Abstract
In specialized fields such as medicine or law, a significant "language gap" exists between the jargon used by experts and the common terminology of laypeople. This paper introduces $\tau$LDA (Topic-Adapted LDA), a probabilistic model that explicitly incorporates "technicality" as a continuous domain factor. By learning dual word distributions for each topic—one for lay language and one for expert jargon—$\tau$LDA aligns semantic themes across the expertise spectrum. Experimental results on a diverse medical corpus demonstrate that $\tau$LDA significantly improves cross-domain information retrieval and provides a more accurate representation of multi-expertise text than standard topic models.
Introduction: The Language Gap Problem
Information accessibility is often hindered by the disparate vocabularies used by different groups to describe identical concepts. For instance, a patient might search for "abnormal tumors," while a medical journal discusses "neoplastic carcinomas." Standard topic models like Latent Dirichlet Allocation (LDA) typically fail in these scenarios; they either segregate jargon and common speech into separate, unlinked topics or conflate them into noisy, incoherent clusters. The authors propose $\tau$LDA to create a "topical bridge" that maintains domain-specific nuances while ensuring semantic alignment.
Methods: The $\tau$LDA Framework
The core innovation of $\tau$LDA is treating technicality as a continuous variable ($\tau \in [0, 1]$) that modulates topic expression.
As shown in the graphical model above, $\tau$LDA employs a Two-Mode Mixture strategy. Instead of a single word distribution per topic, it learns two "extreme" distributions: $\beta^0$ (lay) and $\beta^1$ (expert). A document's specific word distribution for a topic is a mixture of these two, weighted by the document's technicality level.
Furthermore, the model incorporates a Technicality Hierarchy. A per-topic parameter ($\pi_k$) determines the likelihood of a topic appearing in expert versus lay contexts. The model utilizes "mild supervision" in the form of document-level technicality stamps (e.g., source metadata) to guide the discovery of these latent structures. Inference is performed via an efficient Variational Bayesian algorithm.
Datasets and Evaluation
The model was evaluated on a large-scale medical corpus spanning five domains with varying technicality:
- Yahoo! Health QA ($\tau=0.0$): Layperson questions.
- WebMD ($\tau=0.3$): Consumer medical info.
- CDC ($\tau=0.7$): Public health info.
- PubMed ($\tau=1.0$): Technical journal articles.
- MeSH ($\tau=1.0$): Expert descriptors.
Key Insights: The Asymmetry of Expertise
A significant finding is the asymmetry of expertise: while experts often use lay terminology to communicate broadly, laypeople rarely utilize high-level jargon. Consequently, $\tau$LDA found that even highly technical topics maintain some overlap with common language, whereas "purely expert" topics are rare. By forcing alignment between $\beta^0$ and $\beta^1$, the model enables cross-domain retrieval, allowing plain-English queries to successfully surface semantically relevant technical documents.
Key Results
$\tau$LDA demonstrates substantial improvements over baseline models in several key areas:
- Model Fit: As illustrated in the figure below, $\tau$LDA consistently achieves higher predictive log-likelihood on unseen data compared to standard LDA and Supervised LDA (sLDA), indicating a more accurate capture of the underlying text structure.
- Information Retrieval: In tasks using lay queries to find technical documents, $\tau$LDA achieved an nDCG@5 of 0.51, significantly outperforming the 0.38 achieved by LDA.
- Domain Identification: The model accurately predicts the technicality level of unseen documents with lower Root Mean Square Error (RMSE) than competing approaches.
Conclusion and Future Work
$\tau$LDA successfully bridges the language gap by modeling technicality as a drifting domain factor. While the current implementation relies on coarse per-domain supervision and is sensitive to initialization, it provides a robust framework for cross-expertise communication. Future research may explore finer-grained, per-document supervision and apply the $\tau$LDA framework to other domain factors such as sentiment or political leaning.
Summary updated