Misogyny is often expressed through figurative language. Some neutral words can assume a negative connotation when functioning as pejorative epithets, and they can be used to express misogyny. Disambiguating the meaning of such terms might help the detection of misogyny. This challenge addresses a) the disambiguation of specific ambiguous words in a given context; b) the detection of misogyny in instances that contain such polysemic words. In particular, framed as a binary classification, our task is divided into two parts. In Task A, the model is asked to define if, given a tweet, the target word is used in pejorative or non-pejorative way. In Task B, the model is asked whether the whole tweet is misogynous or not.
CEUR ceur-ws.org
This CALAMITA challenge [1] addresses the task of disambiguating pejorative language to detect forms of misogyny that are masked within ambiguous and contextdependent expressions. Pejorative language refers to a word or phrase that has negative connotations and is intended to disparage or belittle.1 An inofensive word becoming pejorative is a form of semantic drift known as pejoration; thus, pejorativity is context-dependent: pejorative words have one primary neutral meaning, and another negatively connotated meaning. In this challenge, our objective is to evaluate large language models (LLM) in Italian by focusing on the disambiguation tation in the training set and a neutral connotation in the test set results in a great number of false positives [5]. This could be overcome by decoder-based LLMs, as they could rely on their implicit knowledge to grasp the meaning of such terms. By asking models to determine whether a term is being used in a pejorative or non-pejorative sense, we challenge the LLMs’ ability to comprehend semantic shifts in Italian. Moreover, asking whether a sentence containing that term is misogynous or not, enables us to comprehend to what extent
through figurative language. We expect models to struggle with this challenge, particularly in sentences with non-standard or regional varieties of Italian, which occur in our corpus.
of pejorative epithets used online to express misogyny. 2. Challenge: Description cial environments where women perceived as violating We introduce pejorative language disambiguation as a patriarchal norms are “kept down” through hostile or preliminary step to detect misogyny. Our goal is to assess benevolent reactions coming from men, other women, whether the disambiguation of potentially pejorative epand social structures [2, 3], in the form of sexual objec- ithets improves the detection of misogynistic language. tification, male privilege, gender discrimination, sexual
harassment, belittling and violence [4].
whose standard meaning refers to the sea mammal, but it is used ofensively to address an overweight woman. Encoder-based models struggle to correctly classify misogyny when sentences contain such terms: the
Fig. 1 shows the pipeline for our tasks. Assume the senlated as Can’t look at that whale with jeans.
occurrence of polysemic words with a pejorative conno- Both tasks are conceived as binary classification tasks. CLiC-it 2024: Tenth Italian Conference on Computational Linguistics, tence Quella balena coi jeans non si può guardare, trans0000-0002-3387-6557 (A. Muti) © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License the meaning of the target word (balena in our example) Attribution 4.0 International (CC BY 4.0).
is pejorative or not. The model should rely on its internal knowledge accumulated during pre-training to under- words refer to the unambiguous words used to define stand whether the term balena (whale) refers to woman polysemic words. We call these words anchors because or cetaceus. Ideally, the model should exploit the context their meaning is univocal and does not change according to perform the disambiguation, as the image of a whale to the context. For instance, the word balena (whale) is with jeans is not plausible. That is why we encourage used to refer to either a sea mammal or an overweight commonsense reasoning for this task. woman. In contrast, the anchor words cetaceo (cetacean)
Task B: In the second step, the model is first informed and grassa (fat) only refer to the animal in the first case with the decision of Task A, whether the target word is and to being overweight in the second case, at least as pejorative or not, and then asked to classify the input far as their use in Twitter is concerned.4 sentence as misogynous or not. 2https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obsceneand-Otherwise-Bad-Words/tree/master, consulted on January 2023. 3Due to their exclusive neutral or negative connotation on Twitter, the following words are discarded: barile, banco, botte, barbona, 4In this case, the word balena has a third anchor word, from the verb facile, gatta morta, passeggiatrice, porca, principessa, privilegiata, balenare, which means ’to flash’. psicopatica, scrofa, somara, travestita. 5https://twarc-project.readthedocs.io
Lexicon. We collect our lexicon by selecting words from three distinct sources.
(1) We ask ten Italian native speakers to provide a list of ofensive words used online to address women. The speakers use social media on a daily basis and their age ranges between 27 and 39 years.
(2) We retrieve the keywords used in the two Italian corpora for the Automatic Misogyny Identification (AMI) shared task [6, 7].
