The development of large language models (LLMs) is becoming increasingly significant, and there is a demand for high-quality, large-scale corpora for their pretraining.The quality of a web corpus is especially essential to improve the performance of LLMs because it accounts for a large proportion of the whole corpus. However, filtering methods for Web corpora have yet to be established.In this paper, we present empirical studies to reveal which filtering methods are indeed effective and analyze why they are.We build classifiers and language models in Japanese that can process large amounts of corpora rapidly enough for pretraining LLMs in limited computational resources. By evaluating these filtering methods based on a Web corpus quality evaluation benchmark, we reveal that the most accurate method is the N-gram language model. Indeed, we empirically present that strong filtering methods can rather lead to lesser performance in downstream tasks.We also report that the proportion of some specific topics in the processed documents decreases significantly during the filtering process.

For languages without natural word boundaries, like Japanese and Chinese, word segmentation is a prerequisite for downstream analysis. For Japanese, segmentation is often done jointly with part of speech tagging, and this process is usually referred to as morphological analysis. Morphological analyzers are trained on data hand-annotated with segmentation boundaries and part of speech tags. A segmentation dictionary or character n-gram information is also provided as additional inputs to the model. Incorporating this extra information makes models large. Modern neural morphological analyzers can consume gigabytes of memory. We propose a compact alternative to these cumbersome approaches which do not rely on any externally provided n-gram or word representations. The model uses only unigram character embeddings, encodes them using either stacked bi-LSTM or a self-attention network, and independently infers both segmentation and part of speech information. The model is trained in an end-to-end and semi-supervised fashion, on labels produced by a state-of-the-art analyzer. We demonstrate that the proposed technique rivals performance of a previous dictionary-based state-of-the-art approach and can even surpass it when training with the combination of human-annotated and automatically-annotated data. Our model itself is significantly smaller than the dictionary-based one: it uses less than 15 megabytes of space.

We present a three-part toolkit for developing morphological analyzers for languages without natural word boundaries. The first part is a C++11/14 lattice-based morphological analysis library that uses a combination of linear and recurrent neural net language models for analysis. The other parts are a tool for exposing problems in the trained model and a partial annotation tool. Our morphological analyzer of Japanese achieves new SOTA on Jumandic-based corpora while being 250 times faster than the previous one. We also perform a small experiment and quantitive analysis and experience of using development tools. All components of the toolkit is open source and available under a permissive Apache 2 License.

Flashcard systems are effective tools for learning words but have their limitations in teaching word usage. To overcome this problem, we propose a novel flashcard system that shows a new example sentence on each repetition. This extension requires high-quality example sentences, automatically extracted from a huge corpus. To do this, we use a Determinantal Point Process which scales well to large data and allows to naturally represent sentence similarity and quality as features. Our human evaluation experiment on Japanese language indicates that the proposed method successfully extracted high-quality example sentences.