and English, fluency and adequacy for each).
The average correlation coefficient for English fluency was 0.72, for English adequacy was 0.78, for Chinese fluency was 0.55, and for Chinese adequacy was 0.64. This indicates that there is a strong correlation between the annotators' scores, validating the judgments, while the English raters exhibited even greater agreement.\footnote{This is likely due to the generated text being both more fluent and more adequate for English than Chinese (see \Cref{ssec:pipe_results,ssec:fine_results}).
% \sw{need to include the section 7 sec results here}
}
Note also that all pair-wise correlation coefficients had a statistically significant p-value (p < 0.05).
|
2502.11973
|
Methodology
| 228,301
| 25
|
22,512,292
| 0
|
\label{sec:baseline}
|
2502.11973
|
Baseline Approach
| 228,302
| 26
|
22,512,293
| 1
|
Given the similarity of UMR to AMR and the prevalence of AMR technologies, we use AMR-to-text generation models
out-of-the-box as a baseline model for sentence-level UMR graphs,
to see how they perform as a zero-shot approach with no exposure to UMR.
We elect to perform the baseline experimentation on sentence-level UMR graphs given that the AMR-to-text generation models are not designed to handle document-level data (as AMR does not contain document-level information).
|
2502.11973
|
Baseline Approach
| 228,302
| 27
|
22,536,846
| 3
|
Further complexity is added when multi-agent systems are integrated into generative AI frameworks. Conventional single-agent systems are frequently constrained by their incapacity to oversee and optimize several tasks at once. Comparatively, multi-agent systems are able to divide up the work and work together to produce more accurate and efficient outcomes. Coordinating several agents, ensuring smooth communication, and keeping the system cohesive are still difficult tasks.
|
2502.13164
|
Introduction
| 229,212
| 4
|
22,536,847
| 4
|
To address these challenges, we propose a novel framework for generative AI that leverages a multi-agent actor-critic model to enhance query resolution and data analysis. Our system consists of three primary generative AI agents: the Actor Generative AI, the Critic Generative AI, and the Expert Analysis Generative AI, each performing distinct roles to ensure accurate and efficient query handling.
|
2502.13164
|
Introduction
| 229,212
| 5
|
22,512,294
| 2
|
\subsection{Models}
We apply this approach to six AMR-to-text generation models:
\begin{itemize}
\addtolength\itemsep{-3mm}
\item[1.] amrlib\footnote{ \href{https://github.com/bjascob/amrlib/tree/master}{amrlib GitHub Repository}}: generates text using a pretrained Sequence-to-Sequence T5 model, trained on AMR3.0 \citep{knight_amr_3.0}.
%\ejm{@emma do you happen to know which dataset this was trained on?}.
\item[2.] AMRBART \citep{bai-etal-2022-graph}: BART-based \citep{lewis-etal-2020-bart} model
% that generates text following a Transformer encoder-decoder architecture,
pretrained on English text and AMR graphs from AMR2.0 \citep{knight_amr_2.0} and AMR3.0, pretrained on 200k silver AMRs parsed by SPRING.
%\ejm{@emma do you happen to know which dataset this was trained on?}.
\item[3-4.] SPRING2 and SPRING3 \citep{bevilacqua-etal-2021-one}: BART-based sequence-to-sequence model that simplifies AMR parsing and generation by treating them as symmetric tasks.
% Trained on AMR2.0, AMR3.0, and BioAMR \citep{garg2016extracting}.
SPRING2 and SPRING3 differ in their training datasets: SPRING2 is trained on AMR2.0 while SPRING3 is trained on AMR3.0.
% \finalversion{we need to talk to rei about this for the final version.}
% is trained on the AMR2.0 dataset, while SPRING3 is trained on the AMR3.0 dataset.
\item[5.] BiBL \citep{cheng-etal-2022-bibl}: utilizes the architectural framework of SPRING to align AMR graphs and text, in order to share information across the parsing and generation tasks trained on AMR2.0 and AMR3.0.
\item[6.] Smelting \citep{ribeiro-etal-2021-smelting}: trained mT5 \citep{xue2021mt5} on gold English AMR graphs and sentences as well as silver machine-translated sentences.
\end{itemize}\begin{itemize}
\addtolength\itemsep{-3mm}
\item[1.] amrlib\footnote{ \href{https://github.com/bjascob/amrlib/tree/master}{amrlib GitHub Repository}}: generates text using a pretrained Sequence-to-Sequence T5 model, trained on AMR3.0 \citep{knight_amr_3.0}.
