Project Athena is a machine learning-based NLP project developed as part of the Human Language Technologies (HLT) course, a.y. 2023-2024.
Its goal is to classify philosophical currents from textual data, inferring the school of thought behind a given sentence.
The project explores:
- Dataset analysis and contrastive comparisons with Gutenberg, Brown, and Simple English Wikipedia corpora
- Multiple modeling approaches: Naive Bayes, RNNs, BERT, DistilBERT, and Zero-shot bart-large-mnli
- Evaluation via macro-averaged F1 score, addressing dataset imbalance
- A hypothesis on the impact of short vs. long sentences in classification accuracy
Full report: HLT_Project.pdf
We used the History of Philosophy dataset curated by Kourosh Alizadeh.
- Contains 360,808 sentences
- Drawn from 59 texts by 36 authors
- Covers 13 philosophical schools of thought
Note: The dataset is highly imbalanced, with some schools being over-represented.
We experimented with three families of models:
- Generative models → Naive Bayes (TF-IDF, Laplace smoothing tuning)
- Discriminative models → Recurrent Neural Networks (LSTMs, GRUs, BiLSTMs with GloVe embeddings)
- Transformer-based models → BERT, DistilBERT, Zero-shot bart-large-mnli
Hypothesis: very short sentences (<15 words) may act like "semantic stopwords" and hinder classification.
We tested this by training and evaluating on both the full dataset and a reduced dataset (sentences ≥ 84 characters).
- BERT and DistilBERT outperformed all other approaches
- Naive Bayes surprisingly strong baseline with low computational cost
- Removing short sentences hurt performance, as they actually contribute useful signals
- Best model: BERT (fine-tuned on full dataset) → F1 ≈ 0.83 (full test set), 0.88 (long sentences only)
- Davide Borghini
- Davide Marchi
- Giordano Scerra
- Andrea Marino
- Yuri Ermes Negri