The first step of NER is to convert a sequence of syllables (or morphemes) into a sequence of numbers in order to process using a computer. A unique number is assigned to every syllable (or morpheme) that appears in the training data. And all syllables (or morphemes) that appear in validation or test data but not appear in training data is given a specific number that denote an OOV syllable (or morpheme).

The numbers that denote a syllable or a morpheme is transformed into vectors so that they can be used in deep learning models. The vectors are computed in a way that they represent various aspects of the meanings of syllables or morphemes. For such a word-embedding vector, there are two methods: using pre-learned values and learning concurrently while training the NER model.

Data expressed using vectors are fed into a neural network such as RNN or convolutional neural network(CNN) and the network compute the probability of each named entity category for current input. The tag of the category with highest computed probability is assigned to the current input.

In the past, Korean NER used methods such as Hidden Markov Model (statistics-based machine learning), conditional random fields (CRFs), and structural support vector machine (SVM). Nowadays, however, deep-learning-based methods are mainly studied. The performance of early studies on deep-learning-based NER was similar to the that of conventional statistics-based methods (through a model using feed-forward neural network) [12].

Another study [13] proposed a method using RNN that is suitable for sequential data.

Figure 1 shows a bidirectional RNN. For the forward calculation, the output of the previous hidden state is used to compute the output of the current hidden state.

Figure 1: 
Bidirectional Recurrent Neural Network

As shown in Equation (1), the current input x_t is multiplied by a weight matrix W_x and the previous hidden state output h_t-1 is multiplied by another weight matrix W_h. These two values and a bias value b is added together and resulting value is inputted to the hyperbolic tangent function to produce h_t. For the backward computation, using the same method, ht' is computed, and the values of both directions are concatenated to output O_t of Equation (2).

RNNs can learn sequential dependencies in data; however, there exists a vanishing gradient issue, in which the learning is not properly performed if the length of the input data increases. Lately, long short-term memory (LSTM) that adds gates to RNN, is used to resolve the vanishing gradient problem.

However, in pure LSTM-based methods, only the context proceeding the current input is considered. To utilize both the proceeding and following contexts of current input, bi-directional LSTM is devised [1]-[3][7][11][14][15]. Furthermore, some studies added CRFs to the bi-LSTM models to reflect dependencies between output tags [1]-[9][11][13]-[15].

In addition, a method of using stacked bi-LSTM that stacks multiple layers of bidirectional LSTM was proposed along with an ensemble model, which learns multiple models and chooses a value having the highest probability among the output values [4][5][15].

For English NER, it is reported that the CNN-CRFs models perform inferior to bi-LSTM-CRFs [14]. However, no study has compared bi-LSTM-CRFs and CNN-CRFs for Korean NER.

In terms of Korean NER, RNN models are dominant models that use deep-learning. As RNN processes input data sequentially, parallel processing is impossible. Therefore, as the length of sentence increases, the processing speed slows down.

Feed-forward neural network has a drawback that local features of input data cannot be extracted. CNN, which is usually used in image processing, is a method of extracting local features of data by performing convolution operations.

As shown in Figure 2, a partial calculation is performed using a filter for input data. Local features can be extracted by dividing the whole data into areas of the same size with the filter and performing the following computation.

Figure 2: 
Convolutional Neural Network

As shown in Equation (3), an element x_i of an area with size h in input data and an element w_j of the filter are multiplied and all values are added. Using this sum and an activation function ReLu, output c_i is obtained.

As shown in Equation (4), the ReLu function outputs 0 for a value less than or equal to 0 and as-is for a value greater than 0.

All output c_i are fed into a pooling layer and the layer produce an output p_k. A pooling layer may be a max pooling or an averaging pooling layer. Max pooling uses the maximum value among the values of c_i as p_k, and average pooling uses an average value.

In general, CNN uses multiple filters of varying sizes to extract various features from input data.

The size of data decreases through a convolution operation because the convolution operation transforms an area of data to a single number using the filter multiplication. To make the size of the output the same as that of the input, we can add paddings which consists of all 0 values to the original data.

Furthermore, 2D or 3D convolution can be used, which can process 2D or 3D data depending on the type of input data. This CNN method is commonly used in image processing.

A CNN-based model is proposed to improve the processing speed of the RNN model which is usually used in Korean NER. The proposed CNN-CRF model is shown in Figure 3. We put a limit to the length of input sentences. Before embedding, we make the lengths of all sentences same by adding paddings to the sentences having shorter length than the maximum length. Sentence-embedding uses a lookup table to convert a sentence into vectors that can properly represent the feature of each syllable. The lookup table is initialized with random numbers and as the learning is performed, the best-fitting values are learned.

Figure 3: 
CNN-CRFs model

Embedded data are inputted in the convolutional layer to extract features. 1D convolution is used for a sentence because calculation is performed in only one direction, including all embedding vectors of a syllable.

When a CNN is used, several types of filter are usually used. However, the proposed model extracts a feature by using only one type of filter having a size 5 to include two syllables, before and after current position.

