《分布式计算实验室》课程教学资源（阅读文献）Enriching Word Embeddings with Domain Knowledge for Readability Assessment

Enriching Word Embeddings with Domain Knowledge for Readability Assessment Zhiwei Jiang and Qing Gu*and Yafeng Yin and Daoxu Chen State Key Laboratory for Novel Software Technology, Nanjing University,Nanjing 210023,China jiangzhiwei@outlook.com,guq,yafeng,cdx@nju.edu.cn Abstract In this paper,we present a method which learns the word embedding for readability assessment. For the existing word embedding models,they typically focus on the syntactic or semantic rela- tions of words,while ignoring the reading difficulty,thus they may not be suitable for readability assessment.Hence,we provide the knowledge-enriched word embedding (KEWE),which en- codes the knowledge on reading difficulty into the representation of words.Specifically,we extract the knowledge on word-level difficulty from three perspectives to construct a knowledge graph,and develop two word embedding models to incorporate the difficulty context derived from the knowledge graph to define the loss functions.Experiments are designed to apply KEWE for readability assessment on both English and Chinese datasets,and the results demonstrate both effectiveness and potential of KEWE. 1 Introduction Readability assessment is a classic problem in natural language processing,which attracts many re- searchers'attention in recent years (Todirascu et al.,2016;Schumacher et al.,2016:Cha et al.,2017). The objective is to evaluate the readability of texts by levels or scores.The majority of recent readabil- ity assessment methods are based on the framework of supervised learning (Schwarm and Ostendorf, 2005)and build classifiers from hand-crafted features extracted from the texts.The performance of these methods depends on designing effective features to build high-quality classifiers. Designing hand-crafted features are essential but labor-intensive.It is desirable to learn representative features from the texts automatically.For document-level readability assessment,an effective feature learning method is to construct the representation of documents by combining the representation of the words contained (Kim,2014).For the representation of word,a useful technique is to learn the word representation as a dense and low-dimensional vector,which is called word embedding.Existing word embedding models(Collobert et al.,2011;Mikolov et al.,2013;Pennington et al.,2014)can be used for readability assessment,but the effectiveness is compromised by the fact that these models typically focus on the syntactic or semantic relations of words,while ignoring the reading difficulty.As a result,words with similar functions or topics,such as"man"and "gentleman",are mapped into close vectors although their reading difficulties are different.It calls for incorporating the knowledge on reading difficulty when training the word embedding. In this paper,we provide the knowledge-enriched word embedding (KEWE)for readability assess- ment,which encodes the knowledge on reading difficulty into the representation of words.Specifically, we define the word-level difficulty from three perspectives,and use the extracted knowledge to construct a knowledge graph.After that,we derive the difficulty context of words from the knowledge graph,and develop two word embedding models to incorporate the difficulty context to define the loss functions. We apply KEWE for document-level readability assessment under the supervised framework.The experiments are conducted on four datasets of either English or Chinese.The results demonstrate that *Corresponding Author This work is licensed under a Creative Commons Attribution 4.0 International Licence.Licence details: http://creativecommons.org/licenses/by/4.0/. 366 Proceedings of the 27th International Conference on Computational Linguistics,pages 366-378 Santa Fe,New Mexico,USA,August 20-26,2018

Proceedings of the 27th International Conference on Computational Linguistics, pages 366–378 Santa Fe, New Mexico, USA, August 20-26, 2018. 366 Enriching Word Embeddings with Domain Knowledge for Readability Assessment Zhiwei Jiang and Qing Gu∗ and Yafeng Yin and Daoxu Chen State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China jiangzhiwei@outlook.com, {guq,yafeng,cdx}@nju.edu.cn Abstract In this paper, we present a method which learns the word embedding for readability assessment. For the existing word embedding models, they typically focus on the syntactic or semantic rela￾tions of words, while ignoring the reading difficulty, thus they may not be suitable for readability assessment. Hence, we provide the knowledge-enriched word embedding (KEWE), which en￾codes the knowledge on reading difficulty into the representation of words. Specifically, we extract the knowledge on word-level difficulty from three perspectives to construct a knowledge graph, and develop two word embedding models to incorporate the difficulty context derived from the knowledge graph to define the loss functions. Experiments are designed to apply KEWE for readability assessment on both English and Chinese datasets, and the results demonstrate both effectiveness and potential of KEWE. 1 Introduction Readability assessment is a classic problem in natural language processing, which attracts many re￾searchers’ attention in recent years (Todirascu et al., 2016; Schumacher et al., 2016; Cha et al., 2017). The objective is to evaluate the readability of texts by levels or scores. The majority of recent readabil￾ity assessment methods are based on the framework of supervised learning (Schwarm and Ostendorf, 2005) and build classifiers from hand-crafted features extracted from the texts. The performance of these methods depends on designing effective features to build high-quality classifiers. Designing hand-crafted features are essential but labor-intensive. It is desirable to learn representative features from the texts automatically. For document-level readability assessment, an effective feature learning method is to construct the representation of documents by combining the representation of the words contained (Kim, 2014). For the representation of word, a useful technique is to learn the word representation as a dense and low-dimensional vector, which is called word embedding. Existing word embedding models (Collobert et al., 2011; Mikolov et al., 2013; Pennington et al., 2014) can be used for readability assessment, but the effectiveness is compromised by the fact that these models typically focus on the syntactic or semantic relations of words, while ignoring the reading difficulty. As a result, words with similar functions or topics, such as “man” and “gentleman”, are mapped into close vectors although their reading difficulties are different. It calls for incorporating the knowledge on reading difficulty when training the word embedding. In this paper, we provide the knowledge-enriched word embedding (KEWE) for readability assess￾ment, which encodes the knowledge on reading difficulty into the representation of words. Specifically, we define the word-level difficulty from three perspectives, and use the extracted knowledge to construct a knowledge graph. After that, we derive the difficulty context of words from the knowledge graph, and develop two word embedding models to incorporate the difficulty context to define the loss functions. We apply KEWE for document-level readability assessment under the supervised framework. The experiments are conducted on four datasets of either English or Chinese. The results demonstrate that ∗Corresponding Author This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http://creativecommons.org/licenses/by/4.0/

