TL;DR
This paper introduces a novel "Deep Modular RNN" (DMRNN) approach for automatically identifying and classifying "ethotic statements" (statements about a speaker's character) in political discourse. By leveraging a rich dataset of UK parliamentary debates, their method achieves state-of-the-art results in discerning positive and negative sentiment towards politicians' ethos, offering a powerful new tool for understanding political dynamics.
What Problem Does This Paper Address?
Understanding how political discourse influences public opinion is crucial, yet complex. While "argument mining" has focused on the logos (the content of arguments), this paper highlights the significant, often overlooked, role of ethos – the perceived character and credibility of the speaker. Politicians frequently leverage ethos to sway audiences, making its automatic extraction a valuable but challenging task. Existing methods for analyzing political text often focus on content or sentiment but don't specifically target the nuanced ways speakers build or attack each other's character. This paper aims to bridge this gap by developing a system that can automatically identify and classify these ethotic statements, providing insights into political alliances and rivalries.
Methods
The core of this paper is the development of a Deep Modular RNN (DMRNN) pipeline for ethos mining. This pipeline integrates several components:
- Raw Text Processing: The system takes raw parliamentary debate transcripts as input.
- Linguistic Feature Extraction: It utilizes Part-of-Speech (POS) tagging and Universal Dependencies (UD) tagging to capture grammatical structure.
- Entity Extraction (EXT): A rule-based system identifies relevant entities (politicians, parties) within sentences, filtering out irrelevant mentions.
- Sentiment Analysis: Standard classifiers (SVM, Naïve Bayes, Logistic Regression) are employed to determine sentence polarity, enhanced by domain-specific lexicons (ethotic and negative words from Hansard).
- Sentiment Presence: A module determines if a sentence actually contains sentiment.
- Polarity (POL): This novel module combines sentiment presence with the sentiment classification to tag sentences as positive, negative, or non-sentiment bearing.
- Deep Modular RNN (DMRNN): This is the central novel component. It's an ensemble of models, each processing different inputs (raw text, POS, UD, EXT, POL) through separate modules (including dense layers and LSTM layers). These modules are then concatenated and fed into a final dense layer for binary classification of ethotic statements (ESE) versus non-ethotic statements (n-ESE). The DMRNN is designed to leverage more data and combine specialized models without overfitting.
- Ethos Analytics: The output of the DMRNN is then fed into a graph-based tool that visualizes relationships between politicians and parties based on their ethotic statements.
Datasets
The primary dataset used is Ethos_Hansard1, a newly constructed corpus of 90,991 tokens from UK parliamentary debates during Margaret Thatcher's premiership (1979-1990). This corpus is publicly available and annotated for ethotic statements, including speaker, target, ethotic support (+ESE), and ethotic attack (-ESE). The dataset was split into 60 training transcripts (with 10% validation) and 30 test transcripts.
Key Insights and Ideas
- Ethos as a Powerful Influence: The paper reinforces the long-held rhetorical principle that a speaker's character (ethos) can be a more potent influence than the content of their arguments (logos).
- DMRNN for Text Classification: The application of a deep modular RNN architecture to text classification, specifically for ethos mining, is a novel contribution. The modularity allows for specialized processing of different linguistic features, leading to improved performance.
- Domain-Specific Lexicons Matter: Incorporating domain-specific lexicons (like the "ethotic word lexicon" and the "HAN (Neg)" lexicon) significantly improves sentiment classification accuracy by capturing nuances of word meaning in political contexts (e.g., "rich" can imply hypocrisy, not just wealth).
- Ethos Analytics for Political Insight: The ability to automatically extract and visualize ethotic relationships provides a powerful lens for analyzing political dynamics, correlating sentiment shifts with real-world events like elections and policy changes.
Key Results
- Ethotic Statement Classification (+/- ESE): The best performing classifier achieved a macro-averaged F1-score of 0.84, outperforming previous work and baselines.
- Ethotic Statement vs. Non-Ethotic Statement Classification (ESE/n-ESE): The DMRNN model achieved the highest F1-score of 0.74 and a macro-averaged F1-score of 0.83, significantly outperforming baselines.
- Combined Classification: The combined approach for classifying both the presence of an ethotic statement and its polarity achieved an F1-score of 0.65, a substantial improvement over previous work.
- Anaphora Resolution: The system achieved an F1-score of 0.76 for correctly tagging target entities in ethotic statements.
- Qualitative Ethos Analytics: Analysis of Margaret Thatcher's career revealed correlations between political events (e.g., appointments, elections) and shifts in ethotic attacks and supports, demonstrating the practical utility of the ethos mining pipeline.
Limitations and Future Work
- Dataset Size: While Ethos_Hansard1 is the largest publicly available corpus of its kind, the authors acknowledge that it is still relatively small, which can make complex models like CNNs prone to overfitting.
- Focus on Logos: The current pipeline primarily focuses on ethos. Future work aims to incorporate logos (argument content) and analyze their interactions.
- Sarcasm and Irony: The pipeline does not currently handle sarcasm or irony, which can significantly alter the perceived ethos. Developing modules for these is a future direction.
- Argument Metrics: The authors plan to explore ethotic metrics similar to argument metrics to identify popular or divisive speakers.
- Module Aggregation: Further research will investigate optimizing the aggregation of the pipeline's individual modules.
Summary updated