2.1

Tibetan syllable-based model

Input vector of the Tibetan syllable-based entity recognition model (Figure 1) is a syllable sequence. There is a defect in this model, that is, the word itself and word order information of Tibetan words cannot be fully utilized.

Figure 1.

Tibetan syllable-based model.

2.2

Tibetan word-based model

The principle of the Tibetan word-based model (Figure 2) is similar that of the Tibetan syllable-based model, except that the input vector to this model is Tibetan word sequences after word segmentation. Therefore, this model will lead to the word segmentation error transmission. Thus it ultimately affects the performance of entity recognition.

Figure 2.

Tibetan word-based model.

2.3

Tibetan Lattice-LSTM-CRF model

Due to the exponential number of word-to-syllable paths in the grid, the Lattice LSTM model is employed to automatically manage the sentence’s information flow from start to finish. The gating unit is used to communicate information about various paths to each syllable in a dynamic manner. Lattice LSTM can learn to automatically identify helpful phrases from the information flow after exercising on the training data set, which enhances the performance of named entity identification (Figure 3). The benefit of the approach suggested in this study over syllable-based and word-based named entity identification systems is the inclusion of explicit vocabulary information for word segmentation rather than merely automatic attention, which reduces word segmentation mistakes.

Figure 3.

Tibetan Lattice-LSTM-CRF model.

2.4

LSTM layer

Although an RNN neural network may theoretically handle sequence information of any length. In practice, the gradient will disappear when the arrays are too long, and then characteristics of long-term dependency cannot be learned. Graves et al. enhanced the recurrent neural network and introduced the LSTM model to address this issue²⁸. The LSTM unit regulates data transport via the input cell, forget cell, and output cell. LSTM is a special RNN that can learn long-term laws. They have been applied very well on various issues and are now widely used. The LSTM encode unit (Figure 4).

Figure 4.

LSTM unit.

The specific calculation process is shown in Equations (1)-(6):

where σ is the sigmoid function and ⊙ is the dot product. At time t, x_t indicates the input status, h_t indicates the hidden state, and o_t indicates the output status, which comprises all effective information from the previous t. The control information goes into the following instant through c_t, which is an update gate. Because f_t indicates a forget gate, that control messages are erased. The output of the concealed state is determined jointly by the two.

3.1

Experiment data

3.2

Annotation strategies and evaluation indicators

General named entity recognition and labeled strategies include BIO mode, BIEO mode, and BIEOS mode. The BIEOS labeled technique is applied in this paper, that B means the start of the entity, I means inside the entity, and E means the entity’s termination. O means non-entity or entity outside, S means single syllable entity. When predicting the boundary of an entity, the entity type needs to be predicted at the same time, so there are 11 different categories of tags that can be predicted are O, S, B-S, I-S, E-S, B-D, I-D, E-D, B-P, I-P, E-P. In the testing process, it is only determined that an entity’s prediction is precision when both its border and type are entirely accurate.

The identification performance evaluation indexes of entities in the field of Tibetan medicine involve precision rate (P), recall rate (R) and composite index F1. The specific calculation method is shown in Equation (7), where Tp is the total number of entities that the model successfully identified, Fp is the total number of unrelated entities that the model correctly identified, and Fn is the total number of related entities that the model failed to identify.

3.3

Experiment environment and super parameter settings

The experiment environment in this study is Python 2.7, and the deep learning framework is Pytorch 0.3.0. post4. The settings of the neural network’s hyperparameter will have an effect on the neural network’s performance. Table 3 displays the settings of neural network parameters in this paper.

Table 3.

Neural network hyper parameter values.

3.4

Experiment design and results

The three categories of entities in the Tibetan medical named entity recognition dataset were utilized to test the recognition performance of the model used in this work, the following two sets of experiments were designed. The evaluation indexes of the experiment are precision (P), recall rate (R) and comprehensive index F1 value.

Experiment 1 Comparison of Tibetan medical named entity recognition results in Traditional Tibetan Medicine clinical text by different models. Table 4 gives the recognition results of the CRF (Conditional Random Field) model and the Lattice LSTM model, respectively.

Table 4.

Performance comparison of various models for Tibetan medical named entity recognition (%).

The findings of the experiment demonstrate that the Lattice LSTM model’s F1 value is 2.55% greater than the CRF model’s F1 value, demonstrating that the deep learning model outperforms the model based on statistics for the entity recognition problem in the field of Tibetan medicine. And it has been demonstrated that the neural network model can decrease the model’s reliance on artificial feature engineering by employing only representation features.

Experiment 2 Lattice LSTM model and CNN-BiLSTM-CRF model based on word vectors were used to contrast the effectiveness of entity recognition in the field of Tibetan medicine. The Lattice LSTM model can concurrently encode word information relating to the sequence and syllable-level sequence information for the model to access. Lattice LSTM adds word information in addition to syllable granularity, which improves the semantic expression and effectively solves the issue of improper word segmentation transfer. Table 5 shows that the Lattice LSTM model can effectively improve Tibetan medical named entity recognition effect.

Table 5.

Comparison of performance between word vector-based model and Lattice LSTM model (%).

Figure 5 shows the change trend of the number of iterations of the neural network and the values of the three entity recognition indexes P, R and F1.

Figure 5.

The number of iterations and the trend of P, R and F1 values.

Table 6 shows the accuracy rate (P), recall rate (R), and comprehensive index F1 value for the three entities SYMPTOM, DISEASE, and PRESCRIPTION.

Table 6.

The recognition outcomes of Tibetan medical named entities of different types.

The predictive accuracy of prescription entities in the above table is low. It is mainly due to the presence of nested, acronyms and other interference information in many Tibetan drug names in prescriptions. So it is easy to predict errors without other sufficient context information. For example, the name of Traditional Tibetan medicine (Ruyi Treasure Pills) is often abbreviated to (Meaningless) in prescriptions. In addition, because there are not many medical records collected at present, there are fewer nested disease entities involved in it. However, as the amount of data increases, the disease names become more and more nested. This condition will have an impact on the accurate identification of disease entities. For example, the both disease entities (Hypertensive heart disease) and (Chronic superficial gastritis) have been nesting phenomena. Based on the above phenomena, we will focus on solving the problems of nesting and abbreviation in Traditional Tibetan medicine entities in our future research work.

3.5

Example of Tibetan medical named entity recognition

Take a sentence in a Tibetan inpatient medical record as an example to demonstrate the Tibetan medical named entity recognition effect of the Tibetan Lattice-LSTM model. Specific examples are shown in Table 7.

Table 7.

Example of Tibetan medical named entity recognition.

Note: Italic words indicate incorrect entities, boldfaced word indicates correct entities.