our method can outperform other well-known readability assessment methods,and the classic text-based word embedding models on all the datasets.By concatenating our knowledge-enriched word embedding with the hand-crafted features,the performance can be further improved. The rest of the paper is organized as follows.Section 2 provides the related work for readability assessment.Section 3 describes the details of KEWE.Section 4 presents the experiments and results. Finally Section 5 concludes the paper with future work. 2 Related Work In this section,we briefly introduce three research topics relevant to our work:readability assessment, word embedding,and graph embedding. Readability Assessment.The researches on readability assessment have a relatively long history from the beginning of last century(Collinsthompson,2014).Early studies mainly focused on designing readability formulas to evaluate the reading scores of texts.Some of the well-known readability formulas include the SMOG formula(McLaughlin,1969),the FK formula (Kincaid et al.,1975),and the Dale- Chall formula(Chall,1995).At the beginning of the 21th century,supervised approaches have been introduced and then explored for readability assessment(Si and Callan,2001;Collins-Thompson and Callan,2004;Schwarm and Ostendorf,2005).Researchers have focused on improving the performance by designing highly effective features (Pitler and Nenkova,2008;Heilman et al.,2008;Feng et al.,2010; Vajjala and Meurers,2012)and employing effective classification models (Heilman et al.,2007;Kate et al.,2010;Ma et al.,2012;Jiang et al.,2015;Cha et al.,2017).While most studies are conducted for English,there are studies for other languages,such as French(Francois and Fairon,2012),German (Hancke et al.,2012),Bangla (Sinha et al.,2014),Basque(Gonzalez-Dios et al.,2014),Chinese (Jiang et al.,2014),and Japanese (Wang and Andersen,2016). Word Embedding.Researchers have proposed various methods on word embedding,which mainly include two broad categories:neural network based methods(Bengio et al.,2003;Collobert et al.,2011; Mikolov et al.,2013)and co-occurrence matrix based methods(Turney and Pantel,2010;Levy and Gold- berg,2014b;Pennington et al.,2014).Neural network based methods learn word embedding through training neural network models,which include NNLM(Bengio et al.,2003),C&W(Collobert and We- ston,2008).and word2vec (Mikolov et al..2013).Co-occurrence matrix based methods learn word embedding based on the co-occurrence matrices,which include LSA (Deerwester,1990),Implicit Ma- trix Factorization (Levy and Goldberg,2014b),and GloVe (Pennington et al.,2014).Besides the general word embedding learning methods,researchers have also proposed methods to learn word embedding to include certain properties (Liu et al.,2015;Shen and Liu,2016)or for certain domains (Tang et al., 2014:Ren et al..2016:Alikaniotis et al..2016:Wu et al.,2017). Graph embedding.Graph embedding aims to learn continuous representations of the nodes or edges based on the structure of a graph.The graph embedding methods can be classified into three categories (Goyal and Ferrara,2017):factorization based (Roweis and Saul,2000;Belkin and Niyogi,2001), random walk based (Perozzi et al.,2014;Grover and Leskovec,2016),and deep learning based(Wang et al.,2016).Among them,the random walk based methods are easy to comprehend and can effectively reserve the centrality and similarity of the nodes.Deepwalks(Perozzi et al.,2014)and node2vec(Grover and Leskovec,2016)are two representatives of the random walk based methods.The basic idea of Deepwalk is viewing random walk paths as sentences,and feeding them to a general word embedding model.node2vec is similar to Deepwalk,although it simulates a biased random walk over graphs,and often provides efficient random walk paths 3 Learning Knowledge-Enriched Word Embedding for Readability Assessment In this section,we present the details of Knowledge-Enriched Word Embedding(KEWE)for readability assessment.By incorporating the word-level readability knowledge,we extend the existing word embed- ding model and design two models with different learning structures.As shown in Figure 1,the above one is the knowledge-only word embedding model(KEWE)which only takes in the domain knowledge, 367

367 our method can outperform other well-known readability assessment methods, and the classic text-based word embedding models on all the datasets. By concatenating our knowledge-enriched word embedding with the hand-crafted features, the performance can be further improved. The rest of the paper is organized as follows. Section 2 provides the related work for readability assessment. Section 3 describes the details of KEWE. Section 4 presents the experiments and results. Finally Section 5 concludes the paper with future work. 2 Related Work In this section, we briefly introduce three research topics relevant to our work: readability assessment, word embedding, and graph embedding. Readability Assessment. The researches on readability assessment have a relatively long history from the beginning of last century (Collinsthompson, 2014). Early studies mainly focused on designing readability formulas to evaluate the reading scores of texts. Some of the well-known readability formulas include the SMOG formula (McLaughlin, 1969), the FK formula (Kincaid et al., 1975), and the Dale￾Chall formula (Chall, 1995). At the beginning of the 21th century, supervised approaches have been introduced and then explored for readability assessment (Si and Callan, 2001; Collins-Thompson and Callan, 2004; Schwarm and Ostendorf, 2005). Researchers have focused on improving the performance by designing highly effective features (Pitler and Nenkova, 2008; Heilman et al., 2008; Feng et al., 2010; Vajjala and Meurers, 2012) and employing effective classification models (Heilman et al., 2007; Kate et al., 2010; Ma et al., 2012; Jiang et al., 2015; Cha et al., 2017). While most studies are conducted for English, there are studies for other languages, such as French (Franc¸ois and Fairon, 2012), German (Hancke et al., 2012), Bangla (Sinha et al., 2014), Basque (Gonzalez-Dios et al., 2014), Chinese (Jiang et al., 2014), and Japanese (Wang and Andersen, 2016). Word Embedding. Researchers have proposed various methods on word embedding, which mainly include two broad categories: neural network based methods (Bengio et al., 2003; Collobert et al., 2011; Mikolov et al., 2013) and co-occurrence matrix based methods (Turney and Pantel, 2010; Levy and Gold￾berg, 2014b; Pennington et al., 2014). Neural network based methods learn word embedding through training neural network models, which include NNLM (Bengio et al., 2003), C&W (Collobert and We￾ston, 2008), and word2vec (Mikolov et al., 2013). Co-occurrence matrix based methods learn word embedding based on the co-occurrence matrices, which include LSA (Deerwester, 1990), Implicit Ma￾trix Factorization (Levy and Goldberg, 2014b), and GloVe (Pennington et al., 2014). Besides the general word embedding learning methods, researchers have also proposed methods to learn word embedding to include certain properties (Liu et al., 2015; Shen and Liu, 2016) or for certain domains (Tang et al., 2014; Ren et al., 2016; Alikaniotis et al., 2016; Wu et al., 2017). Graph embedding. Graph embedding aims to learn continuous representations of the nodes or edges based on the structure of a graph. The graph embedding methods can be classified into three categories (Goyal and Ferrara, 2017): factorization based (Roweis and Saul, 2000; Belkin and Niyogi, 2001), random walk based (Perozzi et al., 2014; Grover and Leskovec, 2016), and deep learning based (Wang et al., 2016). Among them, the random walk based methods are easy to comprehend and can effectively reserve the centrality and similarity of the nodes. Deepwalks (Perozzi et al., 2014) and node2vec (Grover and Leskovec, 2016) are two representatives of the random walk based methods. The basic idea of Deepwalk is viewing random walk paths as sentences, and feeding them to a general word embedding model. node2vec is similar to Deepwalk, although it simulates a biased random walk over graphs, and often provides efficient random walk paths. 3 Learning Knowledge-Enriched Word Embedding for Readability Assessment In this section, we present the details of Knowledge-Enriched Word Embedding (KEWE) for readability assessment. By incorporating the word-level readability knowledge, we extend the existing word embed￾ding model and design two models with different learning structures. As shown in Figure 1, the above one is the knowledge-only word embedding model (KEWEk) which only takes in the domain knowledge

Knowledge Graph Random Wall mple positive contexts on Sim_edge by sliding window Knowledge Difficulty KEWE Base Context Sim_edge Sampling negative contexts on Dissim_edge Dissim_edge...... Sentences Sliding window Hybrid two types Text on sentences Text of contexts Corpus Context KEWE Figure 1:Illustration of the knowledge-enriched word embedding models.KEWE is based on the difficulty context,while KEWEh is based on both the difficulty and text contexts. the other is the hybrid word embedding model(KEWE)which compensates the domain knowledge with text corpus 3.1 The Knowledge-only Word Embedding Model (KEWE) In the classic word embedding models,such as C&W,CBOW,and Skip-Gram,the context of a word is represented by its surrounding words in the text corpus.Levy and Goldberg (2014a)have incorporated the syntactic context from the dependency parse-trees and found that the trained word embedding could capture more functional and less topical similarity.For readability assessment,reading difficulty other than function or topic becomes more important.Hence,we introduce a kind of difficulty context,and try to learn a difficulty-focusing word embedding,which leads to KEWEk.In the following,we describe this model in three steps:domain knowledge extraction,knowledge graph construction,and graph-based word embedding learning.The former two steps focus on modeling the relationship among words on reading difficulty,and the final step on deriving the difficulty context and learning the word embedding 3.1.1 Domain Knowledge Extraction To model the relationship among words on reading difficulty,we first introduce how to extract the knowl- edge on word-level difficulty from different perspectives.Specifically,we consider three types of word- level difficulty:acquisition difficulty,usage difficulty,and structure difficulty. Acquisition difficulty.Word acquisition refers to the temporal stage at which children learn the meaning of new words.Researchers have shown that the information on word acquisition is useful for readability assessment (Kidwell et al.,2009;Schumacher et al.,2016).Generally,the words acquired at primary school are easier than those acquired at high school.We call the reading difficulty reflected by word acquisition as the acquisition difficulty.Formally,given a word w,its acquisition difficulty is described by a distribution K over the age-of-acquisition (AoA)(Kidwell et al.,2009). Since the rating on AoA is an unsolved problem in cognitive science(Brysbaert and Biemiller,2016) and not available for many languages,we explore extra materials to describe the acquisition difficulty.In particular,we collect three kinds of knowledge teaching materials,i.e.,in-class teaching material,extra- curricular teaching material,and proficiency test material.These materials are arranged as lists of words, each of which contains words learned in the same time period and hence corresponds to a certain level of acquisition difficulty.For example,given a lists of words,we can defineKR,whereK1 if a word w belongs to the list i,andK=0otherwise. Usage difficulty.Researchers used to count the usage frequency to measure the difficulty of words (Dale and Chall,1948),which can separate the words which are frequently used from those rarely used. We call the difficulty reflected by usage preference as the usage difficulty.Formally,given a word w,its usage difficulty is described by a distribution K over the usage preferences. We provide two ways to measure the usage difficulty.One way is estimating the level of words' 368

