Overview
Online discussion forums are rich repositories of user-generated content, yet modeling them effectively remains a challenge. Traditional topic modeling, such as Latent Dirichlet Allocation (LDA), primarily focuses on "aspects"—the nouns or entities being discussed. However, in a functional environment like a support forum, the subject alone (e.g., "iPhone") does not reveal the user's objective. This paper introduces a novel unsupervised generative model that identifies user intent by jointly modeling aspects (the "what") and actions (the "how"). By analyzing the relationships between these pairs, the model distinguishes between different goals, such as technical troubleshooting versus general support.
Methodology: The AS-AC Model
The authors propose a multi-stage pipeline to transform raw forum text into a structured representation of user intent.
The process begins with Extraction, using Part-of-Speech (POS) tagging to isolate noun groups as aspects and verb groups as actions. The core of the system is the Joint Aspect-Action Topic Model (AS-AC). Unlike standard LDA, AS-AC assumes each post is generated by a mixture of both aspect and action topics, linked by a latent variable ($Z$) that captures their functional relationship.
Following model training, an Aspect-Action Relationship Graph is constructed. Nodes represent topics, and edges represent their co-occurrence within threads. To distill the most significant intents, the authors apply K-core decomposition, pruning loosely connected nodes to reveal a "core" of highly interconnected concepts that define the primary objectives of the discussion.
Experimental Evaluation
The model was tested on a massive dataset from the Apple Discussion Forums, comprising over 3.3 million posts across 340,000 threads. The data spanned product categories like iPad and MacBook, categorized into functional topics such as "Wi-Fi & Network" or "Camera & Photos."
Key Results
The AS-AC model demonstrated significant improvements over baseline methods like LDA and Multi-Grain LDA (MG-LDA).
- Lower Perplexity: As shown in the performance metrics, AS-AC consistently achieved lower perplexity scores across various topic counts, indicating a better fit for the data and higher predictive power for unseen posts.
- Higher Precision: In ranking tasks, AS-AC achieved a Precision@10 of 0.96 in the "Wi-Fi & Network" category, vastly outperforming MG-LDA (0.70) and LDA (0.51).
- Intent Discrimination: The model successfully used actions as discriminators. For instance, while two threads might both discuss "iPhone internet," the AS-AC model could distinguish a "Technical" intent (actions like reboot, fix) from a "General" intent (actions like tell, help).
Conclusion and Future Work
By treating intent as a joint distribution of aspects and actions, the AS-AC model provides a more nuanced understanding of online discussions than traditional methods. While the current approach relies on POS tagging—which can be sensitive to informal grammar—it offers a robust framework for large-scale forum analysis. Future work may involve modeling the temporal evolution of intent within long-running threads or applying the model to improve automated help-desk systems and semantic search engines.
Summary updated