GateMABSA: Gated Fusion for MABSA

• The paper introduces a gated multimodal architecture using Fuse-mLSTM, Syn-mLSTM, and Sem-mLSTM to selectively fuse multimodal signals and filter out noise.
• It leverages dependency-based and semantic gates to integrate syntactic structure and aspect relevance, improving aspect-level sentiment classification.
• Empirical results on Twitter-15 and Twitter-17 datasets demonstrate competitive accuracy and macro-F1, validating the model's robust design.

Searching arXiv for the specified paper to verify bibliographic details. Search query: "GateMABSA: Aspect-Image Gated Fusion for Multimodal Aspect-based Sentiment Analysis" GateMABSA is a multimodal aspect-based sentiment analysis (MABSA) architecture introduced in "GateMABSA: Aspect-Image Gated Fusion for Multimodal Aspect-based Sentiment Analysis" (Lawan et al., 29 Sep 2025). It addresses a setting in which user-generated content combines text and images, and the task is to predict the sentiment polarity of a chosen target aspect from the label set $\{\text{positive},\text{negative},\text{neutral}\}$. The model is defined as a gated multimodal architecture that integrates syntactic, semantic, and fusion-aware mLSTM, with three specialized modules: Fuse-mLSTM, Syn-mLSTM, and Sem-mLSTM. Its stated purpose is to filter noisy visual signals and align aspects with opinion-bearing content across modalities (Lawan et al., 29 Sep 2025).

1. Problem setting and formalization

In multimodal aspect-based sentiment analysis (MABSA) the input consists of a sentence $S = [w_{1}, w_{2}, \dots, w_{n}]$ of $n$ tokens, an associated image $V$, and a set of aspect terms $\{A_{1},\dots,A_{r}\}$, where each $A_{i}$ may be explicitly mentioned or implicitly inferred (Lawan et al., 29 Sep 2025). For a chosen target aspect $A$, the objective is to predict its sentiment polarity,

$$y \;\in\; \{\text{positive},\;\text{negative},\;\text{neutral}\}.$$

The representational setup uses three aligned feature tensors. $\mathbf{H}_{S}\in\mathbb R^{n\times d}$ denotes the contextual token representations from a text encoder (RoBERTa). $\mathbf{H}_{A}\in\mathbb R^{r\times d}$ denotes the aspect representations, described as mean-pooled and repeated to length $S = [w_{1}, w_{2}, \dots, w_{n}]$0. $S = [w_{1}, w_{2}, \dots, w_{n}]$1 denotes the image features, constructed by ResNet-152 $S = [w_{1}, w_{2}, \dots, w_{n}]$2 linear $S = [w_{1}, w_{2}, \dots, w_{n}]$3 mean-pooled $S = [w_{1}, w_{2}, \dots, w_{n}]$4 repeated (Lawan et al., 29 Sep 2025).

This formulation places GateMABSA in the aspect-level classification regime rather than whole-post sentiment prediction. A plausible implication is that the model is intended for inputs containing multiple potentially sentiment-bearing targets, since the formulation explicitly distinguishes a chosen target aspect $S = [w_{1}, w_{2}, \dots, w_{n}]$5 from the larger set $S = [w_{1}, w_{2}, \dots, w_{n}]$6.

2. Architectural composition

GateMABSA stacks three specialized mLSTM modules on top of unimodal encoders: Fuse-mLSTM, Syn-mLSTM, and Sem-mLSTM (Lawan et al., 29 Sep 2025). Their roles are stated directly:

• Fuse-mLSTM: performs selective fusion of text and image features.
• Syn-mLSTM: injects syntactic structure (dependency graphs) via a graph gate.
• Sem-mLSTM: enforces semantic relevance between aspect and context through a semantic gate.

The final Sem-mLSTM outputs $S = [w_{1}, w_{2}, \dots, w_{n}]$7, which is mean-pooled and passed through a linear+softmax for classification (Lawan et al., 29 Sep 2025).

The architecture is sequential in the sense that multimodal interaction precedes syntax-aware processing, and syntax-aware processing precedes semantic refinement. This ordering reflects the workflow described in the model overview: first fuse text and image, then inject dependency structure, then sharpen aspect-semantic relevance. This suggests a progressive constraint strategy in which the model first builds a shared multimodal state and then regularizes it with linguistic and aspect-specific priors.

3. Aspect-image gated fusion

Before syntax and semantics, GateMABSA fuses text and image representations at the token level. At each token $S = [w_{1}, w_{2}, \dots, w_{n}]$8, let $S = [w_{1}, w_{2}, \dots, w_{n}]$9 and $n$0. The model computes a fusion gate

$n$1

and then forms the fused state

$n$2

The selective mechanism is described in operational terms: irrelevant visual signals are suppressed when $n$3 is low, and complementary image cues are injected when $n$4 is high (Lawan et al., 29 Sep 2025). Within the stated motivation of the model, this module addresses the difficulty that existing MABSA models struggle to filter noisy visual signals.

The fusion stage is central to the model’s multimodal character. Because image features are repeated to length $n$5, the gate is applied in a token-aligned fashion rather than through a single global fusion decision. A plausible implication is that the architecture attempts to localize cross-modal relevance with respect to individual textual positions.

4. Syntax-aware and semantic-aware gating

Syn-mLSTM

Syn-mLSTM consumes fused hidden states $n$6. At each time step $n$7 it forms the standard LSTM gates

$n$8

where $n$9 are linear projections of $V$0 (Lawan et al., 29 Sep 2025).