368 KEWEk Knowledge Base Sentences gggggg Text Corpus KEWEh Knowledge Graph Dissim_edge Sim_edge Difficulty Context Sampling negative contexts on Dissim_edge Sample positive contexts by sliding window Paths ggg Random Walk on Sim_edge Text Context Sliding window on sentences Hybrid two types of contexts Figure 1: Illustration of the knowledge-enriched word embedding models. KEWEk is based on the difficulty context, while KEWEh is based on both the difficulty and text contexts. the other is the hybrid word embedding model (KEWEh) which compensates the domain knowledge with text corpus. 3.1 The Knowledge-only Word Embedding Model (KEWEk) In the classic word embedding models, such as C&W, CBOW, and Skip-Gram, the context of a word is represented by its surrounding words in the text corpus. Levy and Goldberg (2014a) have incorporated the syntactic context from the dependency parse-trees and found that the trained word embedding could capture more functional and less topical similarity. For readability assessment, reading difficulty other than function or topic becomes more important. Hence, we introduce a kind of difficulty context, and try to learn a difficulty-focusing word embedding, which leads to KEWEk. In the following, we describe this model in three steps: domain knowledge extraction, knowledge graph construction, and graph-based word embedding learning. The former two steps focus on modeling the relationship among words on reading difficulty, and the final step on deriving the difficulty context and learning the word embedding. 3.1.1 Domain Knowledge Extraction To model the relationship among words on reading difficulty, we first introduce how to extract the knowl￾edge on word-level difficulty from different perspectives. Specifically, we consider three types of word￾level difficulty: acquisition difficulty, usage difficulty, and structure difficulty. Acquisition difficulty. Word acquisition refers to the temporal stage at which children learn the meaning of new words. Researchers have shown that the information on word acquisition is useful for readability assessment (Kidwell et al., 2009; Schumacher et al., 2016). Generally, the words acquired at primary school are easier than those acquired at high school. We call the reading difficulty reflected by word acquisition as the acquisition difficulty. Formally, given a word w, its acquisition difficulty is described by a distribution KA w over the age-of-acquisition (AoA) (Kidwell et al., 2009). Since the rating on AoA is an unsolved problem in cognitive science (Brysbaert and Biemiller, 2016) and not available for many languages, we explore extra materials to describe the acquisition difficulty. In particular, we collect three kinds of knowledge teaching materials, i.e., in-class teaching material, extra￾curricular teaching material, and proficiency test material. These materials are arranged as lists of words, each of which contains words learned in the same time period and hence corresponds to a certain level of acquisition difficulty. For example, given a lists of words, we can define KA w ∈ R a , where KA w,i = 1 if a word w belongs to the list i, and KA w,i = 0 otherwise. Usage difficulty. Researchers used to count the usage frequency to measure the difficulty of words (Dale and Chall, 1948), which can separate the words which are frequently used from those rarely used. We call the difficulty reflected by usage preference as the usage difficulty. Formally, given a word w, its usage difficulty is described by a distribution KU w over the usage preferences. We provide two ways to measure the usage difficulty. One way is estimating the level of words’

usage frequency by counting the word frequency lists from the text corpus.The other way is estimating the probability distribution of words over the sentence-level difficulties,which is motivated by Jiang et al.(2015).Usage difficulty is defined on both.By discretizing the range of word frequency into b intervals of equal size,the usage frequency level of a word w is i,if its frequency resides in the ith intervals.By estimating the probability distribution vector P from sentence-level difficulties,we can define K andKi,Pl Structure difficulty.When building readability formulas,researchers have found that the structure of words could imply its difficulty (Flesch,1948;Gunning,1952;McLaughlin,1969).For example, words with more syllables are usually more difficult than words with less syllables.We call the difficulty reflected by structure of words as the structure difficulty.Formally,given a word w,its structure difficulty can be described by a distributionK over the word structures. Words in different languages may have their own special structural characteristics.For example,in English,the structural characteristics of words relate to syllables,characters,affixes,and subwords. Whereas in Chinese,the structural characteristics of words relate to strokes and radicals of Chinese characters.Here we use the number of syllables(strokes for Chinese)and characters in a word w to describe its structure difficulty.By discretizing the range of each number into intervals,Kis obtained by counting the interval in which w resides,respectively. 3.1.2 Knowledge Graph Construction After extracting the domain knowledge on word-level difficulty,we then quantitatively represent the knowledge by a graph.We define the knowledge graph as an undirected graph G=(V,E),where V is the set of vertices,each of which represents a word,and E is the set of edges,each of which represents the relation (i.e.,similarity)between two words on difficulty.Each edge e EE is a vertex pair (wi,wj) and is associated with a weight zij,which indicates the strength of the relation.If no edge exists between wi and wj,the weight zij =0.We define two edge types in the graph:Sim-edge and Dissim-edge.The former indicates that its end words have similar difficulty and is associated with a positive weight.The latter indicates that its end words have significant different difficulty and is associated with a negative weight.We derived the edges from the similarities computed between pairs of the words'knowledge vectors.Formally,given the extracted knowledge vector K=[K of a word w,E can be constructed using the similarity between pairs of words (w;,w;)as follows: sim(Kw:,Kw;) wj∈Np(w) sim(K:,Kw） wj∈Wn(u) (1) 0 otherwise where sim()is a similarity function (e.g.,cosine similarity),Np(wi)refers to the set of k most similar (i.e.,greatest similarity)neighbors of wi,and An(wi)refers to the set of k most dissimilar (i.e.,least similarity)neighbors of wi. 3.1.3 Knowledge Graph-based Word Embedding After constructing the knowledge graph,which models the relationship among words on difficulty,we can derive the difficulty context from the graph and train the word embedding focused on reading diffi- culty.For the graph-based difficulty context,given a word w,we define its difficulty context as the set of other words that have relevance to w on difficulty.Specifically,we define two types of difficulty context, positive context and negative context,corresponding to the two types of edges in the knowledge graph (i.e.,Sim_edge and Dissim-edge). Unlike the context defined on texts,which can be sampled by sliding windows over consecutive words, the context defined on a graph requires special sampling strategies.Different sampling strategies may define the context differently.For difficulty context,we design two relatively intuitive strategies,the ran- dom walk strategy and the immediate neighbors strategy,for the sampling of either positive or negative context. From the type Sim_edge,we sample the positive target-context pairs where the target word and the context words are similar on difficulty.Since the similarity is generally transitive,we adopt the random 369

