[Paper Review] Analysis of Morphology in Topic Modeling.

Chandler May, Ryan Cotterell+1 more|arXiv (Cornell University)|Aug 13, 2016

Topic Modeling24 references4 citations

TL;DR

This paper investigates how lemmatization improves the interpretability of topic models by reducing morphological variation in text. Using a word intrusion evaluation on Wikipedia articles in a morphologically rich language, it demonstrates that lemmatization significantly enhances the quality of topic representations in LDA models, leading to more coherent and meaningful top-word lists.

ABSTRACT

Topic models make strong assumptions about their data. In particular, different words are implicitly assumed to have different meanings: topic models are often used as human-interpretable dimensionality reductions and a proliferation of words with identical meanings would undermine the utility of the top-$m$ word list representation of a topic. Though a number of authors have added preprocessing steps such as lemmatization to better accommodate these assumptions, the effects of such data massaging have not been publicly studied. We make first steps toward elucidating the role of morphology in topic modeling by testing the effect of lemmatization on the interpretability of a latent Dirichlet allocation (LDA) model. Using a word intrusion evaluation, we quantitatively demonstrate that lemmatization provides a significant benefit to the interpretability of a model learned on Wikipedia articles in a morphologically rich language.

Motivation & Objective

• To examine how morphological variation affects topic model interpretability in latent Dirichlet allocation (LDA).
• To assess whether lemmatization improves the quality of topic representations by reducing synonymy through word form normalization.
• To empirically evaluate the effect of preprocessing with lemmatization on the interpretability of topic models using a quantitative evaluation method.
• To provide evidence on whether morphological normalization enhances the coherence and human interpretability of top-word lists in topics.

Proposed method

• Lemmatization is applied to Wikipedia articles in a morphologically rich language to reduce inflected word forms to their base forms.
• A latent Dirichlet allocation (LDA) model is trained on both raw and lemmatized text to compare topic representations.
• A word intrusion evaluation is used to measure interpretability by assessing how well human evaluators can detect a target word in a topic's top-word list.
• The evaluation compares model performance on raw text versus lemmatized text, with higher intrusion detection accuracy indicating better interpretability.
• Statistical analysis is used to determine whether the improvement in interpretability due to lemmatization is significant.

Experimental results

Research questions

• RQ1 Does lemmatization improve the interpretability of topic models in morphologically rich languages?
• RQ2 How does morphological variation affect the coherence of top-word lists in LDA topics?
• RQ3 To what extent does lemmatization enhance the quality of topic representations as measured by word intrusion tasks?
• RQ4 Is the improvement in interpretability statistically significant when lemmatization is applied to Wikipedia text?

Key findings

• Lemmatization significantly improves the interpretability of LDA models on Wikipedia articles in a morphologically rich language.
• The word intrusion evaluation shows higher detection accuracy for target words in lemmatized text, indicating better topic coherence.
• The improvement in interpretability is quantitatively measurable and statistically significant.
• The results suggest that morphological normalization enhances the utility of topic models as human-interpretable dimensionality reduction tools.