To inject syntactic priors, GateMABSA introduces a graph (syntax) gate

$V$1

where $V$2 is the dependency adjacency, $V$3 the pairwise query similarity, and $V$4 a syntax-aware aspect embedding. The cell state is then updated in a log-domain cumulative fashion, effectively adding $V$5 as an extra decay term. The hidden state is

$V$6

where $V$7 has absorbed the graph gate (Lawan et al., 29 Sep 2025).

The stated role of this graph gate is to enforce dependency-based aspect–opinion connections. In the model’s logic, syntactic structure supplies relational priors that constrain which tokens should influence the aspect-level sentiment representation.

Sem-mLSTM

Sem-mLSTM takes $V$8 as input and again computes

$V$9

It adds a semantic gate

$\{A_{1},\dots,A_{r}\}$0

where $\{A_{1},\dots,A_{r}\}$1 is the semantic aspect embedding, $\{A_{1},\dots,A_{r}\}$2 the token-aspect distance, and $\{A_{1},\dots,A_{r}\}$3 scaling factors. As before, $\{A_{1},\dots,A_{r}\}$4 augments the cumulative forget term and refines the retrieval of $\{A_{1},\dots,A_{r}\}$5 (Lawan et al., 29 Sep 2025).

The semantic gate is described as sharpening the focus on semantically and positionally relevant tokens. This couples aspect relevance with token-aspect distance, making the semantic stage both content-sensitive and position-sensitive. A plausible implication is that GateMABSA treats semantic relevance as a dynamic retrieval problem rather than as a fixed attention score.

5. Optimization and empirical results

After Sem-mLSTM, the model mean-pools $\{A_{1},\dots,A_{r}\}$6 to $\{A_{1},\dots,A_{r}\}$7 and computes class probabilities

$\{A_{1},\dots,A_{r}\}$8

Training uses the standard aspect-level cross-entropy

$\{A_{1},\dots,A_{r}\}$9

The reported experiments use two benchmark Twitter datasets: Twitter-15 (3.2 K tweets) and Twitter-17 (3.5 K tweets). The baselines are AE-LSTM, IAN, MGAN, BERT, MIMN, TomBERT, ESAFN, EF-CapTrBERT, ITM, HIMT, AMIFN, GLFFCA, KAHGCN, DMIN, REF, and GAS. Evaluation uses Accuracy and macro-F1 at the aspect level (Lawan et al., 29 Sep 2025).

Dataset	GateMABSA	Prior best reported in the data
Twitter-15	79.96 % Acc. / 75.67 % F1	79.14 % / 74.93 % by GAS
Twitter-17	74.82 % Acc. / 71.54 % F1	73.72 % / 72.79 %

The reported gains are attributed to three factors: Fuse-mLSTM’s selective gating, which filters noisy images and aligns tokens to aspect; Syn-mLSTM’s graph gate, which enforces dependency-based aspect–opinion connections; and Sem-mLSTM’s semantic gate, which sharpens the focus on semantically and positionally relevant tokens (Lawan et al., 29 Sep 2025).

A notable detail is that the result pattern is not uniform across all reported metrics. On Twitter-17, GateMABSA reaches 74.82 % Accuracy against a prior best 73.72 %, while the listed macro-F1 value is 71.54 % against a prior best 72.79 %. This is relevant when interpreting the statement that the model outperforms several baselines: the data support strong overall competitiveness, but they do not indicate dominance on every metric in every dataset.

6. Interpretation, scope, and common points of clarification

GateMABSA is characterized in the source as a robust, interpretable pipeline for filtering visual noise and aligning multimodal signals around target aspects (Lawan et al., 29 Sep 2025). The notion of interpretability in this description is tied to the explicit gating structure: fusion-aware gating for cross-modal selection, graph gating for dependency-informed structure, and semantic gating for aspect-sensitive contextual refinement.

One common point of clarification concerns the role of images. GateMABSA does not treat the image as uniformly beneficial contextual evidence. Its fusion formulation explicitly allows image information to be suppressed when irrelevant visual signals are present. This distinguishes multimodal inclusion from unconditional multimodal reliance.

A second clarification concerns the role of syntax. The architecture does not merely append dependency information as a separate feature vector. Instead, the dependency adjacency $A_{i}$0 and pairwise query similarity $A_{i}$1 appear inside a graph gate that affects the cumulative memory dynamics of Syn-mLSTM. This suggests a tighter integration of syntactic priors with recurrent state updates than a late-fusion design would provide.

A third clarification concerns aspect conditioning. Aspect information enters the model in multiple forms: as repeated aspect representations $A_{i}$2, as a syntax-aware aspect embedding $A_{i}$3, and as a semantic aspect embedding $A_{i}$4. The design therefore addresses aspect relevance at more than one stage. A plausible implication is that GateMABSA is intended to reduce failure modes in which multimodal sentiment cues are detected but not correctly assigned to the target aspect.

Within the scope defined by the reported experiments, GateMABSA is a benchmarked MABSA architecture on Twitter-15 and Twitter-17 with a three-stage gated mLSTM design. Its technical identity lies in the combination of selective aspect-image fusion, dependency-aware recurrent gating, and aspect-semantic recurrent gating, all optimized under standard aspect-level cross-entropy for three-way sentiment classification (Lawan et al., 29 Sep 2025).