ABSTRACT
This project addresses the problem of extracting fine-grained insights from multilingual and
code-mixed OTT reviews to enable personalized recommendations and content optimization.
It focuses on analyzing reviews from OTT platforms like Netflix and Hotstar, which are often
written in multiple or mixed languages. Instead of just saying whether a review is positive or
negative, the system looks deeper to understand what people are talking about — such as the
story, acting, pacing, or music — and whether their opinion is good or bad for each aspect. In
the first stage, simple machine learning methods are used to find basic sentiment and themes.
Later, advanced AI models that understand multiple languages are applied to handle sarcasm,
detect actors or show names, and create short summaries of reviews. Finally, the results are
shown in a dashboard with scores, trends, and summaries, helping OTT platforms get clear
insights into audience feedback and provide better recommendations.
INTRODUCTION
OTT platforms like Netflix and Hotstar receive thousands of multilingual and code-mixed
reviews, reflecting audience opinions on story, acting, pacing, music, and more. Manually
analyzing these reviews is difficult and time-consuming. OTT Review Intelligence aims to
automatically extract aspect-level sentiment and generate concise summaries, helping content
teams understand viewer feedback and provide personalized recommendations. The project starts
with a prototype using TF-IDF and SVM for basic sentiment analysis and topic modeling, then
advances to transformer-based models like mBERT/XLM-R for fine-grained sentiment, sarcasm
detection, entity recognition, and summarization, all accessible via an interactive dashboard.
Challenges include sarcasm, noisy text, and limited labeled data.
PROBLEM STATEMENT
OTT platforms generate a huge volume of multilingual and code-mixed user reviews, containing
valuable insights about story, acting, pacing, music, and other aspects. Manually analyzing these
reviews is time-consuming and often inaccurate. The challenge is to automatically identify
aspect-level sentiments, handle sarcasm and code-mixed language, generate concise summaries,
and provide actionable insights for content teams. This enables fine-grained recommendations to
viewers and helps platforms improve content strategies. Key difficulties include imbalanced
aspect data, noisy text, and limited labeled datasets, requiring advanced natural language
processing techniques for effective analysis.
OBJECTIVES
Analyze aspect-level sentiment (story, acting, pacing, music) from user reviews
Handle multilingual and code-mixed text efficiently
Detect sarcasm and indirect opinions for accurate sentiment
Generate concise summaries of audience feedback
Provide fine-grained personalized content recommendations
Display insights via an interactive dashboard with trends and scores
Odd Semester (Prototype)
Dataset: TRAI/FCC complaints + Hinglish/Tenglish synthetic samples.
Preprocessing: Tokenization, transliteration normalization.
Baseline Classifier: TF-IDF + Logistic Regression/SVM.
Regex Extraction: Extract account IDs and recharge amounts.
Even Semester (Product)
Transformer Fine-Tuning: mBERT/XLM-R for multilingual classification.
NER Extraction: Identify location, plan type, recharge details.
Semantic Similarity Fallback: Handle ambiguous cases.
Explainability: Use SHAP for model interpretation.
Deployment: FastAPI with interactive dashboard.
Challenges & Bottlenecks
Handling code-mix slang.
Imbalanced complaint categories.
Risk of routing errors.
Ensuring model explainability for regulators.
Expected Outcomes
Accurate and faster routing of telecom complaints.
Transparent system with SHAP explanations.
Deployment-ready solution with dashboard.
Future Work
Expand support to more Indian languages.
Integrate voice-to-text complaints from call centers.
Implement continuous learning from new complaint data.
Conclusion
This project presents an NLP-driven solution to automate telecom complaint routing. By
moving from a TF-IDF baseline to advanced transformer models, it addresses misrouting
caused by code-mixing and slang, improving customer satisfaction and regulatory
compliance.
References
1. TRAI Public Complaint Data
2. FCC Complaint Database
3. HuggingFace Transformers Documentation
4. SHAP Explainability Framework