Cross-lingual Opinions and Emotions Mining in Comparable Documents
Motaz Saad, David Langlois, Kamel Smaïli · 2025
Read Cross-lingual Opinions and Emotions Mining in Comparable Documents by Motaz Saad, David Langlois, Kamel Smaïli (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-03112.txt.
Introduction
Motaz Saad, David Langlois, and Kamel Smaili study how news in two tongues can share a topic yet differ in tone, in work from twenty twenty-five. They work with paired English and Arabic reports that cover the same events but are not direct translations. The goal is to mine views and feelings across the pair.
The team marks each text for view type and for six basic feelings such as joy, fear, and anger. They built an Arabic word list for feelings by hand from an English source. Pairs from the same news source matched in tone far more than pairs from mixed sources. The method needs no machine translation and can extend to other tongue pairs.
Catalog record
| CREATOR | Motaz Saad, David Langlois, Kamel Smaïli |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-03112
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Cross-lingual Opinions and Emotions Mining in Comparable Documents Motaz Saad, David Langlois, Kamel Smaïli Abstract Comparable texts are a set of topic aligned documents in multiple languages, which are not necessarily translations of each other. These documents are informative because they can tell what is being said about a topic in different languages. The aim of this research is to study differences in cross-lingual comparable documents in terms of sentiments and emotions. In this work, we focus on English-Arabic language pair. Before comparing sentiments and emotions in comparable documents, texts must be annotated first with sentiment and emotions labels. For that, we use a cross-lingual method to annotate source and target texts with opinion labels (subjective and objective). The advantage of this method is that it produces resources in multiple languages without the need of machine translation system. To be able to annotate the English-Arabic document with emotion labels (anger, disgust, fear, joy, sadness, and surprise), we manually translate the WordNet-Affect (WNA) emotion lexicon from English into Arabic. Then we use these English-Arabic emotion lexicons to annotate English-Arabic comparable documents. Finally, we use a statistical measure to compare the agreement of sentiments and emotions in the source and the target pair of the comparable documents. This comparison is interesting, especially when the source and the target documents come from different sources. To our knowledge, this has not addressed in the literature. Our method inspects the agreement for each pair (source and target) of documents. In our experiment, we study the agreement between English and Arabic document pairs collected from Euronews, British Broadcast Corporation (BBC), and Al-Jazeera (JSC) news websites. The methods presented in this research are language independent and they can be applied to any language pair. Result shows that sentiments and emotions in comparable converge (agree) when the inspected news articles are from the same news agencies, and diverge (disagree) when they are from different news agencies. \setcode utf8 Keywords : text mining; natural language processing; comparable corpus; cross-lingual projection; sentiment analysis 1 Introduction Retrieving opinions and emotions from comparable multilingual documents can be very useful. For example, a journalist wants to know what is being said around the world about war in Syria. For that, he needs to access and analyze documents about this topic in several languages. This is necessary because the opinions and sentiments expressed in documents depend strongly on the nationality of the source. This is due to cultural differences and geopolitical aspects. Indeed, during the Egyptian revolution in 2011, many non-Arabic speaking journalists were interested in what was being said in Arabic posts in the social media. During that time, automatic translation services were provided by Twitter and other organizations to understand these posts [ Arthur, 2011 , Diehn, 2013 ] . Two documents are called comparable because they contain approximately the same information, i.e., they are related to the same topic. This is the case, for example, for Wikipedia pages about the same entry. Figure 1 shows an example of English 1 1 1 https://en.wikipedia.org/wiki/T._E._Lawrence and French 2 2 2 https://fr.wikipedia.org/wiki/Thomas_Edward_Lawrence Wikipedia comparable documents related to a biography of a person. It can be noted from the figure that the first paragraph in the English document is longer than the French one, it also provides more information about the person. Even if the documents differ in terms of facts, they remain both objective. (a) The English document (b) The French document Figure 1: English-French comparable documents from Wikipedia Another example of comparable documents is the English 3 3 3 http://www.euronews.com/2013/12/03/transparency-international-warns-23rds-of-states-corrupt/ and the French 4 4 4 http://fr.euronews.com/2013/12/03/corruption-l-espagne-recule-dans-l-index-de-transparency-international/ news articles from Euronews shown in Figure 2 . The news articles are related to a report about transparency and corruption in the world. Despite the documents are related to the same news story, the translations of their titles are different. Furthermore, the first paragraph in the English and in the French documents are different. The French document also has an additional paragraph that describes the position of France in this report. This paragraph does not exist in the English document. Nevertheless the both document remain objective and relate facts. Transparency International warns 2/3rds of states “corrupt” 03/12/13 16:59 CET The fairtrade watchdog Transparency International reports in its latest global roundup that two-thirds of the 177 countries it surveys are below average in the corruption stakes. The report highlights continuing political pressure resulting in market distortions caused by bribery, cronyism, a lack of accountability and inadequate legal systems. “We have two-thirds of all countries, 177 countries in total, where we can see that they score under 50 points, which means they are below the average, and corruption in those countries, two thirds of the 177, are still to be seen in a very critical situation,” says Transparency’s German head Edda Muller. Top of the class is Denmark, followed by New Zealand and Finland. The USA was only 19th, while among the major European economies France, (22), again lagged behind Germany, (12), and the UK, (14). Greece’s position improved, but it is still a lowly 80th, a potential turn-off for investors and the worst in Europe. Somalia, Afghanistan and North Korea tied for last place, unchanged from last year. (a) The English news document Corruption : l’Espagne recule dans l’index de Transparency International 03/12/13 16:59 CET Selon l’indice de la corruption dans le monde, calculé par l’ONG Transparency International et publié mardi, le Danemark est le pays le plus honnête, à égalité avec la Nouvelle Zélande. Ces deux pays jugé les moins corrompus en 2013, étaient déjà les deux meilleurs de la liste de 177 pays établie par l’ONG basée en Allemagne. “Nous avons deux tiers des 177 pays qui sont sous les 50 points, explique Edda Muller, la présidente de Transparency International, ce qui veut dire qu’ils sont sous la note moyenne. La corruption dans ces pays : les deux tiers des 177, est dans une situation très critique”. Les trois premiers du classement sont le Danemark, la Nouvelle Zélande et la Finlande. La Grèce gagne quelques places mais reste 80ème et dernier pays européen. La surprise c’est l’Espagne en proie aux affaires et qui passe de la trentième à la quarantième place. En règle générale les pays du Nord de l’Europe font figure d‘élèves modèles. L’enquête de Transparency International n’est pas un classement selon le niveau de corruption mais elle mesure la perception de la corruption en raison du secret qui entoure les pratiques les moins avouables. La France, 22ème, n’occupe que le 10ème rang en Europe. L’ONG considère que le bilan des lois votées en France en 2013 en matière de transparence et de lutte contre la corruption est “globalement positif” mais s’interroge sur leur mise en oeuvre effective. (b) The French news document Figure 2: English-French comparable news article from EURONEWS But a document can contain subjective paragraphs or sentences. Comparable documents about a political or polemic topic (war, societal topic) may contain very contrasting opinions and sentiments. This can be the case in social medias or blogs. Another example is a review about a product. To illustrate this fact, let’s use comments about a holiday home in Avignon (France) posted on a famous commercial web service. For the same house, one have