Paper deep dive
DuConTE: Dual-Granularity Text Encoder with Topology-Constrained Attention for Text-attributed Graphs
Lexuan Liang, Tao Zou, Xuxiang Ta, Zekun Qiu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/27/2026, 7:20:49 AM
Summary
DuConTE (Dual-Granularity Text Encoder with Topology-Constrained Attention) is a novel framework designed for Text-Attributed Graphs (TAGs). It addresses the limitations of existing hybrid LM-GNN models by implementing a dual-granularity encoding process: first at the word-token level and then at the node level. The model utilizes a topology-constrained attention mechanism that uses an attention masking strategy to incorporate graph structural information (connectivity) directly into the Transformer's self-attention layers. Additionally, it features a Node Representation Composer that distinguishes between center-node and neighborhood contexts to effectively aggregate word-token embeddings into node-level representations.
Entities (6)
Relation Signals (4)
DuConTE → contains → Node Representation Composer
confidence 100% · Furthermore, we design a node representation composer that assesses the importance of individual word tokens
DuConTE → processes → Text-attributed graph
confidence 100% · DuConTE: Dual-Granularity Text Encoder with Topology-Constrained Attention for Text-attributed Graphs
DuConTE → uses → Topology-constrained attention
confidence 100% · The model employs a cascaded architecture of two pretrained LMs... DuConTE employs a topology-constrained attention mechanism
Language Model → ispartof → DuConTE
confidence 90% · The model employs a cascaded architecture of two pretrained LMs
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Text-attributed graphs integrate semantic information of node texts with topological structure, offering significant value in various applications such as document classification and information extraction. Existing approaches typically encode textual content using language models (LMs), followed by graph neural networks (GNNs) to process structural information. However, during the LM-based text encoding phase, most methods not only perform semantic interaction solely at the word-token granularity, but also neglect the structural dependencies among texts from different nodes. In this work, we propose DuConTE, a dual-granularity text encoder with topology-constrained attention. The model employs a cascaded architecture of two pretrained LMs, encoding semantics first at the word-token granularity and then at the node granularity. During the self-attention computation in each LM, we dynamically adjust the attention mask matrix based on node connectivity, guiding the model to learn semantic correlations informed by the graph structure. Furthermore, when composing node representations from word-token embeddings, we separately evaluate the importance of tokens under the center-node context and the neighborhood context, enabling the capture of more contextually relevant semantic information. Extensive experiments on multiple benchmark datasets demonstrate that DuConTE achieves state-of-the-art performance on the majority of them.
Tags
Links
- Source: https://arxiv.org/abs/2604.17411v1
- Canonical: https://arxiv.org/abs/2604.17411v1
Trouble viewing inline? Open PDF directly →
Full Text
85,286 characters extracted from source content.
Expand or collapse full text
DuConTE: Dual-Granularity Text Encoder with Topology-Constrained Attention for Text-attributed Graphs Lexuan Liang School of Computer Science and Engineering Beihang University Beijing, China 100191 23373299@buaa.edu.cn &Tao Zou School of Computer Science and Engineering Beihang University Beijing, China 100191 zoutao@buaa.edu.cn &Xuxiang Ta School of Computer Science and Engineering Beihang University Beijing, China 100191 taxuxiang@buaa.edu.cn &Zekun Qiu School of Computer Science and Engineering Beihang University Beijing, China 100191 qzk@buaa.edu.cn Abstract Text-attributed graphs integrate semantic information of node texts with topological structure, offering significant value in various applications such as document classification and information extraction. Existing approaches typically encode textual content using language models (LMs), followed by graph neural networks (GNNs) to process structural information. However, during the LM-based text encoding phase, most methods not only perform semantic interaction solely at the word-token granularity, but also neglect the structural dependencies among texts from different nodes. In this work, we propose DuConTE, a dual-granularity text encoder with topology-constrained attention. The model employs a cascaded architecture of two pretrained LMs, encoding semantics first at the word-token granularity and then at the node granularity. During the self-attention computation in each LM, we dynamically adjust the attention mask matrix based on node connectivity, guiding the model to learn semantic correlations informed by the graph structure. Furthermore, when composing node representations from word-token embeddings, we separately evaluate the importance of tokens under the center-node context and the neighborhood context, enabling the capture of more contextually relevant semantic information. Extensive experiments on multiple benchmark datasets demonstrate that DuConTE achieves state-of-the-art performance on the majority of them. 1 Introduction Text-attributed graphs (Yang et al., 2021; Seo et al., 2024) have emerged as an increasingly significant research domain, with substantial applications in real-world scenarios such as social media analysis (Seo et al., 2024), academic citation systems (Wang et al., 2025), and knowledge base construction (Zhang et al., 2024). In such graphs, each node is associated with a piece of textual content, resulting in richly structured data that encapsulates both semantic text information and topological structure. Learning high-quality representations that effectively capture both the textual and structural characteristics of nodes is crucial for downstream tasks such as node classification (Zhao et al., 2024). Recently, a growing body of research (Chen et al., 2023; Chien et al., 2021; Zhu et al., 2024) has begun leveraging Transformer-based language models (LMs) to model textual information in text-attributed graphs, aiming to enhance graph neural networks (GNNs). Thanks to their strong pre-trained understanding of natural language, LMs can produce highly expressive representations of textual content. For example, GraphBridge(Wang et al., 2024) attempts to combine the text from the center-node and its neighbors into the LM, enabling the model to jointly encode the central text and its contextual information from neighboring nodes. Current approaches (Zhu et al., 2024; He et al., 2023; Jin et al., 2023) that jointly employ GNNs and LMs largely follow a common paradigm: the LM is responsible for encoding textual features, while the GNN focuses on capturing structural information. Figure 1: Overview of the text-attributed graph learning pipeline (top) and comparison between existing methods and the proposed DuConTE (bottom). However, existing approaches typically perform semantic interaction only at the word-token granularity when using LMs for text encoding, failing to capture meaningful node-granularity semantic interactions—where the textual content of different nodes is treated as holistic units and interacts across the graph. Moreover, current methods either do not incorporate structural information into the LM at all, or the injected structural signals are insufficient to guide the encoding process effectively. Additionally, existing methods lack an effective mechanism for composing node representations from word-token embeddings. To address these limitations, we propose DuConTE, a dual-granularity text encoder with topology-constrained attention for text-attributed graphs. As illustrated in the top panel of Figure 1, the text-attributed graph learning pipeline consists of three stages, with DuConTE acting as a plug-and-play text encoder module. It takes as input the text of each node and its sampled neighborhood structure (e.g., from random walks or k-hop sampling), obtained through upstream preprocessing, and outputs enriched node representations for downstream GNN models. DuConTE performs dual-granularity semantic encoding, in which two pretrained LMs sequentially encode