Unified Representation Learning for Relation Extraction in Visually-Rich Documents

1. Introduction

In recent years, a wide range of industries—spanning healthcare, insurance, e-commerce, and beyond—have increasingly turned to digitization and artificial intelligence to harness valuable information embedded in documents. This trend has given rise to the field of Visually-rich Document Understanding (VrDU) [11,14,21,24], which focuses on the extraction of structured knowledge from scanned or digital documents for practical downstream applications. VrDU encompasses several critical sub-fields such as Named-Entity Recognition (NER) [2], layout comprehension [7], and document classification [22]. Among these, relation extraction (RE) plays a pivotal role by establishing and identifying meaningful connections between recognized entities. Typically, RE is framed in a question-answer (Q/A) paradigm where the task is to design a function that determines whether a pair of entities in a document share a relevant relationship [11,23].

Parallel to these developments, advances in multimodal deep learning have revolutionized several fields including medical imaging [16], neurotechnology [4], and early diagnosis of neurological disorders such as Alzheimer’s disease [17]. The availability of robust optical character recognition (OCR) systems—such as AWS Textract1, Microsoft Read API2, and PyTesseract3—has further facilitated the extraction of detailed textual and spatial information from visually complex documents. These capabilities have catalyzed the development of a wide array of multimodal architectures that aim to fuse textual, geometric, and visual features for enriched document understanding [12,14,19,21,22,24]. In particular, these approaches promote end-to-end learning of joint representations to capture the full spectrum of information inherent in documents. Although transformer-based architectures have become highly popular in this domain [13,23], alternative methodologies, such as graph neural networks [3,6], have also shown promise in optimizing RE tasks by exploiting complex structural dependencies.

Despite significant progress in the development of multimodal systems for VrDU, there remains considerable ambiguity regarding the relative contributions of the different modalities. While text is intuitively expected to be the dominant predictor for entity relationships, the true extent of its superiority over geometric layout or visual data—and the interplay among these modalities—has not been definitively characterized. Notably, even studies that incorporate ablation tests rarely examine scenarios where text representations are completely omitted [9]. This gap in the literature motivates our systematic investigation into the predictive capacities of each modality, as well as their combined impact on relation extraction performance.

To address these challenges, we conduct a rigorous set of experiments using a variety of multimodal and unimodal configurations. Our study focuses on the RE task because it represents a fundamental yet unresolved problem in information extraction, with significant implications for industry applications. In our work, we leverage the novel DocFusion framework, which embodies an innovative approach to joint representation training. Through extensive empirical evaluations, we aim to (1) validate the benefits of learning joint representations for VrDU-based relation extraction, (2) quantify the asymmetric predictive capacities of text, layout, and visual modalities—thereby confirming that textual data holds a dominant role while also recognizing the supportive influence of layout and visual cues, and (3) introduce an optimized and simplified classifier for RE tasks that is built upon the principles of the DocFusion classification head (formerly derived from the LayoutXLM head) [23].

Beyond these primary contributions, our work also explores additional dimensions of document understanding by considering various forms of data interaction and feature fusion. For instance, our experiments examine the sensitivity of RE performance to different weighting schemes in the joint representation space. In certain configurations, we model the interaction between modalities using a weighted sum approach:

$$R=\alpha ·T+\beta ·L+\gamma ·V,$$

where R denotes the final representation, T, L, and V represent the text, layout, and visual features respectively, and $\alpha$, $\beta$, and $\gamma$ are hyperparameters that balance their relative contributions. Such formulations underscore the nuanced manner in which various modalities can be combined to yield a more robust feature set, which is critical for accurately inferring relations among document entities.

Furthermore, our investigation delves into the scalability and generalization capabilities of the DocFusion approach. We analyze its performance across diverse document types and domains, reflecting the heterogeneous nature of real-world documents. Through detailed comparative studies, we establish that while the text/layout combination often offers optimal performance, scenarios exist in which the inclusion of visual information can mitigate ambiguities introduced by noisy or sparse textual data.

In summary, this work contributes to the growing body of research on multimodal document analysis by presenting a comprehensive study on the integration of text, layout, and visual cues for relation extraction. The remainder of this paper is organized as follows. Section 2 reviews the literature related to multimodal approaches for document understanding and discusses current insights into the relative importance of different modalities. Section 3 outlines our methodology, detailing the dataset, model architecture, and experimental procedures. Section 4 reports our experimental results and performance comparisons. Section 5 discusses the limitations of our study and proposes directions for future research. Finally, Section 6 offers concluding remarks and reflections on the broader implications of our findings.

