ChatGPT Personalization in Writing: Are Arabic and English Equal?

ChatGPT offers personalized texts in English and Arabic, promoting HCI and usability. It is ideal for language assistance with careful consideration of grammar and translation.

By Nur Masarwa * March 19, 2023 * 4 minutes read

HCI (Human Computer Interaction) and usability are major concepts linked to system development and the effectiveness of meeting users' needs (Issa & Isaias, 2022). Therefore, language knowledge may pose an obstacle to system acceptance or rejection (Kyriakoullis & Zaphiris, 2016). To overcome this obstacle, localization of the user interface (UI) and personalization are important to match the system's characteristics with the target country (Kyriakoullis & Zaphiris, 2016). This personalization exists also in digital media and technology like AI.

Artificial Intelligence (AI) has been a hot topic in education, with people enrolling in courses that use AI, as well as in different industries and computer science. AI is commonly referred to as the use of computers that simulate intelligent behavior with little human intervention (Hamet & Tremblay, 2017). AI is at the forefront of the fields of science and engineering that involve creating machines and algorithms with the ability to perform intelligent tasks that people can do, such as speech assistants like Siri and Alexa, or automated emails that approve online purchases (Hamet & Tremblay, 2017; Riedl, 2019). AI doesn't stop there. It can also adapt to behaviors based on the collected data and cookies, like what we see in commercial ads on platforms like YouTube or TikTok (Riedl, 2019).

ChatGPT

An example of AI that has attracted a lot of hype and attention is ChatGPT (Generative Pre-trained Transformer) (Kasneci et al., 2023). ChatGPT is the first large language model (LLM) that was released by OpenAI in 2018, as part of the GPT 3-5 series, and is recognized as a question-and-answer dialogue system (Kasneci et al., 2023). The AI system uses a vast amount of textual data and creates new content by analyzing patterns from the collected material. Therefore, the generated responses can be extremely convincing, making it challenging to differentiate them from human-written content (Borji, 2023). The system can produce sophisticated written works, such as essays and poems, as well as generating functional code and building websites and charts based on textual descriptions, often without the need for significant human input (Borji, 2023). According to Nakano et al. (2021), users must give commands like "Search," "Proofread," or "Quote," so that the model can use and utilize various web pages to create and craft an answer.

Despite the great power that ChatGPT holds, there are still some shortcomings in the generated output that we need to acknowledge to assess the outcome and performance. Although ChatGPT does an excellent job at answering any question, it lacks credibility. ChatGPT uses any public information from various reliable and unreliable sources. In addition, ChatGPT does not include any source or citation from where this information is coming from harming trustworthiness. However, based on my experience, users may ask ChatGPT to provide sources that may discuss the targeted subject.

The focus of this article is showcasing the performance of Chat GPT in relation to language and multilingualism, specifically Arabic and English. Based on the study of Bang et al. (2023), Chat GPT performs better in high-resource and medium-resource languages like English and lacks understanding in low-resource languages. Unlike English- a high-resource language- Arabic is considered a low-resource language with limitations in conversational systems and pre-trained models like Chat GPT (Naous et al., 2021).

Let's put Chat GPT to the test in Arabic and English.

Test #1: English and Arabic Dialects and Personalization

Arabic is a spoken language by over 330 million people from the Arabian/Persian Gulf in the east to the Atlantic Ocean in the west. Apart from Modern Standard Arabic (MSA), which is formal Arabic, there are spoken dialects like Levantine, Gulf, North African, and Egyptian Arabic (Diab & Habash, 2007). For this test, I used the spoken language of Arabic and English with abbreviations.

