4.5 Error Analysis To better understand our method,we perform error analysis on the classification results.We mainly describe two sources of errors:word representation failure and word order capturing failure.From the perspective of word representation,the classification error is caused by the fact that the difficulty infor- mation may not be properly encoded into the word embedding.Table 4 shows that KEWE performs better than the classical text-based word embeddings in encoding difficulty information into represen- tation,but in the final results there still exists the failure of word representation in KEWE.From the perspective of word order capturing,the classification error is caused by the fact that the order among words may be neglected,so that the syntactic difficulty,pragmatic difficulty,and discourse difficulty of documents are ignored during the process of readability assessment.Table 3 shows that KEWEn can be further improved by combining with the features related to the word order(i.e.,HCF+KEWE),which means that there exists the failure of word order capturing in KEWE.These two kinds of error sources re- veal the limitation of our method.In future work,the neural network accompanied with word embedding is a good alternative,which can produce better representation of documents. 5 Conclusion In this paper,we propose the knowledge-enriched word embedding(KEWE)for readability assessment. We extract the domain knowledge on word-level difficulty from three different perspectives and construct a knowledge graph.Based on the difficulty context derived from the knowledge graph,we develop two word embedding models(i.e.,KEWEk and KEWE).The experimental results on English and Chinese datasets demonstrate that KEWE can outperform other well-known readability assessment methods,and the classic text-based word embedding models.Future work is planned to involve extra datasets and additional word embedding strategies so that the soundness of KEWE can be further approved. Acknowledgements This work is supported by the National Key R&D Program of China under Grant No.2018YFB1003800; National Natural Science Foundation of China under Grant Nos.61373012.61321491.91218302:the Fundamental Research Funds for the Central Universities under Grant No.020214380049.This work is partially supported by the Collaborative Innovation Center of Novel Software Technology and Industri- alization. References Dimitrios Alikaniotis,Helen Yannakoudakis,and Marek Rei.2016.Automatic text scoring using neural networks In Proceedings of the 54th Annual Meeting of the Association for ComputationalLinguistics,pages 715-725. Mikhail Belkin and Partha Niyogi.2001.Laplacian eigenmaps and spectral techniques for embedding and clus- tering.Advances in Neural Information Processing Systems,14(6). Yoshua Bengio,Rejean Ducharme,Pascal Vincent,and Christian Jauvin.2003.A neural probabilistic language model.Journal of machine learning research,3(Feb):1137-1155. Marc Brysbaert and Andrew Biemiller.2016.Test-based age-of-acquisition norms for 44 thousand english word meanings.Behavior Research Methods,pages 1-4. Miriam Cha,Youngjune Gwon,and H.T.Kung.2017.Language modeling by clustering with word embeddings for text readability assessment.In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management,CIKM 2017,pages 2003-2006. Jeanne Sternlicht Chall.1995.Readability revisited:The new Dale-Chall readability formula,volume 118. Brookline Books Cambridge.MA. Kevyn Collins-Thompson and James P Callan.2004.A language modeling approach to predicting reading diffi- culty.In Proceedings of the 2004 Conference of the North American Chapter of the Association for Computa tional Linguistics:Human Language Technologies,pages 193-200. 375375 4.5 Error Analysis To better understand our method, we perform error analysis on the classification results. We mainly describe two sources of errors: word representation failure and word order capturing failure. From the perspective of word representation, the classification error is caused by the fact that the difficulty infor￾mation may not be properly encoded into the word embedding. Table 4 shows that KEWE performs better than the classical text-based word embeddings in encoding difficulty information into represen￾tation, but in the final results there still exists the failure of word representation in KEWE. From the perspective of word order capturing, the classification error is caused by the fact that the order among words may be neglected, so that the syntactic difficulty, pragmatic difficulty, and discourse difficulty of documents are ignored during the process of readability assessment. Table 3 shows that KEWEh can be further improved by combining with the features related to the word order (i.e., HCF+KEWEh), which means that there exists the failure of word order capturing in KEWE. These two kinds of error sources re￾veal the limitation of our method. In future work, the neural network accompanied with word embedding is a good alternative, which can produce better representation of documents. 5 Conclusion In this paper, we propose the knowledge-enriched word embedding (KEWE) for readability assessment. We extract the domain knowledge on word-level difficulty from three different perspectives and construct a knowledge graph. Based on the difficulty context derived from the knowledge graph, we develop two word embedding models (i.e., KEWEk and KEWEh). The experimental results on English and Chinese datasets demonstrate that KEWE can outperform other well-known readability assessment methods, and the classic text-based word embedding models. Future work is planned to involve extra datasets and additional word embedding strategies so that the soundness of KEWE can be further approved. Acknowledgements This work is supported by the National Key R&D Program of China under Grant No. 2018YFB1003800; National Natural Science Foundation of China under Grant Nos. 61373012, 61321491, 91218302; the Fundamental Research Funds for the Central Universities under Grant No. 020214380049. This work is partially supported by the Collaborative Innovation Center of Novel Software Technology and Industri￾alization. References Dimitrios Alikaniotis, Helen Yannakoudakis, and Marek Rei. 2016. Automatic text scoring using neural networks. In Proceedings of the 54th Annual Meeting of the Association for ComputationalLinguistics, pages 715–725. Mikhail Belkin and Partha Niyogi. 2001. Laplacian eigenmaps and spectral techniques for embedding and clus￾tering. Advances in Neural Information Processing Systems, 14(6). Yoshua Bengio, Rejean Ducharme, Pascal Vincent, and Christian Jauvin. 2003. A neural probabilistic language ´ model. Journal of machine learning research, 3(Feb):1137–1155. Marc Brysbaert and Andrew Biemiller. 2016. Test-based age-of-acquisition norms for 44 thousand english word meanings. Behavior Research Methods, pages 1–4. Miriam Cha, Youngjune Gwon, and H. T. Kung. 2017. Language modeling by clustering with word embeddings for text readability assessment. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, CIKM 2017, pages 2003–2006. Jeanne Sternlicht Chall. 1995. Readability revisited: The new Dale-Chall readability formula, volume 118. Brookline Books Cambridge, MA. Kevyn Collins-Thompson and James P Callan. 2004. A language modeling approach to predicting reading diffi- culty. In Proceedings of the 2004 Conference of the North American Chapter of the Association for Computa￾tional Linguistics: Human Language Technologies, pages 193–200