Incorporating external knowledge for text matching model

Computer Speech & Language, Год журнала: 2024, Номер 87, С. 101638 - 101638

Опубликована: Март 12, 2024

Язык: Английский

---

Prompt Defect Response Via Machine Reading Comprehension Using a Hybrid Large Language Model Approach

Опубликована: Янв. 1, 2024

This study proposes a hybrid method using two types of large language models (LLM) for prompt response to defect complaints by exploring rapidly potential causes and repair methods via machine reading comprehension (MRC) tasks. Although numerous past maintenance records guidelines offer valuable insights into or newly reported defects, manually reviewing all data is impractical due the significant time effort required. MRC natural processing (NLP) task that trains read extensive texts answer questions. While recent state-of-the-art (SOTA) LLMs, as they are, exhibit high performance general questions, falter in specialized domains require fine-tuning. However, generating question-answer (QA) datasets fine-tuning time-consuming, taking over 200 days with crowdsourcing. Furthermore, many companies restrict LLM usage daily tasks leakage risks. To mitigate these challenges, this introduces approach wherein Bidirectional Encoder Representations from Transformers (BERT) fine-tuned QA datasets, automatically generated Generative Pre-trained Transformer (GPT) publicly available construction guidelines. The GPT-applied part proposed 2,548 pairs seven half hours, significantly reducing dataset generation time. For MRC, BERT achieved competitive highest F1 score 88.0%, outperforming Korean benchmark's (68.5%). contributes reduced cost resources constructing domain-specific performing efficient complaint within data-secure environment.

Язык: Английский

---

Weight Averaging and re-adjustment ensemble for QRCD

Journal of King Saud University - Computer and Information Sciences, Год журнала: 2024, Номер 36(4), С. 102037 - 102037

Опубликована: Апрель 1, 2024

Question Answering (QA) is a prominent task in the field of Natural Language Processing (NLP) with extensive applications. Recently, there has been notable surge research interest concerning development QA systems for Holy Qur'an, an Islamic religious text. The Qur'an Reading Comprehension Dataset (QRCD) Malhas and Elsayed (2020) highly commendable effort this respect. It stands as first benchmark dataset specifically designed set directly answerable questions from Qur'an. Each question meticulously labeled all potential answers sourced From our perspective, main challenge QRCD stems limited volume training data it offers. As solution we propose innovative approach to build Deep Neural Network (DNN) ensemble, centered around Ara-Electra model (Antoun et al., 2021), that called Weight Averaging Re-adjustment (WAR) model. constructed by computing running averages states evolve during single session ensuring weights are readjusted prior each epoch, order hold back over fitting data. scheme results standalone exhibits benefits multi-model ensembles. distinguished other ensembles proposed accumulates outputs multiple expert models employs classic techniques like hard voting or score averaging on output probabilities unified results. costs individual time compute resources. WAR outperforms existing improved generalization unseen achieves F1, partial Reciprocal Rank (pRR), exact-match (EM) scores 0.567, 0.60 0.29 respectively, exceeding best reported 3%, 1.5% 0.69% respectively. Notably, we're comparing top different models, highlighting model's consistent performance across three metrics.

Язык: Английский

---

Curriculum Learning Driven Domain Adaptation for Low-Resource Machine Reading Comprehension

IEEE Signal Processing Letters, Год журнала: 2024, Номер 31, С. 2650 - 2654

Опубликована: Янв. 1, 2024

Язык: Английский

---

Maximizing discrimination masking for faithful question answering with machine reading

Information Processing & Management, Год журнала: 2024, Номер 62(1), С. 103915 - 103915

Опубликована: Окт. 9, 2024

Язык: Английский

---