Large Language Models Analyze Complex Political Opinions: An Automated Public Opinion Analysis Framework Based on Target-Stance Extraction

This study focuses on using target-stance extraction (TSE) technology of large language models (LLMs) to address the problem of difficulty in capturing fine-grained stance structures in traditional political text analysis. By structuring complex political opinions, it provides new methods for computational social science research, public opinion monitoring, and other scenarios, and discusses technical advantages, experimental design, application value, challenges, and future directions.

Political text analysis is an important direction in computational social science, but traditional sentiment analysis or topic modeling has shortcomings: it is difficult to capture fine-grained opinions with multiple stances, implicit attitudes, and composite targets (e.g., a tweet that supports immigration policy but criticizes its implementation method), which easily leads to public opinion misunderstandings. The target-stance extraction (TSE) task emerged to identify target entities/issues in text and their corresponding stance tendencies.

LLMs (such as GPT series, Llama, etc.) have three major advantages over traditional methods:

1. Context Understanding: Handle complex language phenomena like irony and metaphor through self-attention mechanisms to accurately identify real stances;
2. Few-shot Learning: Quickly adapt to the TSE task via prompt engineering, solving the problem of limited labeled data in the political domain;
3. Structured Output: Directly output structured results like JSON, simplifying system architecture.

This study constructs a TSE dataset covering multiple domains such as immigration and economy, with annotations including target entities, stance polarity (support/oppose/neutral/no stance), and confidence. Evaluation uses precision, recall, F1-score, stance consistency, and adversarial testing. Experiments compare BERT baseline, GPT-3.5 prompt learning, and domain-fine-tuned Llama models. Results show that the optimized LLM method improves F1-score by 15-20 percentage points, especially performing better in handling implicit stances and composite targets.

Application scenarios of TSE technology include:

1. Election Public Opinion Monitoring: Real-time analysis of opinion distribution in social media and news to reveal differentiated stances of different groups on issues;
2. Policy Effect Evaluation: Track changes in public opinion before and after policy introduction, and quantify the spatiotemporal distribution of support;
3. Disinformation Detection: Identify false opinions through stance inconsistency;
4. Cross-Language Analysis: Use the transfer capability of multilingual LLMs to support non-English political text analysis.

Technical Challenges: Ambiguous target boundaries (nesting, coreference resolution, issue drift), insufficient stance intensity modeling, temporal dynamics (delayed model knowledge update). Ethical Considerations: Need to prevent the amplification of biases in training data and establish a fairness evaluation framework; also pay attention to privacy protection and data usage compliance.

Future development directions include:

1. Multimodal Fusion: Integrate text, image, and video information to build a cross-modal stance analysis framework;
2. Causal Inference: Shift from correlation to identifying key factors in stance formation;
3. Interactive Systems: Develop human-machine collaboration tools to support expert calibration of models;
4. Real-Time Stream Processing: Build large-scale real-time data stream processing systems to support instant public opinion analysis.

Large language models open up new possibilities for fine-grained analysis of political texts. Through the TSE framework, structured opinion data can be extracted from massive unstructured discourse to support social science research and public decision-making. With the improvement of model capabilities and methodology, this technology is expected to play a greater role in understanding complex social phenomena, promoting democratic participation, and addressing information manipulation.