369 usage frequency by counting the word frequency lists from the text corpus. The other way is estimating the probability distribution of words over the sentence-level difficulties, which is motivated by Jiang et al. (2015). Usage difficulty is defined on both. By discretizing the range of word frequency into b intervals of equal size, the usage frequency level of a word w is i, if its frequency resides in the ith intervals. By estimating the probability distribution vector Pw from sentence-level difficulties, we can define KU w ∈ R 1+|Pw| , and KU w,i = [i, Pw]. Structure difficulty. When building readability formulas, researchers have found that the structure of words could imply its difficulty (Flesch, 1948; Gunning, 1952; McLaughlin, 1969). For example, words with more syllables are usually more difficult than words with less syllables. We call the difficulty reflected by structure of words as the structure difficulty. Formally, given a word w, its structure difficulty can be described by a distribution KS w over the word structures. Words in different languages may have their own special structural characteristics. For example, in English, the structural characteristics of words relate to syllables, characters, affixes, and subwords. Whereas in Chinese, the structural characteristics of words relate to strokes and radicals of Chinese characters. Here we use the number of syllables (strokes for Chinese) and characters in a word w to describe its structure difficulty. By discretizing the range of each number into intervals, KS w is obtained by counting the interval in which w resides, respectively. 3.1.2 Knowledge Graph Construction After extracting the domain knowledge on word-level difficulty, we then quantitatively represent the knowledge by a graph. We define the knowledge graph as an undirected graph G = (V, E), where V is the set of vertices, each of which represents a word, and E is the set of edges, each of which represents the relation (i.e., similarity) between two words on difficulty. Each edge e ∈ E is a vertex pair (wi , wj ) and is associated with a weight zij , which indicates the strength of the relation. If no edge exists between wi and wj , the weight zij = 0. We define two edge types in the graph: Sim edge and Dissim edge. The former indicates that its end words have similar difficulty and is associated with a positive weight. The latter indicates that its end words have significant different difficulty and is associated with a negative weight. We derived the edges from the similarities computed between pairs of the words’ knowledge vectors. Formally, given the extracted knowledge vector Kw = [KA w , KU w , KS w] of a word w, E can be constructed using the similarity between pairs of words (wi , wj ) as follows: zij =    sim(Kwi , Kwj ) wj ∈ Np(wi) −sim(Kwi , Kwj ) wj ∈ Nn(wi) 0 otherwise (1) where sim() is a similarity function (e.g., cosine similarity), Np(wi) refers to the set of k most similar (i.e., greatest similarity) neighbors of wi , and Nn(wi) refers to the set of k most dissimilar (i.e., least similarity) neighbors of wi . 3.1.3 Knowledge Graph-based Word Embedding After constructing the knowledge graph, which models the relationship among words on difficulty, we can derive the difficulty context from the graph and train the word embedding focused on reading diffi- culty. For the graph-based difficulty context, given a word w, we define its difficulty context as the set of other words that have relevance to w on difficulty. Specifically, we define two types of difficulty context, positive context and negative context, corresponding to the two types of edges in the knowledge graph (i.e., Sim edge and Dissim edge). Unlike the context defined on texts, which can be sampled by sliding windows over consecutive words, the context defined on a graph requires special sampling strategies. Different sampling strategies may define the context differently. For difficulty context, we design two relatively intuitive strategies, the ran￾dom walk strategy and the immediate neighbors strategy, for the sampling of either positive or negative context. From the type Sim edge, we sample the positive target-context pairs where the target word and the context words are similar on difficulty. Since the similarity is generally transitive, we adopt the random

walk strategy to sample the positive context.Following the idea of node2vec (Grover and Leskovec, 2016),we sample the positive contexts of words by simulating a 2d order random walk on the knowledge graph with only Sim_edge.After that,by applying a sliding window of fixed length s over the sampled random walk paths,we can get the positive target-context pairs {(wt,we). From the type Dissim_edge,we sample the negative target-context pairs where the target word and the context words are dissimilar on difficulty.Since dissimilarity is generally not transitive,we adopt the immediate neighbor strategy to sample the negative context.Specifically,on the knowledge graph with only Dissim_edge,we collect the negative context from the immediate neighbors of the target node wt and get the negative context list Cn(wt). By replacing the text-based linear context with our graph-based difficulty context,we can train the word embedding using the classic word embedding models,such as C&W,CBOW,and Skip-Gram. Here we use the Skip-Gram model with Negative Sampling (SGNS)proposed by Mikolov et al.(2013). Specifically,given N positive target-context pairs (wt,we)and the negative context list of the target word Cn(wt),the objective of KEWEk is to minimize the loss function Ck,which is defined as follows: 4=∑ loga(uweVw)+EwECn()loga(-uw,Vw) (2) （t,e） where v and u are the"input"and"output"vector representation of w,and o is the sigmoid function defined as a()=.This loss function enables the positive context (e.g.,wc)to be distinguished from the negative context (e.g.,wi). 3.2 The Hybrid Word Embedding Model(KEWE) The classic text-based word embedding models yield word embedding focusing on syntactic and seman- tic contexts,while ignoring the word difficulty.By contrast,KEWEk trains the word embedding focusing on the word difficulty,while leaving out the syntactic and semantic information.Since readability may also relate to both syntax and semantics,we develop a hybrid word embedding model (KEWE),to incorporate both domain knowledge and text corpus.The loss function of the hybrid model Ch can be expressed as follows: Ch=λCk+(1-入)Ct (3) where Ck is the loss of predicting the knowledge graph-based difficulty contexts,Ct is the loss of pre- dicting the text-based syntactic and semantic contexts,and A [0,1]is a weighting factor.Clearly,the case of A =1 reduces the hybrid model to the knowledge-only model. As there are many text-based word embedding models,the text-based loss Ct can be defined in various ways.To be consistent with KEWEk,we formalize Ct based on the Skip-Gram model.Given a text corpus,the Skip-Gram model aims to find word representations that are good at predicting the context words.Specifically,given a sequence of training words,denoted as w1,w2,...,wr,the objective of Skip-Gram model is to minimize the log loss of predicting the context using target word embedding, which can be expressed as follows: T Ct=- 1 logp(wt+ilwt) (4) t=1-s≤j≤sj≠0 where s is the window size of the context sampling.Since the full softmax function used to define p(wwt)is computationally expensive,we employ the negative sampling strategy (Mikolov et al., 2013)and replace every log p(welwt)in Ct by the following formula: k logp(welwt)=loga(uwev)+>EwP.(w)loga(-uw,Vw) (5) =1 where v,u,and o are of the same meanings as in Eq.2,k is the number of negative samples,and P(w)is the noise distribution.This strategy enables the actual context we to be distinguished from the noise context wi drawn from the noise distribution Pn(w). 370