textual semantics at the word-token and node granularities, respectively. This design aligns with the inherent multi-granular nature of text-attributed graphs, allowing for a more complete capture of textual semantics. During the encoding process, DuConTE employs a topology-constrained attention mechanism to leverage graph structural information for enhanced text encoding. This is achieved through an attention masking strategy specifically designed for TAG, motivated by the homophily analysis in Section A, enabling pretrained LMs to better process graph-structured textual data without architectural modification. Furthermore, we design a node representation composer that assesses the importance of individual word tokens under both center-node and neighborhood semantic contexts. This enables the model to capture salient semantic information more effectively when composing node representations from word-token embeddings. • We propose DuConTE, a dual-granularity text encoder with topology-constrained attention for text-attributed graphs. It performs dual-granularity semantic encoding to model textual semantics at both the word-token granularity and node granularity, capturing a comprehensive, multi-scale understanding of the text-attributed graph. • We introduce a topology-constrained attention mechanism that leverages an attention masking strategy, specifically designed for TAGs and grounded in the homophily analysis in Section A, to effectively incorporate structural guidance into the textual encoding process. • We design a node representation composer that distinctly models token importance under center-node and neighborhood contexts, enabling effective fusion of word-token embeddings into comprehensive node representations. 2 Related Work 2.1 Text-attributed graph learning Learning on text-attributed graphs has evolved from employing simple text features like Bag-of-Words (Zhang et al., 2010) to sophisticated methods centered on language models (LMs) (Chen et al., 2023; Chien et al., 2021; Zhu et al., 2024). These modern approaches generally follow two main paradigms. The first relies on a single, powerful LM to jointly process text and structure. For instance, LLaGA (Chen et al., 2024) injects structural information by mapping it into the LM’s token space and relies solely on the LM to generate predictions. While conceptually unified, this paradigm is often computationally demanding, suffers from poor scalability, and achieves limited effectiveness in leveraging structural information. The second, more common paradigm, employs a hybrid LM-GNN pipeline where an LM first serves as a text encoder, and a subsequent GNN performs the downstream task using the resulting node embeddings. Representative works like GraphBridge (Wang et al., 2024) enrich node text with neighbor semantics before encoding, whereas Engine (Zhu et al., 2024) uses a GNN to process features from multiple LM layers. A critical limitation across most hybrid models is that the LM encoding process remains largely unaware of the graph topology. This decoupled approach hinders the deep fusion of structural and semantic information, a key challenge we address in this work. 2.2 Transformers for Modeling Structured Data In recent years, numerous studies have leveraged Transformers to process graph-structured data (Shehzad et al., 2024). An early effort in this direction is Graph-BERT (Zhang et al., 2020), which applies a BERT-style Transformer to sampled subgraphs without relying on message passing. More recent approaches further enhance structural awareness: Graphormer (Ying et al., 2021) enhances the Transformer’s understanding of graph structures by introducing spatial encoding and degree encoding. Another work NeuralWalker (Chen et al., 2025) generates serialized representations of graphs through random walks to exploit the self-attention mechanism of Transformers for modeling purposes. Edge-augmented methods (Rampášek et al., 2022; Satorras et al., 2021) explicitly model edge features to enhance the Transformer’s sensitivity towards different edge types. Masked Graph Modeling (Hou et al., 2023; Tian et al., 2024) employs a masking strategy to learn structural information by predicting masked node or edge features. Notably, another strategy enhances structural awareness by using attention masks to explicitly control token interactions. K-BERT (Liu et al., 2020) employs a visibility mask to prevent injected knowledge tokens from attending to irrelevant input positions, preserving original semantics. UniD2T (Li et al., 2024) constructs attention masks based on the connectivity of a unified graph derived from structured data (e.g., tables, knowledge graphs) to enforce structure-aware interactions during pre-training. In this work, based on the homophily analysis in Section A, we design a TAG-specific attention masking strategy to inject structural information at both word-token and node granularities. 3 Preliminaries 3.1 Problem Formulation Definition 1. Text-Attributed Graph. A text-attributed graph (TAG) is formally defined as a triplet =(,ℰ,)G=(V,E,T). Here, =v1,v2,…,vNV=\v_1,v_2,…,v_N\ is the set of N nodes, and ℰ⊆×E ×V is the set of edges describing the graph’s topological structure, which can be represented by an adjacency matrix ∈0,1N×NA∈\0,1\^N× N. Each node vi∈v_i is associated with a text description iw_i, and =1,2,…,NT=\w_1,w_2,…,w_N\ denotes the collection of all node-associated text descriptions, where each i=(wi1,wi2,…,wiLi)w_i=(w_i1,w_i2,…,w_iL_i) is a sequence of word tokens of length LiL_i. Definition 2. Node Classification in Text-Attributed Graphs. Given a text-attributed graph G and a set of K predefined classes =c1,c2,…,cKC=\c_1,c_2,…,c_K\, the task of node classification aims to learn a mapping function f:→f:V . The objective of this function is to predict the correct label yi∈y_i for every node vi∈v_i by jointly considering the graph structure ℰE and the semantic information T. 3.2 Transformer and Self-Attention with Masking The Transformer architecture utilizes self-attention to capture dependencies within sequences. Given input ∈ℝn×d X ^n× d, query, key, and value projections are computed as =Q Q= X W_Q, =K K= X W_K, =V V= X W_V. The process is: Attention(,,)=softmax(⊤dk+),Attention( Q, K, V)=softmax ( Q K d_k+ M ) V, (1) where M is derived from a binary mask matrix mask∈0,1n×n M_mask∈\0,1\^n× n: valid attention positions are marked as 11 in mask M_mask, and their corresponding entries in M are set to 0; invalid positions are marked as 0 in mask M_mask, and their entries in M are set to −∞-∞. This mechanism enables the model to selectively attend to semantic interactions between specific tokens, a property that we leverage to design our topology-constrained attention mechanism. 4 Method In this section, we propose DuConTE illustrated in Figure 2, a dual-granularity text encoder with topology-constrained attention. It employs two language models as a word-token encoder ℳLM_L and a node encoder ℳNM_N respectively, both incorporating topology-constrained attention mechanisms. Given a target node viv_i and its neighborhood (vi)N(v_i), DuConTE first concatenates the textual content of viv_i and all nodes in (vi)N(v_i), and applies ℳLM_L to this combined sequence to generate word-token representations. A node representation composer then aggregates these into first-stage node representations. Subsequently, ℳNM_N encodes the sequence of first-stage node representations to produce a second-stage node representation for viv_i. The final representation i o_i is obtained through a weighted fusion of the node’s first-stage and second-stage representations. Figure 2: Overview of DuConTE with the dual-granularity cascaded architecture (middle), the topology-constrained attention mechanism (left), and the target node representation construction process in the node representation composer (right). The node representation composer is denoted as Composer in the figure. 