Overview
The sentiment data originates in nine Arabic-language news websites from Israel:
1. alittihad44.com
2. arab48.com
3. ashams.com
4. kul-alarab.com
5. bokra.net
6. mawteni48.com
7. nasapp.net
8. panet.com
9. sonara.net
I chose the first four sites based on my past research in the Arab community in Israel. I used Claude AI and ChatGPT to suggest similarly prominent sites to them in order to identify the latter five sites.
Limitations
There are two main limitations to this data. (1) This selection is intended to represent a common set of Arabic-language media in Israel that might discuss the war. So, local or non-news sites were not included. It is also distinct from Hebrew media, though Arab individuals might read Hebrew language Israeli media, or Arab media from other countries or English media. So, it is not representative of public opinion in Arab community of Israel generally or even of the readers of these sites but rather only of how the target terms are discussed in these outlets. How the terms are discussed in these outlets is of interest because of media can influence public opinion and also be influenced by it and these outlets are more likely to have a relationship with the Arab community of Israel than others. Yet, the relationship is not a simple one. (2) Moreover, each source (and these source collectively) may have particular views of people, countries, and events that sets a baseline valence for a term. So, the absolute level of valence is less informative than the changes over time because the latter are likely in response to current events.
Methods
A Python script written with the assistance of ChatGPT and Claude AI has scraped the sites daily since April 25, 2026, beginning with each homepage and following any links to other pages within the same site (domain) and successive links on those pages. If the script estimates a page to be older than the preceding day, it ignores it. Data older than April 25, 2026, was gathered several days before that date by storing all pages found within each site regardless of estimated date. While the data is sufficient to calculate valence going back to 2006, they are necessarily a smaller sample of the data actually created on those dates than those scraped in real time because some old pages gradually go offline or become more difficult to reach from the current homepage. This means older dates must be understood as having greater statistical error.
I chose the target words (those which appear in the viewer) by beginning with places and leaders relevant to the main actors in the war: the US, Iran, and Israel, and gathering further suggested terms from Claude AI. For each day, for each target word, all terms appearing within five words from each instance of the target word are identified. A positive pointwise mutual information (PPMI) score for identified co-occurring word is calculated which measures how much more often the co-occurring word appears with the target word compared to chance. The Arabic Emoticon Lexicon (Salameh, Mohammed, and Kiritchenko (2015)), which measures the co-occurrance of many Arabic terms in Twitter data with positive and negative emoticons, is used to identify whether a given co-occuring word is positive or negative. The target word’s score is then its PPMI with positive words minus its PPMI with negative words.
Lexicon Source
Salameh, Mohammed, Saif M. Mohammed, and Svetlana Kiritchenko. 2015, June. “Sentiment After Translation: A Case-Study on Arabic Social Media Posts.” In Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL-2015). Denver, Colorado.
Salameh, Mohammed, Saif M. Mohammed, and Svetlana Kiritchenko. 2016, January. “How Translation Alters Sentiment.” Journal of Artificial Intelligence Research 55: 95-130.
Salameh, Mohammed, Saif M. Mohammed, and Svetlana Kiritchenko. 2016, May. “Sentiment Lexicons for Arabic Social Media.” In Proceedings of the 10th edition of the Language Resources and Evaluation Conference. Portorož (Slovenia).