Recently, large language models (LLMs) have attracted considerable attention due to their remarkable capabilities. However, LLMs’ generation of biased or hallucinatory content raised significant concerns, posing major challenges for their practical application. Many studies have dedicated efforts to address these critical issues, adopting various approaches to mitigate bias and hallucinations in LLM-generated content. Remarkably, no review papers have synthesized insights on these two primary problems. Addressing this gap, this paper aims to conduct a simultaneous and dual-focused review of the current landscape of research. The discussions encompass widely used and newly proposed benchmarks and evaluation methods on bias and hallucination in LLMs. This paper also investigates advanced mitigation methods and present a taxonomy based on different mitigation strategies. Moreover, a comparative analysis of the sources, mitigation methods, and evaluation methods for bias and hallucination is included. In the end, this paper provides a synthesis of current research trends and suggests potential directions for future research to address bias and hallucination in LLMs, considering the ongoing challenges in this field.
The web is being loaded daily with a huge volume of data, mainly unstructured textual data, which increases the need for information extraction and NLP systems significantly. Named‐entity recognition task is a key step towards efficiently understanding text data and saving time and effort. Being a widely used language globally, English is taking over most of the research conducted in this field, especially in the biomedical domain. Unlike other languages, Arabic suffers from lack of resources. This work presents a BERT‐based model to identify biomedical named entities in the Arabic text data (specifically disease and treatment named entities) that investigates the effectiveness of pretraining a monolingual BERT model with a small‐scale biomedical dataset on enhancing the model understanding of Arabic biomedical text. The model performance was compared with two state‐of‐the‐art models (namely, AraBERT and multilingual BERT cased), and it outperformed both models with 85% F1‐score.
ChatGPT has attracted extension attention of academia and industry. This paper aims to evaluate ChatGPT in Chinese language understanding capability on 6 tasks using 11 datasets. Experiments indicate that ChatGPT achieved competitive results in sentiment analysis, summary, and reading comprehension in Chinese, while it is prone to factual errors in closed-book QA. Further, on two more difficult Chinese understanding tasks, that is, idiom fill-in-the-blank and cants understanding, we found that a simple chain-of-thought prompt can improve the accuracy of ChatGPT in complex reasoning. This paper further analyses the possible risks of using ChatGPT based on the results. Finally, we briefly describe the research and development progress of our ChatBIT.
1 April 2014 in International Journal of Automation and Computing
Context. Social media platforms such as Facebook and Twitter carry a big load of people’s opinions about politics and leaders, which makes them a good source of information for researchers to exploit different tasks that include election predictions. Objective. Identify, categorize, and present a comprehensive overview of the approaches, techniques, and tools used in election predictions on Twitter. Method. Conducted a systematic mapping study (SMS) on election predictions on Twitter and provided empirical evidence for the work published between January 2010 and January 2021. Results. This research identified 787 studies related to election predictions on Twitter. 98 primary studies were selected after defining and implementing several inclusion/exclusion criteria. The results show that most of the studies implemented sentiment analysis (SA) followed by volume‐based and social network analysis (SNA) approaches. The majority of the studies employed supervised learning techniques, subsequently, lexicon‐based approach SA, volume‐based, and unsupervised learning. Besides this, 18 types of dictionaries were identified. Elections of 28 countries were analyzed, mainly USA (28%) and Indian (25%) elections. Furthermore, the results revealed that 50% of the primary studies used English tweets. The demographic data showed that academic organizations and conference venues are the most active. Conclusion. The evolution of the work published in the past 11 years shows that most of the studies employed SA. The implementation of SNA techniques is lower as compared to SA. Appropriate political labelled datasets are not available, especially in languages other than English. Deep learning needs to be employed in this domain to get better predictions.