(3) We consult the ’List of Dirty Naughty Obscene Bad Words’.2
We only keep polysemic words whose primary meaning is neutral and that are frequently used on Twitter with both pejorative and neutral connotations. To ensure the quality of our vocabulary, we qualitatively verify that such words are used with both connotations by manually searching them on Twitter.3
Table 1 shows our lexicon of 24 words. For each word, we report the English translation of its literal and pejorative meaning, and their anchors in Italian. Anchor Tweets. We use Twarc5 to retrieve tweets from December 2022 to February 2023 containing words in our lexicon. We select 50 tweets for each word in our lexicon, resulting in 1,200 tweets. We maintain a balance of pejorative and neutral use of lexicon words, although an equal distribution for each word could not be guaranteed. We choose tweets as source of data for three reasons. First, Twitter is a prominent platform for expressing opinions, where language is varied, conversational, and often informal, which makes it suitable to analyze misogyny conveyed through figurative language. Second, at the time of data collection, Twitter API was public and free, which facilitated our data collection process. Third, the character limit on tweets encourages condensed language, limiting the context of expression. Choosing tweets allows us to challenge LLMs in disambiguating pejorative language for misogyny detection within the constraints of limited or lack of context.
tics, gender studies, cognitive sciences, and NLP to label our corpus for pejorative word disambiguation (wordlevel) and misogyny detection (sentence-level).
We first devise a pilot annotation study to explore the complexity of the task. For this purpose, we follow a descriptive annotation paradigm [8], which encourages annotator subjectivity by not providing guidelines. We split the annotators into two groups and assign 50 tweets each for labeling. Each group is composed of two women and one man with ages ranging between 27 and 39 years old. We use Krippendorf’s alpha [ 9] to measure the inter- 3.3. Detailed data statistics annotator agreement (IAA). The IAA of the first group is moderate for both pejorativity (0.48) and misogyny (0.50), Table 3 shows the statistics of our corpus. The Pearson whereas the IAA of the second group is fair for pejorativ- correlation between misogyny and pejorativity labels is ity (0.33) and moderate for misogyny (0.50). We observe 0.70, which is in line with our expectations. The tweets that, in terms of gender diferences, men tend to consider for which misogyny and pejorativity are not aligned are sexual objectifying compliments as non-pejorative. More mainly reported speech or men-related ofensive landetails about the annotation process, including the dis- guage. It is worth noting that some sentences are annocussion of edge cases, can be found in Muti et al. [10]. tated as misogynous, although they do not express any After the pilot studies, we annotate our collected cor- form of hate towards women. However, they contain pus of 1,200 tweets. Only one person carries out the subtle sexist language, which we consider misogynous whole annotation process. We select the annotator with according to the definition provided in Section 1. For the most interdisciplinary background, who is an expert instance, the sentence “che bella bambola ciao tesoro”6 in gender studies, linguistics and NLP, who has been a does not express hate, but perpetuates the objectification target of misogyny. This setting is considered among of women by addressing the target of the tweet as a doll, the best practices for the annotation of phenomena like falling into the category of benevolent sexism [12]. misogyny [11].
Data are collected in an Excel file and published at https://github.com/arimuti/PejorativITy. Each row contains the ID of the tweet, the tweet, the target word, the annotation for pejorativity at word level and the annotation for misogyny at sentence level. Table 2 shows examples.
ity disambiguation at word-level and misogyny detection at sentence-level. We adopt a zero-shot approach, although participants are encouraged to experiment with diferent prompting techniques.
asks whether the target {word} in a given {sentence} is pejorative or not. We prompt the LLMs for each instance in our corpus. The prompt is given as follows:
La parola {word} è utilizzata in modo peggiorativo nel testo in input? Rispondi solo con 1 per sì, 0 per no. Testo: {tweet}
without having the decision for Task A to inform the decision for Task B.
