Sentiment Analysis: AI-Powered Review Classification

MOR Software partnered with a leading Japan-based retail company to build an AI-powered sentiment analysis system capable of automatically classifying large volumes of Japanese-language customer reviews in real time. Delivered within a 4-month timeline and requiring 10 man-months (10MM) of engineering effort, the solution combined Natural Language Processing (NLP), deep learning, and cloud infrastructure to transform unstructured customer feedback into actionable business insight, enabling the retailer to continuously improve customer experience based on real, data-driven signals rather than manual review sampling.

• A Japan-based retail enterprise operating a large-scale e-commerce platform serving thousands of daily transactions.
• Collects a high volume of user-generated product and service reviews in Japanese across multiple product categories.
• Relies heavily on customer feedback to guide product improvement, merchandising decisions, and service quality initiatives.
• Sought to modernize its feedback analysis process to keep pace with the growing scale of its online marketplace.

• High volume of unstructured review data: Thousands of Japanese-language reviews generated daily made manual analysis slow, inconsistent, and impossible to scale.
• Linguistic complexity of Japanese text: Unlike English, Japanese has no explicit spacing between words, requiring specialized word segmentation (tokenization) to accurately break sentences into meaningful units.
• Context and nuance sensitivity: Japanese customer reviews often contain indirect expressions, polite/formal phrasing, and subtle emotional cues that generic sentiment models fail to capture accurately.
• Need for real-time processing: The business required a system that could classify sentiment as reviews came in, rather than through periodic batch reporting.
• Business risk of inaction: Without accurate sentiment insights, the retailer risked missing early signals of product or service dissatisfaction, directly impacting customer retention and brand reputation.

MOR Software designed and deployed an end-to-end NLP pipeline tailored specifically to the linguistic characteristics of Japanese text, built for scalability and real-time inference.

NLP Processing Pipeline

• Data Exploration: Conducted in-depth analysis of the review dataset to understand vocabulary distribution, review length, sentiment imbalance, and domain-specific terminology.
• Data Cleansing: Removed noise, duplicates, and irrelevant content while standardizing text formatting to improve model input quality.
  Japanese Tokenization with MeCab: Applied MeCab, a purpose-built morphological analyzer, to accurately segment Japanese text into words and grammatical units — a critical step absent in English-centric NLP workflows.

Word Embedding Layer

• Built custom Word2Vec embeddings to represent tokenized Japanese words as dense vectors capturing semantic relationships.
• Combined and benchmarked embedding approaches using Gensim, FastText, and BERT, allowing the team to evaluate trade-offs between training speed, subword handling (important for Japanese morphology), and contextual accuracy.

Sentiment Classification Model

• Developed a hybrid CNN-LSTM architecture:
  • CNN layers extracted local n-gram-level text features and key sentiment-bearing phrases.
  • LSTM layers captured sequential context and word order, preserving the meaning that depends on sentence structure.
• Performed Hyperparameter Optimization (HPO) to fine-tune model depth, embedding dimensions, learning rate, and dropout for optimal accuracy.
• Conducted rigorous model testing and validation using held-out review datasets to ensure generalization across product categories.

Technology Stack & Infrastructure

• Language/Framework: Python 3, TensorFlow
• NLP Tools: MeCab, Gensim, FastText, BERT
• Cloud Infrastructure: Deployed and scaled on AWS, ensuring reliable real-time inference and elastic capacity during peak review traffic

• Delivered a production-ready sentiment analysis system capable of automatically classifying incoming Japanese-language reviews in real time.
• Achieved high classification accuracy by combining specialized Japanese tokenization with a CNN-LSTM hybrid model tuned for the language's linguistic nuances.
• Enabled the retail client to continuously monitor customer sentiment trends across products and categories without manual review.
• Empowered business and product teams to make faster, data-backed decisions to enhance customer experience and address dissatisfaction early.
• Established a scalable AI foundation that can be extended to additional languages, review sources, or downstream analytics use cases.