Abstract
Traditional recommendation systems often treat relevance as a universal metric, overlooking the linguistic and cultural contexts that shape user interests. This paper introduces LaSER (Language-Specific Event Recommendation), a framework designed to recommend events tailored to specific language communities. By integrating language-specific knowledge graph embeddings with spatio-temporal features into a Learning to Rank (LTR) model, LaSER demonstrates significant improvements in recommendation accuracy. Evaluated against state-of-the-art baselines, LaSER achieves an average nDCG@10 of 0.957 and improves user-perceived relevance by up to 33 percentage points.
Introduction
When a user searches for a global phenomenon like the "Coronavirus pandemic," their specific information needs are often dictated by their location and language. A German speaker may prioritize European lockdown policies, while a Spanish speaker might seek updates on the situation in South America. Despite this, most current event recommendation systems provide a one-size-fits-all list of global events.
Two primary obstacles hinder the development of more nuanced systems: contextual neglect, where models ignore the latent properties of events specific to a language, and reproducibility, caused by a reliance on proprietary search logs from major tech companies. LaSER addresses these challenges by utilizing publicly available datasets and incorporating language-specific latent representations to capture cultural relevance.
The LaSER Framework
The LaSER pipeline consists of three distinct phases designed to transform raw knowledge graph data into a prioritized list of relevant events.
1. Language-Specific Embeddings
To capture how different language communities connect concepts, the system employs DeepWalk on language-specific Wikipedia link graphs. This process generates latent representations (embeddings) for entities where the "neighborhood" of an entity reflects the specific associations made within that language's Wikipedia edition.
2. Multi-Faceted Feature Extraction
For every query entity and candidate event, LaSER extracts ten features across four categories:
- Spatial Features: Includes a novel "Language Distance" metric, measuring the physical distance between an event's location and countries where the target language holds official status.
- Temporal Features: Evaluates the overlap and proximity between the time intervals of the query and the event.
- Link-based Features: Leverages Wikipedia’s internal structure, including incoming/outgoing links and the Milne-Witten relatedness score.
- Embedding-based Features: Calculates the cosine similarity between the language-specific embeddings generated in the first phase.
3. Learning to Rank (LTR)
These features are processed by a LambdaMART model (implemented via XGBoost). The model is trained to predict event relevance based on actual user navigation patterns, effectively fine-tuning the final ranking for maximum precision.
Datasets and Reproducibility
A core contribution of this work is its reliance on open-source data, ensuring the results are reproducible by the broader research community. The study utilizes:
- EventKG: A multilingual knowledge graph providing the foundational entities, events, and spatio-temporal attributes.
- Wikipedia Clickstream: Used as the "ground truth" for training, this dataset tracks how users navigate between articles in specific language editions.
- EventKG+Click: An extended dataset created by the authors that normalizes click counts across languages to establish a standardized language-specific relevance score.
Experimental Results
LaSER was evaluated against several strong baselines, including the Milne-Witten relatedness score and SuperGAT (a graph attention network), across German, French, and Russian language sets.
Automated Metrics
In automated testing, LaSER achieved an average nDCG@10 of 0.957 and a MAP@10 of 0.970. These results outperformed the best-performing baselines by 8 and 17 percentage points, respectively. Furthermore, the language-specific DeepWalk embeddings proved highly effective for candidate generation, achieving a recall of 0.364, compared to just 0.008 for non-language-specific embeddings like TransE.
User Study
A study involving 17 researchers confirmed the model's real-world utility. Participants judged LaSER’s recommendations to be significantly more relevant to specific language audiences than those of the baselines. Specifically, LaSER outperformed baselines by 33 percentage points (MAP@5) for Russian and 17 points for French.
Discussion and Key Insights
The research yielded several critical insights into the nature of cross-lingual information retrieval:
- Language as a Spatial Proxy: The "Language Distance" feature proved to be a powerful predictor; events occurring in countries where the query language is official are statistically more likely to be relevant to that user base.
- Feature Importance: Analysis revealed that link-based features—how articles are interconnected within Wikipedia—contributed most significantly to the model's ranking performance.
- Information Asymmetry: The study highlighted that different Wikipedia editions contain vastly different information. For instance, certain events present in the English Wikipedia were entirely absent from the Russian version, reinforcing the need for tools that can bridge these linguistic gaps.
Conclusion and Future Work
LaSER demonstrates that incorporating language-specific context significantly enhances the relevance of event recommendations. While the model currently faces challenges with data sparsity (only 48% of entities in EventKG have spatial data), it provides a robust framework for future development. Future iterations may expand beyond Wikipedia to include real-time data from social media and explore how these recommendations can assist journalists and historians in identifying cultural viewpoints on global conflicts.
Summary updated