
Sentiment Bundle Benchmark
TF-IDF, embeddings, fine-tuned encoders and a local language model, evaluated on fixed splits across six sentiment datasets. The successor to GURA.
AI/ML
Spin-off · Master's thesisText Preprocessing in Neural Text Classifiers
An unofficial, reproducible re-run of a BlackboxNLP 2018 paper's IMDb experiments: four ways to preprocess the text, two neural models, every run traceable. It started as research for my master's thesis in 2019. That version had three bugs; this one fixes them and says where they were.

Does it matter whether you lowercase, lemmatize or group multiword expressions before training a neural text classifier? Camacho-Collados and Pilehvar measured it in On the Role of Text Preprocessing in Neural Network Architectures (BlackboxNLP 2018). This repository re-implements the IMDb part of that study: pick one of four preprocessing variants (vanilla, lowercased, lemmatized, multiword) and one of two models (CNN or CNN+LSTM), and preproc-imdb trains on the 25,000 training reviews and reports the accuracy on the 25,000 test reviews. Every run leaves a folder with its settings, library versions, training history and every prediction. It started in 2019 as a Python 3 port of the authors' sensecnn code for my master's research, and was rewritten in 2026 on Keras 3.15 and TensorFlow 2.21, fixing three bugs in the 2019 port.
My master's research needed to know whether lowercasing, lemmatizing or joining multiword phrases changes what a neural sentiment model learns, so I re-ran a paper's IMDb experiments. In the rerun, changing the random seed moved accuracy more than changing the preprocessing did.
For anyone wondering whether lowercasing or lemmatizing is worth it: pick a variant and a model, get an accuracy traceable to a folder, and compare it with the paper. Your own preprocessing needs four text files, one command and patience: a CNN+LSTM run takes 13 to 19 minutes on CPU.
Simplified on purpose. The real thing has more arrows and at least one config file nobody understands.

TF-IDF, embeddings, fine-tuned encoders and a local language model, evaluated on fixed splits across six sentiment datasets. The successor to GURA.

The earlier BiGRU + self-attention sentiment experiment. Preserved for reference; current comparisons live in Sentiment Bundle Benchmark.

Led the sentiment engine behind a company's social-listening products, reading more than three billion crawled posts. Lesson learned: the internet is mostly complaints.