Palestinian Arabic

Australian Dialect

The Results:

Chat GPT was unable to generate Arabic dialects or any other dialect in English. It appears that Chat GPT uses modern and official English and Arabic by default, unless users ask to translate a specific sentence using a specific dialect. Using the official and modern language makes the generated texts formal and suitable for formal platforms, which may increase their trustworthiness. Additionally, the fact that the chat can transform specific sentences into a different language may affect positively the personalization. According to Nahai (2012) in the book "Humanizing the Web: Change and Social Innovation,", she discusses the importance of personalization in creating an engaging online experience. She believes that by customizing content and experiences to individual users based on their preferences and behaviors, websites can create a more meaningful and relevant experience. However, it may be a disadvantage for users who do not understand official Arabic or the differences in terminology between dialects.

Test #2: Linguistics and Translations

arabic-English Translation

Arabic and English are very different languages in various aspects.

• Writing System: Arabic is written from right to left whereas English is written from left to right.

• Phonology: Arabic has consonant letters that do not exist in English (Farghaly & Shaalan, 2009). Additionally, it has a system called "Tashkeel" that affects how the letters are pronounced (apart from vowels), which English does not have.

• In Arabic, morphology is a crucial component due to its highly structured and derivational nature. The shape of Arabic letters can vary based on where they appear in a word, and many letters are added to the word to indicate the gender and number. This contextual shaping is an important feature of Arabic calligraphy and typography (Farghaly & Shaalan, 2009).

ChatGPT was asked to translate different sentences from English and write a short text to test its ability to understand the linguistic and grammatical differences.

Translation

Arabic-English Translation

Arabic-English Translation

The Results:

Chat GPT did a great job translating a full paragraph from both English and Arabic. The system took into consideration number and gender agreement and made the appropriate additions to the words to match the pronoun. However, the system struggled in translating short sentences from English to Arabic. It was not able to detect the gender correctly to make the necessary changes to the words. On the other hand, the translation to English was accurate, but the system had to include parentheses to indicate if the word is for a female or male.

Short Story

Arabic

English

The Results:

When it comes to short story writing, the system did well with grammar, linguistics, and number/gender agreement even though there were slight mistakes with number agreement. In addition, the system lacked some of the "Tashkeel" system. Without this system, the pronunciation and the tense of the verbs may be affected. If the user has no prior knowledge in Arabic, I believe they will read it without those vowels, which may change the context and tenses. When it comes to English, there were no mistakes.

In short,

with the lack of grammar and linguistics in Arabic, Chat GPT has shown efficiency in both Arabic and English languages placing personalization and HCI at the top priority of the company. Chat GPT can be used in different languages, but users need to consider rechecking the language itself since it also may lack reasoning.

• References

  Attwood, A. I. (2020). Changing social learning theory through reliance on the Internet of Things and artificial intelligence. Journal of Social Change, 12, 103–111. https://doi.org/10.5590/JOSC.2020.12.1.08

  Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., ... & Fung, P. (2023). A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023.

  Borji, A. (2023). A categorical archive of chatgpt failures. arXiv preprint arXiv:2302.03494.

  Diab, M., & Habash, N. (2007, April). Arabic dialect processing tutorial. In Proceedings of the human language technology conference of the NAACL, companion volume: tutorial abstracts (pp. 5-6).

  Farghaly, A., & Shaalan, K. (2009). Arabic natural language processing: Challenges and solutions. ACM Transactions on Asian Language Information Processing (TALIP), 8(4), 1-22.

  Hamet, P., & Tremblay, J. (2017). Artificial intelligence in medicine. Metabolism, 69, S36-S40.

  Kasneci, E., Seßler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., ... & Kasneci, G. (2023). ChatGPT for good? On opportunities and challenges of large language models for education.

  Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., ... & Schulman, J. (2021). Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332.

  Naous, T., Antoun, W., Mahmoud, R. A., & Hajj, H. (2021). Empathetic BERT2BERT conversational model: learning Arabic language generation with little data. arXiv preprint arXiv:2103.04353.

  Riedl, M. O. (2019). Human‐centered artificial intelligence and machine learning. Human Behavior and Emerging Technologies, 1(1), 33-36

  Wenger, E. (2009). A social theory of learning. In Contemporary theories of learning (pp. 217-240). Routledge