4.1 Dual-Granularity Semantic Encoding To capture semantics at the word-token and node granularities, which naturally exist in text graphs, we propose a dual-granularity cascaded architecture, illustrated in the middle of Figure 2. This architecture employs the word-token encoder ℳLM_L for the word-token granularity and the node encoder ℳNM_N for the node granularity, in a sequential manner. Word-Token Granularity Encoding. Given a target node vi∈v_i and its neighborhood (vi)⊆N(v_i) , let S(i)=vi∪(vi)S^(i)=\v_i\ (v_i) denote the set consisting of the target node and its neighbors. For each node vj∈S(i)v_j∈ S^(i), we obtain its associated word-token sequence j=(wj1,…,wjLj)∈w_j=(w_j1,…,w_jL_j) . These sequences are concatenated with [SEP] tokens inserted between adjacent nodes to form a unified neighborhood input: (i)=[j1;[SEP];⋯;j|(vi)|;[SEP];i]∈ℝL×dL,W^(i)=[w_j_1; [SEP];·s;w_j_|N(v_i)|; [SEP];w_i] ^L× d_L, (2) where vj1,…,vj|(vi)|∈(vi)v_j_1,…,v_j_|N(v_i)| (v_i). The word-token encoder ℳLM_L (a pre-trained LM) processes (i)W^(i) to perform semantic interaction at the word-token granularity, producing word-token embeddings (i)∈ℝL×dL H^(i) ^L× d_L: (i)=ℳL((i))=[j1(i);SEP1(i);…;i(i)], H^(i)=M_L(W^(i))= [ h_j_1^(i); h_SEP_1^(i);…; h_i^(i) ], (3) where j(i)∈ℝLj×dL h_j^(i) ^L_j× d_L is the embedding matrix for the tokens of node vjv_j after such interaction, SEPk(i) h_SEP_k^(i) denotes the embedding of the k-th [SEP] token, and dLd_L is the hidden dimension of ℳLM_L. To distill these fine-grained word-token features into node semantics, we employ a node representation composer f, detailed in Section 4.3. This function maps (i) H^(i) to a sequence of first-stage node representations (i) Z^(i): (i) Z^(i) =f((i)), =f ( H^(i) ), (4) (i) Z^(i) =[j1(i);…;j|(vi)|(i);i(i)], =[ z_j_1^(i);…; z_j_|N(v_i)|^(i); z_i^(i)], (5) where each j(i)∈ℝdL z_j^(i) ^d_L denotes the first-stage node representation of vjv_j. Node Granularity Encoding. To further model semantic interactions at the node granularity, we feed (i) Z^(i) into node encoder ℳNM_N(another pre-trained LM), to produce a sequence of second-stage node representations (i) E^(i): (i) E^(i) =ℳN((i))∈ℝ(k+1)×dL, =M_N( Z^(i)) ^(k+1)× d_L, (6) (i) E^(i) =[j1(i);…;j|(vi)|(i);i(i)], =[ e_j_1^(i);…; e_j_|N(v_i)|^(i); e_i^(i)], (7) where each j(i)∈ℝdL e_j^(i) ^d_L denotes the second-stage node representation of vjv_j. Note that for vj∈(vi)v_j (v_i), j(i) z_j^(i) and j(i) e_j^(i) are computed within the context of target node viv_i, and thus represents a context-dependent, neighbor-oriented encoding—distinct from the representation obtained when vjv_j is treated as a target node. Dual-Granularity Representation Fusion. To integrate complementary semantic information from both granularities, we compute the final representation of the target node viv_i through a weighted combination of its first-stage and second-stage node representations: i=α⋅i(i)+(1−α)⋅i(i), o_i=α· z_i^(i)+(1-α)· e_i^(i), (8) where α∈[0,1]α∈[0,1] is a fixed fusion coefficient. 4.2 Topology-constrained attention mechanism To endow our dual-granularity encoders with topological awareness, we transform their standard self-attention mechanism into a topology-constrained variant, as illustrated on the left in Figure 2. This is achieved through an attention masking strategy specifically designed for TAG. Informed by the homophily analysis in Section A, it constructs masks based on node connectivity, applied at every layer and attention head to restrict attention exclusively between structurally connected word-tokens or nodes. The approach seamlessly integrates graph information without altering the core Transformer architecture. Word-Token Mask Construction. For the word-token encoder ℳLM_L processing sequence (i)∈ℝL×dLW^(i) ^L× d_L, we allow attention only between pairs of word-tokens within the same node or in connected nodes. Additionally, attention between [SEP] tokens and any word-token is always allowed to preserve a basic awareness of inter-node boundaries at the word-token granularity. Accordingly, the attention mask matrix maskword M^word_mask is constructed as follows: for any two tokens at positions p and q in (i)W^(i), if neither token is a [SEP] token, let v(p)v(p) and v(q)v(q) denote their associated nodes in the graph. The entry p,qword∈0,1L×L M_p,q^word∈\0,1\^L× L is defined as: p,qword=1if the token at p or q is [SEP],1if v(p)=v(q) or (v(p),v(q))∈ℰ,0otherwise. M_p,q^word= cases1&if the token at p or q is [SEP],\\ 1&if v(p)=v(q) or (v(p),v(q)) ,\\ 0&otherwise. cases (9) Node Mask Construction. For the node encoder ℳNM_N processing the sequence (i)∈ℝ(k+1)×dL Z^(i) ^(k+1)× d_L, we allow attention only between node representations that correspond to the same node or to connected nodes in the graph. Accordingly, the attention mask matrix masknode M^node_mask is constructed as follows: for any two positions m and n in (i) Z^(i), let v(m)v(m) and v(n)v(n) denote the corresponding nodes in the graph. The entry m,nnode∈0,1(k+1)×(k+1) M^node_m,n∈\0,1\^(k+1)×(k+1) is defined as: m,nnode=1if v(m)=v(n) or (v(m),v(n))∈ℰ,0otherwise. M^node_m,n= cases1&if v(m)=v(n) or (v(m),v(n)) ,\\ 0&otherwise. cases (10) 4.3 Node Representation Composer To effectively fuse the word-token embeddings (i) H^(i) into high-quality first-stage node representations, we design a Node Representation Composer f. The composer employs two distinct modules: a more sophisticated module f1f_1 to compute the representation of the target node viv_i, and a lightweight module f2f_2 to independently encode each neighbor node vj∈(i)v_j (i). This asymmetric design enables the target node to capture rich contextual information while ensuring efficient and undisturbed representation learning for neighbors. Target Node Representation Construction. To capture the most salient semantics of the target node viv_i under both center-node and neighborhood context—and to explicitly balance their relative influence—we design f1f_1 to assess word-token significance from dual perspectives, as shown on the right in Figure 2. Specifically, f1f_1 employs a specialized attention mechanism to compute the importance of each word-token in the target node’s text iw_i. With learnable projection matrices Q,K∈ℝdL×dL W_Q, W_K ^d_L× d_L, we compute the queries (i) Q^(i) as the projected embeddings of all word-tokens in the neighborhood, and the keys (i) K^(i) as the projected embeddings of the target node’s word-tokens: (i) Q^(i) =(i)Q∈ℝL×dL, = H^(i) W_Q ^L× d_L, (11) (i) K^(i) =i(i)K∈ℝLi×dL. = h_i^(i) W_K ^L_i× d_L. (12) As defined in 3.1, wjpw_jp is the p-th word-token in node vjv_j. The attention weight aj,p,q(i)a_j,p,q^(i) from wjpw_jp to wiqw_iq is computed using the scaled dot-product attention mechanism, with softmax normalization over all queries attending to wiqw_iq. The total importance of wiqw_iq is decomposed into two components: • Importance under center-node context: αqcen=∑p=1Liai,p,q(i) _q^cen= _p=1^L_ia_i,p,q^(i); • Importance under neighborhood context: αqneigh=∑vj∈(i)∑p=1Ljaj,p,q(i) _q^neigh= _v_j (i) _p=1^L_ja_j,p,q^(i). Each component is independently normalized via softmax to obtain μqcen _q^cen and μqneigh _q^neigh, which are fused into the final importance score μq _q using a fixed coefficient β∈[0,1]β∈[0,1]: μq=β⋅μqcen+(1−β)⋅μqneigh. _q=β· _q^cen+(1-β)· _q^neigh. (13) The final representation i(i) z_i^(i) is a weighted sum over the target node’s word-token embeddings: i(i)=∑q=1Liμqi,q(i). z_i^(i)= _q=1^L_i _q h_i,q^(i). (14) Neighbor Node Representation Construction. To enable efficient encoding while preserving each neighbor’s intrinsic semantic content, we design a lightweight module f2f_2 that employs local attention pooling. Given a neighbor node vj∈(i)v_j (i), an importance score sj,ps_j,p is computed for each word-token embedding j,p(i) h_j,p^(i) via a learnable projection vector a∈ℝdL w_a ^d_L. After softmax normalization to obtain weights πj,p _j,p, the first-stage representation of vjv_j is computed as a weighted sum: j(i)=∑p=1Ljπj,pj,p(i). z_j^(i)= _p=1^L_j _j,p h_j,p^(i). (15) 4.4 Two-stage training procedure We train DuConTE using a two-stage procedure. We first train ℳLM_L and f1f_1 to learn high-quality first-stage node representations, then train ℳNM_N and f2f_2 based on these representations. The full training procedure is detailed in Appendix H. 5 Experiments 5.1 Datasets In this paper, we evaluate DuConTE for node classification on five widely-used datasets: Cora (Sen et al., 2008), CiteSeer (Giles et al., 1998), WikiCS (Mernyei and Cangea, 2007), ArXiv-2023 (He et al., 2023) and OGBN-Products (Hu et al., 2020). For detailed descriptions of each dataset, please refer to Appendix L. 5.2 Baselines To evaluate the effectiveness of our proposed model, we employ several baseline models for comparison. For a detailed description of all baseline models, please refer to Appendix I. These baselines can be categorized into three main types: • Graph-Specific Models: Models specifically designed and trained from scratch for graph-structured data, e.g., NodeFormer (Wu et al., 2022), GraphFormers (Yang et al., 2021). • Pure LMs: Language models that perform inference solely based on node texts while completely ignoring the graph structure, e.g., BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019). • Recent TAG Methods: Leading approaches that have demonstrated strong performance on text-attributed graph benchmarks, e.g., GraphBridge (Wang et al., 2024), ENGINE (Zhu et al., 2024). Table 1: Experiment results: Mean accuracy and standard deviation over 10 runs with different random seeds. Bold indicates the best performance, underlined denotes the second-best, and ‘–’ signifies that the method is not applicable to the dataset.