2. Related Work

In the realm of Visually-rich Document Understanding (VrDU) for relation extraction (RE), two primary benchmark datasets have emerged as indispensable resources. These are the FUNSD dataset [11] and the XFUND dataset [23]. Both datasets provide detailed annotations, where each instance includes a specification of linked entities represented by pairs of entity IDs, or an empty array in cases where no relationship is identified. This comprehensive annotation strategy enables researchers to delve deeply into the multimodal aspects of document analysis.

The availability of multiple data types—namely text, geometric layout, and document images—in these datasets naturally motivates the development of advanced multimodal methodologies. Early pioneering work in this area includes the LayoutLM family of models [21,22,23], which ingeniously combine textual embeddings with spatial (position) embeddings to capture the intrinsic layout information present in documents. LayoutLMv2 further enhances this approach by fully integrating visual cues extracted directly from document images, thereby providing a richer and more nuanced representation.

In parallel, the method presented in [18] employs a similar trifecta of inputs—text, layout, and visual data—to tackle tasks on both the FUNSD and MedForm datasets. Additionally, research conducted by [1] illustrates the effectiveness of leveraging a multimodal framework for document classification by fusing text and image modalities. Other notable contributions in the literature have focused on harnessing the synergistic effects of text and layout representations exclusively [12,13,15], thereby underscoring the critical role these modalities play in extracting structural and relational information from documents.

A central question that has arisen in the literature is: What are the relative effects and contributions of these different data types? Although several studies report ablation experiments, the precise impact of individual modalities is often difficult to ascertain, particularly because textual information is typically preserved across all experimental configurations [9,15]. This ambiguity is especially pertinent in industry applications, where the increased training and inference costs of large-scale multimodal models must be justified by commensurate improvements in performance. Notably, the original XFUND paper [23] does not offer a detailed breakdown of the contributions from text, layout, and visual information—a gap that has motivated further research in this direction.

In response to these challenges, our work introduces a novel framework named UniFusion. UniFusion is designed to harmonize the strengths of various modalities while maintaining computational efficiency. By systematically integrating text, layout, and visual features, UniFusion not only builds upon the foundational ideas present in earlier models but also provides a more balanced and cost-effective approach to multimodal representation learning. Although previous studies have largely treated each modality in isolation or in fixed combinations, UniFusion employs adaptive fusion strategies that dynamically adjust the contributions of each modality according to the complexity and quality of the input data.

Moreover, the literature reveals that the interplay among different modalities is not merely additive but often involves complex interactions that can either enhance or detract from overall system performance. For example, certain studies have observed that the presence of strong textual signals can sometimes overshadow the contribution of layout and visual features, leading to imbalanced models. Conversely, in scenarios where textual data is sparse or noisy, the complementary information provided by geometric and visual cues becomes critically important. These insights further validate the need for a framework like UniFusion, which is engineered to optimize the integration of heterogeneous data sources.

In summary, the extensive body of related work not only highlights the rapid evolution of multimodal approaches in document understanding but also points to significant gaps in our understanding of modality-specific contributions. Our proposed UniFusion framework aims to address these challenges by providing a unified, adaptive, and efficient method for multimodal representation learning. The insights gleaned from previous studies serve as a robust foundation for our work, setting the stage for a more thorough and balanced evaluation of multimodal strategies in the subsequent sections.

3. Methodology

In this section, we present in exhaustive detail the methodological framework underpinning our experiments on relation extraction (RE) using the XFUND dataset. Our approach leverages a novel multimodal architecture, termed UniFusion, which is an evolution of earlier transformer-based models. UniFusion is designed to integrate text, layout, and visual modalities into a unified representation for RE tasks. In what follows, we describe the dataset and preprocessing steps, elaborate on the architecture of UniFusion with all relevant technical and mathematical formulations, and detail the experimental procedures and training protocols employed.