370 walk strategy to sample the positive context. Following the idea of node2vec (Grover and Leskovec, 2016), we sample the positive contexts of words by simulating a 2 nd order random walk on the knowledge graph with only Sim edge. After that, by applying a sliding window of fixed length s over the sampled random walk paths, we can get the positive target-context pairs {(wt , wc)}. From the type Dissim edge, we sample the negative target-context pairs where the target word and the context words are dissimilar on difficulty. Since dissimilarity is generally not transitive, we adopt the immediate neighbor strategy to sample the negative context. Specifically, on the knowledge graph with only Dissim edge, we collect the negative context from the immediate neighbors of the target node wt and get the negative context list Cn(wt). By replacing the text-based linear context with our graph-based difficulty context, we can train the word embedding using the classic word embedding models, such as C&W, CBOW, and Skip-Gram. Here we use the Skip-Gram model with Negative Sampling (SGNS) proposed by Mikolov et al. (2013). Specifically, given N positive target-context pairs (wt , wc) and the negative context list of the target word Cn(wt), the objective of KEWEk is to minimize the loss function Lk, which is defined as follows: Lk = − 1 N X (wt,wc) h log σ(u > wc vwt ) + Ewi∈Cn(wt) log σ(−u > wi vwt ) i (2) where vw and uw are the “input” and “output” vector representation of w, and σ is the sigmoid function defined as σ(x) = 1 (1+e−x) . This loss function enables the positive context (e.g., wc) to be distinguished from the negative context (e.g., wi). 3.2 The Hybrid Word Embedding Model (KEWEh) The classic text-based word embedding models yield word embedding focusing on syntactic and seman￾tic contexts, while ignoring the word difficulty. By contrast, KEWEk trains the word embedding focusing on the word difficulty, while leaving out the syntactic and semantic information. Since readability may also relate to both syntax and semantics, we develop a hybrid word embedding model (KEWEh), to incorporate both domain knowledge and text corpus. The loss function of the hybrid model Lh can be expressed as follows: Lh = λLk + (1 − λ)Lt (3) where Lk is the loss of predicting the knowledge graph-based difficulty contexts, Lt is the loss of pre￾dicting the text-based syntactic and semantic contexts, and λ ∈ [0, 1] is a weighting factor. Clearly, the case of λ = 1 reduces the hybrid model to the knowledge-only model. As there are many text-based word embedding models, the text-based loss Lt can be defined in various ways. To be consistent with KEWEk, we formalize Lt based on the Skip-Gram model. Given a text corpus, the Skip-Gram model aims to find word representations that are good at predicting the context words. Specifically, given a sequence of training words, denoted as w1, w2, · · · , wT , the objective of Skip-Gram model is to minimize the log loss of predicting the context using target word embedding, which can be expressed as follows: Lt = − 1 T X T t=1 X −s≤j≤s,j6=0 log p(wt+j |wt) (4) where s is the window size of the context sampling. Since the full softmax function used to define p(wt+j |wt) is computationally expensive, we employ the negative sampling strategy (Mikolov et al., 2013) and replace every log p(wc|wt) in Lt by the following formula: log p(wc|wt) = log σ(u > wc vwt ) +X k i=1 Ewi∼Pn(w) log σ(−u > wi vwt ) (5) where vw, uw, and σ are of the same meanings as in Eq. 2, k is the number of negative samples, and Pn(w) is the noise distribution. This strategy enables the actual context wc to be distinguished from the noise context wi drawn from the noise distribution Pn(w)

3.3 Model Training We adopt the stochastic gradient descent to train our models.Specifically,for the hybrid model (KEWE),we adopt the mini-batch mode used in (Yang et al.,2016).Firstly,we sample a batch of random walk paths of size N and take a gradient step to optimize the loss function C.Secondly,we sample a batch of text sentences of size N2 and take a gradient step to optimize the loss function Ct.We repeat the above procedures until the model converged or the predefined number of iterations is reached. The ratio is used to approximate the weighting factorin E3.For training the knowledge-only model,the process is the same without Ct and A. 3.4 Readability Assessment We apply KEWE for document-level readability assessment under a supervised learning framework pro- posed by Schwarm and Ostendorf(2005).The classifier for readability assessment is built on documents with annotated reading levels.Instead of using the hand-crafted features,we use the word embedding to produce the features of documents. To extract high level features from documents using word embedding,we design the max layer sim- ilar to the one used in the convolutional sentence approach proposed by Collobert et al.(Collobert et al.,2011).The max layer is used to generate a fix-size feature vector from variant length sequences. Specifically,given a document represented by a matrix M E Rmx",where the kth column is the word embedding of the kth word in the document,the max layer output a vector fmax(M): [fmax(M)li=max[M]i,t1≤i≤m (6) where t ={1,2,...,n}represents"time"of a sequence,and m is the dimension of embedding vectors. Besides the max layer,the min and average layers are also used to extract features.By concatenating all three feature vectors,we get the final feature vector f(M)of the document M as follows,which can be fed to the classifier for readability assessment. f(M)=[fmaz(M),fmin(M),favg(M)] (7) 4 Experiments In this section,we conduct experiments based on four datasets of two languages,to investigate the following two research questions: RQ1:Whether KEWE is effective for readability assessment,compared with other well-known read- ability assessment methods,and other word embedding models? RQ2:What are the effects of the quality of input (i.e.,the quality of the knowledge base and text corpus)and the hybrid ratio(i.e.,the weighting factor A)on the prediction performance of KEWE? 4.1 The Datasets and Knowledge Base The experiments are conducted on four datasets,including two English datasets:ENCT and EHCT, two Chinese datasets:CPT and CPC.ENCT,CPT (Jiang et al.,2015),and EHCT are extracted from textbooks,where the documents have already been leveled into grades;CPC is extracted from the students'compositions2 written by Chinese primary school students,where the documents are leveled by the six grades of authors.EHCT is collected from the English textbooks of Chinese high schools and colleges,which contains 4 levels corresponding to the 3 grades of high school plus undergraduates.The details of the four datasets are shown in Table 1. Since the experiments are conducted on two languages,we collect the knowledge bases for both En- glish and Chinese,which are used for extracting domain knowledge and constructing the knowledge graphs,as described in Section 3.1.The details are shown in Table 2. http://www.dzkbw.com http://www.eduxiao.com 371

371 3.3 Model Training We adopt the stochastic gradient descent to train our models. Specifically, for the hybrid model (KEWEh), we adopt the mini-batch mode used in (Yang et al., 2016). Firstly, we sample a batch of random walk paths of size N1 and take a gradient step to optimize the loss function Lk. Secondly, we sample a batch of text sentences of size N2 and take a gradient step to optimize the loss function Lt . We repeat the above procedures until the model converged or the predefined number of iterations is reached. The ratio N1 N1+N2 is used to approximate the weighting factor λ in Eq. 3. For training the knowledge-only model, the process is the same without Lt and λ. 3.4 Readability Assessment We apply KEWE for document-level readability assessment under a supervised learning framework pro￾posed by Schwarm and Ostendorf (2005). The classifier for readability assessment is built on documents with annotated reading levels. Instead of using the hand-crafted features, we use the word embedding to produce the features of documents. To extract high level features from documents using word embedding, we design the max layer sim￾ilar to the one used in the convolutional sentence approach proposed by Collobert et al. (Collobert et al., 2011). The max layer is used to generate a fix-size feature vector from variant length sequences. Specifically, given a document represented by a matrix M ∈ R m×n , where the kth column is the word embedding of the kth word in the document, the max layer output a vector fmax(M): [fmax(M)]i = max t [M]i,t 1 ≤ i ≤ m (6) where t = {1, 2, . . . , n} represents ”time” of a sequence, and m is the dimension of embedding vectors. Besides the max layer, the min and average layers are also used to extract features. By concatenating all three feature vectors, we get the final feature vector f(M) of the document M as follows, which can be fed to the classifier for readability assessment. f(M) = [fmax(M), fmin(M), favg(M)] (7) 4 Experiments In this section, we conduct experiments based on four datasets of two languages, to investigate the following two research questions: RQ1: Whether KEWE is effective for readability assessment, compared with other well-known read￾ability assessment methods, and other word embedding models? RQ2: What are the effects of the quality of input (i.e., the quality of the knowledge base and text corpus) and the hybrid ratio (i.e., the weighting factor λ) on the prediction performance of KEWE? 4.1 The Datasets and Knowledge Base The experiments are conducted on four datasets, including two English datasets: ENCT and EHCT, two Chinese datasets: CPT and CPC. ENCT, CPT (Jiang et al., 2015), and EHCT are extracted from textbooks 1 , where the documents have already been leveled into grades; CPC is extracted from the students’ compositions 2 written by Chinese primary school students, where the documents are leveled by the six grades of authors. EHCT is collected from the English textbooks of Chinese high schools and colleges, which contains 4 levels corresponding to the 3 grades of high school plus undergraduates. The details of the four datasets are shown in Table 1. Since the experiments are conducted on two languages, we collect the knowledge bases for both En￾glish and Chinese, which are used for extracting domain knowledge and constructing the knowledge graphs, as described in Section 3.1. The details are shown in Table 2. 1 http://www.dzkbw.com 2 http://www.eduxiao.com

