Abstract
Linking authors across noisy, informal microblogging platforms like Twitter is a significant challenge for natural language processing due to brevity, slang, and temporal shifts in language use. This paper introduces SoulMate, a novel framework that embeds textual content alongside multi-aspect temporal contexts (hours, days, and seasons). By dynamically extracting latent concepts and learning how word relationships evolve over time, SoulMate identifies highly correlated user communities with significantly higher precision than traditional text-mining models.
Introduction
Author linking in short-text environments faces four primary hurdles: mismatched informal content, temporal skewness (where word associations change over time), conceptual gaps between users, and the bias of external knowledge bases. Traditional models often fail to account for the fact that a word pair like "Cold" and "Drink" has a much higher co-occurrence probability in summer than in winter. SoulMate addresses these issues by treating time as a core feature rather than a nuisance, utilizing a multi-aspect temporal-textual embedding approach.
Methodology
The SoulMate framework is divided into two distinct phases: an offline embedding phase and an online linking phase.
Offline Phase: Temporal-Textual Foundation
- Temporal Slabs: The system employs Hierarchical Agglomerative Clustering (HAC) to group similar time periods into "slabs" (e.g., grouping weekdays vs. weekends) based on user behavior consistency.
- TCBOW (Temporal Continuous Bag of Words): An extension of the CBOW architecture, TCBOW learns word vectors specific to different temporal slabs, capturing the evolution of word meanings across different hours, days, and seasons.
- Concept Discovery: Rather than relying on static external databases like Wikipedia, SoulMate uses unsupervised clustering (DBScan or K-Medoids) on tweet vectors to identify latent concepts directly from the corpus vernacular.
Online Phase: Author Linking and Subgraph Mining
- Contextual Similarity: The system generates author "context vectors" that integrate both textual content and the discovered latent concepts.
- SW-MST (Stack-Wise Maximum Spanning Tree): A specialized graph-cutting algorithm extracts the most tightly connected subgraphs of authors from a similarity graph to identify correlated communities.
Experimental Evaluation
The researchers evaluated SoulMate using a dataset of 8 million English tweets collected in Australia over one year. A filtered subset of 1 million geo-tagged tweets from 4,000 users was used for core experiments.
Key Results
- Precision: The integrated "SoulMateJoint" model achieved a precision of 0.67 for identifying high textual and conceptual similarity, vastly outperforming TF-IDF (0.21) and Exact Matching (0.39).
- Semantic Recovery: In cases where authors used entirely different vocabularies (low textual similarity), the concept-based module maintained 30% precision in detecting correlations, whereas traditional models dropped to near zero.
- Embedding Quality: TCBOW outperformed standard CBOW, Skip-gram, and GloVe in word analogy tests, demonstrating superior robustness against microblog noise.
Discussion: The Importance of Temporal Context
A core insight of this research is that word co-occurrence is highly dependent on temporal facets. By analyzing how word pair probabilities shift across different dimensions, the model provides a more accurate representation of user intent and community behavior.
Conclusion
SoulMate demonstrates that integrating multi-aspect temporal data with textual embeddings significantly improves author linking in short-text environments. By avoiding external knowledge base bias and focusing on the dynamic nature of language, the framework successfully bridges conceptual gaps between users. Future work may involve incorporating tweet "weight" (influence/popularity) into the concept extraction process to further refine community detection.
Summary updated