%\ejm{@emma do you happen to know which dataset this was trained on?}.
\item[2.] AMRBART \citep{bai-etal-2022-graph}: BART-based \citep{lewis-etal-2020-bart} model
% that generates text following a Transformer encoder-decoder architecture,
pretrained on English text and AMR graphs from AMR2.0 \citep{knight_amr_2.0} and AMR3.0, pretrained on 200k silver AMRs parsed by SPRING.
%\ejm{@emma do you happen to know which dataset this was trained on?}.
\item[3-4.] SPRING2 and SPRING3 \citep{bevilacqua-etal-2021-one}: BART-based sequence-to-sequence model that simplifies AMR parsing and generation by treating them as symmetric tasks.
% Trained on AMR2.0, AMR3.0, and BioAMR \citep{garg2016extracting}.
SPRING2 and SPRING3 differ in their training datasets: SPRING2 is trained on AMR2.0 while SPRING3 is trained on AMR3.0.
% \finalversion{we need to talk to rei about this for the final version.}
% is trained on the AMR2.0 dataset, while SPRING3 is trained on the AMR3.0 dataset.
\item[5.] BiBL \citep{cheng-etal-2022-bibl}: utilizes the architectural framework of SPRING to align AMR graphs and text, in order to share information across the parsing and generation tasks trained on AMR2.0 and AMR3.0.
\item[6.] Smelting \citep{ribeiro-etal-2021-smelting}: trained mT5 \citep{xue2021mt5} on gold English AMR graphs and sentences as well as silver machine-translated sentences.
\end{itemize}
\addtolength\itemsep{-3mm}-3mm
\item[1.] amrlib\footnote{ \href{https://github.com/bjascob/amrlib/tree/master}{amrlib GitHub Repository}}: generates text using a pretrained Sequence-to-Sequence T5 model, trained on AMR3.0 \citep{knight_amr_3.0}.
\item[2.] AMRBART \citep{bai-etal-2022-graph}: BART-based \citep{lewis-etal-2020-bart} model
pretrained on English text and AMR graphs from AMR2.0 \citep{knight_amr_2.0} and AMR3.0, pretrained on 200k silver AMRs parsed by SPRING.
\item[3-4.] SPRING2 and SPRING3 \citep{bevilacqua-etal-2021-one}: BART-based sequence-to-sequence model that simplifies AMR parsing and generation by treating them as symmetric tasks.
SPRING2 and SPRING3 differ in their training datasets: SPRING2 is trained on AMR2.0 while SPRING3 is trained on AMR3.0.
\item[5.] BiBL \citep{cheng-etal-2022-bibl}: utilizes the architectural framework of SPRING to align AMR graphs and text, in order to share information across the parsing and generation tasks trained on AMR2.0 and AMR3.0.
\item[6.] Smelting \citep{ribeiro-etal-2021-smelting}: trained mT5 \citep{xue2021mt5} on gold English AMR graphs and sentences as well as silver machine-translated sentences.
|
2502.11973
|
Baseline Approach
| 228,302
| 28
|
22,512,296
| 3
|
Amrlib, AMRBART, SPRING2, SPRING3 and BiBL are all trained to generate English text, while Smelting can produce Spanish, Italian, German, and Chinese text.
Thus, as a baseline, we run the first five models on the English UMR graphs (to generate into English), and run Smelting on the Chinese UMR graphs (to generate into Chinese).
|
2502.11973
|
Baseline Approach
| 228,302
| 29
|
22,512,297
| 4
|
\subsection{Results}
\label{ssec:base_results}
|
2502.11973
|
Baseline Approach
| 228,302
| 30
|
22,512,298
| 5
|
The automatic metric results for the English and Chinese baseline models were highly comparable, a promising indication of the utility of AMR-to-text generation tools for UMR-to-text generation.
|
2502.11973
|
Baseline Approach
| 228,302
| 31
|
22,512,299
| 6
|
Out-of-the-box, the English results from the five baseline models (\Cref{tab:auto-eval-english}) all achieved multilingual BERTscores of around 0.7 (ranging from 0.681-0.704). The BLEU and METEOR scores were noticeably lower.