Langauage Dataset #Level #doc #sent/doc #word/doc V #doc in reading levels ENCT 4 276 16.92 7747 English 266.62 [72,96.60,48 EHCT 252 40.82 646.70 10485 [67,67,58.60] Chinese CPT 6 637 25.47 448.16 19254 96,110,106.108.113,104] CPC 6 3 14.91 305.16 9459 [50.50.50.50.50.50] Table 1:Statistics of the Four Datasets Perspective Type of Material English Chinese AoA rating AoA norms for over 50,000 English Words In-class teaching New word list (primary,junior,high) New word list(primary,junior,high) Acqusition Extra-curricular teaching New Concept English vocabulary Overseas Chinese Language vocabulary Proficiency test CET(4,6),PETS(1,2,5)vocabulary HSK(1-6)vocabulary Frequency List Usage Word Frequency List of American English Chinese High Frequency Word List Sentence Corpus English Wikipedia Chinese Wikipedia Structure Dictionary Syllabary Stroke dictionary Table 2:Details of the collected knowledge bases 4.2 Experiment Settings For readability assessment,we design experiments on the four datasets using the hold-out validation, which randomly divides a dataset into training and test sets by stratified sampling.The test ratio is set as 0.3.To reduce randomness,under each case,100 rounds of hold-out validations are performed,and the average results are reported.To tune the hyper-parameters,we randomly choose three-tenths of the train- ing set as development set.We choose three widely used metrics (Jiang et al.,2014):Accuracy (Acc), Adjacent Accuracy (+Acc)and Pearson's Correlation Coefficient(PCC),to measure the performance on readability assessment. For the training of word embedding,we use all the sentences in the target dataset as the training corpus (denote as the internal corpus),to ensure sufficient word coverage.To avoid mixing the level information into the internal corpus,we shuffle the sentences in it before feeding to the model.For the hyper-parameters of text-based word embedding,we set the dimension as 300,the window size as 5,the number of negative samples as 5,and the iteration number as 5.For KEWE,the default length of random walk path is set as 80(Grover and Leskovec,2016).k and A are tuned using the development set. 4.3 Comparison with Other Methods To address RQ1,we firstly compare KEWE with the following readability assessment methods: SMOG(McLaughlin,1969)and FK(Kincaid et al.,1975)are two widely-used readability formulas. .SUM (Collins-Thompson and Callan,2004)is a smoothed unigram model. HCF is the hand-crafted feature based method.For English and Chinese,we employ the state-of- the-art feature set proposed by Vajjala and Meurers(2012)and Jiang et al.(2014)respectively. BOW refers to the bag-of-words model. For the feature-based methods(i.e.,HCF,BOW,and KEWE),we use both Logistic Regression(LR)and Random Forest(RF)as the classifiers for readability assessment.Table 3 lists the performance measures of all these methods on four datasets.The value marked in bold in each column refers to the maximum (best)measures acquired by the methods on each dataset by certain metric. From Table 3,we can see that the performances of readability formulas (SMOG and FK)are not good on all the datasets,except the adjacent accuracies on ENCT.The smoothed unigram model(SUM) outperforms SOMG and FK on all the datasets,and on EHCT,its accuracy is only slightly inferior to KEWEh.HCF performs the best among methods other than KEWE.Even compared with KEWE,it achieves the best performance on CPC and the best adjacent accuracy on ENCT.KEWE is only slightly inferior to HCF in four columns,but outperforms all the other methods in the other eight columns 372

372 Langauage Dataset #Level #doc #sent/doc #word/doc |V| #doc in reading levels English ENCT 4 276 16.92 266.62 7747 [72, 96, 60, 48] EHCT 4 252 40.82 646.70 10485 [67, 67, 58, 60] Chinese CPT 6 637 25.47 448.16 19254 [96, 110, 106, 108, 113, 104] CPC 6 300 14.91 305.16 9459 [50, 50, 50, 50, 50, 50] Table 1: Statistics of the Four Datasets Perspective Type of Material English Chinese Acqusition AoA rating AoA norms for over 50,000 English Words − In-class teaching New word list (primary, junior, high) New word list (primary, junior, high) Extra-curricular teaching New Concept English vocabulary Overseas Chinese Language vocabulary Proficiency test CET(4,6), PETS(1,2,5) vocabulary HSK(1-6) vocabulary Usage Frequency List Word Frequency List of American English Chinese High Frequency Word List Sentence Corpus English Wikipedia Chinese Wikipedia Structure Dictionary Syllabary Stroke dictionary Table 2: Details of the collected knowledge bases 4.2 Experiment Settings For readability assessment, we design experiments on the four datasets using the hold-out validation, which randomly divides a dataset into training and test sets by stratified sampling. The test ratio is set as 0.3. To reduce randomness, under each case, 100 rounds of hold-out validations are performed, and the average results are reported. To tune the hyper-parameters, we randomly choose three-tenths of the train￾ing set as development set. We choose three widely used metrics (Jiang et al., 2014): Accuracy (Acc), Adjacent Accuracy (±Acc) and Pearson’s Correlation Coefficient (PCC), to measure the performance on readability assessment. For the training of word embedding, we use all the sentences in the target dataset as the training corpus (denote as the internal corpus), to ensure sufficient word coverage. To avoid mixing the level information into the internal corpus, we shuffle the sentences in it before feeding to the model. For the hyper-parameters of text-based word embedding, we set the dimension as 300, the window size as 5, the number of negative samples as 5, and the iteration number as 5. For KEWE, the default length of random walk path is set as 80 (Grover and Leskovec, 2016). k and λ are tuned using the development set. 4.3 Comparison with Other Methods To address RQ1, we firstly compare KEWE with the following readability assessment methods: • SMOG (McLaughlin, 1969) and FK (Kincaid et al., 1975) are two widely-used readability formulas. • SUM (Collins-Thompson and Callan, 2004) is a smoothed unigram model. • HCF is the hand-crafted feature based method. For English and Chinese, we employ the state-of￾the-art feature set proposed by Vajjala and Meurers (2012) and Jiang et al. (2014) respectively. • BOW refers to the bag-of-words model. For the feature-based methods (i.e., HCF, BOW, and KEWE), we use both Logistic Regression (LR) and Random Forest (RF) as the classifiers for readability assessment . Table 3 lists the performance measures of all these methods on four datasets. The value marked in bold in each column refers to the maximum (best) measures acquired by the methods on each dataset by certain metric. From Table 3, we can see that the performances of readability formulas (SMOG and FK) are not good on all the datasets, except the adjacent accuracies on ENCT. The smoothed unigram model (SUM) outperforms SOMG and FK on all the datasets, and on EHCT, its accuracy is only slightly inferior to KEWEh. HCF performs the best among methods other than KEWE. Even compared with KEWE, it achieves the best performance on CPC and the best adjacent accuracy on ENCT. KEWE is only slightly inferior to HCF in four columns, but outperforms all the other methods in the other eight columns