3.1. Dataset and Preprocessing

We utilize the XFUND dataset4 for our experiments on multimodal RE [23]. XFUND comprises a large collection of document images specifically curated for form understanding across seven languages. Each document in the dataset is annotated with a unique identifier, a class label, bounding box coordinates

$$\{x_{left},y_{top},x_{right},y_{bottom}\},$$

the extracted textual content, and a linking indicator. The linking indicator is instrumental in defining relationships between entities, thereby facilitating the application of the dataset to the VrDU RE task. In this context, entities are organized as key-value pairs corresponding to the questions and answers found in the forms. For a comprehensive description of the data collection and annotation process, readers are referred to [23]. It is noteworthy that the dataset statistics we employ differ slightly from those originally reported in [23]; hence, we provide the revised statistics in Table 1. The language codes ZH, JA, ES, FR, IT, DE and PT correspond respectively to Chinese, Japanese, Spanish, French, Italian, German, and Portuguese.

Prior to feeding the data into our model, we perform several preprocessing steps. These include normalization of the bounding box coordinates to ensure consistency across different document scales, tokenization of the textual content using sub-word techniques, and resizing of document images to a fixed resolution. Furthermore, we apply language-specific preprocessing pipelines to handle the multilingual nature of the dataset.

3.2. UniFusion Architecture for Relation Extraction

The core of our proposed methodology is the UniFusion architecture, which builds upon the conceptual foundations of transformer-based models while introducing key innovations to enhance multimodal feature integration. UniFusion is a pretrained transformer that simultaneously ingests text, layout, and visual inputs to generate a rich joint representation suitable for RE tasks.

3.2.1. Feature Extraction and Embedding Layers

UniFusion begins by encoding each modality using dedicated embedding layers:

• Textual Embedding: Each word token is embedded into a high-dimensional space via a learned embedding matrix $E_t\in {\mathbb{R}}^{V\times d}$, where V is the vocabulary size and d is the embedding dimension.
• Layout Embedding: The geometric layout of the document is represented using normalized bounding box coordinates. These coordinates are transformed via a fully-connected layer into an embedding $E_l\in {\mathbb{R}}^{4\times d_l}$, where $d_l$ denotes the dimensionality specific to spatial features.
• Visual Embedding: Document images are processed by a visual backbone based on ResNeXt 101-FPN [20], which outputs visual feature maps. These maps are further encoded into a vector $E_v\in {\mathbb{R}}^{d_v}$ using convolutional and pooling layers.

For each token, the final multimodal representation is computed as a summation of the modality-specific embeddings along with positional encodings. To account for the two-dimensional nature of document layouts, UniFusion incorporates both one-dimensional (1D) and two-dimensional (2D) positional embeddings:

$$E=E_t+E_{p1D}+E_l+E_{p2D}+E_v,$$

where $E_{p1D}$ and $E_{p2D}$ denote the 1D and 2D positional embeddings respectively.

3.2.2. Multimodal Fusion Mechanism

A distinguishing feature of UniFusion is its adaptive multimodal fusion strategy. Instead of statically concatenating modalities, UniFusion learns to weight the contribution of each modality dynamically. The fusion process is formulated as:

$$F={\lambda }_T·\varphi \left(T\right)+{\lambda }_L·\varphi \left(L\right)+{\lambda }_V·\varphi \left(V\right),$$

where $\varphi (·)$ represents a transformation function (implemented as a feed-forward network) for the text $\left(T\right)$, layout $\left(L\right)$, and visual $\left(V\right)$ features, and ${\lambda }_T$, ${\lambda }_L$, and ${\lambda }_V$ are learnable parameters that modulate the influence of each modality. In practice, these weights are normalized using a softmax function to ensure that

$${\lambda }_T+{\lambda }_L+{\lambda }_V=1.$$

Additionally, the fusion layer is enhanced by a self-attention mechanism that captures the interactions between different modalities. The attention operation is given by:

$$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}\left({\displaystyle \frac{Q{K}^{\top }}{\sqrt{d_k}}}\right)V,$$

where Q, K, and V denote the query, key, and value matrices respectively, and $d_k$ is the dimensionality of the key vectors.

3.2.3. Customized Classification Head

For the relation extraction task, UniFusion incorporates a specialized classification head that builds upon the joint representation. Unlike the original LayoutXLM design, our classification head has been streamlined to reduce computational overhead while preserving performance. Specifically, the classification head comprises:

• A single fully-connected layer that projects the fused representation F to a lower-dimensional space.
• A leaky ReLU activation function to introduce non-linearity:

  $$\mathrm{LeakyReLU}\left(x\right)=max(\alpha x,x),$$

  where $\alpha$ is a small positive constant (set to 0.01 in our experiments).
• A dropout layer with a dropout probability $p=0.2$ to mitigate overfitting.

