In the example sentence (A), the predicate "배당된(baedang-doen, be assigned)" has the meaning group information of "award" in the Sejong case frame dictionary. Therefore, the phrase 'to me' can be used feature as a predicate and relationship argument having the meaning of 'award'. The meaning of 'to me' was analyzed as' ARG3 (goal) 'because the information available before the semantic group information was used was only type information of particle like ‘-게’, but' ARG2 (benefactive) can be analyzed correctly.

The role of the argument varies according to the predicate information. It is hard to know the exact meaning of the word vector or cluster information. The following is an example in which the predicate of the same vocabulary changes the role of the argument according to the predicate meaning.

(B) and (C) are the same as '달리다', but  '달리다' of (B) has the meaning of 'hanging or tying things at a certain place' as the passive word of '달다03'. and  '달리다' of (C) means a 'car or boat moves quickly'. In (B), the subject is 'lamp', and in (C) the subject is 'car'. However, the meaning of predicate of the sentences used in the two sentences is different, and the voice is also different depending on the meaning. As a result, "lamp" and "car" is both subject, but "lamp" becomes the patient and "car" becomes the agent.

(D) and (E) are also the predicates that change the state of voice to passive and active by meaning. '보이다' of (D) is a passive voice of '보다01', and '보이다' of (E) has the same meaning of '보다01' but has the causative.

In addition to this, there is an example in which the role of semantics is not the opposite as in the above example, but it can be confirmed that the arguments used together are different.

In the above example, the subject of (F) is '방세(Bangse, the rent)', and the subject of (G) is '길동(Gildong, Gildong)'. In the examples of (F) and (G), all of the subject are 'theme(patient)'. However, in the case of '밀리다' of (G) is used the passive voice of the predicate '밀다01'. so 'agent' is '철수(Cheolsu)'. As such, the predicates such as (F) and (G) '밀다' are different cases in which the arguments used together are different. These predicates can help to distinguish the roles of arguments accurately more than when using only syntactic information and investigation information, since they can be used to understand the arguments used in the arguments and the arguments used in the semantics.

As shown in the above examples, the ambiguity of meaning leads to a change in the analysis result even though it is the same vocabulary. Therefore, it is necessary to more clearly determine the semantic information of the predicate that plays an important role determining the semantic role. [5] proved that predicate senses information is useful for semantic decision in neural network model which is designed by using the existing FrameNet semantic group information and PropBank predicate senses information together. In this paper, we experimented to investigate the effect of morphological semantic information in Korean semantic role determination.

The composition of this paper is introduced in Chapter 2, and in Chapter 3, experimental method and experimental results are analyzed and explained. Finally, Section 4 describes the conclusion.

There have been studies using various feature that can express meaning for Korean semantic role determination. [3,4] used semantic dictionary information from the Sejong Electronic Dictionary to conduct semantic role labeling. The case frame dictionary shows the type of noun and 'josa' information used together with the predicate. [3] applied the self-training algorithm using the case frame dictionary. [4] conducted experiments to improve semantic role labeling performance by selecting a specific 'josa' with a high ambiguity in semantic role decision. high ambiguity 'josa' are  '에','로', '에서', and '에게'. [2] applied the Deep Learning method to determine the Korean semantic role labeling.

In order to utilize the semantic representation information of words in English, various types of deep learning algorithms are presented as word vectors generated by word embedding. [5] used the information of FrameNet and the information of PropBank as learning information of neural network. As a result of the experiment, [5] proved that the performance is improved when using the information of PropBank 's predicate senses rather than using only the information of FrameNet. This proves that the word senses can be solved to help the semantic role labeling. [6] is an improved version of the one-step model that has been studied previously. One-step model is the same structure as SENNA [7] and is labeled using IOB tag instead of IOBES tag. Two-step model was developed by developing one-step model. In the first step, IOBES tags are classified, and in the second step, the arguments are classified. [8] measured the semantic role labeling performance using various neural network models. [9] compared the performance of both CNN and LSTM models. However, all of these methods did not use semantic information directly, so we could not confirm the effect of semantic information.

