Blame Attribution and Moral Evaluation Signals in Thai Political and Entertainment News: A Rule Based Content Analysis of the ThaiSum Dataset

This project compares linguistic signals of blame attribution and moral evaluation in Thai political and entertainment news from the ThaiSum dataset. The source files were audited before sampling to prevent duplicate records from being counted. The training split contained 358,868 unique news URLs, while all 11,000 validation records and all 11,000 test records duplicated entries in the training split. The analysis therefore used only the training split. A stratified sample was drawn from Thairath and Thai PBS, with 200 political and 200 entertainment articles from each source, producing 800 unique articles. The random seed was 20261031.Articles were classified with deterministic Thai phrase rules at the document level. Blame attribution was measured with strict, core, and broad rules. Moral evaluation was divided into negative and positive signals. Differences in proportions were assessed with stratified bootstrap confidence intervals, within-source permutation tests, and logistic regression adjusted for news source and article length.Under the core rule, blame signals appeared in 179 of 400 political articles, or 44.75 percent, and in 72 of 400 entertainment articles, or 18.00 percent. The difference was 26.75 percentage points. The direction remained unchanged under the strict and broad rules. After adjustment for source and article length, the odds of detecting a blame signal in entertainment news were about one quarter of those in political news, with an adjusted odds ratio of 0.256 and a 95 percent confidence interval from 0.183 to 0.358.Evaluative language of any kind appeared in 46.25 percent of entertainment articles and 26.50 percent of political articles. This difference was driven mainly by positive language, which appeared in 40.75 percent of entertainment articles and 20.00 percent of political articles. Negative signals appeared in 7.50 percent and 10.50 percent, respectively, and the difference was not statistically clear. The results therefore support the conclusion that political news in this sample contained more blame-related language, while entertainment news contained more evaluative language because it used more praise and encouragement, not because it used more condemnation. From a media-psychology perspective, examining these signals at the text level can help illuminate how news discourse may cue moral judgment and social perception; however, the study does not directly measure audience cognition, attitudes, or behavior.All findings are phrase-rule signals rather than judgments about authorial intent or results from human coding. Full news text was kept within the research environment and was not redistributed. The report presents only aggregate statistics and retains source URLs for traceability.