Furthermore, the classification layer employs a bi-affine transformation to model interactions between entity pairs. Given two entity representations $h_i$ and $h_j$, the score for a potential relation is computed as:

$$s_{ij}=h_i^{\top }W{h}_j+U(h_i+h_j)+b,$$

where W, U, and b are learnable parameters.

3.2.4. Loss Function and Optimization

The overall training objective of UniFusion for the RE task is defined by a composite loss function that includes a cross-entropy loss for classification and an auxiliary regularization term to enforce smoothness in the multimodal fusion weights. The loss function is defined as:

$$\mathcal{L}={\mathcal{L}}_{CE}+\beta {\mathcal{L}}_{reg},$$

where

$${\mathcal{L}}_{CE}=-\sum _{i=1}^N{y}_{i}log\left({\widehat{y}}_i\right)$$

and

$${\mathcal{L}}_{reg}={∥\nabla \lambda -\overline{\lambda }∥}_2^2.$$

Here, ${\widehat{y}}_i$ is the predicted probability for the i-th sample, $y_i$ is the corresponding ground truth label, $\lambda$ denotes the set of fusion weights, $\overline{\lambda }$ is the mean value of the weights, and $\beta$ is a regularization hyperparameter.

3.3. Experimental Procedures

Our experiments are designed to rigorously evaluate the effectiveness of UniFusion by testing six distinct multimodal and unimodal configurations on the multilingual XFUND dataset. In total, experiments are performed separately for each of the seven language subsets, and each configuration is fine-tuned from the pretrained UniFusion model. The six configurations are as follows:

1.

Multimodal (MM): Incorporates text, layout, and visual information.

2.

Bimodal Text and Layout (text/layout): Visual components are entirely removed.

3.

Bimodal Text and Visual (text/visual): Layout information, including 2D position embeddings, is omitted.

4.

Bimodal Layout and Visual (layout/visual): Textual data and corresponding embeddings are excluded.

5.

Unimodal Layout (layout): Only layout information is used.

6.

Unimodal Text (text): Only text data is leveraged.

3.3.1. Configuration-Specific Model Modifications

For the MM configuration, the full UniFusion architecture is employed without modifications. In the text/layout setting, we excise all visual components, including the visual backbone and associated embeddings. Conversely, in the text/visual configuration, the layout-specific 2D position embeddings are removed from both the primary stream and the visual branch. In the layout/visual configuration, the textual stream is completely omitted by discarding tokenized text and the corresponding 1D and 2D positional embeddings. The unimodal experiments (configurations 5 and 6) are derived from the aforementioned bimodal settings by applying the relevant exclusions. It is important to note that many prior studies have neglected to consider the absence of text data entirely [9,15], despite its critical role in multimodal fusion.

3.3.2. Training Protocol and Hyperparameter Optimization

Given the inherent differences in data characteristics across modalities, we hypothesized that the optimal learning rate might vary between experiments. Consequently, learning rate was the sole hyperparameter subjected to optimization. A grid search was conducted over three candidate values: $5\times {10}^{-5}$, $1\times {10}^{-5}$, and $5\times {10}^{-6}$. All other hyperparameters remained constant across experiments. In our training regimen, the following parameters were employed:

• Batch Size: 2.
• Number of Epochs: 50.
• Optimizer: Adam with default momentum parameters.
• Learning Rate Scheduler: A cosine annealing scheduler was applied to gradually decrease the learning rate over the training period.

To monitor the training process and prevent overfitting, we also implemented early stopping based on the validation loss. Moreover, we computed additional evaluation metrics such as precision, recall, and F1 score, defined as:

$$\mathrm{Precision}={\displaystyle \frac{TP}{TP+FP}},\mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}},\mathrm{F}1\mathrm{Score}={\displaystyle \frac{2\times \mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}},$$

where $TP$, $FP$, and $FN$ denote the true positives, false positives, and false negatives, respectively.

3.3.3. Additional Technical Considerations

To further enhance the stability and performance of the UniFusion model, several technical innovations were incorporated:

1.

Gradient Clipping: A threshold of 1.0 was imposed on gradient norms to prevent exploding gradients.

2.

Layer Normalization: Each transformer block within UniFusion is equipped with layer normalization to stabilize training dynamics.

3.