To make the length of the output data the same as the maximum length of a sentence, two padding symbols are added at the beginning and end of the sentence. ReLu is used as the activation function, and the pooling layer is not used (as the input and output sizes are identical).

The output of the convolution layer is fed into a fully connected neural network which produce the probabilities of all categories for each input symbol. The shape of the output of the fully connected neural network is the length of the sentence times the number of categories. The output of the fully connected layer is converted to probability values between 0 and 1 using the Softmax function.

Softmax is as shown in Equation (5). The exponential function of an input value a_k is divided by the sum of exponential function of all input values, thereby outputting y_k. As the sum of all output y_k is the same as 1, this value can be used as the probability value.

This output value is the given to the CRFs, and a tag sequence corresponding to the whole sentence is finally obtained as the output. The CRFs are trained by considering the forward and backward dependency of the output tags, and the Viterbi algorithm is used to output the optimal tag sequence.

Experiments were conducted using NER corpora of Electronics and Telecommunication Research Institute (ETRI), Korean Information Processing System Competition (KIPSC) [16], and Korea Maritime and Ocean University (KMOU) [17]. The corpora were converted in syllable units, and the maximum length was set to 180 to conduct the experiments in as similar environments as possible. To tag named entities, BIO tags were used to mark the beginning of named entity as B (Begin), the middle as I (Inside), and a syllable that is not a part of a named entity was marked as O (Outside).

Table 1 shows the details of each corpus. The ETRI’s corpus (A) and the KMOU’s corpus (C) consist of 10 categories: personal names (PER), location names (LOC), organization names (ORG), and others (POH); date (DAT), time (TIM), and duration (DUR) for time expressions; currency (MNY), percentage (PNT), and other numeral expressions (NOH) for numeral expressions.

The corpus distributed at the KIPSC (B) consists of five categories: personal names (PS), location names (LC), organization name (OG), date (DT), and time (TI). The corpus distributed at the KIPSC has the least number of sentences and contains relatively small number of personal names and large number of time expressions. Next, the corpus of ETRI consists of similar volumes with KIPSC corpus. The most recently distributed corpus of KMOU has the largest volume.

The experiments were performed by dividing each corpus into training dataset (80%), validation dataset (10%), and test dataset (10%).

Table 2 shows the equipment used in the experiments. Each model was implemented using Tensorflow, an open-source deep learning framework provided by Google. And the five models were compared for each corpus. The word-embedding size of models was 128 and the dropout rate was 0.2, which were all identical. The other parameters are shown in Table 3.

Table 4 shows the experimental results. In the experiment using ETRI corpus, the performance of the bi-LSTM model was 0.5% better than the bi-LSTM-CRFs model. The bi-LSTM-CRFs (83.4%) performed better (83.4%) with the validation data than with the test dataset (82.7%). We believe the reason as the amount of learning data being insufficient. The corpus distributed by KIPSC showed a low performance overall. While analyzing the prediction results, the prediction results of KIPSC corpus showed a performance of approximately 90% for the O tag whereas those of the other corpuses showed a performance of more than 95% for the O tag. The models trained using the KIPSC corpus could not distinguish named entities and non-named entities properly because the volume of corpus was small.

The KMOU corpus having the largest amount of data showed the most reliable results. In the experiments using the KMOU corpus, the performance improved 0.8% when the bi-LSTM model and CRFs were combined; however, the speed declined (approximately 2.7 times). The CNN model was fast; however, its performance was 13.2% worse than bi-LSTM’s. The CNN-CRFs model using 64 filters showed a performance of 83%, which was similar to that of bi-LSTM-CRFs. Meanwhile, the CNN-CRFs-2 model using 256 filters showed a performance of 83.6%, which was 0.4% higher, and 27.5% faster than bi-LSTM-CRFs.

In experimental results of CNN-CRFs-2 model using the test dataset of KMOU, the performances of ORG and LOC, which could exhibit ambiguity issues, were low. Furthermore, there were many cases of incorrectly predicting the named entity category in the POH category, which contained many OOV words. In addition, low performances were exhibited by DUR, PNT, and MNY.

H. Y. Yu and Y. J. Ko [6] used the KIPSC corpus and demonstrated a performance of F1 85.49%. The study conducted experiments based on morphemes, whereas the proposed model is based on syllables. Furthermore, unlike the proposed model that did not use additional resources, the study utilized additional resources, such as pre-trained word-embedding and named entity dictionary.

M. A. Cheon et al. [10] used multi-head attention for the ETRI corpus and demonstrated a performance of F1 84.84%. The study used only three categories of personal names, location names, and organization names.

The proposed study demonstrated a performance of F1 83.6% in the experiment with the KMOU corpus. Our method tags NER based on syllables. However, the experimental method was different from that of studies [6] and [10]. As Korean NER is studied using several kinds of corpus, categories are occasionally different. Furthermore, direct performance comparison is difficult because additional resources such as morpheme-based learning method, pre-learned embedding, and named entity dictionary are used to improve the performance.