Search

The complete semantic representation of a Tibetan sentence is mainly determined by the addition of a specific functional word. The choice of Tibetan functional words is mainly influenced (both explicitly and implicitly) by the sequence of Tibetan suffixes. In this article, we propose an RNN-based Tibetan radical suffix unit (TRSU) to consider this relationship. Specifically, for the Tibetan radical suffix unit-explicit (TRSU-E) method, the fixed suffix in Tibetan is used to determine the virtual functional words. For the Tibetan radical suffix unit-implicit (TRSU-I) method, the decision is assisted by adding a specific suffix. To test the method, we design a standard Tibetan corpus, which consists of different genres. Our experimental results show that the complexity of our method is reduced by up to 22.2% relative to the best baseline. Furthermore, with the hidden semantic information and implicit suffix, TRSU-I outperforms TRSU-E by reducing the perplexity (PPL) by 3%. Moreover, good results are achieved on the English Penn Treebank data set.

The Tibetan language model (TLM) is the key to Tibetan natural language processing. In this paper, we first observe that, different from widely used languages, Tibetan contains many morphological verbs that rarely appear in natural sentences but play a key role in accurate text prediction. This property is usually ignored by existing methods and makes traditional training strategies less effective in constructing accurate and robust TLMs. Hence, we propose a morphological verb-aware TLM by offline learning via a character frequency reweighting strategy and online tuning of discriminative weights conditioned on morphological verbs. However, because of the influence of morphological verbs on the tense and semantics of sentences, it is necessary to consider the morphological verbs in Tibetan. As a result, compared with state-of-the-art methods, our method not only reduces the perplexity but also improves the character error on tasks of the text prediction and automatic speech recognition (ASR).

In virtue of the superiority of handling the sequence data and the effectiveness of preserving long-distance information, recurrent neural network language model (RNNLM) has prevailed in a range of tasks in recent years. However, a large quantities of data are required for language modelling with good performance, which poses the difficulties of modeling for low-resource languages. To address this issue, Tibetan as one of minority languages is instantiated, and its radicals (components of Tibetan characters) are explored for constructing language model. Motivated by the inherent structure of Tibetan, a novel construction of Tibetan character embedding is exploited to RNNLM. The fusion of individual radical embedding is enhanced by three ways, including using uniform weight (TRU), different weights (TRD) and radical combination (TRC). This structure, especially combining with the radicals, can extend the capability to capture long-term context dependencies and solve the low-resource problem to some extent. The experimental results suggest that this proposed structure obtained a better performance than standard RNNLM, yielding 7.4%, 12.7% and 13.5% relative perplexity reduction by using TRU, TRD and TRC respectively.