“DuConTE” refers to the pipeline instance using DuConTE as the text encoder, as described in Section 5.3. Methods Cora CiteSeer WikiCS ArXiv-2023 OGBN-Products GraphFormers 80.29±1.7480.29± 1.74 71.84±1.2371.84± 1.23 71.37±0.3571.37± 0.35 63.14±0.5963.14± 0.59 68.09±0.5768.09± 0.57 NodeFormer 88.24±0.3488.24± 0.34 74.96±0.6174.96± 0.61 75.56±0.5175.56± 0.51 67.68±0.4767.68± 0.47 67.37±0.8367.37± 0.83 GraphSAGE 87.42±1.3187.42± 1.31 72.26±1.2172.26± 1.21 76.91±0.7776.91± 0.77 68.56±0.5368.56± 0.53 70.56±0.2770.56± 0.27 BERT 79.63±1.8179.63± 1.81 71.27±1.1171.27± 1.11 77.96±0.5777.96± 0.57 76.84±0.0976.84± 0.09 76.45±0.1676.45± 0.16 Sentence-BERT 78.94±1.4378.94± 1.43 72.93±1.8472.93± 1.84 77.84±0.0677.84± 0.06 77.41±0.5577.41± 0.55 74.98±0.1574.98± 0.15 RoBERTa-base 78.37±1.2978.37± 1.29 71.76±1.2371.76± 1.23 76.86±0.5276.86± 0.52 77.24±0.1977.24± 0.19 76.03±0.1276.03± 0.12 RoBERTa-large 79.81±1.3779.81± 1.37 72.31±1.7472.31± 1.74 77.64±0.9577.64± 0.95 77.81±0.4377.81± 0.43 76.24±0.3576.24± 0.35 GLEM 87.59±0.1787.59± 0.17 77.42±0.6877.42± 0.68 78.23±0.5678.23± 0.56 79.23±0.1779.23± 0.17 76.04±0.3476.04± 0.34 TAPE 87.48±0.7687.48± 0.76 – – 80.04±0.3180.04± 0.31 79.23±0.13 79.23± 0.13 SimTeG 86.74±1.7186.74± 1.71 78.51±1.0478.51± 1.04 79.73±0.8479.73± 0.84 79.45±0.5379.45± 0.53 76.43±0.4976.43± 0.49 ENGINE 87.61±1.3487.61± 1.34 76.84±1.4176.84± 1.41 77.92±0.8977.92± 0.89 78.57±0.1978.57± 0.19 77.68±1.3177.68± 1.31 GraphBridge 93.60±0.98¯ 93.60± 0.98 88.62±0.76¯ 88.62± 0.76 80.47±0.26¯ 80.47± 0.26 86.43±0.29¯ 86.43± 0.29 77.92±0.2777.92± 0.27 DuConTE 95.24±0.79 95.24± 0.79 89.45±1.22 89.45± 1.22 81.09±0.43 81.09± 0.43 90.31±0.35 90.31± 0.35 78.80±0.10¯ 78.80± 0.10 5.3 Experimental Settings Evaluation Task and Metric. In this study, we focus on node classification as the downstream task for text-attributed graphs, and adopt classification accuracy as the evaluation metric. Implementation Details. We instantiate a text-attributed graph learning pipeline, as illustrated in the top panel of Figure 1. DuConTE serves as the text encoder in this pipeline, implemented with two RoBERTa-base models serving as the word-token encoder and node encoder respectively. In the downstream phase, a two-layer GraphSAGE with a hidden dimension of 64 is employed as the GNN component. All methods are evaluated under a unified experimental protocol to ensure a fair comparison. Detailed configurations for model hyperparameters, upstream preprocessing, implementation settings of baseline methods, and training procedures are provided in Appendix J. 5.4 Performance Comparison and Discussions We compare the performance of various models on text-attributed graph node classification, with results reported in Table 1. DuConTE achieves state-of-the-art performance on most datasets, outperforming the second-best method by 2.7% on ArXiv-2023 and 1.6% on Cora. The results demonstrate DuConTE’s ability to produce high-quality, semantically rich node representations that effectively support downstream GNN models. 6 Analysis 6.1 Sensitivity Analysis Figure 3: Sensitive analysis of the fusion coefficient α Figure 4: Sensitive analysis of the fusion coefficient β. We analyze the sensitivity of DuConTE to the fusion coefficients α and β over the range [0,1][0,1]. The performance trends are shown in Figure 4 and Figure 4. Across all experiments, the performance variation remains within 1%, demonstrating the model’s robustness to these hyperparameters. For α, which controls the fusion of dual-granularity semantic representations, the optimal performance on Cora and CiteSeer falls within the range [0.7,0.9][0.7,0.9]. This indicates a clear fusion pattern: word-token granularity semantics provide stable and reliable information, while node granularity semantics contribute complementary yet essential signals—consistent with their role as more abstract, high-level features. For β, which balances the influence of center-node and neighborhood contexts in word-token importance assessment, the performance trend varies across datasets, indicating that the relative importance of these two contexts is dataset-dependent. On Cora and CiteSeer, strong performance is observed within [0.4,0.7][0.4,0.7] and [0.2,0.8][0.2,0.8], respectively, confirming that both contexts contribute meaningfully. Notably, the optimal values consistently fall within [0.6,0.8][0.6,0.8], suggesting that the center-node context exerts a stronger influence—aligning with the intuition that a token’s relevance is primarily shaped by the target node itself. 6.2 Ablation Study We conduct ablation studies to evaluate the three key innovations in DuConTE. The variants are defined in Appendix M, including NoDual, NoMask-T/D/Both, and MeanPool/Center-Only/Neigh-Only/UnifiedContext. All variants are evaluated under the same experimental setup. As shown in Table 2, DuConTE outperforms all variants, confirming the effectiveness of its three key designs: (1) DuConTE surpasses NoDual by +0.8% on Cora and OGBN-Products, verifying that dual-granularity encoding aligns with the inherent semantic granularity of text-structured graphs and thus better captures rich semantic information. (2) Performance drops in NoMask-T/D/Both confirm that topology-constrained attention effectively injects structural information at both word-token and node granularities; notably, NoMask-D consistently outperforms NoMask-T, suggesting that structural information is critical even at the finest semantic granularity. (3) The lower performance of MeanPool further validates that importance-based weighted fusion captures key semantic information more effectively than uniform averaging. Gains over Center-Only, Neigh-Only, and UnifiedContext demonstrate that both center-node and neighborhood contexts are important for assessing word-token importance, and explicitly differentiating their distinct influences leads to more accurate semantic weighting. Table 2: Ablation results on Cora, CiteSeer, and OGBN-Products Methods Cora CiteSeer OGBN-Products NoDual 94.46±0.7694.46± 0.76 89.22±1.3489.22± 1.34 77.98±0.3877.98± 0.38 NoMask-T 94.23±0.7694.23± 0.76 88.84±1.2888.84± 1.28 78.19±0.1378.19± 0.13 NoMask-D 94.59±0.5894.59± 0.58 88.86±1.2788.86± 1.27 78.52±0.1578.52± 0.15 NoMask-Both 94.10±0.8594.10± 0.85 89.04±0.9989.04± 0.99 78.40±0.1778.40± 0.17 MeanPool 94.43±0.9494.43± 0.94 88.57±0.9588.57± 0.95 78.27±0.1278.27± 0.12 Center-Only 95.13±0.8095.13± 0.80 88.46±1.2088.46± 1.20 78.17±0.1878.17± 0.18 Neigh-Only 95.09±0.7495.09± 0.74 88.71±1.4088.71± 1.40 78.36±0.1578.36± 0.15 UnifiedContext 95.09±0.8695.09± 0.86 88.98±1.1088.98± 1.10 78.56±0.2378.56± 0.23 DuConTE 95.24±0.7995.24± 0.79 89.45±1.2289.45± 1.22 78.80±0.1078.80± 0.10 7 Conclusion In this paper, we introduce DuConTE, a dual-granularity text encoder with topology-constrained attention for text-attributed graphs. DuConTE encodes node semantics at both word-token and node granularity to capture the inherent dual-granularity semantic structure of text-attributed graphs. Our topology-constrained attention mechanism utilizes an attention masking strategy specifically designed for TAG, offering an effective and architecture-preserving approach to adapt LMs to graph-structured data. In the node representation composer, the contexts of the center node and its neighborhood are separately considered to more effectively assess the semantic importance of word-tokens in the target node. Extensive experiments on multiple benchmark datasets show that DuConTE achieves state-of-the-art performance on the majority of them. References D. Chen, T. H. Schulz, and K. M. Borgwardt (2025) Learning long range dependencies on graphs via random walks. