Lexical simplification (LS) is the process of replacing complex words in a sentence with simpler alternatives to improve readability and understandability. However, LS is a challenging task because the target population may have varying literacy levels, resulting in different needs. This study aims to enhance the SimpLex architecture, a LS framework proposed in the literature. Several techniques have been explored to achieve this goal. First, we added new features and trained two additional machine learning models: LightGBM and XGBoost. Additionally, multiple transformer models were examined, including DistilBERT, RoBERTa Large, Facebook BART, and Microsoft DeBERTa. The results demonstrated the effectiveness of LightGBM and XGBoost in predicting complex words. Furthermore, DistilBERT and RoBERTa Large achieved higher SARI scores and lower perplexity compared to the other pre-trained models, except BERT, which had a slightly lower perplexity than the aforementioned models.

Evaluating the quality and effectiveness of text summarization techniques is essential to ensure accurate and informative summaries. In this paper we present a novel method for evaluating Arabic text summarization based on measuring the similarity between embedding keyphrases. We utilize a pretrained BERT model to represent each keyphrase as a dense vector. Then, we quantify the similarity between keyphrases in the system and the human summary. We tested several lengths of keyphrases. After, we built a predictive model by combining different keyphrases based features with diverse linguistic quality features. To evaluate the effectiveness of our method, we compare our embedding keyphrase similarity method to traditional metrics such as ROUGE. Our experiments demonstrate that embedding-based evaluation provides a more nuanced assessment of the summarization quality, considering the semantic similarities between keyphrases. With the incorporation of semantic similarities, our approach captures the contextual relevance of keyphrases and enhances the evaluation process.

In this paper, we propose a hybrid method for the spoken Tunisian dialect understanding within a limited task. This method couples a discriminative statistical method with a domain ontology. The statistical method is based on conditional random field (CRF) models learned from a little size corpus to perform conceptual labeling task. These models are able to detect the semantic dependency between words. However, the domain ontology is used to add prior knowledge about the task. Our experiments are based on a real spoken Tunisian dialect corpus. The obtained results show that the proposed method is able to improve the performance of CRF models for speech understanding by the integration of the domain ontology. Our method can be exploited for under-resourced languages and Arabic dialects to overcome the lack of linguistic resources .