TL;DR
This paper investigates the evaluative meaning of German personal name compounds (PNCs) like "Willkommens-Merkel" (Welcome-Merkel), showing they carry distinct positive or negative connotations compared to the full name. Using both traditional valence norms and modern pre-trained language models (PLMs), the authors demonstrate that PNCs are indeed evaluative, with factors like domain and political affiliation influencing their meaning.
What Problem Does This Paper Address?
Personal name compounds (PNCs) are a common linguistic phenomenon in German, appearing frequently in news and social media. These compounds, formed by a modifier and a personal name (e.g., "Willkommens-Merkel" for Angela Merkel), are hypothesized to carry an evaluative function, meaning they can evoke a more positive or negative perception of the person than their full name. However, this phenomenon has been largely under-investigated from both theoretical and computational perspectives. Understanding this evaluative function is crucial for tasks like sentiment analysis and emotion detection in discourse. This paper aims to fill this gap by computationally modeling and analyzing the evaluative meaning of PNCs.
Methods
The authors employ two primary computational approaches to assess the evaluative nature of PNCs:
- Valence Norms: They utilize automatically generated valence norms (Köper & Schulte im Walde, 2016) that assign a score from 0 (low valence) to 10 (high valence) to words. The valence of a PNC or a full name is calculated as the normalized mean valence of its context words.
- Pre-trained Language Models (PLMs): They leverage several PLMs (XLM-ROBERTa, BERT-based models) fine-tuned for sentiment analysis. Sentiment predictions (negative, neutral, positive) are mapped onto a valence scale to derive a weighted valence score for PNCs and their corresponding full names.
Additionally, the study uses regression analysis to explore factors influencing PNC evaluation, including personal information (age, gender, nationality), domain-specific information (domain, political party), and extra-linguistic factors (modifier valence, event frames).
Datasets
- Personal Name Compounds (PNCs): The study starts with 770 eventive PNCs provided by Belosevic (2022), further filtered to 321 PNCs with at least one context and 217 with at least five contexts for modeling.
- Context Corpora:
- Twitter: Tweets containing PNCs or full names were collected using the Twitter Academic API. This resulted in 9,145 tweets for PNCs and 24,688 for full names.
- Deutscher Wortschatz (DW): Sentences from the Leipzig Corpora Collection were used, yielding 170 sentences for PNCs and 233,477 for full names.
- Human Evaluation Data: A subset of 30 PNCs was annotated by five human annotators for sentiment (positive, negative, neutral) to compare with computational model predictions.
Key Insights and Ideas
- PNCs are inherently evaluative: The core finding is that PNCs consistently carry a distinct evaluative meaning compared to their full names, often leaning towards a more negative evaluation.
- Domain matters significantly: PNCs from the "politics" domain tend to be more negatively evaluated, while those from "sports" and "show business" are generally more positive.
- Modifier valence is a strong predictor: The valence of the modifier word in a PNC plays a crucial role in the overall evaluation of the compound, sometimes even more so than the name itself.
- PLMs offer a powerful, albeit sometimes more negative, perspective: PLMs, especially multilingual ones, can predict PNC valence with high accuracy, often showing a stronger negative bias than valence norm-based approaches.
- Personal and political factors influence meaning: A person's age, political party affiliation, and even birthplace can significantly influence how a PNC referring to them is evaluated.
Key Results
- Valence Norms: 56% of PNCs were found to be more negatively evaluative than their corresponding full names. A moderately positive correlation (p = 0.43) was observed between name and PNC valence.
- PLMs: Up to 93.55% of PNCs were predicted to be more negatively evaluative than their full names by PLMs. XLM-based models showed a slightly more positive interpretation of PNCs compared to German BERT-based models.
- Regression Analysis: The best model explaining the variance in PNC evaluation included all variables except name valence, achieving an adjusted R² of 0.96. Key predictors included PNC valence, modifier valence, age, political party, and birthplace.
- Human Evaluation: While generally agreeing with the negative evaluative trend, human annotators found the task difficult, suggesting nuances not fully captured by computational models. They also noted domain-specific differences, with athletes being evaluated more positively than politicians.
Limitations and Future Work
- Language Specificity: The study focuses solely on German PNCs. Adapting the approach to other languages might require generating language-specific valence norms or using multilingual PLMs with potential adapter-based fine-tuning for under-represented languages.
- Corpus Comparison: The comparison between Twitter and news corpora is limited by the number of full-name contexts available from Twitter. Future work could explore how social media platforms foster or change PNC usage compared to professional news outlets.
- PLM Bias: PLMs, while powerful, can exhibit biases (e.g., a tendency towards negative sentiment). Further research could focus on mitigating these biases or developing more nuanced sentiment analysis methods.
- Discourse Context: While context is used, the study acknowledges that deeper investigation into predicting sentiment towards specific targets or incorporating more sophisticated discourse analysis might be beneficial.
- Human Annotation Difficulty: The self-reported difficulty by annotators suggests that PNC evaluation is complex and may require more refined annotation guidelines or expert judgment.
Summary updated