Assessing Public Energy Discussions Through Label-Free Aspect and Sentiment Discovery
This study addresses the challenge of converting large volumes of unstructured online energy discussions into interpretable aspect-level knowledge without manual annotation.
Indonesian public energy discussions provide useful information about issue salience and evaluative judgments. However, they are often informal, noisy, code-mixed, and difficult to analyze manually. Existing label-free methods typically operate at the topic level and often lack aspect-level outputs that are sufficiently grounded in local evidence and faithful to the text.
This study proposes a two-stage label-free knowledge discovery pipeline. The first stage extracts aspect terms by combining LDA-based topic representation, multilingual sentence embeddings, autoencoder compression, clustering, evidence-based candidate selection, and contextual part-of-speech filtering. The second stage assigns aspect-level sentiment using opinion lexicons, dependency-distance scoring, negation and intensity handling, pragmatic adjustment, and segment–aspect term pairs. The evaluation includes ablation analysis, baseline comparison, human validation of aspect-term outputs, and evidence-based sentiment auditing.
This study presents a label-free pipeline for extracting aspect terms and mapping aspect-level sentiment from unstructured public discussion. It distinguishes between mere mentions and those supported by explicit evaluative evidence, thereby capturing both issue salience and evaluative sentiment.
The findings show that effective aspect-term extraction requires a balance among cluster separation, representation quality, textual faithfulness, and linguistic validity. Human validation showed that most sampled aspect-term outputs were contextually valid. The pipeline produced 17,004 aspect-term occurrences, with non-evaluative aspect mentions forming the largest sentiment category at 35.40%. This result indicates that public attention to energy issues often appears without explicit evaluation.
Practitioners in energy communication, policy analysis, and public opinion monitoring can use this pipeline to identify which energy issues are merely mentioned and which are explicitly evaluated. The outputs can also help identify aspects that receive support or opposition and monitor changes in public attention and sentiment over time.
Researchers should develop label-free text analytics that preserve local evidence, support evidence-based labeling, and distinguish mentioned aspect terms from evaluated aspect terms. Future work should strengthen text-grounded aspect-term representations for knowledge discovery.
This study supports the assessment of public energy discussions by transforming unstructured online conversations into evidence-based knowledge. It also enables a more informed interpretation of digital public discourse in the energy domain.
Future research could improve this framework by incorporating adaptive domain filtering, enhancing the handling of sarcasm and implicit sentiment, conducting multilingual or cross-domain validation, and integrating human-in-the-loop evaluation.
Back