In  this paper, to capture the semantic role of arguments according to the  predicate meaning, we suggest new features, such as word sense Information.

In  this paper, we use the corpus which is supplemented from the Korean semantic  role corpus of Ulsan University for learning and evaluation. The features used in the paper  were the feature used in [1] and the word senses information extracted from the  corpus. Table 1 shows the feature used in the baseline system to verify the performance of meaning information  feature. The model was generated using CRFs and evaluated by 5-cross fold.  11,468 sentences were used for the learning and 2,867 sentences were used for the evaluation. Table  2 shows the results of the baseline experiment using only the feature of Table  1. P means precision, and r means recall.

Table  3 shows the experimental results of adding the meaning group information of the  case frame dictionary, the meaning number of the morphological semantic  analysis sense information corpus, and the cluster information using the  word vector generated by Word2Vector in the baseline. The cluster information  is generated by a skip-gram word vector, then 200 clusters are generated by the  k-means algorithm, and the group number including the word is used as a  feature. When k was set to 100, 200, and 300, the best performance was 200 and  the k value was set to 200, which was then applied to the experiment.

Experimental results show that all semantic  expression features are useful for semantic role labeling. Among them,  morphological semantic analysis information provided the greatest help in  determining semantic role. In the experiment using all semantic expression  feature, the average performance was 77.25%. In addition, we confirmed that the  performance is improved when the word senses information is used only for the  predicate.

In  this paper, we confirmed that the role of the argument changes according to the  meaning of the word, and experimented to confirm the performance changes of semantic  role labeling according to the use of the word semantic expression information.  To confirm this, we used morphological semantic analysis corpus, and we used  word semantic numbers as learning feature of CRF. In addition, according to the  meaning of the predicate, the change of the voice and the change of the case  frame were classified, and the role of the argument used together and the  presence or absence of use was observed. The performance was measured by  5-cross fold using 11,468 sentences for training and 2,867 sentences for  evaluation, and the average performance was 77.25%.    Future  work will be done to analyze the exact reason why performance is improved when  using the semantic number only in the descriptor. We also plan to design a  structure for syntactic analysis, predicate semantic determination, and  semantic decision making for AMR (Abstract Meaning Representation) model  generation.

[1] Tae-Ho Park, Jeong-Won Cha. "Feature Selection for Korean Sematic Role Labeling Using CRFs." Communications of the Korean Institute of Information Scientists and Engineers, 34.8 (2016.8): 37-41.

[2] Jangseong Bae, Changki Lee and Soojong Lim, “Korean Semantic Role Labeling using Deep Learning”, Korea Computer Congress 2015 , 2015.06, 690-692, 2015.

[3] Byoung-Soo Kim, Yong-Hun Lee and Jong-Hyeok Lee, “Unsupervised Semantic Role Labeling for Korean Adverbial Case”, Journal of KISS : Software and Applications – 2007.6 34(2), 112-122

[4] Hyun-Ki Jung and Yu-Seop Kim, “Semantic Role Labeling of Korean Adverbial Arguments by using the Expanded Case Frame Dictionary”, Journal of Korean Institute of Information Technology 9(10), 167-176, 2011.

[5] FitzGerald, Nicholas, et al. "Semantic Role Labeling with Neural Network Factors." EMNLP. 2015.

[6] Fonseca, Erick R., and Joao Luis G. Rosa. "A two-step convolutional neural network approach for semantic role labeling." Neural Networks (IJCNN), The 2013 International Joint Conference on. IEEE, 2013.

[7] Collobert, Ronan, and Jason Weston. "A unified architecture for natural language processing: Deep neural networks with multitask learning." Proceedings of the 25th international conference on Machine learning. ACM, 2008.

[8] Wang, Zhen, et al. "Chinese Semantic Role Labeling with Bidirectional Recurrent Neural Networks." EMNLP. 2015.

[9] Zhou, Jie, and Wei Xu. "End-to-end learning of semantic role labeling using recurrent neural networks." ACL (1). 2015.