In recent years, the large-scale open Chinese and English question answering (QA) corpora have provided important support for the application of deep learning in the Chinese and English QA systems. However, for low-resource languages, such as Tibetan, it is difficult to construct satisfactory QA systems, owing to the lack of large-scale Tibetan QA corpora. To solve this problem, this paper proposes a QA corpus generation model, called QuGAN. This model combines Quasi-Recurrent Neural Networks and Reinforcement Learning. The Quasi-Recurrent Neural Networks model is used as a generator for Generative Adversarial Network, which speeds up the generation of text. At the same time, the reward strategy and Monte Carlo search strategy are optimized to effectively update the generator network. Finally, we use the Bidirectional Encoder Representations from Transformers model to correct the generated questions at the grammatical level. The experimental results show that our model can generate a certain amount of effective Tibetan QA corpus, and the BLEU-2 value increases by 13.07% than baseline. Moreover, the speed of the model has been greatly improved.

Currently, research on question answering (QA) with deep learning methods is a hotspot in natural language processing. And most of the research mainly focused on English or Chinese since there are large-scale open corpora, such as WikiQA or DoubanQA. However, how to use deep learning methods to QA of the low resource languages like Tibetan becomes a challenge. In this paper, we propose a hybrid network model for the Tibetan QA, which combines the convolutional neural network (CNN) and long short memory network (LSTM) to extract effective features from small-scale corpora. Meanwhile, since the strong grammar rules of Tibetan, we use the language model to decode the output of LSTM layer which makes the answer more accurate and smoother. In addition, we add the batch normalization to accelerate deep network training and prevent overfitting. Finally, the experiments show that the ACC@1 value of the proposed model in Tibetan QA is 126.2.

The task of relation extraction is classifying the relations between two entities in a sentence. Distant supervision relation extraction can automatically align entities in texts based on Knowledge Base without labeled training data. For low-resource language relation extraction, such as Tibetan, the main problem is the lack of labeled training data. In this paper, we propose an improved distant supervised relation extraction model based on Piecewise Convolutional Neural Network (PCNN) to expand the Tibetan corpus. We add self-attention mechanism and soft-label method to decrease wrong labels, and use Embeddings from Language Models (ELMo) to solve the semantic ambiguity problem. Meanwhile, according to the Tibetan characteristics, we combine the word vector and part of speech vector to extract deeply feature of words. Finally, the experimental results show that P@avg value increases by 14.4% than baseline.