The Chinese baseline results (\Cref{tab:auto-eval-chinese}) were similarly high, with the baseline Smelting model achieving a multilingual BERTscore of 0.716.
|
2502.11973
|
Baseline Approach
| 228,302
| 32
|
22,512,300
| 7
|
Our qualitative analysis of the baseline models revealed that it was very common to see the inclusion of UMR-specific terms such as \texttt{refer-number singular}, \texttt{full-affirmative}, \texttt{umr-unknown}, and \texttt{3rd person} in the text output. Examples of this can be seen in the following generated sentence:
\noindent \texttt{Full-affirmative, though, the first singular thought that the second singular would not see the apron at first, is a state of 'full affirmative'.}
The only UMR graphs that were converted to highly fluent and adequate text were one-word sentences such as ``ok...'' or ``anyway...''.
We also noted that the generated output from the UMR graphs tended to be much longer than the reference output, which was likely due to the inclusion of the UMR-specific terms mentioned previously.\footnote{The average sentence length for each model was 11.7 words for amrlib, 13.6 for AMRBART, 14.6 for SPRING3, 14.7 for SPRING2, and 11.8 for BiBL.}
|
2502.11973
|
Baseline Approach
| 228,302
| 33
|
22,512,301
| 8
|
\caption{English automatic evaluation Results.
BERT: multilingual BERTscore, EN: fine-tuned only on English, Sent: sentence-level, Doc: document-level, All: trained on all languages. Bolded entries were selected for the human evaluation. Due to having trained over 50 individual models, here we show a representative sample of the best performers from each approach.
%\ejm{@emma to do fill in these scores, check all existing scores, and change all to 3 significant digits}
}
\label{tab:auto-eval-english}
|
2502.11973
|
Baseline Approach
| 228,302
| 34
|
22,512,302
| 9
|
\caption{Chinese - Automatic Evaluation Results. BERT: multilingual BERTscore, ZH: Chinese fine-tuning, Sent: sentence-level, Doc: document-level, All: trained on all languages. Bolded entries were selected for the human evaluation, Smelt: Smelting model.
Due to the large number of models trained, here we show a representative sample of the best performers from each approach.
%\ejm{@emma to do fill in these scores, check all existing scores, and change all to 3 significant digits}
}
\label{tab:auto-eval-chinese}
|
2502.11973
|
Baseline Approach
| 228,302
| 35
|
22,512,303
| 10
|
Based on the shorter output length and higher perceived fluency in our initial qualitative analysis, we determined that BiBL was the best baseline English model, and Smelting was our only baseline Chinese model.
Thus, we included the baseline BiBL and Smelting models in the human evaluation survey.
The human evaluation scores (\Cref{tab:human-eval-english} and \Cref{tab:human-chinese}) were low with regard to both fluency and adequacy, in spite of the fairly high BERTscore values. Baseline BiBL for English had a fluency score of 1.44 (scale of 1-4) and an adequacy score of 1.37, while baseline Smelting for Chinese received a fluency score of 1.69 and an adequacy score of 1.48. The English and Chinese reference sentences received a fluency score of 3.35 and 3.27, respectively, showing that even the ground-truth sentences were not perceived as perfectly fluent by the adjudicators, perhaps due to the narrative structure of the UMR data.
|
2502.11973
|
Baseline Approach
| 228,302
| 36
|
22,512,305
| 1
|
Our next approach again leveraged existing AMR-to-text generation models, but in this case first converting the UMR graph into an AMR.
|
2502.11973
|
Pipelined Approach
| 228,303
| 38
|
22,512,306
| 2
|
\subsection{Conversion Process}
In order to convert UMR graphs into AMR, we designed a rule-based conversion process.
When converting UMR graphs into AMR graphs, some elements of UMR graphs do not appear in AMR (such as aspect and mode) and were thus simply removed.
Other changes included converting split roles, renaming roles, and introducing additional roles. This process was informed by \citet{bonn-etal-2023-mapping} as well as the AMR guidelines \citep{AMRGuidelines} and the UMR guidelines
\citep{UMRGuidelines}.
The ``person'' concepts (\texttt{:refer-number} or \texttt{refer-person}) and pronouns generally are handled in a more complicated way in UMR than AMR, so in order to convert these concepts into their AMR counterparts, we changed all ``person'' concepts to a single node of the equivalent English pronoun.
We converted third-person pronouns, both singular and plural, to ``they,''
opting for a gender-neutral approach.
Additionally, if a graph referred to the first person but did not include a refer-number tag, we defaulted to using the singular pronoun ``I.''
