PejorativITy: Disambiguating Pejorative Epithets to Improve Misogyny Detection in Italian Tweets

Wed, 01 May 2024 00:00:00 +0000

Research setting

Misogyny detection is a text classification task: a model reads a social media post and decides whether it expresses hate toward women. Misogyny is often figurative, and many neutral words become insults when they are aimed at a woman. The paper calls such words pejorative epithets: a word with a primary neutral meaning and a second, negative one. In Italian, balena means whale, but it can also insult a woman as fat. Only the context of the tweet tells the two senses apart.

One lexicon word, two senses. An anchor is a word with a single meaning that stands for one sense. Example from Table 1 of the paper (balena has a second neutral anchor, balenare, to flash).

Word sense disambiguation is the task of picking the right sense of a word in its context. Here it is a binary decision: is this word, in this tweet, used in a pejorative or in a neutral way?

Motivation

When a word such as balena appears with its insulting sense in the training data, misogyny classifiers tend to flag every tweet that contains it. A tweet about a real whale then becomes a false positive, that is, a non-misogynous tweet labeled as misogynous. Earlier work studied pejorative words or misogyny, but it never used the sense of a word to help misogyny detection.

Work	Italian data	Word-level labels for ambiguous words	Targets neutral words used as insults	Word sense informs misogyny detection
Pejorative lexicon and tweets (Dinu et al., 2021)	partially	yes	yes	no
SWAD swear word dataset (Pamungkas et al., 2023)	no	no	no	no
AMI misogyny datasets (Fersini et al., 2018, 2020)	yes	no	not addressed	no
This paper (PejorativITy) This work	yes	yes	yes	yes

How PejorativITy relates to the prior work discussed in the paper (Sections 1, 2, and 5).

Objective. Find out which epithets are used against women in Italian tweets, test whether disambiguating them lowers the error rate of misogyny detection, and check whether language models can tell the two senses apart from context.

Approach

The authors first built a lexicon of 24 Italian words that tweets use with both a neutral and a pejorative sense, each with its neutral and pejorative anchors. They then collected 50 tweets per word, from December 2022 to February 2023, and an expert annotator labeled each tweet for misogyny and each lexicon word for pejorativity. On this corpus, called PejorativITy, a first AlBERTo model (a BERT model pre-trained on Italian tweets), named model_pej, learns to label the lexicon word as pejorative or neutral. Its answer then enriches the tweet before a second AlBERTo model, named model_mis, classifies misogyny.

Corpus 1,200 tweets with 24 ambiguous words, labeled for pejorativity and misogyny.
model_pej Labels the lexicon word as pejorative or neutral.
Inject the sense Two ways to add the answer to the tweet.
- CONCAT append the label
- SUBST swap in the anchor
model_mis Classifies the enriched tweet as misogynous or not.

The PejorativITy pipeline (Sections 3 to 5 and Figure 1 of the paper).

Class	Training	Test	Total
Misogynous	369	28	397
of which with a pejorative word	363	28	391
Non-misogynous	735	68	803
of which with a pejorative word	172	18	190

The PejorativITy corpus (Table 3 of the paper). Some non-misogynous tweets use the word pejoratively, for example in reported speech or against men.

Results

The models are scored with macro-F1, the F1 score averaged over the two classes, from 0 to 1 and averaged over three runs. Misogyny detection is tested on the PejorativITy test set (96 tweets) and on the test sets of AMI-2018 and AMI-2020 (1,000 tweets each), the two Italian misogyny benchmarks. The AMI tweets have no pejorativity labels, so model_pej predicts them. Each approach is compared with a baseline: AlBERTo fine-tuned on the original tweets.

0.82macro-F1 of model_pej in telling pejorative from neutral uses
+9 pointsmisogyny macro-F1 on PejorativITy with SUBST, from 0.68 to 0.77
25 → 16false positives on the PejorativITy test set with CONCAT

Macro-F1 on PejorativITy (higher is better)

Baseline 0.68
CONCAT, predicted This work 0.75
SUBST, predicted This work 0.77
CONCAT, gold This work 0.83
SUBST, gold This work 0.87

Misogyny detection on the PejorativITy test set. The gold rows use the annotators’ pejorativity labels instead of model_pej predictions, an upper bound for the pipeline. Source: Table 5 of the paper.

Show as table

System	Macro-F1 on PejorativITy
Baseline	0.68
CONCAT, predicted (this work)	0.75
SUBST, predicted (this work)	0.77
CONCAT, gold (this work)	0.83
SUBST, gold (this work)	0.87

Test set	Baseline	CONCAT	SUBST
AMI-2018, tweets with a lexicon word (34)	0.79	0.82	0.79
AMI-2018, whole test set	0.86	0.86	0.86
AMI-2020, tweets with a lexicon word (192)	0.77	0.81	0.77
AMI-2020, whole test set	0.82	0.83	0.82

Macro-F1 on the AMI benchmarks. Source: Table 7 of the paper.

Both ways of injecting the word sense improve misogyny detection on PejorativITy, and with gold labels the gain reaches +19 points. On the AMI tweets that contain a lexicon word, CONCAT gains 3 points on AMI-2018 and 4 points on AMI-2020. SUBST brings no gain there, possibly because some anchors fit the tweet poorly. The whole AMI test sets barely change, since few of their tweets contain a lexicon word.

Fine-tuning also moves the AlBERTo representation of a lexicon word toward the anchor of its sense (Section 7). The average cosine similarity with pejorative anchors becomes 0.39 in pejorative tweets and 0.20 in neutral ones, against 0.34 and 0.29 before fine-tuning. Among three open LLMs prompted without examples, Mistral 7B explained the sense of lexicon words best (Section 8). It still missed some pejorative uses and reported speech, and the two LLaMA-based models mostly answered that the word means itself.

Takeaways