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, Cited by: §2.2. R. Chen, T. Zhao, A. Jaiswal, N. Shah, and Z. Wang (2024) Llaga: large language and graph assistant. arXiv preprint arXiv:2402.08170. Cited by: §2.1. Z. Chen, H. Mao, H. Wen, H. Han, W. Jin, H. Zhang, H. Liu, and J. Tang (2023) Label-free node classification on graphs with large language models (llms). arXiv preprint arXiv:2310.04668. Cited by: §1, §2.1. E. Chien, W. Chang, C. Hsieh, H. Yu, J. Zhang, O. Milenkovic, and I. S. Dhillon (2021) Node feature extraction by self-supervised multi-scale neighborhood prediction. arXiv preprint arXiv:2111.00064. Cited by: §1, §2.1. J. Devlin, M. Chang, K. Lee, and K. Toutanova (2019) Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), p. 4171–4186. Cited by: Appendix I, 2nd item. K. Duan, Q. Liu, T. Chua, S. Yan, W. T. Ooi, Q. Xie, and J. He (2023) Simteg: a frustratingly simple approach improves textual graph learning. arXiv preprint arXiv:2308.02565. Cited by: Appendix I. C. L. Giles, K. D. Bollacker, and S. Lawrence (1998) CiteSeer: an automatic citation indexing system. In Proceedings of the third ACM conference on Digital libraries, p. 89–98. Cited by: Appendix L, Appendix B, §5.1. W. L. Hamilton, Z. Ying, and J. Leskovec (2017) Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, p. 1024–1034. Cited by: Appendix I. X. He, X. Bresson, T. Laurent, A. Perold, Y. LeCun, and B. Hooi (2023) Harnessing explanations: llm-to-lm interpreter for enhanced text-attributed graph representation learning. arXiv preprint arXiv:2305.19523. Cited by: §J.2, Appendix L, Appendix L, Appendix B, Appendix I, §1, §5.1. Z. Hou, Y. He, Y. Cen, X. Liu, Y. Dong, E. Kharlamov, and J. Tang (2023) Graphmae2: a decoding-enhanced masked self-supervised graph learner. In Proceedings of the ACM web conference 2023, p. 737–746. Cited by: §2.2. W. Hu, M. Fey, M. Zitnik, Y. Dong, H. Ren, B. Liu, M. Catasta, and J. Leskovec (2020) Open graph benchmark: datasets for machine learning on graphs. Advances in neural information processing systems 33, p. 22118–22133. Cited by: §J.2, Appendix L, Appendix B, §5.1. B. Jin, Y. Zhang, Q. Zhu, and J. Han (2023) Heterformer: transformer-based deep node representation learning on heterogeneous text-rich networks. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long Beach, CA, USA, August 6-10, 2023, A. K. Singh, Y. Sun, L. Akoglu, D. Gunopulos, X. Yan, R. Kumar, F. Ozcan, and J. Ye (Eds.), p. 1020–1031. Cited by: §1. S. Li, L. Li, R. Geng, M. Yang, B. Li, G. Yuan, W. He, S. Yuan, C. Ma, F. Huang, et al. (2024) Unifying structured data as graph for data-to-text pre-training. Transactions of the Association for Computational Linguistics 12, p. 210–228. Cited by: §2.2. W. Liu, P. Zhou, Z. Zhao, Z. Wang, Q. Ju, H. Deng, and P. Wang (2020) K-bert: enabling language representation with knowledge graph. In Proceedings of the AAAI conference on artificial intelligence, Vol. 34, p. 2901–2908. Cited by: §2.2. Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov (2019) Roberta: a robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Cited by: Appendix I, 2nd item. P. Mernyei and C. W. Cangea (2007) A wikipedia-based benchmark for graph neural networks. arxiv 2020. arXiv preprint arXiv:2007.02901. Cited by: §J.2, Appendix L, Appendix B, §5.1. L. Rampášek, M. Galkin, V. P. Dwivedi, A. T. Luu, G. Wolf, and D. Beaini (2022) Recipe for a general, powerful, scalable graph transformer. Advances in Neural Information Processing Systems 35, p. 14501–14515. Cited by: §2.2. N. Reimers and I. Gurevych (2019) Sentence-bert: sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084. Cited by: Appendix I. V. G. Satorras, E. Hoogeboom, and M. Welling (2021) E (n) equivariant graph neural networks. In International conference on machine learning, p. 9323–9332. Cited by: §2.2. P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Galligher, and T. Eliassi-Rad (2008) Collective classification in network data. AI magazine 29 (3), p. 93–93. Cited by: Appendix L, Appendix B, §5.1. H. Seo, T. Kim, J. Y. Yang, and E. Yang (2024) Unleashing the potential of text-attributed graphs: automatic relation decomposition via large language models. CoRR abs/2405.18581. Cited by: §1. A. Shehzad, F. Xia, S. Abid, C. Peng, S. Yu, D. Zhang, and K. Verspoor (2024) Graph transformers: A survey. CoRR abs/2407.09777. Cited by: §2.2. Y. Tian, C. Zhang, Z. Kou, Z. Liu, X. Zhang, and N. V. Chawla (2024) Ugmae: a unified framework for graph masked autoencoders. arXiv preprint arXiv:2402.08023. Cited by: §2.2. Y. Wang, Y. Zhu, W. Zhang, Y. Zhuang, L. Liyunfei, and S. Tang (2024) Bridging local details and global context in text-attributed graphs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 12-16, 2024, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), p. 14830–14841. Cited by: §J.4, Appendix I, §1, §2.1, 3rd item. Z. Wang, S. Liu, Z. Zhang, T. Ma, C. Zhang, and Y. Ye (2025) Can llms convert graphs to text-attributed graphs?. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2025 - Volume 1: Long Papers, Albuquerque, New Mexico, USA, April 29 - May 4, 2025, L. Chiruzzo, A. Ritter, and L. Wang (Eds.), p. 1412–1432. Cited by: §1. Q. Wu, W. Zhao, Z. Li, D. P. Wipf, and J. Yan (2022) NodeFormer: A scalable graph structure learning transformer for node classification. In NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, Cited by: Appendix I, 1st item. J. Yang, Z. Liu, S. Xiao, C. Li, D. Lian, S. Agrawal, A. Singh, G. Sun, and X. Xie (2021) GraphFormers: gnn-nested transformers for representation learning on textual graph. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, M. Ranzato, A. Beygelzimer, Y. N. Dauphin, P. Liang, and J. W. Vaughan (Eds.), p. 28798–28810. Cited by: Appendix I, §1, 1st item. C. Ying, T. Cai, S. Luo, S. Zheng, G. Ke, D. He, Y. Shen, and T. Liu (2021) Do transformers really perform badly for graph representation?. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, M. Ranzato, A. Beygelzimer, Y. N. Dauphin, P. Liang, and J. W. Vaughan (Eds.), p. 28877–28888. Cited by: §2.2. D. C. Zhang, M. Yang, R. Ying, and H. W. Lauw (2024) Text-attributed graph representation learning: methods, applications, and challenges. In Companion Proceedings of the ACM on Web Conference 2024, W 2024, Singapore, Singapore, May 13-17, 2024, T. Chua, C. Ngo, R. K. Lee, R. Kumar, and H. W. Lauw (Eds.), p. 1298–1301. Cited by: §1. J. Zhang, H. Zhang, C. Xia, and L. Sun (2020) Graph-bert: only attention is needed for learning graph representations. arXiv preprint arXiv:2001.05140. Cited by: §2.2. Y. Zhang, R. Jin, and Z. Zhou (2010) Understanding bag-of-words model: a statistical framework. Int. J. Mach. Learn. Cybern. 1 (1-4), p. 43–52. Cited by: §2.1. H. Zhao, B. Yang, Y. Cen, J. Ren, C. Zhang, Y. Dong, E. Kharlamov, S. Zhao, and J. Tang (2024) Pre-training and prompting for few-shot node classification on text-attributed graphs. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, R. Baeza-Yates and F. Bonchi (Eds.), p. 4467–4478. Cited by: §1. J. Zhao, M. Qu, C. Li, H. Yan, Q. Liu, R. Li, X. Xie, and J. Tang (2022) Learning on large-scale text-attributed graphs via variational inference. arXiv preprint arXiv:2210.14709. Cited by: Appendix I. Y. Zhu, Y. Wang, H. Shi, and S. Tang (2024) Efficient tuning and inference for large language models on textual graphs. