Relating Romanized Comments to News Articles by Inferring Multi-Glyphic Topical Correspondence

Goutham Tholpadi, Mrinal Kanti Das, Trapit Bansal, Chiranjib BhattacharyyaAAAI 2015 · OpenReview

Abstract

Commenting is a popular facility provided by news sites. Analyzing such user-generated content has recently attracted research interest. However, in multilingual societies such as India, analyzing such user-generated content is hard due to several reasons: (1) There are more than 20 official languages but linguistic resources are available mainly for Hindi. It is observed that people frequently use romanized text as it is easy and quick using an English keyboard, resulting in multi-glyphic comments, where the texts are in the same language but in different scripts. Such romanized texts are almost unexplored in machine learning so far. (2) In many cases, comments are made on a specific part of the article rather than the topic of the entire article. Off-the-shelf methods such as correspondence LDA are insufficient to model such relationships between articles and comments. In this paper, we extend the notion of correspondence to model multi-lingual, multi-script, and inter-lingual topics in a unified probabilistic model called the Multi-glyphic Correspondence Topic Model (MCTM). Using several metrics, we verify our approach and show that it improves over the state-of-the-art.

Introduction

In multilingual societies like India, news platforms face a unique linguistic challenge. While articles are typically written in standard scripts (e.g., Devanagari for Hindi), user comments often appear in a mix of English, native tongues, and—most complexly—romanized vernacular. Because most users utilize English keyboards, they type native words phonetically (e.g., "Rohin" for Romanized Hindi).

Standard NLP tools often fail in these environments due to script variance, the lack of standard romanization rules, and the fact that comments often target specific segments of an article rather than the whole. Furthermore, a high volume of comments are "robotic" spam or off-topic meta-talk. This paper introduces the Multi-glyphic Correspondence Topic Model (MCTM) to bridge these gaps.

The Multi-glyphic Correspondence Topic Model (MCTM)

The MCTM framework extends traditional topic modeling to handle multi-glyphic environments. Unlike models that assume a global relationship between a document and its metadata, MCTM uses a "stick-breaking prior" to allow different segments (paragraphs) of an article to have their own topic distributions. This enables the model to identify when a comment corresponds to a specific quote or paragraph.

Conceptual overview of the MCTM framework showing how comments in different scripts relate to article segments or null topics.

Key technical innovations include:

  • Multi-glyphic Correspondence: The model assumes that while words differ across scripts (e.g., Hindi vs. Rohin), the underlying topics are shared.
  • Null Correspondence: To handle noise, the model identifies topics unique to the comment section (like spam or expletives) that do not correspond to the article text.
  • Joint Topic-Word Sparsity: By forcing each word to belong to only a few topics, the model maintains clarity in short, noisy comments.

Evaluation and Results

The researchers evaluated MCTM using datasets from Dainik Jagran (Hindi) and Kendasampige (Kannada). These datasets contained significant portions of romanized text (46% and 23% respectively). The full model, incorporating sparsity and null correspondence (MCTM-DSGNP), significantly outperformed baselines.

Comparison of model fit and perplexity across different numbers of topics, showing MCTM-DSGNP's superior performance in negative log-likelihood and held-out data perplexity.

Key Findings

  • Superior Model Fit: As shown in the performance metrics, MCTM-DSGNP achieved lower perplexity and better log-likelihood than simpler variants, indicating a more accurate representation of the data.
  • Robust Spam Detection: The model achieved up to a 589% improvement in F1-score for detecting "robotic" comments by shunting them into the "null correspondence" category.
  • Segment-Level Precision: The model showed a 15% improvement in precision for mapping comments to specific article segments compared to global correspondence models.

Conclusion

The MCTM provides a unified probabilistic framework for relating messy, multi-script user comments to formal news articles. By leveraging sparsity and modeling segment-level correspondence, the approach successfully extracts meaningful topical relationships without the need for traditional machine translation. This work represents a significant step forward in analyzing user-generated content in linguistically diverse digital landscapes.

Reference: Tholpadi, G., Das, M. K., Bansal, T., & Bhattacharyya, C. (2015). Relating Romanized Comments to News Articles by Inferring Multi-Glyphic Topical Correspondence. Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence.

Summary updated

Abstract

Commenting is a popular facility provided by news sites. Analyzing such user-generated content has recently attracted research interest. However, in multilingual societies such as India, analyzing such user-generated content is hard due to several reasons: (1) There are more than 20 official languages but linguistic resources are available mainly for Hindi. It is observed that people frequently use romanized text as it is easy and quick using an English keyboard, resulting in multi-glyphic comments, where the texts are in the same language but in different scripts. Such romanized texts are almost unexplored in machine learning so far. (2) In many cases, comments are made on a specific part of the article rather than the topic of the entire article. Off-the-shelf methods such as correspondence LDA are insufficient to model such relationships between articles and comments. In this paper, we extend the notion of correspondence to model multi-lingual, multi-script, and inter-lingual topics in a unified probabilistic model called the Multi-glyphic Correspondence Topic Model (MCTM). Using several metrics, we verify our approach and show that it improves over the state-of-the-art.
BibTeX
@inproceedings{tholpadi2015relating,
  author = {Goutham Tholpadi and Mrinal Kanti Das and Trapit Bansal and Chiranjib Bhattacharyya},
  title = {{Relating Romanized Comments to News Articles by Inferring Multi-Glyphic Topical Correspondence}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  url = {https://openreview.net/forum?id=rkWXppx_ZH},
}
Figure