Dataset Method ENCT EHCT CPT CPC Acc 土Acc PCC Acc 士AcC PCC Acc 土AcC PCC Acc 士Acc PCC SMOG 46.0494.130.744128.4784.05 0.551730.9570.210.606319.58 60.100.4076 FK 50.98 97.67 0.787025.17 76.16 0.341725.15 64.61 0.4912 16.93 50.49 0.0808 SUM 66.6497.190.834570.4191.410.725733.6172.04 0.586626.71 56.590.2734 LR87.2498.01 0.913653.6989.850.693743.7181.69 0.765227.08 61.54 0.4275 HCF RF90.961000.959260.3691.240.742147.9787.990.815935.09 72.490.5880 BOW LR81.5799.280.904965.8889.610.725734.8274.67 0.659329.01 56.710.3083 RF78.8994.710.829459.0389.960.738439.5680.730.748631.1362.740.5247 LR91.1299.720.954564.0391.430.774552.6987.350.809130.5961.320.5272 KEWEk RF92.3499.71 0.960669.1395.28 0.825152.9488.390.816730.73 66.140.5278 KEWE LR93.6799.580.965465.2593.400.812954.3388.030.823334.8165.110.5507 RF94.4899.720.970571.2096.120.844954.8388.940.828733.26 65.630.5111 LR87.3798.070.915353.9690.050.697645.9283.530.787627.3161.840.4300 HCF+KEWE RF95.8799.890.979470.0595.85 0.838060.2392.280.877635.5270.240.5801 Table 3:Performance comparison between KEWE and other readability assessment methods Dataset Model ENCT EHCT CPT CPC Acc 士AcC PCC Acc 土Acc PCC Acc 土Acc PCC Acc 土Acc PCC Random 31.73 75.57 0.0834 24.96 61.35 0.0786 16.96 45.96 0.0590 16.81 45.59 0.0743 NNLM 79.80 98.66 0.8843 58.43 88.36 0.704643.47 85.22 0.7806 30.68 65.07 0.5195 C&W 82.35 99.05 0.9039 59.05 90.06 0.716045.59 84.70 0.7874 30.60 64.83 0.5023 GloVe 82.66 99.52 0.9129 58.15 90.53 0.7302 43.63 85.80 0.7882 30.33 64.48 0.4933 CBOW 73.94 96.86 0.829962.40 91.29 0.751840.49 82.83 0.756427.40 59.48 0.3929 SG 82.2098.60 0.900162.7362.900.756344.3185.540.7869 31.9263.04 0.4527 KV 89.1098.710.929564.1590.850.748350.6384.840.780628.1961.90 0.4658 SG+KV 84.71 98.81 0.913666.00 92.73 0.7833 46.49 86.08 0.7921 32.03 63.34 0.4729 KEWEL92.3499.710.960669.1395.280.825152.9488.390.816730.7366.14 0.5278 KEWEh 94.48 99.72 0.970571.2096.12 0.844954.83 88.94 0.8287 33.26 65.63 0.5111 Table 4:Performance comparison between KEWE and other word embedding models Overall,KEWE is competitive for readability assessment.In addition,by combining KEWE with the hand-crafted feature set of HCF,the performance can be further improved in many columns. Secondly,we compare KEWE with other text-based word embedding models for readability assess- ment.These models include NNLM(Bengio et al.,2003),C&W(Collobert and Weston,2008),GloVe (Pennington et al.,2014),CBOW and Skip-Gram (SG)(Mikolov et al.,2013).Random embedding (Random)and knowledge vectors from the three perspectives(KV)are also used as baselines.All the text-based models mentioned are trained on the four datasets respectively.Table 4 lists the results of applying word embedding for readability assessment using Random Forest as the classifier.From Table 4,the model KEWE gets the best performance among all the embedding baselines,including SG+KV, which also takes in both knowledge and text corpus.The results demonstrate the superiority of hybridiz- ing the knowledge graph and text corpus to learn the representation of words. 4.4 Model Analysis To address RQ2,we analyze the effects of the knowledge graph,the text corpus,and the hybrid ratio on the performance of KEWE.Firstly,we study the impacts of the knowledge graph on the performance of KEWE.Specifically,we study the following three parameters of the knowledge graph:the knowledge vectors generated from the three word-level difficulties (i.e.,acquisition difficulty,usage difficulty,and structure difficulty),the similarity function (i.e.,sim()),and the number of neighbors of each node in the knowledge graph (i.e.,k).Figure 2 shows the performance measures of using each of the three 373

373 Method Dataset ENCT EHCT CPT CPC Acc ±Acc PCC Acc ±Acc PCC Acc ±Acc PCC Acc ±Acc PCC SMOG 46.04 94.13 0.7441 28.47 84.05 0.5517 30.95 70.21 0.6063 19.58 60.10 0.4076 FK 50.98 97.67 0.7870 25.17 76.16 0.3417 25.15 64.61 0.4912 16.93 50.49 0.0808 SUM 66.64 97.19 0.8345 70.41 91.41 0.7257 33.61 72.04 0.5866 26.71 56.59 0.2734 HCF LR 87.24 98.01 0.9136 53.69 89.85 0.6937 43.71 81.69 0.7652 27.08 61.54 0.4275 RF 90.96 100 0.9592 60.36 91.24 0.7421 47.97 87.99 0.8159 35.09 72.49 0.5880 BOW LR 81.57 99.28 0.9049 65.88 89.61 0.7257 34.82 74.67 0.6593 29.01 56.71 0.3083 RF 78.89 94.71 0.8294 59.03 89.96 0.7384 39.56 80.73 0.7486 31.13 62.74 0.5247 KEWEk LR 91.12 99.72 0.9545 64.03 91.43 0.7745 52.69 87.35 0.8091 30.59 61.32 0.5272 RF 92.34 99.71 0.9606 69.13 95.28 0.8251 52.94 88.39 0.8167 30.73 66.14 0.5278 KEWEh LR 93.67 99.58 0.9654 65.25 93.40 0.8129 54.33 88.03 0.8233 34.81 65.11 0.5507 RF 94.48 99.72 0.9705 71.20 96.12 0.8449 54.83 88.94 0.8287 33.26 65.63 0.5111 HCF+KEWEh LR 87.37 98.07 0.9153 53.96 90.05 0.6976 45.92 83.53 0.7876 27.31 61.84 0.4300 RF 95.87 99.89 0.9794 70.05 95.85 0.8380 60.23 92.28 0.8776 35.52 70.24 0.5801 Table 3: Performance comparison between KEWE and other readability assessment methods Model Dataset ENCT EHCT CPT CPC Acc ±Acc PCC Acc ±Acc PCC Acc ±Acc PCC Acc ±Acc PCC Random 31.73 75.57 0.0834 24.96 61.35 0.0786 16.96 45.96 0.0590 16.81 45.59 0.0743 NNLM 79.80 98.66 0.8843 58.43 88.36 0.7046 43.47 85.22 0.7806 30.68 65.07 0.5195 C&W 82.35 99.05 0.9039 59.05 90.06 0.7160 45.59 84.70 0.7874 30.60 64.83 0.5023 GloVe 82.66 99.52 0.9129 58.15 90.53 0.7302 43.63 85.80 0.7882 30.33 64.48 0.4933 CBOW 73.94 96.86 0.8299 62.40 91.29 0.7518 40.49 82.83 0.7564 27.40 59.48 0.3929 SG 82.20 98.60 0.9001 62.73 62.90 0.7563 44.31 85.54 0.7869 31.92 63.04 0.4527 KV 89.10 98.71 0.9295 64.15 90.85 0.7483 50.63 84.84 0.7806 28.19 61.90 0.4658 SG+KV 84.71 98.81 0.9136 66.00 92.73 0.7833 46.49 86.08 0.7921 32.03 63.34 0.4729 KEWEk 92.34 99.71 0.9606 69.13 95.28 0.8251 52.94 88.39 0.8167 30.73 66.14 0.5278 KEWEh 94.48 99.72 0.9705 71.20 96.12 0.8449 54.83 88.94 0.8287 33.26 65.63 0.5111 Table 4: Performance comparison between KEWE and other word embedding models Overall, KEWE is competitive for readability assessment. In addition, by combining KEWE with the hand-crafted feature set of HCF, the performance can be further improved in many columns. Secondly, we compare KEWE with other text-based word embedding models for readability assess￾ment. These models include NNLM (Bengio et al., 2003), C&W (Collobert and Weston, 2008), GloVe (Pennington et al., 2014), CBOW and Skip-Gram (SG) (Mikolov et al., 2013). Random embedding (Random) and knowledge vectors from the three perspectives (KV) are also used as baselines. All the text-based models mentioned are trained on the four datasets respectively. Table 4 lists the results of applying word embedding for readability assessment using Random Forest as the classifier. From Table 4, the model KEWEh gets the best performance among all the embedding baselines, including SG+KV, which also takes in both knowledge and text corpus. The results demonstrate the superiority of hybridiz￾ing the knowledge graph and text corpus to learn the representation of words. 4.4 Model Analysis To address RQ2, we analyze the effects of the knowledge graph, the text corpus, and the hybrid ratio on the performance of KEWE. Firstly, we study the impacts of the knowledge graph on the performance of KEWEk. Specifically, we study the following three parameters of the knowledge graph: the knowledge vectors generated from the three word-level difficulties (i.e., acquisition difficulty, usage difficulty, and structure difficulty), the similarity function (i.e., sim()), and the number of neighbors of each node in the knowledge graph (i.e., k). Figure 2 shows the performance measures of using each of the three