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI 2024, Jeju, South Korea, August 3-9, 2024, p. 5734–5742. Cited by: Appendix I, §1, §2.1, 3rd item. Appendix A Why Topology-Constrained Attention Works: A Homophily Perspective In this subsection, we analyze the effectiveness of topology-constrained attention from the perspective of the homophily assumption, which posits that connected nodes in a graph are more likely to share similar semantic properties. To the best of our knowledge, this assumption is well-supported by most widely used text-attributed graph benchmarks, where adjacent nodes are more likely to belong to the same class.This is further supported by the homophily statistics reported in Appendix B. In the topology-constrained attention mechanism, the masks masktoken M^token_mask and masknode M^node_mask are injected into the attention layers of the word-token encoder and the node encoder, respectively. As a result, cross-node attention interactions are constrained to occur between semantic information from connected nodes at both granularities. Under the homophily assumption, such information is more likely to be semantically related, thereby enabling mutually complementary interactions. This allows the model to effectively leverage the graph structure to learn higher-quality representations. Appendix B Homophily Analysis In this section, we analyze the homophily of the five datasets used in our experiments: Cora (Sen et al., 2008), CiteSeer (Giles et al., 1998), WikiCS (Mernyei and Cangea, 2007), ArXiv-2023 (He et al., 2023) and OGBN-Products (subset) (Hu et al., 2020). Specifically, we compute the label homophily ratio H, defined as: H=1|ℰ|∑(i,j)∈ℰ(yi=yj),H= 1|E| _(i,j) I(y_i=y_j), (16) where ℰE denotes the set of edges, yiy_i is the class label of node i, and (⋅)I(·) is the indicator function that equals 1 if the condition is true and 0 otherwise. This metric measures the proportion of edges connecting nodes with identical labels; a higher value indicates stronger homophily. The results are summarized in Table 3. Table 3: Label Homophily Ratios Across Datasets Dataset Cora CiteSeer WikiCS ArXiv-2023 OGBN-Products (subset) Homophily (H) 0.8100 0.7451 0.6547 0.6465 0.7950 According to the results, all datasets exhibit homophily ratios above 0.6, indicating a relatively high level of homophily. Appendix C Additional Evaluation on Link Prediction To assess the general applicability of DuConTE beyond node classification, we conduct link prediction experiments on the Cora, CiteSeer, and ArXiv-2023 datasets, using AUC as the evaluation metric. Detailed configurations and training procedures are provided in Appendix K. According to Table 4, DuConTE consistently outperforms baseline methods on the link prediction task, indicating that it is highly effective at representation learning on text-attributed graphs. This result further highlights the versatility of DuConTE and its potential for broader applications across diverse TAG-based tasks. Table 4: Experimental results on link prediction Methods Cora CiteSeer ArXiv-2023 GraphSAGE 97.10±0.4397.10± 0.43 87.29±1.2287.29± 1.22 91.81±0.2691.81± 0.26 SimTeG 97.86±0.4497.86± 0.44 90.06±1.3490.06± 1.34 93.12±0.4693.12± 0.46 GraphBridge 98.07±0.7798.07± 0.77 91.86±1.0391.86± 1.03 94.35±0.6594.35± 0.65 DuConTE 99.13±0.1999.13± 0.19 93.29±0.7593.29± 0.75 95.40±0.3395.40± 0.33 Appendix D Parameter Efficiency Analysis To evaluate the parameter efficiency of DuConTE, we replace the LM backbone in baseline methods with RoBERTa-large (340M parameters) while keeping other configurations unchanged. We then compare their performance against DuConTE using two RoBERTa-base models (150M parameters each) as its LM backbones. In this setup, every baseline has a larger total parameter count than DuConTE. TAPE is excluded from the comparison as it relies on a large language model. As shown in Table 5, DuConTE achieves the best performance despite using fewer parameters, highlighting its parameter efficiency. This suggests a novel parameter-efficient scaling paradigm: rather than improving performance by scaling up a single large LM, DuConTE achieves greater gains with fewer total parameters by leveraging two smaller LMs. Table 5: Experiment results: Subscript (large) indicates the use of RoBERTa-large as the LM backbone, while (base) indicates RoBERTa-base. Methods Cora CiteSeer WikiCS ArXiv-2023 OGBN-Products GLEM(large) 89.07±0.2589.07± 0.25 78.04±0.3678.04± 0.36 78.14±0.8178.14± 0.81 78.94±0.4578.94± 0.45 78.37±0.2978.37± 0.29 SimTeG(large) 88.64±0.8988.64± 0.89 79.89±1.2379.89± 1.23 80.16±0.6580.16± 0.65 80.69±0.4980.69± 0.49 78.31±0.6178.31± 0.61 ENGINE(large) 88.57±1.2588.57± 1.25 78.14±0.7478.14± 0.74 80.36±0.2480.36± 0.24 77.37±0.4377.37± 0.43 78.44±0.5778.44± 0.57 GraphBridge(large) 94.06±0.94¯ 94.06± 0.94 88.91±0.98¯ 88.91± 0.98 80.96±0.57¯ 80.96± 0.57 87.14±0.36¯ 87.14± 0.36 78.51±0.6878.51± 0.68 DuConTE(base) 95.24±0.7995.24± 0.79 89.45±1.2289.45± 1.22 81.09±0.4381.09± 0.43 90.31±0.3590.31± 0.35 78.80±0.1078.80± 0.10 Appendix E Computational Overhead of the Node Representation Composer We measure the training and inference time of DuConTE and its ablation variant MeanPool on Cora and CiteSeer. As reported in Appendix F, the Node Representation Composer introduces an average overhead of 23.8% in training time and 19.9% in inference time. This cost is generally acceptable, and further acceleration is possible by reducing the dimensionality of keys and queries in f1f_1 to lower computational load. A key direction for future work is to design methods that convert word-token embeddings into node representations with both higher performance and lower computational cost. This is crucial for TAG representation learning but remains underexplored. Appendix F Computational Overhead Statistics We report the total training time (over 8 epochs) and single-pass inference time on the full dataset for DuConTE and its ablation variant MeanPool across Cora and CiteSeer. All timing measurements were conducted on a system equipped with four NVIDIA GeForce RTX 4090 GPUs, each with 24GB of memory. Table 6: Total Training Time (seconds) Method Cora(training) CiteSeer(training) DuConTE 1054 434 MeanPool 880 326 Table 7: Total Inference Time (seconds) Method Cora(inference) CiteSeer(inference) DuConTE 185 62 MeanPool 163 49 Appendix G Reproducibility Statement Dataset description. We provide a detailed description of the datasets, including information on their sources, in Appendix L. We describe the dataset splitting strategy in Appendix J.2. Baseline description. We provide a detailed description of the baseline models we used and include links to their source code in Appendix I. Implementation details. We provide a detailed description of the model hyperparameter settings and training configurations in Appendix J to facilitate reproducibility. Open access to code. The source code of DuConTE is included as a ZIP file in the supplementary materials. We will release it publicly via an open-source repository upon publication. Appendix H Two-Stage Training Procedure of DuConTE We train DuConTE using a two-stage procedure: the word-token encoder is trained first to learn high-quality representations, and the node encoder is then trained based on these representations. Stage 1: Word-Token Encoder Training. We first train the word-token encoder ℳLM_L and the aggregator f1f_1, while the node encoder ℳNM_N and the aggregator f2f_2 are not involved in this stage. The first-stage representation of the target node, i(i) z_i^(i), serves as input to a learnable linear classifier cls(1)W_cls^(1). The objective is to minimize the standard cross-entropy loss over the training set trainV_train: ℒ1=−∑i∈traini⊤log(softmax(cls(1)i(i))).L_1=- _i _train y_i (softmax(W_cls^(1) z_i^(i))). (17) Stage 2: Node Encoder Training. We then fix ℳLM_L and f1f_1, and train the node encoder ℳNM_N and the aggregator f2f_2. The final node representation i o_i is fed to a new learnable classifier cls(2)W_cls^(2) for prediction. The objective is to minimize the cross-entropy loss: ℒ2=−∑i∈traini⊤log(softmax(cls(2)i)).L_2=- _i _train y_i (softmax(W_cls^(2) o_i)). (18) Appendix I Baseline Graph-Specific Models: We adopt two graph transformers: GraphFormers (Yang et al., 2021)[Code] and NodeFormer (Wu et al., 2022)[Code]. We also adopt GraphSAGE (Hamilton et al., 2017)[Code], a Graph Neural Network, which also serves as the GNN backbone for other baseline models. Pure LMs: We adopt four commonly used pre-trained language models: BERT (Devlin et al., 2019)[Code], Sentence-BERT (Reimers and Gurevych, 2019)[Code], and two versions of RoBERTa (Liu et al., 2019): RoBERTa-base[Code] and RoBERTa-large[Code]. Recent TAG Methods: GLEM (Zhao et al., 2022)[Code], is a framework that integrates language models and GNNs during training using a variational EM approach. TAPE (He et al., 2023)[Code], leverages large language models such as ChatGPT to generate pseudo labels and explanations for textual nodes. These are then used to fine-tune pre-trained language models alongside the original texts. SimTeG (Duan et al., 2023)[Code] uses a cascading structure specifically designed for textual graphs. It employs a two-stage training paradigm: first, it fine-tunes language models, and then it trains GNNs. ENGINE (Zhu et al., 2024)[Code] is an efficient fine-tuning and inference framework for text-attributed