Detecting Anchors' Opinion in Hinghlish News Delivery

Siddharth Sadhwani, Nishant Grover, Md. Shad Akhtar, Tanmoy ChakrabortyarXiv 2022 · DOI · arXiv

Introduction

In the landscape of modern journalism, the ideal of objective, fact-based reporting often clashes with the high-pressure environment of live television. This is particularly evident in Indian news debates, where anchors frequently transition from neutral moderators to opinionated participants. These exchanges often occur in "Hinglish"—a code-mixed blend of Hindi and English—which adds a layer of linguistic complexity to automated analysis.

Table 1: Examples of anchor (A) and speaker (S) utterances with corresponding opinion labels in the ODIN dataset.

This research introduces the task of detecting personal opinions voiced by news anchors in Hinglish debates. To address this, the authors present ODIN, a specialized dataset, and DetONADe, an interactive attention-based deep learning framework designed to identify these subjective shifts.

The ODIN Dataset

The ODIN (Opinion Detection In News) dataset was constructed from 46 debate videos across three major Indian news channels: ABP News, Aaj Tak, and Zee News. The dataset comprises 4,490 total utterances, with a specific focus on the 2,054 utterances delivered by anchors.

Table 2: Comprehensive features and statistics of the ODIN dataset, highlighting the distribution of anchor utterances.

The annotation process involved manual transcription and labeling by two experts, achieving a high inter-annotator agreement (Cohen’s kappa of 0.88). Approximately 30% (597) of anchor utterances were identified as opinionated, providing a realistic representation of news delivery where neutral reporting remains the baseline.

The DetONADe Framework

The proposed model, DetONADe (Detecting Opinion in News Anchor Delivery), utilizes a multi-layered architecture to handle the nuances of code-mixed speech and conversational context:

  • HIT Embeddings: Hierarchically Attentive Transformer embeddings operate at the sub-word level. This makes the model resilient to the non-standard spellings and phonetic variations typical of transcribed Hinglish (e.g., "laiv" for "live").
  • Sequence Learning: A bidirectional LSTM (biLSTM) layer captures the temporal dynamics of the debate, acknowledging that an utterance's meaning is often contingent on the preceding dialogue.
  • Interactive Attention: This component allows the model to "attend" to specific parts of the prior conversation. By treating the target utterance as a context vector, the model identifies triggers in the previous five utterances that might prompt an anchor to express an opinion.

Key Insights and Results

The study revealed several behavioral patterns in news anchors. Anchors tend to become more opinionated as debates progress, with peaks in subjectivity occurring during the middle (20-40%) and final (80-100%) segments of a broadcast. Linguistically, opinionated statements are often marked by specific Hinglish fillers like bilkul (certainly) and matlab (means), and a higher frequency of interrogative words (1.45 per utterance vs. 0.9 in neutral ones).

In performance benchmarks, DetONADe achieved a weighted-F1 score of 0.703, outperforming strong multilingual baselines such as ML-BERT (0.686) and XLM-RoBERTa (0.669). While the model excels at identifying neutral statements (0.778 F1), detecting subtle opinions remains a challenge (0.510 F1), highlighting the inherent difficulty of parsing context-heavy subjective speech.

Conclusion and Future Directions

While ODIN and DetONADe provide a robust foundation for opinion detection in code-mixed environments, several avenues for improvement remain. The current reliance on text-only transcripts misses critical non-verbal cues such as tone of voice and facial expressions. Future work aims to incorporate multi-modal data (audio and video) and explore dynamic context windows to better capture the fluid nature of televised debates.

Summary updated

BibTeX
@misc{sadhwani2022detecting,
  author = {Siddharth Sadhwani and Nishant Grover and Md. Shad Akhtar and Tanmoy Chakraborty},
  title = {{Detecting Anchors' Opinion in Hinghlish News Delivery}},
  year = {2022},
  eprint = {2204.02155},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2204.02155},
}
Figure