The Sanapaná data also included a number of instances of \texttt{:wiki}, which we removed.\footnote{We also chose to map \texttt{:material} to \texttt{:consist-of}, \texttt{:concessive-condition} to \texttt{:condition}, and \texttt{have-91} to \texttt{have-03}.}
|
2502.11973
|
Pipelined Approach
| 228,303
| 39
|
22,512,307
| 3
|
In order to verify the validity of the conversion process itself, we leveraged the 66 held-out English UMRs which contained equivalent AMRs in the AMR3.0 dataset (detailed in \Cref{ssec:data}), and compared the generated graphs to their equivalent human-annotated AMR graphs using SMATCH. The average Smatch score for our conversion process was 0.63, indicating a fairly accurate conversion process. One of the ways in which our rule-based approach to conversion struggled to meet AMR norms was that substantial additional information is contained in the UMR, which is not removed during conversion, such as in the case of the sentence ``Pleasure:''\\
|
2502.11973
|
Pipelined Approach
| 228,303
| 40
|
22,512,308
| 4
|
\noindent Original AMR:
\begin{small}
\vspace{-1mm}
\begin{verbatim}
(p / pleasure)
\end{verbatim}
\end{small}\begin{small}
\vspace{-1mm}
\begin{verbatim}
(p / pleasure)
\end{verbatim}
\end{small}
\vspace{-1mm}
|
2502.11973
|
Pipelined Approach
| 228,303
| 41
|
22,512,309
| 5
|
\vspace{13mm}
\noindent Original UMR:
\begin{small}
\vspace{-1mm}
\begin{verbatim}
(s29s / say-01
:ARG0 (s29p / person)
:ARG1 (s29h / have-experience-91
:ARG1 s29p
:ARG2 (s29p3 / pleasure)
:ARG3 (s29t / thing)
:aspect state)
:ARG2 (s29p2 / person)
:aspect performance)
\end{verbatim}
\end{small}\begin{small}
\vspace{-1mm}
\begin{verbatim}
(s29s / say-01
:ARG0 (s29p / person)
:ARG1 (s29h / have-experience-91
:ARG1 s29p
:ARG2 (s29p3 / pleasure)
:ARG3 (s29t / thing)
:aspect state)
:ARG2 (s29p2 / person)
:aspect performance)
\end{verbatim}
\end{small}
\vspace{-1mm}
|
2502.11973
|
Pipelined Approach
| 228,303
| 42
|
22,512,310
| 6
|
\noindent Converted AMR:
\begin{small}
\vspace{-1mm}
\begin{verbatim}
(s29s / say-01
:ARG0 (s29p / person)
:ARG1 (s29h / have-experience-91
:ARG1 s29p
:ARG2 (s29p3 / pleasure)
:ARG3 (s29t / thing))
:ARG2 (s29p2 / person))
\end{verbatim}
\end{small}\begin{small}
\vspace{-1mm}
\begin{verbatim}
(s29s / say-01
:ARG0 (s29p / person)
:ARG1 (s29h / have-experience-91
:ARG1 s29p
:ARG2 (s29p3 / pleasure)
:ARG3 (s29t / thing))
:ARG2 (s29p2 / person))
\end{verbatim}
\end{small}
\vspace{-1mm}
|
2502.11973
|
Pipelined Approach
| 228,303
| 43
|
22,512,311
| 7
|
\noindent Generated text from BiBL of the converted AMR: ``People said it was a pleasurable experience.'' \\
|
2502.11973
|
Pipelined Approach
| 228,303
| 44
|
22,512,312
| 8
|
\caption{English - Human Evaluation Results. EN: English fine-tuning, Sent: sentence-level, Doc: document-level, All: trained on all languages in our datase}
\label{tab:human-eval-english}
|
2502.11973
|
Pipelined Approach
| 228,303
| 45
|
22,512,313
| 9
|
\caption{Chinese - Human Evaluation Results. ZH: Chinese fine-tuning, Sent: sentence-level, Doc: document-level, All: trained on all languages in our dataset, Smelt: Smelting model.}
\label{tab:human-chinese}
|
2502.11973
|
Pipelined Approach
| 228,303
| 46
|
22,512,315
| 11
|
Our pipelined approach outperformed the baseline approach for both English and Chinese.
