When Words Hurt

The problem

Online misogyny is rarely just a list of insults. Often it is carried by neutral words used in a pejorative context, by sarcasm, or by a threat phrased without a single vulgar word. The question of the thesis was simple: how well, and with which limits, can a text classifier recognise it in Italian social media language?

I used the dataset of the AMI task at EVALITA 2020: 13,115 tweets annotated by hand. I kept only the binary label (misogynistic or not): 58.4% of the tweets are non-misogynistic and 41.6% misogynistic, a mild imbalance that makes macro-F1 a better yardstick than accuracy.

Cleaning a very noisy language

Tweets are full of slang, typos, emoji and abbreviations, so most of the work went into the preprocessing pipeline:

• Cleaning: URLs and mentions replaced by placeholders, emoji collapsed into a single token, punctuation removed except ! and ?, which carry emphasis.
• Normalisation: lowercasing, a dictionary that expands common abbreviations (xkè → perché), and elongated words shortened (ciaooo → ciao).
• Lemmatisation with spaCy, so that all the forms of a verb or noun share one token.
• Disambiguation: the step I found most interesting. Words like cagna mean “female dog” in one context and are a slur in another. For a lexicon of such words I defined two textual “anchors”, one neutral and one pejorative, and used a zero-shot NLI model (mDeBERTa) to decide which anchor fits the tweet. The ambiguous word is then replaced by that anchor, so the downstream models see the meaning, not just the word.

The pipeline cut the vocabulary from 20,007 to 13,342 unique words (−33%) and the share of out-of-vocabulary tokens, measured against the AlBERTo vocabulary, from 18.3% to 10.8%.

Two families of models

Classical models on TF-IDF features. Multinomial Naive Bayes, logistic regression, linear SVM, decision tree and random forest, each with a feature-selection step and tuned by grid search with 5-fold cross-validation on the training set.

A CNN on word embeddings. Pretrained GloVe vectors learned on tweets (100 dimensions), kept trainable so they can adapt to the jargon of the dataset, followed by a 1D convolution, global max pooling, dropout and a sigmoid output. I deliberately kept it small to limit overfitting, and trained it with early stopping on a validation set.

The data was split once, stratified, into 60% training, 20% validation and 20% test; the test set was used only for the final evaluation.

Multinomial Naive Bayes 79.2 % Multinomial Naive Bayes 79.2 %

Linear SVM 80.4 % Linear SVM 80.4 %

Logistic regression 80.5 % Logistic regression 80.5 %

Decision tree 82.4 % Decision tree 82.4 %

Random forest 82.9 % Random forest 82.9 % · best classical

CNN + GloVe 83.0 % CNN + GloVe 83.0 % · best overall

75 % 77.5 % 80 % 82.5 % 85 %

Scores ×100. The random forest also has the highest AUC of all models (0.933 vs. 0.927 for the CNN).

Show as table

Approach	Value	Note
Multinomial Naive Bayes	79.2 %
Linear SVM	80.4 %
Logistic regression	80.5 %
Decision tree	82.4 %
Random forest	82.9 %	best classical
CNN + GloVe	83.0 %	best overall

Same score, different mistakes

The random forest’s most important features are explicit insults and words that objectify the female body. That makes it an excellent detector of open abuse, and a fragile one everywhere else:

• “Beh, non mi sento sottomessa o malmenata” (“Well, I don’t feel subjugated or beaten up”) is flagged as misogynistic: the model reacts to malmenata and ignores the non in front of it.
• “Ti apro in due” (“I’ll split you in two”) is missed: no single word is offensive, the violence is in the construction.

The CNN reads short word sequences through its convolutions, so it can partly learn that a negation flips the meaning of the verb after it, and its fine-tuned embeddings pick up the dataset’s slang. It is not immune to these errors, but it is better placed to handle them.

What I took away

• A strong baseline is a real result. With careful preprocessing and feature selection, a random forest was within 0.001 macro-F1 of the neural model. Any claim that a model is “better” needs that baseline next to it.
• Look at the errors, not only the score. Two models with the same F1 failed on very different tweets, and that tells you much more about what they learned.
• Context is the hard part. Negation, sarcasm and implicit threats are exactly where both models struggled. The natural next step is contextual representations: using the hidden states of a pretrained language model as features, or fine-tuning one directly — the direction I then followed in my master’s.