graphs. It co-trains large language models and GNNs using a ladder-side approach to optimize both memory and time efficiency. For inference, ENGINE utilizes an early exit strategy to further accelerate the process. GraphBridge (Wang et al., 2024)[Code] first encodes both local and global text information using a language model, by incorporating neighboring nodes’ text. A GNN is then applied to further refine node representations. Appendix J Node Classifiction: Implementation and Experimental Details J.1 Computational Resources In our experiments, we use four NVIDIA GeForce RTX 3090 GPUs, each with 24 GB of VRAM. The LM components are trained and run on these four GPUs, while the GNN module is executed on a single GPU. J.2 Dataset Split For Cora and CiteSeer, we use a random node split with 60% of nodes for training, 20% for validation, and 20% for testing. For WikiCS, ArXiv-2023, and OGBN-Products, we adopt the official training, validation, and test splits (Mernyei and Cangea, 2007; He et al., 2023; Hu et al., 2020). J.3 Baseline Model Deployment Settings Graph-Specific Models: For NodeFormer and GraphSAGE, we use the raw node features from each dataset, constructed via one-hot encoding. For GraphFormers, we implement the model using its official source code. Pure LMs: For BERT, Sentence-BERT, and RoBERTa-base, we perform full-parameter fine-tuning using the raw texts of each node. For RoBERTa-large, we employ Low-Rank Adaptation (LoRA) with a rank of 8. Recent TAG Methods: We use RoBERTa-base as the language model backbone and a two-layer GraphSAGE with hidden size 64 as the GNN backbone. This configuration is consistent with that of DuConTE to ensure a fair comparison. We implement these models using their official source code, and the training epochs as well as learning rates for both the LM and GNN components are kept consistent with DuConTE. J.4 Implementation Details of our Pipeline Instance : We provide a comprehensive overview of the configuration and training parameters adopted by the pipeline instantiated in Section 5.3. Upstream Preprocessing Configurations. We adopt 2-hop neighborhood sampling with a maximum of 39 neighbors per node. This means that for any node vi∈v_i , the sampled neighborhood (vi)N(v_i) satisfies |(vi)|≤39|N(v_i)|≤ 39, and we denote S(i)=vi∪(vi)S^(i)=\v_i\ (v_i) with |S(i)|≤40|S^(i)|≤ 40. The text of each node is processed using a reduction module (Wang et al., 2024) to fit the input length limit of the LM. This module, introduced in the GraphBridge framework, is a token selector pre-trained on the training set that assigns importance scores to word tokens within each node’s text. Given that the RoBERTa-base model has a maximum context length of 512 tokens, we enforce a uniform token budget across all nodes in S(i)S^(i). Specifically, let B=⌊512|S(i)|⌋−1B= 512|S^(i)| -1 be the per-node token budget (excluding the [SEP] token). For any node vj∈S(i)v_j∈ S^(i) whose original token sequence jw_j exceeds B tokens, we retain only the top-B most important tokens as ranked by the reduction module, preserving their original order. The resulting truncated sequences are then concatenated with [SEP] separators to form the unified input (i)W^(i). Hyperparameter Settings of DuConTE. For the internal hyperparameters α and β of DuConTE, we perform a grid search over the range [0,1][0,1] with a step size of 0.1, selecting the best combination based on performance on the validation set. The selected hyperparameter values for each dataset are reported in Table 8. Table 8: Hyperparameter settings of α and β in the experiments. Hyperparameter Cora CiteSeer WikiCS ArXiv-2023 OGBN-Products α 0.70.7 0.90.9 0.90.9 0.60.6 0.80.8 β 0.70.7 0.80.8 0.80.8 0.90.9 0.90.9 Training Setup for DuConTE. DuConTE uses two pre-trained RoBERTa-base models for ℳLM_L and ℳNM_N. ℳLM_L has positional encoding enabled. ℳNM_N takes (i) H^(i) as input directly, bypassing the token embedding layer, with positional encoding kept on. The detailed two-stage training procedure of DuConTE is described in Section H. In both Stage 1 and Stage 2, the learning rate is set to − 5e-5, and the number of training epochs is specified in Table 9. Training Setup for the Downstream GNN. We adopt a two-layer GraphSAGE with a hidden dimension of 64 as the GNN backbone in the downstream task. The model is trained using the final node representations generated by DuConTE as input features. We employ a learning rate of − 1e\!\!-\!\!2, train for up to 500 epochs, and apply early stopping with a patience of 20 epochs based on validation performance. Table 9: Training Epochs in Stage 1 and Stage 2 Stage Cora CiteSeer WikiCS ArXiv-2023 OGBN-Products Stage 1 8 8 16 8 8 Stage 2 8 8 16 8 8 Appendix K Link Prediction: Implementation and Experimental Details K.1 Dataset Split For Cora, CiteSeer, and ArXiv-2023, we randomly split edges into training, validation, and test sets in a 6:2:2 ratio. K.2 Baseline Model Deployment Settings GraphSAGE: We use a one-layer GraphSAGE with hidden dimension 16 and a two-layer MLP link predictor. Recent TAG Methods: We use RoBERTa-base as the language model backbone and a one-layer GraphSAGE with hidden dimension 16 as the GNN backbone, paired with a two-layer MLP link predictor. This configuration matches that of DuConTE to ensure a fair comparison. We implement these models using their official source code, and the training epochs as well as learning rates for both the LM and GNN components are kept consistent with DuConTE. K.3 Implementation Details of our Pipeline Instance : We instantiate a text-attributed graph learning pipeline for link prediction, with DuConTE serving as the text encoder. In the downstream phase, we use a one-layer GraphSAGE with hidden dimension 16 and a two-layer MLP link predictor. Upstream Preprocessing Configurations. We use the same upstream preprocessing configuration as in J.4. Hyperparameter Settings of DuConTE. The values of the internal hyperparameters α and β are set as in Table 8. Training Setup for DuConTE. The training configuration of DuConTE follows that in J.4.The detailed training procedure is described in K.4. Training Setup for the Downstream GNN. We adopt a one-layer GraphSAGE with hidden dimension 16 as the downstream GNN, followed by a two-layer MLP link predictor, using the final node representations from DuConTE as input features. The model is trained with a learning rate of − 1e\!\!-\!\!2, up to 500 epochs, and early stopping (patience = 20) based on validation performance. K.4 Two-Stage Training Procedure of DuConTE We train DuConTE using a two-stage procedure tailored for link prediction. In both stages, link scores are computed as the dot product of node representations, and the model is optimized using binary cross-entropy loss on positive and negative edges. Stage 1: Word-Token Encoder Training. We train the word-token encoder ℳLM_L and the composer f1f_1, while ℳNM_N and f2f_2 remain frozen. For each training edge (i,j)∈ℰtrain(i,j) _train, we compute the dot-product score between first-stage representations: sij(1)=(i(i))⊤j(j).s^(1)_ij=( z_i^(i)) z_j^(j). A corresponding negative edge (i,k)(i,k) is sampled by replacing j with a uniformly random node k. The loss is computed as: ℒ1=∑(i,j)∈ℰtrain[ℓ(sij(1),1)+ℓ(sik(1),0)],L_1= _(i,j) _train [ (s^(1)_ij,1)+ (s^(1)_ik,0) ], (19) where ℓ(y^,y)=BCEWithLogits(y^,y) ( y,y)=BCEWithLogits( y,y). Stage 2: Node Encoder Training. We freeze ℳLM_L and f1f_1, and train ℳNM_N together with f2f_2. The final representations i o_i and j o_j are scored analogously: sij(2)=i⊤j.s^(2)_ij= o_i o_j. Using the same positive/negative edge sampling strategy, the second-stage loss is: ℒ2=∑(i,j)∈ℰtrain[ℓ(sij(2),1)+ℓ(sik(2),0)].L_2= _(i,j) _train [ (s^(2)_ij,1)+ (s^(2)_ik,0) ]. (20) Appendix L Dataset Descriptions The experiments are conducted on five benchmark text-attributed graph datasets, widely adopted in graph representation learning. Below we provide a brief overview of each. For detailed statistics, including the number of nodes, edges, classes, and average token count per node, please refer to Table 10. Cora (Sen et al., 2008) The Cora dataset contains 2,708 scientific publications divided into seven classes: case-based reasoning, genetic algorithms, neural networks, probabilistic methods, reinforcement learning, rule learning, and theory. The papers form a citation network with 5,429 undirected edges, where each node has