B We enrich the prompt for Task B with the output from the previous task;
Il testo in input è misogino? Rispondi solo con 1 per sì, 0 per no. Testo: {tweet}
Model AlBERTo llama-3-8b llama-3-8b-Instruct
disambiguation. The fine-tuned AlBERTo model reaches a macro F1-measure of 0.82 ± 0.03, as reported in [10]. When it comes to decoder-based models, llama3-8bInstruct shows a lower score, with a diference of 0.18 points, showing room for improvement in the prompt design. However, those scores are not comparable as the testing partitions difer. Llama-3-8b fails to complete the task, since it only repeats the prompt without providing an answer. For this reason, we discard llama-3-8b in the next task. It should be noted that llama has undergone a safety tuning process, preventing the model from always providing an answer, responding I cannot provide a response that condones hate speech. We excluded such cases from the evaluation. Of the 174 excluded instances, 123 were pejorative and 51 were not pejorative according to the gold standard. Although the fine-tuned version of AlBERTo achieves a higher performance (in a smaller subset of instances), llama aids in explainability by deliberately adding explanations of why it considers the target word to be pejorative or not. We will explore the plausibility of such explanations in future work.
Il testo in input è misogino? Rispondi solo con 1 per sì, 0 per no. Considera l'informazione sulla parola {word}: {response from Task A} Testo: {tweet}
misogyny detection at sentence level.
In Setting A, where the model is not informed of the 4. Evaluation of Baselines output for Task A, AlBERTo scores are much lower compared to Task A, achieving 0.68 ± 0.03. Llama performs In order to evaluate our tasks, we employ Macro F1score. better in Task B compared to Task A, overcoming AlWe select two models as baselines: the fine-tuned Al- BERTo by just 0.01 point. However, the fact that all BERTo model [13] and llama-3-8b (both the standard and answers were provided in Task B (unlike in the previous misogyny detection. We have designed two tasks as binary classification problems: A) pejorative language disambiguation at word level and B) misogyny detection at sentence level. Our preliminary experiments show that a Transformer-based fine-tuned model performs better than llama-3-8b-Instruct in detecting pejorative words, while llama-3-8b-Instruct performs slightly better than the Transformer-based model in misogyny detection. In the future, we plan to explore how the unrequested explanations provided by llama-3-8b-Instruct about the pejorativity of a target word impact the classification of misogynous sentences.
that can serve as pejorative epithets for women, it is not an exhaustive list, as we have discarded all the terms task) plays a role and does not necessarily imply that that are not polysemic and that are used only with one misogyny detection is an easier task than pejorativity connotation (either positively or negatively) on Twitter. disambiguation for llama. Moreover, only 100 tweets are annotated by six an
In Setting B, the model is informed of the decision on notators, while the remaining 1,100 are labelled by only pejorativity of the target word. While for llama the infor- one annotator. Although we select an expert with an inmation about pejorativity can be injected in the prompt, terdisciplinary background in linguistics, gender studies with AlBERTo we have adopted two approaches: i) we and NLP to carry out all the annotations, their personal concatenate the information about the pejorativity of biases, opinions, or interpretations can lead to skewed or the target word at the end of the tweet or ii) we substi- one-sided data. tute the ambiguous word with its corresponding anchor Finally, our corpus is characterized by the presence word from our lexicon. Fig. 2 shows the pipeline. We ob- of sarcasm, abbreviations, and non-standard varieties of serve a notable improvement over the baseline model for Italian, which might make the semantics of our instances concatenation (+7 absolute points) and substitution (+9 hard to be captured by current language models. absolute points) when using the predictions for Task A. Another limitation of our study concerns the substitu
On the other hand, llama does not benefit from the tion approach. First of all, some words have more than injection of knowledge about pejorative words, with a one neutral anchor words. This is the case of balena, drop of 0.09 points. This could be due to the noisy re- which has two neutral anchors: balenare (to flash) and sponse from Task A, including the refusal to answer, and cetaceo (sea mammal). In neutral examples, we substitute possible wrong explanations of why the target word is balena with both anchors. This process may alter the used pejoratively or not. semantic meaning of the tweet since only one anchor is suitable for substitution. Moreover, in some cases, we Setting Model Macro F1 replace a lexicon word with anchors that do not have A AlBERTo 0.68 the same meaning. For instance, the neutral anchor of BB__scuonbsctat AAllBBEERRTToo 00..7775 acida is aspra (sour ). However, expressions like sour beer A llama-3-8b-Instruct 0.69 or sour cream do not have a valid anchor replacement. B llama-3-8b-Instruct 0.60 Therefore, replacing aspra with acida is not an appropriate substitution.
and privacy policies. As this research involves the analy5. Conclusion sis of publicly available tweets, we do not seek explicit consent from individual users. Nevertheless, we make We have presented a new challenge for CALAMITA: pe- every efort to protect the anonymity of all individuals jorative word disambiguation as a preliminary step for mentioned. However, the exposure to misogynistic content still poses a mental health risk for researchers and annotators.