TriAligner: Multi-Source Alignment

• TriAligner is a framework combining dual-encoder architectures for crosslingual retrieval with tensor-based methods for higher-order network alignment.
• The system employs symmetric contrastive loss and LLM-driven data augmentation to optimize semantic matching and improve retrieval metrics.
• In network analysis, TriAligner maximizes motif conservation using tensor eigenvector methods, enhancing functional and structural alignment.

TriAligner refers to complementary families of algorithms and systems designed for higher-order alignment tasks—spanning both large-scale network matching in bioinformatics and crosslingual retrieval in natural language processing—linked by their use of multi-source or multi-view representations and alignment objectives that extend beyond pairwise similarity. In the context of crosslingual retrieval, TriAligner is a system for matching social-media posts with previously fact-checked claims, leveraging native and English representations with contrastive learning in a dual-encoder architecture (Abootorabi et al., 24 Dec 2025). In higher-order network analysis, TriAligner (as deployed in TAME) refers to tensor-based methods that maximize conservation of motifs (notably triangles) between networks, crucial for applications such as comparative interactomics (Mohammadi et al., 2015). The unifying principle is alignment via fusion of multiple sources or modalities, whether embeddings or topological motifs.

1. Dual-Encoder Multi-Source Pipeline in Crosslingual Retrieval

TriAligner implements a dual-encoder (“two-tower”) architecture, processing native and English modalities of posts and claims in parallel (Abootorabi et al., 24 Dec 2025). Each input—post_native, post_english, fact_native, fact_english—is embedded (typically with BGE-M3 or LaBSE pretrained backbones), encoded through linear layers with batch normalization, ReLU activations, dropout, and further fused via concatenation. The architecture yields three 256- or 512-dimensional representations spaces: fused (concatenated), English-only, and native-only. Cosine similarity matrices $S_1, S_2, S_3$ are formed for each modality-pair. The final score matrix $X$ aggregates these, learning scalar weights $\lambda_1, \lambda_2, \lambda_3$ and scaling factors $S_1, S_2, S_3$ such that: $$x_{i,j} = \lambda_1 e^{S_1} A_{i,j} + \lambda_2 e^{S_2} B_{i,j} + \lambda_3 e^{S_3} C_{i,j}$$ where $A_{i,j}, B_{i,j}, C_{i,j}$ denote the fused, English, and native similarity matrices respectively (see Equation 1 in (Abootorabi et al., 24 Dec 2025)).

This weighted fusion mechanism affords complementary semantic coverage, addressing translation loss and representation mismatch. The system is trained end-to-end to optimize these parameters for optimal separation of true vs. false post–claim pairs.

2. Contrastive Symmetric Loss for Pairwise Alignment

TriAligner’s training is governed by a symmetric contrastive loss, maximizing similarity of correct pairs and minimizing incorrect ones. For batch size $N$, the row-softmax and column-softmax probabilities are: $$P_{ij} = \frac{\exp(x_{ij})}{\sum_{k=1}^N \exp(x_{ik})}, \qquad Q_{ij} = \frac{\exp(x_{ij})}{\sum_{k=1}^N \exp(x_{kj})}$$ The loss function is expressed as (Eqn 2 in (Abootorabi et al., 24 Dec 2025)): $$\mathcal{L} = -\frac{1}{2N} \sum_{i=1}^N (\log P_{ii} + \log Q_{ii})$$ This bidirectional cross-entropy is similar to InfoNCE but with implicit temperature control via the learned scale factors. True pairs ($x_{ii}$) are maximized; all off-diagonal ($x_{ij},\ i\neq j$) combinations act as negatives, with hard negative mining further sharpening discriminability.

3. Data Preprocessing and Augmentation via LLMs

Robustness is enhanced by multi-stage preprocessing and augmentation (Abootorabi et al., 24 Dec 2025). Titles are merged with OCR text for posts (and claim text for facts); extraneous tokens are removed and abbreviations expanded. Sparse or noisy social media inputs are augmented with GPT-4o: each post’s text+OCR is rewritten into a unified narrative (≥15 words) preserving original meaning. Hard negative sampling is injected at batch preparation: embeddings are indexed and semantically similar but irrelevant claims retrieved, improving contrastive learning by enforcing fine-grained distinction among near-duplicates. The main pipeline is outlined in Table 1 below.

