TL;DR
This paper introduces MIN_PT, a novel European Portuguese lexicon specifically designed to identify terms related to minorities, including both offensive and neutral language. This resource addresses the limitations of existing lexicons by focusing on European Portuguese and incorporating neutral terms, making it more comprehensive for hate speech detection and analysis.
What Problem Does This Paper Address?
The automatic detection of hate speech is a critical area in natural language processing, often relying on lexicons – lists of words and phrases associated with specific topics. However, existing Portuguese lexicons for hate speech detection suffer from two major drawbacks:
- Lack of European Portuguese specificity: Many resources conflate European and Brazilian Portuguese, despite significant ethnographic and linguistic differences. This can lead to misinterpretations, as certain terms (e.g., "bicha" for homosexual) have distinct meanings in each variant.
- Absence of neutral terms: Most lexicons focus solely on offensive language. This overlooks neutral terms that refer to minority groups, which are crucial for a broader understanding of online discourse, including data collection for a wider range of content and bias detection.
This paper aims to fill this gap by creating a dedicated European Portuguese lexicon that includes both offensive and neutral terms related to minorities.
Methods
The creation of the MIN_PT lexicon followed a qualitative, iterative approach involving expert annotators. The process involved:
- Initial Data Source: The process began with a subset of the Hatebase lexicon for Portuguese terms, which provided an initial list of 319 terms but lacked European Portuguese specificity and neutral terms.
- Data Curation and Enrichment: Two annotators, native European Portuguese speakers with experience in social context and annotation, worked individually and collectively. They:
- Enumerated terms: Manually expanded the list by adding synonyms, ensuring variations in singular/plural and masculine/feminine forms, and including all known terms for minority groups.
- Removed ambiguous terms: Filtered out words with dual meanings where the most common usage did not relate to minorities (e.g., "preto" as a color vs. an insult).
- Removed non-European Portuguese terms: Excluded words specific to Brazilian Portuguese.
- Data Annotation: All curated terms were classified into specific minority groups (Roma, LGBT, migrants, women, people based on religion, people based on ethnicity, and refugees). Each term was also labeled as either an insult (1) or not an insult (0). Crucially, terms that could be used both as insults and neutrally were classified as non-insults.
Datasets
- Hatebase (Portuguese subset): Used as the initial seed for the lexicon.
- MIN_PT Lexicon: A newly introduced benchmark lexicon containing 155 carefully curated terms for European Portuguese minority-related terms, classified by minority group and insult status.
- Portuguese 2021 Presidential Elections Tweets: A dataset of 35,101 tweets from September 2nd, 2020, to November 22nd, 2020, used for a case study application of the MIN_PT lexicon.
Key Insights and Ideas
- The importance of linguistic and ethnographic specificity: The paper strongly emphasizes that for nuanced social phenomena like hate speech, distinguishing between language variants (European vs. Brazilian Portuguese) is not just a technical detail but a necessity for accurate analysis.
- Neutral terms are vital for comprehensive analysis: Including neutral terms significantly broadens the scope of analysis, enabling the identification of a wider range of minority-related content and facilitating bias detection.
- Human annotation is indispensable: Despite the potential for automatic methods, the paper highlights the critical role of human annotators in ensuring the quality, accuracy, and contextual relevance of lexicon entries.
- Lexicons are context-dependent: The authors acknowledge that even their carefully curated lexicon has limitations, as terms can be neutral or offensive depending on the specific context and author.
Key Results
- The MIN_PT lexicon contains 155 carefully curated terms related to 7 minority groups.
- Of these terms, only 23% overlap with the Hurtlex lexicon, indicating the novelty and specificity of MIN_PT.
- A case study analyzing tweets from the Portuguese 2021 presidential elections demonstrated the lexicon's utility. It revealed that while candidates did not explicitly use the offensive terms from the lexicon, these terms were present in the audience's replies, highlighting a disconnect between political discourse and public sentiment.
- The analysis showed that different candidates focused on distinct minority topics, with some candidates addressing a broader range of groups than others.
Limitations and Future Work
- Static nature of lexicons: Lexicons are static and cannot fully capture the dynamic and context-dependent nature of language. Terms can shift in meaning or be used neutrally or offensively depending on the situation.
- Ambiguity removal: While efforts were made to remove ambiguous terms, future users are cautioned to validate terms against their specific datasets and contexts.
- Limited annotator pool: The lexicon was created by two annotators. Future work could benefit from a larger, more diverse group of annotators to increase vocabulary richness.
- Keyword matching limitations: The authors caution against using the lexicon for direct hate speech classification solely through keyword matching, as context is paramount.
Future work could involve expanding the lexicon with more annotators and exploring its application in other datasets and contexts to further refine hate speech detection and minority discourse analysis.
Summary updated