Building machine reading comprehension model from scratch

Abstract

In this paper, we introduce a machine reading comprehension model and how we built this model from scratch. Reading comprehension is a crucial requisite for artificial intelligence applications, such as Question-Answering systems, chatbots, virtual assistants etc. Reading comprehension task requires the highest complexity of natural language processing methods. In recent years, the transformer neural architecture could achieve the ability to solve high complexity tasks. To make these applications available in Hungarian as well it is inevitable to design a Hungarian corpus of reading comprehension so that the pretrained models can be fine-tuned on this dataset.In our research, we have created the HuRC (Hungarian Reading Comprehension) corpus, which is the first dataset in Hungarian aiming to train, test and evaluate language models on a reading comprehension task. We built such a dataset based on the English ReCoRD corpus. This is a dataset of 120,000 examples consisting of news articles containing a passage and a close-style query, where a named entity is masked and the reference answer has to be found in a list.Using the evaluated dataset and transformers' question-answering library, we have built the first neural machine reading comprehension models in commonsense reasoning task for Hungarian.

Improve Performance of Fine-tuning Language Models with Prompting

“…As usual, these corpora are divided into training, validation and test sets. The subcorpora of HuLU are either translated datasets (Hungarian Choice of Plausible Alternatives Corpus -translated from CoPA [12] -, Hungarian Recognizing Textual Entailment dataset -translated from the RTE1, RTE2, RTE3 and RTE5 datasets [13], [14], [15], [16] -, Hungarian version of the Stanford Sentiment Treebanksentences translated from the SST5 dataset [17] -, Anaphora resolution datasets for Hungarian as an inference task [18] -the examples translated from the Winograd schemata and the WNLI dataset [19], [10]) or datasets created from scratch the design of which follows some English datasets (Hungarian CommitmentBank Corpus -designed based on Commitment-Bank [20] -, Hungarian Corpus of Linguistic Acceptability -designed based on COLA [21] -, Hungarian Corpus for Reading Comprehension with Commonsense Reasoning [22] -designed based on ReCoRD [23]).…”