Multi-Head Self-Attention: The self-attention mechanism in UniFusion is implemented with multiple heads, each head performing independent attention operations. The overall attention output is then concatenated and projected back into the model’s hidden space.

4.

Residual Connections: To facilitate the flow of gradients, residual connections are employed throughout the network.

The fusion of these advanced techniques ensures that UniFusion is capable of robustly capturing the intricate interactions among text, layout, and visual modalities, ultimately leading to superior performance on the RE task.

In summary, our experimental procedures are meticulously designed to investigate the impact of each modality on relation extraction performance. The comprehensive analysis across multiple configurations and languages not only validates the efficacy of the proposed UniFusion framework but also provides critical insights into the nuanced contributions of each data type in visually-rich document understanding.

4. Experimental Outcomes and Analysis

In this section, we present a detailed discussion of the experimental results obtained using the UniFusion framework on the XFUND relation extraction task. Our evaluation is primarily based on the F1 score, with additional metrics such as precision and recall also reported. Although multimodal results have been previously presented for the XFUND RE task [23], our results are reported independently due to variations in network configuration, dataset statistics, and training procedures.

4.1. Bimodal Integration Outperforms Trimodal Fusion

Table 2 reports the F1 scores achieved by models trained under six different configurations. Supplementary recall and precision metrics are provided in Table 3 and Table 4, respectively. Overall, the experimental findings strongly support the advantage of learning joint representations for the VrDU RE task. In particular, the bimodal configuration that combines text and layout information achieves a mean F1 score of 0.6843, which is higher than the trimodal configuration (text, layout, and image) that obtains a mean F1 of 0.6728.

This result reinforces previous observations that the integration of text and layout can capture the essential document features required for accurate relation extraction, sometimes even outperforming configurations that include additional visual data. In our experiments, the text/layout model outperforms the trimodal model by margins of approximately 8.08% and 6.93% when compared to the next best alternative, as detailed in Table 2. Additionally, the other bimodal configurations—text/visual and layout/visual—yield F1 scores of 0.604 and 0.558, respectively, which further emphasizes the hierarchical predictive strength of these modalities. Text appears to be the most dominant signal, followed by layout, with visual information contributing marginally in the current setup.

To quantify the performance more systematically, we also consider a weighted F1 score metric defined as:

$${\mathrm{F}1}_{\mathrm{weighted}}={\displaystyle \frac{{\sum }_{i=1}^N{n}_i·{\mathrm{F}1}_i}{{\sum }_{i=1}^N{n}_i}},$$

where $n_i$ denotes the number of samples for language i and ${\mathrm{F}1}_i$ is the corresponding F1 score. This formulation confirms that the text/layout configuration consistently outperforms other modalities across the multilingual dataset.

4.2. Text Remains the Central Component in Relation Extraction

Our results clearly indicate that the exclusion of text data leads to a significant drop in performance. Among the six configurations evaluated, those that do not include textual information (namely, the layout/visual and unimodal layout setups) are the worst performing. Notably, the unimodal text model achieves considerably better classification results than the bimodal layout/visual model, which highlights the pivotal role of text in the RE task.

However, the inclusion of additional modalities still offers benefits. When combined with text, both layout and visual information contribute to improved performance. For example, the addition of layout features not only enhances the performance in the highest scoring text/layout configuration but also substantially improves the unimodal layout score (from 0.471 to 0.558 in the layout/visual setting). This observation can be mathematically expressed as:

$$\Delta \mathrm{F}1={\mathrm{F}1}_{\mathrm{bimodal}}-{\mathrm{F}1}_{\mathrm{unimodal}},$$

demonstrating the incremental benefit derived from multimodal fusion. The experimental outcomes suggest that while text is indispensable, its optimal performance is achieved when supplemented by layout information, particularly in scenarios where text quality may be compromised.

4.3. Training Dynamics and Data-Dependent Variability

The training dynamics further emphasize the crucial role of text in the learning process. Loss curves (not shown here) indicate that model configurations incorporating text experience a rapid and steady decline in training loss, converging reliably within 50 epochs. In contrast, models that rely solely on layout or a combination of layout and visual data display more gradual loss reductions and greater variability across different language sets.