For English, BiBL achieved the highest BERTscore, 0.784, and
for Chinese, Smelting achieved a BERTscore of 0.767.
|
2502.11973
|
Pipelined Approach
| 228,303
| 48
|
22,512,316
| 12
|
This approach proved to be very effective in reducing the amount of UMR terms that appeared in the output, which led to more comprehensible sentences.
The perceived fluency of this approach was validated by the adjudicators of the human evaluation survey. For English, we saw that this approach led to the most fluent output, with a score of 3.59 and an adequacy score of 2.71, which is a major improvement from the baseline approach. In addition, we saw that Chinese also had increased scores of 2.19 for fluency and 1.69 for adequacy. It is evident that this approach was more successful for English, but we did observe quantitative and qualitative improvements for both languages.
|
2502.11973
|
Pipelined Approach
| 228,303
| 49
|
22,512,317
| 13
|
While the pipelined approach makes use of AMR graphs as the input of the well-developed ecosystem of AMR technologies, it is also worth noting that UMR contains \emph{more} information than AMR does.
This would suggest that converting from UMR to AMR may result in less accurate generated text (with regard to the UMR-specific content, such as tense and aspect), than generating text directly from the UMR itself.
|
2502.11973
|
Pipelined Approach
| 228,303
| 50
|
22,512,319
| 1
|
\subsection{Methods}
We determined which AMR-to-text generation models to fine-tune based on the results from the \Cref{sec:baseline} and \Cref{sec:conversion}, picking the best-performing models for the baseline and pipelined approaches.
Accordingly, the first model we fine-tuned was amrlib, also
following \citet{wein-2022-spanish} (which fine-tuned the t5wtense AMR-to-English generation model for Spanish generation).
Next, we fine-tuned SPRING2 and Smelting, given their quantitative performances for English and Chinese, respectively.
Finally, we fine-tuned BiBL due to its high scores, relative brevity, and perceived fluency
(as discussed in \Cref{ssec:base_results}).
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 52
|
22,512,321
| 3
|
After having selected these models, we fine-tuned each model using the UMR data.
In order to determine whether document-level information and UMR data of other languages benefit UMR-to-text generation via fine-tuning, we created 8 different datasets:
(1)~sentence-level English data, (2)~sentence-level Chinese data, (3)~sentence-level English and Chinese data, (4)~sentence-level all languages, (5)~document-level English data, (6)~document-level Chinese data, (7)~document-level English and Chinese data, and (8)~document-level all languages.
This culminates in 8 fine-tuned versions of each of the 7 models; we highlight the key findings in \Cref{tab:auto-eval-english}.
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 54
|
22,512,322
| 4
|
For reproducibility, details of our hyperparameters follow.
For BiBL and SPRING, training occurred over 30 epochs with a constant learning rate of 0.0001, incorporating gradient clipping at 2.5 and dropout regularization at 0.25. The data is processed in batches of 8 with gradient accumulation over 16 steps. To maintain model stability and performance, we implemented a weight decay of 0.004 and restricted sequence length to 1024 tokens, while generation tasks utilize a beam size of 100 and a maximum generation length of 500 tokens.
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 55
|
22,512,323
| 5
|
We fine-tuned mBART-large-50 and MT5 models for 15 epochs (effectively 30 epochs since each training instance is processed twice through silversent and silveramr files) using Adafactor optimization (learning rate=0.0001), batch size 8, and 2-step gradient accumulation, with beam search generation and 1024 token limits. Similarly, we fine-tuned Gemma2-2b using the same optimization parameters but with mixed-precision training for memory efficiency, gradient checkpointing, and extended output sequence length (2048 tokens).
\subsection{Results}
\label{ssec:fine_results}
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 56
|
22,512,324
| 6
|
While the fine-tuning model scores varied more widely than the scores of the individual baseline and pipelined approaches, the best-performing fine-tuning models outperformed the baseline and the pipeline methods.
For English, BiBL and SPRING2 trained on English sentence-level data, as well as BiBL trained on English and Chinese sentence-level data, all received the highest BERTscores of 0.825.
SPRING2 trained on English document-level data achieved the highest BLEU score of 0.358, and BiBL trained on English and Chinese sentence-level data had a METEOR score of 0.601.
For Chinese, SPRING2 trained on Chinese document-level data achieved the highest BERTscore of 0.882. BiBL trained on English and Chinese sentence-level data had the highest BLEU score at 0.250 and the highest METEOR score of 0.602.
