Global Spelling Correction in Context using Language Models: Application to the Arabic Language

Laaroussi, Saida; Yousfi, Abdellah; Aouragh, Si Lhoussain; Alaoui, Said Ouatik El

doi:10.12785/ijcds/130129

Cited by 3 publications

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Recently, a few papers introduced deep learning techniques and embedding representations in Arabic QAS like in [14], [15] and [16]. In other research [17], authors use a probabilistic model, and algorithms such as Viterbi to a contextual spelling correction for the Arabic text.…”

Section: B Arabic Question Answering Systemsmentioning

confidence: 99%

Exploring Sentence Embedding Representation For Arabic Question Answering

Lahbari,

Ouatik El Alaoui

2024

IJCDS

View full text Add to dashboard Cite

Question Answering Systems (QAS) are made to automatically provide accurate response to user questions that are phrased in natural language. Most of the existing QAS adopting traditional representations like word embedding and bag-of-words, have shown promising results. However, only a few works take into account the contextual information and meaning within texts to extract answers from huge sources of information. Moreover, dealing with Arabic open-domain question-answering systems is still challenging due to its rich morphology and ambiguity of words. To address these limitations, we introduce, in this study, a novel QA approach for the Arabic language that is based on passage retrieval and Sentence Embedding (SE) representation. It consists of three steps: (1) Question classification and query formulation, (2) Documents and passages retrieval, and (3) Answers extraction. In this work, we make use of the AraBERT transformer model to compute vector representation. This allows for considering both implicit semantics and the words' context within the text. Furthermore, in order to collect potential passages for user questions, we investigate a method for retrieving Arabic passages using the BM25 model, a query expansion process, and SE representation. The final answer is generated by fine-tuning AraBERT parameters and ranking passages so that the most relevant ones can be extracted. To assess our approach, we carried out several experiments on CLEF and TREC datasets using two different taxonomies. The obtained results show that the proposed method achieves 92% in terms of F1-score.

show abstract