at least one citation link. CiteSeer (Giles et al., 1998) The CiteSeer dataset consists of 3,186 scientific documents categorized into six areas: Agents, Machine Learning, Information Retrieval, Databases, Human–Computer Interaction, and Artificial Intelligence. Each document is represented by its title and abstract, and the task is to classify papers based on this text and the citation structure. WikiCS (Mernyei and Cangea, 2007) WikiCS is a Wikipedia-based dataset for evaluating graph neural networks. It includes 10 classes corresponding to computer science topics and exhibits high connectivity. Node features are obtained from the text of the corresponding Wikipedia articles. ArXiv-2023 (He et al., 2023) ArXiv-2023 is a directed citation network introduced in TAPE, containing computer science papers from arXiv published in 2023 or later. Nodes represent papers, and directed edges represent citations. The task is to classify each paper into one of 40 subject areas, such as cs.AI, cs.LG, and cs.OS, using labels provided by authors and arXiv moderators. OGBN-Products (Hu et al., 2020) OGBN-Products is a dataset of Amazon products with co-purchase relations. The full version has over 2 million nodes and 61 million edges. The subset used here, created via node sampling in TAPE (He et al., 2023), contains 54,000 nodes and 74,000 edges. Each node corresponds to a product and is labeled with one of 47 top-level categories. Table 10: Dataset statistics. Nodes, Edges, Classes and Avg.degrees mean the number of nodes, edges, classes and average degrees for each dataset, respectively. Avg.tokens represents the average number of tokens per node in each dataset when using the RoBERTa-base’s tokenizer. Dataset Nodes Edges classes Avg.degrees Avg.tokens Cora 27082708 54925492 77 3.903.90 194194 CiteSeer 31863186 42774277 66 1.341.34 196196 WikiCS 1170111701 215863215863 1010 36.7036.70 545545 ArXiv-2023 4619846198 7854378543 4040 3.903.90 194194 OGBN-Products(subset) 5402554025 7442074420 4747 2.682.68 163163 Appendix M Ablation Variants In this section, we detail the design of each ablation variant used in our experiments. NoDual It encodes semantic information only at the word-token granularity, achieved by setting the hyperparameter α=0α=0. NoMask-T It uses the vanilla self-attention mechanism in every attention layer of the word-token encoder. NoMask-D It uses the vanilla self-attention mechanism in every attention layer of the node encoder. NoMask-Both It uses the vanilla self-attention mechanism in every attention layer of both encoders. MeanPool It directly converts word-token embeddings into node representations using mean pooling. Center-Only Its node representation composer evaluates word-token importance only in the center-node semantic context, with the hyperparameter β set to 1. Neigh-Only Its node representation composer evaluates word-token importance only in the neighborhood semantic context, with the hyperparameter β set to 0. UnifiedContext Its node representation composer evaluates word-token importance in a shared context, without differentiating the contextual influence from the center-node and its neighborhood. The unnormalized importance of token wiqw_iq is computed as: μq′=∑p=1Liai,p,q(i)+∑vj∈(i)∑p=1Ljaj,p,q(i), _q = _p=1^L_ia_i,p,q^(i)+ _v_j (i) _p=1^L_ja_j,p,q^(i), (21) and the final importance score μq _q is obtained by applying softmax normalization over all word-tokens in viv_i. NeurIPS Paper Checklist 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The abstract and introduction accurately state our three main contributions. Guidelines: • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper. • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers. • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: The limitations of our work are discussed in Appendix E. Guidelines: • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper. • The authors are encouraged to create a separate “Limitations” section in their paper. • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [N/A] Justification: This paper does not include theoretical results. Guidelines: • The answer [N/A] means that the paper does not include theoretical results. • All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. • All assumptions should be clearly stated or referenced in the statement of any theorems. • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. • Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: All experimental configurations and implementation details are provided in the main paper and appendix for reproducing the main results. Guidelines: • The answer [N/A] means that the paper does not include experiments. • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The open access to data and code is described in detail in Appendix G. Guidelines: • The answer [N/A] means that paper does not include experiments requiring code. • Please see the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results? Answer: [Yes] Justification: Necessary experimental settings are provided in the experimental section of the main paper, with full details given in the Appendix. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. • The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: The relevant information is described in detail in the experimental section. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). • It should be clear whether the error bar is the standard deviation or the standard error of the mean. • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates). • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The compute resource information is described in detail in Appendix J. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper). 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.c/public/EthicsGuidelines? Answer: [Yes] Justification: This research conforms to the NeurIPS Code of Ethics. Guidelines: • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics. • If the authors answer [No] , they should explain the special circumstances that require a deviation from the Code of Ethics. • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: This work can be applied to social domains related to text-attributed graphs, with potential positive impacts on recommendation systems and knowledge discovery. Guidelines: • The answer [N/A] means that there is no societal impact of the work performed. • If the authors answer [N/A] or [No] , they should explain why their work has no societal impact or why the paper does not address societal impact. • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)? Answer: [N/A] Justification: This paper does not release models or datasets with high risk for misuse. Guidelines: • The answer [N/A] means that the paper poses no such risks. • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: All datasets used are properly cited with their original papers and licenses listed in Appendix L. Guidelines: • The answer [N/A] means that the paper does not use existing assets. • The authors should cite the original paper that produced the code package or dataset. • The authors should state which version of the asset is used and, if possible, include a URL. • The name of the license (e.g., C-BY 4.0) should be included for each asset. • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. • If this information is not available online, the authors are encouraged to reach out to the asset’s creators. 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: The code is provided in the supplementary material, and the training pipeline is described in detail in the Appendix. Guidelines: • The answer [N/A] means that the paper does not release new assets. • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. • The paper should discuss whether and how consent was obtained from people whose asset is used. • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [N/A] Justification: This paper does not involve crowdsourcing or research with human subjects. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. 15. Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [N/A] Justification: This paper does not involve research with human subjects. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required. Answer: [N/A] Justification: The core method development in this research does not involve LLMs. Guidelines: • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.