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 57
|
22,512,325
| 7
|
As suggested by the quantitative results, the fine-tuning approach resulted in the generation of highly fluent and accurate sentences, while also revealing unique patterns across all the models. After fine-tuning, amrlib was very successful in generating comprehensible English text along with producing the script for the four indigenous languages; however, it was unable to generate any Chinese text at all. BiBL produced remarkably intelligible text in English, the script of the four indigenous languages, and adequate but not fully fluent Chinese text. An interesting pattern found in BiBL's generated output was that it included ellipses and pauses. Because the models were trained on descriptions of a silent film and narrative data, the data contained breaks in dialogue that were annotated as ``--'' or ``.. '', which were replicated by BiBL.
SPRING produced exceptionally fluent and adequate text in English, replicated the four indigenous languages, and produced the most adequate Chinese text while being slightly more fluent than BiBL's Chinese output. SPRING was also among the models that captured pauses, but to a much lesser extent than BiBL.
The Smelting model was unable to produce meaningful output in English or the four indigenous languages, instead generating repetitive characters and words in various non-English languages like Spanish.
Smelting repeated characters and words for these languages, leading to further diminished fluency and accuracy.
However, after fine-tuning, it was more successful in producing exclusively Chinese script for Chinese UMRs.
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 58
|
22,512,326
| 8
|
The fine-tuned generation models consistently outperformed the fine-tuned LLMs across all three automatic metrics. Gemma's performance notably differed from the other two LLMs due to its architecture as a causal language model rather than a sequence-to-sequence model. The model would continue generating nonsensical tokens until it reached the pre-specified maximum length of 1024 tokens. This architectural difference helps explain its lower performance scores compared to mT5 and mBART. The fine-tuned models generally performed better at a specific language when they were only trained on that language's data versus being trained on all languages. This may be due to the ``curse of multilinguality:'' exposure to additional languages can degrade monolingual abilities \citep{conneau-etal-2020-unsupervised}.
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 59
|
22,512,327
| 9
|
The Chinese and English human evaluation scores for the fine-tuned models were also much higher than the baseline scores and reflect that fine-tuning the models can improve both fluency and adequacy. For English, the best fluency score for this approach was 3.44, which was achieved by SPRING being trained on English document-level data. Additionally, SPRING trained on English document-level data, achieved the highest adequacy score of any English model.
These scores revealed a notable finding: the inclusion of document-level information does improve the generation task, which is a crucial difference in UMR from AMR.
For Chinese, the best model for fluency using this approach was MT5, which was trained on sentence-level information for all six languages. For adequacy, SPRING trained on Chinese document-level data resulted in the highest adequacy score for Chinese across all models.
Overall, our fine-tuning approach on AMR-to-text generation models was the best performing of all approaches introduced in this work, and the models which were exposed to both sentence-level and document-level information, trained on just the language of generation language, showed the most success.
|
2502.11973
|
Fine-tuning Processes
| 228,304
| 60
|
22,512,328
| 0
|
In this work, we developed tools to generate text from Uniform Meaning Representation, introducing three distinct approaches that make use of existing Abstract Meaning Representation technologies to varying degrees.
|
2502.11973
|
Conclusion
| 228,305
| 61
|
22,512,329
| 1
|
While the UMR-to-AMR conversion pipeline successfully leverages existing AMR tools, our results indicated that the fine-tuned AMR-to-text generation models were stronger for both English and Chinese.
Further, the best performing models were fine-tuned exclusively on the same language UMR data, suggesting that UMR data in other languages may not be helpful in enhancing monolingual UMR-to-text generation models.
We also find that, while we only generate a single sentence at a time, the document-level information seems to improve model performance.
Finally, we find that the inclusion of AMR in the language model performance fine-tuning also resulted in better model performance, as the AMR-to-text generation models fine-tuned on UMR data produced more adequate output than the fine-tuned LLMs (with no exposure to AMR, only UMR).
|
2502.11973
|
Conclusion
| 228,305
| 62
|
22,512,330
| 2
|
Future work may explore advancements toward producing text in other languages from UMR, perhaps leveraging additional external resources.
The annotation and release of additional UMR data will support this effort, as well as improved performance in English and Chinese.
|
2502.11973
|
Conclusion
| 228,305
| 63
|
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 13
Size of downloaded dataset files:
1.25 GB
Size of the auto-converted Parquet files:
1.25 GB
Number of rows:
5,140,865