We define the training loss function as:

$$\mathcal{L}\left(t\right)={\displaystyle \frac{1}{N}}\sum _{i=1}^N\ell \left(y_i,{\widehat{y}}_i\left(t\right)\right),$$

where $\ell (·)$ represents the per-sample loss, $y_i$ is the ground truth label, and ${\widehat{y}}_i\left(t\right)$ is the predicted output at epoch t. The steeper decline in $\mathcal{L}\left(t\right)$ for models utilizing text confirms that textual features provide a robust signal that facilitates effective optimization.

Moreover, the observed higher variance in training loss for configurations that exclude text suggests that layout and visual modalities are more sensitive to dataset-specific characteristics. This is further reflected in the optimal learning rates determined by grid search: while the full multimodal and text/layout configurations were trained with a learning rate of $5\times {10}^{-5}$, the models excluding text required lower learning rates (either $1\times {10}^{-5}$ or $5\times {10}^{-6}$). Such differences underscore the importance of modality-specific hyperparameter tuning in multimodal systems.

In summary, the experimental outcomes validate the efficacy of the UniFusion framework, highlighting that text is the foundational component in relation extraction tasks within visually-rich documents. The combination of text with layout information yields the best overall performance, while visual information plays a more supplementary role. These insights are crucial for developing efficient and effective multimodal deep learning systems for document analysis.

5. Conclusions and Future Directions

5.1. Conclusions

In this work, we explored the complex trade-offs that arise when employing multimodal approaches in industrial applications, particularly for relation extraction within visually-rich document understanding (VrDU). Our investigation involved training a multimodal transformer—renamed here as UniFusion—under several distinct data configurations to evaluate the effectiveness of joint representation learning. The empirical results demonstrate that the bimodal configuration incorporating both text and layout information consistently outperforms the full trimodal setup, even though the latter integrates visual cues as well.

More specifically, our findings indicate that textual information serves as the primary driver in predictive tasks, with unimodal text achieving a higher mean F1 score compared to the bimodal layout/visual configuration. This suggests that while text is indispensable for capturing the semantic essence of documents, layout information plays a crucial complementary role. In scenarios where visual data is incorporated, its contribution appears to be supplemental, enhancing performance in particular conditions such as when the quality of textual data is compromised. This observation not only corroborates previous studies but also reinforces the design philosophy behind UniFusion, where the strategic integration of modalities is key to enhanced performance.

Moreover, our analysis showed that training dynamics and convergence behaviors are highly dependent on the inclusion or exclusion of specific modalities. The experiments reveal that when text is omitted, the performance degradation is significant, underscoring its pivotal role in the learning process. Overall, the results provide robust evidence that while each modality has its own merits, the most effective approach for the RE task is to combine text with layout data, thereby harnessing the strengths of both semantic and structural information.

5.2. Limitations and Future Work

Despite the promising outcomes, several limitations in the current study warrant discussion. One major constraint is the relative scarcity of available data for the VrDU RE task. With FUNSD [11] being the only other comparable dataset, the limited diversity and volume of samples restrict our ability to generalize the results. Future research should aim to expand the dataset repertoire and increase the number of samples, which could help better understand how data diversity influences model robustness.

Another limitation is the focus on a single multimodal architecture—UniFusion (previously LayoutXLM)—which may not capture the full spectrum of architectural innovations available in the literature. Other frameworks with alternative strategies for encoding and fusing modalities, as explored in recent studies [8,10], could offer additional insights. A systematic comparison involving a wider range of architectures is essential to further delineate the relative contributions of text, layout, and visual information.

Future work should also consider extending the scope of analysis to include other document understanding tasks such as document classification, semantic entity recognition, and key information extraction. A large-scale, cross-task evaluation would help to determine whether the observed performance hierarchy—text as the primary modality, followed by layout and then visual data—holds across different applications. Additionally, more advanced ablation studies, possibly incorporating a cost-performance trade-off analysis, could provide valuable insights into the efficiency and practicality of deploying such multimodal systems in high-volume business environments.

Furthermore, the exploration of dynamic fusion strategies remains a promising avenue for research. Future studies might investigate adaptive weighting schemes that adjust the contributions of different modalities on a per-sample basis. Such mechanisms could further optimize joint representation learning and improve overall system resilience against variations in document quality.

In summary, while the current study demonstrates the efficacy of UniFusion for the VrDU RE task, there remains considerable scope for refining multimodal fusion strategies and broadening the evaluation framework to encompass diverse document analysis challenges. Addressing these limitations will be critical for advancing the state-of-the-art in document understanding and achieving higher levels of performance in real-world applications.