Stage	Description	Technique
Preprocessing	Title/text fusion, cleaning	Regex, OCR
Augmentation	Narrative rewriting of posts	GPT-4o, LLM
Negative Sampling	Retrieval of close but non-matching facts	BGE-M3, kNN

4. Training Procedure and Implementation Details

Training proceeds on a single NVIDIA P100 GPU with large batch sizes (10,000 pairs). The AdamW optimizer is used at $6 \times 10^{-4}$ learning rate, controlled by cosine annealing with warm restarts. Early stopping monitors Recall@10 on the development set, clipping patience at 5 epochs. Training typically completes in 20–30 epochs. Implementation is based on PyTorch Lightning and HuggingFace Transformers (Abootorabi et al., 24 Dec 2025).

5. Evaluation, Benchmarking, and Empirical Results

TriAligner is evaluated on the MultiClaim dataset, comprising $\sim 206,000$ fact-checks in 39 languages and $\sim 28,000$ social posts in 27 languages. Principal metrics are Success@K (fraction of queries retrieving ≥1 relevant item in top $K$) and Recall@K (fraction of relevant items found in top $K$ divided by total). Monolingual and crosslingual retrieval accuracy is reported as follows:

Stage	Monolingual ($R@10/S@10$)	Crosslingual ($R@10/S@10$)
BGE-M3 (baseline)	0.776 / 0.794	0.473 / –
ConcatEnc (fused only)	0.816 / –	0.680 / –
MultiSim (native+Eng)	0.741 / –	0.651 / –
TriAligner	0.837 / 0.848	0.687 / 0.707
+Augmentation	0.860 / –	0.702 / –
+Re-ranker	– / 0.881	– / 0.748

TriAligner consistently outperforms baselines, with substantial gains observed in crosslingual settings. Language-specific tables confirm improvements across multiple scripts and linguistic families. On the test set, TriAligner with no reranker achieves 0.808 monolingual $S@10$ compared to the winning system’s 0.960 (Abootorabi et al., 24 Dec 2025).

6. Higher-Order Network Alignment via Tensor Methods

The TriAligner class also refers to tensor-based higher-order network alignment under the Triangular AlignMEnt (TAME) framework (Mohammadi et al., 2015). Classical pairwise graph alignment maximizes edge overlap, which is NP-hard; TAME generalizes to motif conservation (triangles and beyond) and recasts the objective as maximizing the number of aligned substructures.

Given graphs $G_1, G_2$, triangle tensors $\Delta_{G_1}, \Delta_{G_2}$ encode all triangles. The alignment objective maximizes: $$\max_{f} \quad \sum_{\Delta \in G_1,\,\Delta' \in G_2} f(i,i')\,f(j,j')\,f(k,k')\qquad\text{subject to 1-1 matching}$$ An NP-hard integer cubic program is relaxed to a tensor eigenvector problem via the Kronecker product $\mathcal{T} = \Delta_{G_2} \otimes \Delta_{G_1}$: $$\max_{x \in \mathbb{R}^n} \mathcal{T} x^3 \quad \text{subject to } \|x\|_2 = 1$$ SS-HOPM (Shifted Symmetric Higher-Order Power Method) solves this efficiently with an implicit kernel on motif sets. Sequence-based priors are integrated by initializing $x^{(0)} \propto w$ (sequence similarity scores). Post-processing applies bipartite matching and local swaps.

Empirical results on NAPAbench and yeast-human PPI networks show TAME achieves up to $+18.6\%$ more conserved triangles than edge-based methods and demonstrates that triangle conservation correlates more significantly with node correctness and functional co-expression than edge conservation (Mohammadi et al., 2015).

7. Analysis, Limitations, and Future Directions

TriAligner’s retrieval gains stem from multi-source alignment, contrastive loss with extensive negatives, LLM-driven augmentation for sparse content, and lightweight reranking. Fusing native and translated embeddings leverages complementary semantic signals; LLM augmentation enriches data, and rerankers further refine results. Limitations include reliance on two backbone encoders, English-centric augmentation, and restricted reranker scale due to GPU constraints.

In higher-order network alignment, motif-based objectives capture richer functional structure (e.g., clustering, modules) than edge-based formulations. Triangle conservation serves as a better proxy for functional and orthological correctness.

Suggested directions include employing more powerful multilingual backbones, expanding to cross-modal claims (e.g., multimodal with images/text), advanced negative sampling, dynamic $\lambda$ weighting conditional on language pair, and integration of emerging LLMs for reranking and augmentation. TAME’s tensor-eigenproblem framework generalizes to arbitrary $k$-motifs for future topology-driven applications in biology and beyond (Mohammadi et al., 2015, Abootorabi et al., 24 Dec 2025).