Dataset of implicit aspects for LADy but in Arabic

[hosseinfani]

[AliAlsalkhadi]

Found this 100k Arabic reviews dataset on Kaggle. The dataset combines reviews from hotels, books, movies, products and a few airlines. It has three classes (Mixed, Negative, and Positive).

@Mahmoud-s-programs check this out:

https://www.kaggle.com/datasets/abedkhooli/arabic-100k-reviews?resource=download

Currently working one finding common stop words that would help me refine and determine the aspect of the review.
Do we care about if the review is positive or negative? I don't think so but let me know if this could help us.

[AliAlsalkhadi]

@Mahmoud-s-programs I was thinking we use this dataset since it has different types of reviews instead of the one you found which is only for books and we then filter it using Python based on the subject and combine the refined dataset based on our needs.

[hosseinfani]

Hi @AliAlsalkhadi pls also check they have aspect term/category labels.
Also, I'm waiting for your reply email :)

[AliAlsalkhadi]

I have sent the reply email thanks for reminding me.
Unfortunately, they don't have aspect labels I will keep looking to see if they have something that best suits our needs but for now I will try to filter out the dataset I have. But I think our dataset will shrink by a huge margin how much data do we need for this I know the more the better but what is our baseline if we have one?

[hosseinfani]

@AliAlsalkhadi
there is no dataset with hidden aspects. Also, those with aspects are around 4-5k. Am I right @Sepideh-Ahmadian ?
Also, in other domains like hidden entity detection, like this (https://github.com/HawreH/Implicit-Entity-Recognition-and-Linking-in-Tweets-Resources-and-Dataset) is about 8k.

[Sepideh-Ahmadian]

Yes, you are correct, @hosseinfani.

Sentiment analysis alone is not sufficient; we need to focus on datasets that explicitly include aspects. Datasets and research categorized under Aspect-Based Sentiment Analysis (ABSA) are closely related to our work and can serve as valuable resources. I recommend starting with this category when searching for relevant works and datasets.

Most datasets in the ABSA domain are in English, with Chinese being the second most common language. However, there are no datasets solely in Arabic, though the SemEval 2016 multilingual dataset does include Arabic.

The first step in the project (similar to the plan for the English version) would be to familiarize ourselves with the format of the SemEval 16 dataset. This includes creating a detailed report on the number of implicit aspects it contains and analyzing them individually to determine whether the aspects can be inferred from the context or if they are more challenging to identify.

Once we have a clearer understanding of the reviews, we can begin our examination and experimentation using large language models (LLMs).

[Sepideh-Ahmadian]

@AliAlsalkhadi This is a link to the SemEval 16 train dataset.
(https://alt.qcri.org/semeval2016/task5/data/uploads/trial-data/hotels_trial_arabic.xml)
The reviews contains target="NULL" are the ones with implicit aspects.
I think the SemEval 16 is a good reference for addressing the challenges in Arabic review analysis.

I have a question regarding preprocessing Arabic reviews. Since spoken Arabic varies across countries, with differences in diction and even additional letters for specific sounds, should we consider the regional variations of the language during preprocessing? As a native speaker, I believe you might have a better understanding of how to approach preprocessing these reviews effectively.

[AliAlsalkhadi]

@Sepideh-Ahmadian I think we should take a look at the dataset first before deciding. Diction and dialects could definitely be an issue since Arabic is such a rich language, but I believe that when people write, they usually stick to Standard Arabic, which is what books and articles are typically written in. I don't think that we should consider the regional variations since that would be almost impossible. I think we should focus on Standard Arabic reviews also since we are using backtranslation those would be more proper.

[Sepideh-Ahmadian]

@AliAlsalkhadi , @Mahmoud-s-programs
The primary question concerns the variability of Arabic dialects in real-world contexts, such as product reviews on platforms like Amazon. It is impractical to restrict dataset solely to Modern Standard Arabic (MSA) when addressing everyday language use, as dialectal Arabic plays a significant role in such scenarios.
In LADy’s pipeline, back-translation is employed as a data augmentation method to help identifying implicit aspects. However, given that the dataset will be used for diverse experiments, other teams might opt for alternative methods like synonym replacement or even non-augmentation approaches. Ensuring the dataset reflects real-world scenarios and aligns with the specific task is critical. In this case, our work, which focuses on aspect detection as part of Aspect-Based Sentiment Analysis (ABSA), emphasizes identifying implicit aspects.
Several resources provide insights into Arabic datasets and their characteristics. For instance, the ASTD: Arabic Sentiment Tweets Dataset, as noted in Table 1 of the paper, includes both MSA and dialectal Arabic, offering a useful reference for Arabic dataset creation. Another example is the dataset discussed in , which has been cited over 100 times and consists of tweets about COVID-19, likely containing a mix of MSA and dialectal Arabic.
Additional resources for Arabic NLP tasks include:
• "Hotel Arabic-Reviews Dataset Construction for Sentiment Analysis Applications" link: This study uses separate classifiers for MSA and dialectal Arabic, demonstrating the value of combining both for a richer dataset.
• "An Annotated Huge Dataset for Standard and Colloquial Arabic Reviews for Subjective Sentiment Analysis" link: This dataset also includes both MSA and dialectal Arabic.
• "Aspect-Based Sentiment Analysis: An Overview in the Use of Arabic Language" link: This paper provides a comprehensive overview of Arabic datasets for ABSA.
For a deeper understanding of available Arabic datasets, conducting a literature review is essential. The tables and references in these papers (For instance the Table 1 in Aspect-Based Sentiment Analysis: An Overview in the Use of Arabic Language) provide a solid foundation for examining the features and characteristics of Arabic datasets relevant to ABSA. This approach will strengthen the arguments and ensure the dataset is both comprehensive and representative of real-world language use.

My review of recent ABSA studies in Arabic language indicates that many of them rely on datasets such as SemEval 2016( article, dataset sample). However, for the Arabic section of SemEval 2016, the original paper does not provide detailed information regarding data collection or annotation processes. To gain further insights, I recommend reviewing other works conducted by the same research group. If sufficient information is not available, reaching out to the authors for additional resources would be a practical next step.
The following articles provide useful references to the datasets employed in Arabic ABSA research. While the focus here is not on the techniques, these works can help track the datasets they utilize:

1. Using long short-term memory deep neural networks for aspect-based sentiment analysis of Arabic reviews. link
2. Arabic aspect-based sentiment analysis using bidirectional GRU-based models. link
3. Enhancing Arabic aspect-based sentiment analysis using deep learning models. link
4. Enhancing Aspect-Based Sentiment Analysis of Arabic Hotels’ Reviews Using Morphological, Syntactic, and Semantic Features. link
   These articles serve as a valuable starting point for identifying Arabic datasets used in ABSA and understanding their characteristics, which could be beneficial for advancing your research.

[Sepideh-Ahmadian]

@AliAlsalkhadi and @Mahmoud-s-programs as discussed with @hosseinfani today, there is no need to be concerned about the preprocessing.