class: center, middle, inverse, title-slide .title[ # Frame? Kein Thema! ] .subtitle[ ## Die Validität der induktiven Frame-Erkennung in Texten mit mehreren Themen ] .author[ ### Chung-hong Chan (GESIS), Rainer Freundenthaler (MZES) ] .date[ ### 2022-10-05 ] --- # Acknowledgement The authors would like to thank the Open Science Office, University of Mannheim for the financial support of this project. Know more: https://www.uni-mannheim.de/open-science/open-science-office/ --- # So, are these frames? .pull-left[  .extrasmall[Walter & Ophir (2019)] ] .pull-right[ <img src="img/lda.png" height = 170> .extrasmall[Ylä-Anttila, Eranti, & Kukkonen. (2022)] <img src="img/lda_ai.png" height = 170> .extrasmall[Zeng, Chan & Schäfer (2019)] ] --- .extralarge["Some studies employ the concept only in a metaphoric sense, whereas others reduce frames to story topics, attributes, or issue position"] Carragee & Roefs (2004) --- **Entmanian Frame:** "Select some aspects of a perceived reality and make them more salient in a communicating text, in such a way as to promote a particular problem definition, causal interpretation, moral evaluation, and/or treatment recommendation for the item described." **Chan's restatement:** Frame is the result of an act of **selecting** certain aspects of a perceived reality by a communicator, whose **intention** is to promote a particular problem definition, causal interpretation, moral evalution, and/or treatment recommendation. --- # What actually is frame detection? .extralarge[ * Evaluate which aspects of a perceived reality are selected? * Guess the communicator's intention? ] --- class: hide-footer  --- class: hide-footer  --- class: hide-footer  --- class: hide-footer # Generic frames .pull-left[ "**transcend** thematic limitations and can be identified in relation to different topics, some even over time and in different cultural contexts" (de Vreese, 2005)  ] .pull-right[  .extrasmall[Semetko & Valkenburg (2000)] ] --- # "Platinum" Standard: Communicative intention is known .pull-left[  ] .pull-right[  ] --- # "Gold standard" .pull-left[ <img src="img/generic.png" height = 420> ] .pull-right[ Two trained MA students coded according to the standard coding scheme .extrasmall[Semetko & Valkenburg (2000)] ] --- .pull-left[ # Inductive methods * K-means with TFIDF (Burscher et al., 2016) * PCA with TF-IDF (Greussing & Boomgaarden, 2017) * LDA (DiMaggio et al., 2013) * STM (Nicholls & Culpepper, 2021) * ANTMN (Walter & Ophir, 2019) ] .pull-right[ # Semi-supervised methods * Seeded-LDA (Watanabe & Zhou, 2020) * keyATM (Eshima et al., 2020) Keywords from two journalism researchers k = 5 ] --- # So many methodological decisions * stemming, lemmatizing, trimming of sparse / frequent words, stopwords, `\(\alpha\)`, ... * Multiverse Analysis (Pipal, Song, Boomgaarden, 2022) .pull-left[ <img src="img/prereg.png" height = 100> ] --- # Not better than random guess <img src="img/km_gg.png" height = 420> --- # Slightly better than random guess <img src="img/pca_gg.png" height = 420> --- # Sometimes better than random guess .pull-left[ <img src="img/stm_gg.png" height = 420> ] .pull-right[ <img src="img/lda_gg.png" height = 420> ] --- # Sometimes better than random guess, but with a large multiverse .pull-left[ <img src="img/antmn_gg.png" height = 420> ] .pull-left[ <img src="img/seeded_gg.png" height = 420> ] --- # Often better than random guess, even with a large multiverse <img src="img/keyatm_gg.png" height = 420> --- # The golden question: Does gold standard work? -- .pull-left[ <img src="img/human_gg.png" height = 420> ] .pull-right[ Always better than random guess, but not superb ] --- # Zum Mitnehmen .pull-left[ * Frame detection is not the same as finding topics (Carragee & Roefs, 2004) * Detection of communicative intention from text is crazily hard, even human can't do it well. But.. * OSF: https://osf.io/gkft5/?view_only=8bf2a34000a64c5fa5d645c3f37f681a (Or contact me) ] .pull-right[ <img src="img/turtle.jpg" height = 420> ]