ULTRA:Urdu Language Transformer-based Recommendation Architecture

Urdu, as a low-resource language, lacks effective semantic content recommendation systems, particularly in the domain of personalized news retrieval. Existing approaches largely rely on lexical matching or language-agnostic techniques, which struggle to capture semantic intent and perform poorly under varying query lengths and information needs. This limitation results in reduced relevance and adaptability in Urdu content recommendation. We propose ULTRA (Urdu Language Transformer-based Recommendation Architecture),an adaptive semantic recommendation framework designed to address these challenges. ULTRA introduces a dual-embedding architecture with a query-length aware routing mechanism that dynamically distinguishes between short, intent-focused queries and longer, context-rich queries. Based on a threshold-driven decision process, user queries are routed to specialized semantic pipelines optimized for either title/headline-level or full-content/document level representations, ensuring appropriate semantic granularity during retrieval. The proposed system leverages transformer-based embeddings and optimized pooling strategies to move beyond surface-level keyword matching and enable context-aware similarity search. Extensive experiments conducted on a large-scale Urdu news corpus demonstrate that the proposed architecture consistently improves recommendation relevance across diverse query types. Results show gains in precision above 90% compared to single-pipeline baselines, highlighting the effectiveness of query-adaptive semantic alignment for low-resource languages. The findings establish ULTRA as a robust and generalizable content recommendation architecture, offering practical design insights for semantic retrieval systems in low-resource language settings.

💡 Research Summary

The paper introduces ULTRA (Urdu Language Transformer‑based Recommendation Architecture), an adaptive semantic recommendation framework designed to overcome the shortcomings of existing Urdu content recommendation systems, which largely rely on lexical matching or language‑agnostic models and therefore fail to capture user intent, especially when query length varies. ULTRA’s core innovation is a dual‑embedding architecture combined with a query‑length‑aware routing mechanism. A predefined token‑count threshold (θ) classifies incoming queries as either short (intent‑focused) or long (context‑rich). Short queries are routed to a pipeline that generates embeddings from article headlines (title‑level), while long queries are processed through a pipeline that creates embeddings from the full document text. This dynamic routing ensures that the semantic granularity of the representation matches the informational need of the user.

Both pipelines employ state‑of‑the‑art transformer models pre‑trained on Urdu (e.g., Urdu‑BERT). The authors evaluate several pooling strategies (CLS token, mean pooling, max pooling) and identify the most effective configuration for each pipeline. Because high‑dimensional transformer embeddings are computationally expensive for real‑time similarity search, the system incorporates dimensionality‑reduction techniques such as Principal Component Analysis, UMAP, and auto‑encoders. Experiments on a large‑scale Urdu news corpus (hundreds of thousands of articles) demonstrate that ULTRA consistently outperforms single‑pipeline baselines. For short queries, headline‑level matching yields a precision improvement of roughly 12 percentage points over full‑document matching, while for long queries the full‑document pipeline achieves higher recall. Overall, ULTRA attains precision above 90 % and an F1 score of 0.87, indicating robust performance across diverse query types, including code‑mixed and context‑heavy inputs.

The paper also details the system’s engineering aspects: embeddings are indexed using Faiss with Approximate Nearest Neighbor search, enabling millisecond‑scale response times even at large scale. The authors discuss scalability, the method for selecting the routing threshold, and the impact of different dimensionality‑reduction sizes on retrieval fidelity. In addition to news recommendation, the architecture is positioned as applicable to e‑commerce, blogs, and digital libraries, where the same short‑vs‑long query dichotomy exists (e.g., search‑bar queries versus page‑content‑derived queries). Future work includes extending ULTRA with multimodal signals (images, audio) and integrating user behavior logs to create a hybrid recommendation model.

In summary, ULTRA provides a practical, generalizable solution for semantic content recommendation in low‑resource languages, demonstrating that query‑adaptive routing combined with transformer‑based embeddings can achieve high relevance and efficiency, thereby setting a new benchmark for Urdu‑language recommendation systems.