Jakarta Hotel Review Sentiment Analysis
PROJECT OVERVIEW & OPERATIONAL CONTEXT
Jakarta Hotel Review Sentiment Analysis is an applied natural language processing system engineered to automatically aggregate, process, and evaluate guest feedback across the hospitality sector in Jakarta, Indonesia. Analyzing unstructured customer reviews at scale poses distinct operational challenges, particularly when subtle negation cues inverted by naive stopword removal distort underlying sentiment. The architecture addresses this through an integrated pipeline combining automated geospatial data extraction via SerpAPI with a customized NLTK preprocessing engine that preserves critical negation tokens during WordNet lemmatization. Feature extraction is governed by a bounded TF-IDF vectorizer feeding into a maximum-margin Linear Support Vector Classifier (SVC), strategically chosen for its mathematical robustness in sparse, high-dimensional textual feature spaces. Serialized into a unified artifact and served via an interactive Streamlit web application, the system delivers instantaneous CPU-efficient sentiment predictions, enabling hotel operators to detect service bottlenecks and prospective guests to rapidly assess hospitality quality.
PROBLEM STATEMENT & SYSTEM CONSTRAINTS
Standard natural language preprocessing pipelines universally discard stopwords, which inadvertently purges negation words such as 'not' and inverts polarized feedback like 'not clean' into false positive classifications. Additionally, hotel reviews exhibit extreme length variance, colloquial vocabulary, and sparse feature representations, causing non-linear models to overfit while manual review auditing fails to scale across thousands of guest entries.
- Automate the extraction of localized Google Maps hotel guest reviews across Jakarta using SerpAPI pagination query loops.
- Eliminate sentiment polarity inversion by engineering an NLTK text normalization pipeline that selectively retains negation terms.
- Reduce vocabulary dimensionality and singleton noise using bounded TF-IDF feature extraction with max_df=0.5 and min_df=2 thresholds.
- Construct an optimal maximum-margin separating hyperplane between binary sentiment classes using a Linear Support Vector Classifier.
- Prevent offline-online feature skew by serializing the vectorizer and classifier together into a single atomic pickle artifact.
- Deploy an accessible, zero-dependency interactive Streamlit web dashboard capable of sub-100ms on-demand review classification.
Negation-Preserving NLP Preprocessor
Sanitizes raw text with regular expressions and filters English stopwords while explicitly retaining the token 'not' to prevent reversing critical sentiment polarity.
Morphological WordNet Lemmatization
Reduces inflected token variants to grammatical base forms using WordNet and OMW corpora, stabilizing vocabulary frequencies across diverse reviewer styles.
Document-Bounded TF-IDF Vectorization
Transforms normalized text into sparse feature matrices using document frequency constraints that discard ubiquitous terms and singleton noise.
Maximum-Margin Linear SVC Engine
Fits an optimal convex separating hyperplane with linear kernel formulation to robustly partition positive and negative sentiment in high-dimensional text space.
Atomic Pipeline Pickle Serialization
Encapsulates both the fitted vectorizer and trained classifier within a unified binary pickle file to guarantee training-serving feature alignment.
Interactive Streamlit Inference Dashboard
Provides a responsive browser GUI featuring asynchronous text input, real-time token transformation, and instant binary sentiment classification displays.
SerpAPI Google Maps crawler + NLTK negation-preserving WordNet lemmatizer + Document-bounded TF-IDF vectorizer + Linear Support Vector Classifier (SVC) + Streamlit web serving runtime
Automated Geospatial Review Extraction
Execute paginated SerpAPI queries against Google Maps search engines targeting hotels in Jakarta with >= 100 reviews to build the raw corpus.
Text Sanitization & Token Normalization
Strip non-alphabetic characters via regex, lowercase characters, and filter standard English stopwords while preserving the negation token 'not'.
Morphological WordNet Lemmatization
Apply NLTK WordNetLemmatizer across filtered tokens to map inflected grammatical variations back to base dictionary lemma representations.
TF-IDF Matrix Assembly & L2 Normalization
Transform cleaned review strings into sparse statistical term-frequency matrices bounded by 50% max document frequency and min count of 2.
Linear SVC Hyperplane Margin Scoring
Evaluate dot-product decision function w^T * x + b on the normalized TF-IDF vector to classify review sentiment into positive or negative polarity.
Streamlit Reactive UI Rendering
Desequence classification output and dynamically render binary prediction state to the client browser through reactive Streamlit components.
SYSTEM EVALUATION & EMPIRICAL VALIDATION
VERIFIED_TELEMETRYThe linear support vector classifier was evaluated on a 70/30 holdout split across 210 harvested Google Maps reviews, comprising 147 training instances and 63 unseen test reviews across seven prominent Jakarta hotel properties. Bounded TF-IDF feature extraction with document frequency constraints filtered ubiquitous terms while preserving discriminative n-grams, enabling the linear kernel to achieve clean convex margin separation across sparse high-dimensional textual coordinates. Explicit preservation of the negation token 'not' directly resolved common baseline false-positive inversions on critical negative feedback, allowing the model to accurately differentiate subtle dissatisfaction from praise. In local deployment benchmarks on a standard CPU workstation, the complete inference pipeline—encompassing regular expression sanitization, NLTK WordNet lemmatization, sparse vector transformation, and SVC hyperplane projection—achieved an average inference latency of under 4.8ms per query, sustaining real-time interactive responsiveness within the Streamlit web dashboard with zero server overhead.
KEY PROJECT DELIVERABLES & MEASURED IMPACT
GitHub Repository
Project source code, architectural configuration, and technical repository documentation.
Technical Documentation & Spec
Primary architectural specification: PRO-15 // AI-02 (Python, scikit-learn, NLTK, Pandas).
Architecture & Quality Protocol
Engineered under strict static type-soundness, modular component architecture, runtime profiling, and deterministic execution standards.
KIP Student Eligibility Prediction
Supervised machine learning decision-support platform leveraging Gaussian Naive Bayes and one-hot categorical encoding to classify student scholarship eligibility for Indonesia's PIP/KIP program from demographic and socioeconomic features, deployed via an interactive real-time Streamlit web dashboard.
Center of Mass and Velocity Calculator
Vectorized scientific computing engine implemented in Python and NumPy to simulate discrete N-body mechanical systems, calculating spatial center-of-mass coordinates and aggregate kinematic velocity vectors with instant 2D/3D scatter and quiver vector field projections in sub-millisecond execution times.