knowledge vectors respectively in bar chart.It can be found that the acquisition difficulty outperforms the other two on all four datasets.By combining the three knowledge vectors,the performance can be further improved.Figure 3 shows the performance measures of using different similarity functions and different values of k in line charts.It can be found that the knowledge graphs achieve a relatively good performance,when the cosine similarity is used and k is close to 50 or 100. a Figure 3:The performance of KEWEk using knowledge graphs with dif- ferent similarity functions and neighbor numbers Figure 2:The perfor- mance of KEWEk using different types of knowl- edge vectors Figure 4:The performance of KEWEh using different volumes of exter- nal corpus Secondly,we study the impacts of using external corpus on the performance of KEWE.To learn text- based word embedding for both languages,we collect the external corpus from both English Wikipedia (Ewiki)and Chinese Wikipedia(Cwiki).After preprocessing,Ewiki contains 7M sentences and 172M words,and Cwiki contains 7.4M sentences and 160M words.We sample different number of sentences from the corpus for training word embedding,and then measure the performance of KEWEh on read- ability assessment.Figure 4 shows the performance measures by using different volumes of external corpus in line charts,with Skip-Gram trained for comparison.Both models are also trained using the internal corpus,and the performance measures(dotted lines)are depicted as baselines.From Figure 4,it can be found that the performance of both Skip-Gram and KEWEh increases as the volume of external corpus increases,and keeps stable when the corpus is large enough.Besides,on English datasets,word embedding trained using external corpus achieves comparable performance with that using internal cor- pus.The above suggests that external corpus is a good substitution for the internal corpus,especially on English datasets,and a relative large volume of corpus is required to achieve stable performance. Finally,we study the impacts of the weighting factor A on the performance of KEWEh.Since A is approximated by Ni/(N1+N2)in our model,we vary X from 0 to 1 by setting Ni+N2=100 and then varying Ni from 0 to 100 stepping by 10.Figure 5 shows the performance of KEWEn with varied A in line charts with error bars,where A=0 and 1 correspond to Skip-Gram and KEWEk respectively.From Figure 5,it can be found that KEWEh can outperform its base models(i.e.,KEWEk and Skip-Gram),by setting A with a suitable value near the base model which performs better.However,it requires further study to find a suitable A for training KEWE. Intemal Corpus(ENCT 4k Corpus(EHCT 10k) mal Corpus(CPT 16k Exemal Corpus(Ewii 1000 nal Corpus(Ewii 1000 enaC0pus( Weigheing Factor 0.) Figure 5:The performance of KEWEh with varied weighting factor A 374

374 knowledge vectors respectively in bar chart. It can be found that the acquisition difficulty outperforms the other two on all four datasets. By combining the three knowledge vectors, the performance can be further improved. Figure 3 shows the performance measures of using different similarity functions and different values of k in line charts. It can be found that the knowledge graphs achieve a relatively good performance, when the cosine similarity is used and k is close to 50 or 100. ENCT EHCT CPT CPC 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Datasets Accuracy Acquisition Usage Structure All Figure 2: The perfor￾mance of KEWEk using different types of knowl￾edge vectors 10 50 100 150 200 0.76 0.8 0.84 0.88 0.92 0.96 Numerber of neighbors in knowledge graph Accuracy (ENCT) Cosine Eclidean 10 50 100 150 200 0.58 0.6 0.62 0.64 0.66 0.68 0.7 0.72 Numerber of neighbors in knowledge graph Accuracy (EHCT) Cosine Eclidean 10 50 100 150 200 0.48 0.49 0.5 0.51 0.52 0.53 0.54 0.55 0.56 Numerber of neighbors in knowledge graph Accuracy (CPT) Cosine Eclidean 10 50 100 150 200 0.28 0.29 0.3 0.31 0.32 0.33 Numerber of neighbors in knowledge graph Accuracy (CPC) Cosine Eclidean Figure 3: The performance of KEWEk using knowledge graphs with dif￾ferent similarity functions and neighbor numbers 10k 100k 1000k 7000k 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Numerber of Wiki sentences Accuracy (ENCT) RSWEh Skip−Gram Internal Corpus 10k 100k 1000k 7000k 0.45 0.5 0.55 0.6 0.65 0.7 0.75 Numerber of Wiki sentences Accuracy (EHCT) RSWEh Skip−Gram Internal Corpus 10k 100k 1000k 7000k 0.35 0.4 0.45 0.5 0.55 Numerber of Wiki sentences Accuracy (CPT) RSWEh Skip−Gram Internal Corpus 10k 100k 1000k 7000k 0.29 0.3 0.31 0.32 0.33 0.34 Numerber of Wiki sentences Accuracy (CPC) RSWEh Skip−Gram Internal Corpus Figure 4: The performance of KEWEh using different volumes of exter￾nal corpus Secondly, we study the impacts of using external corpus on the performance of KEWEh. To learn text￾based word embedding for both languages, we collect the external corpus from both English Wikipedia (Ewiki) and Chinese Wikipedia (Cwiki). After preprocessing, Ewiki contains 7M sentences and 172M words, and Cwiki contains 7.4M sentences and 160M words. We sample different number of sentences from the corpus for training word embedding, and then measure the performance of KEWEh on read￾ability assessment. Figure 4 shows the performance measures by using different volumes of external corpus in line charts, with Skip-Gram trained for comparison. Both models are also trained using the internal corpus, and the performance measures (dotted lines) are depicted as baselines. From Figure 4, it can be found that the performance of both Skip-Gram and KEWEh increases as the volume of external corpus increases, and keeps stable when the corpus is large enough. Besides, on English datasets, word embedding trained using external corpus achieves comparable performance with that using internal cor￾pus. The above suggests that external corpus is a good substitution for the internal corpus, especially on English datasets, and a relative large volume of corpus is required to achieve stable performance. Finally, we study the impacts of the weighting factor λ on the performance of KEWEh. Since λ is approximated by N1/(N1 + N2) in our model, we vary λ from 0 to 1 by setting N1+N2 = 100 and then varying N1 from 0 to 100 stepping by 10. Figure 5 shows the performance of KEWEh with varied λ in line charts with error bars, where λ = 0 and 1 correspond to Skip-Gram and KEWEk respectively. From Figure 5, it can be found that KEWEh can outperform its base models (i.e., KEWEk and Skip-Gram), by setting λ with a suitable value near the base model which performs better. However, it requires further study to find a suitable λ for training KEWEh. 0 0.2 0.4 0.6 0.8 1 0.75 0.8 0.85 0.9 0.95 1 Weighting Factor (λ) Accuracy (ENCT) Internal Corpus(ENCT 4k) External Corpus(Ewiki 1000k) 0 0.2 0.4 0.6 0.8 1 0.55 0.6 0.65 0.7 0.75 Weighting Factor (λ) Accuracy (EHCT) Internal Corpus(EHCT 10k) External Corpus(Ewiki 1000k) 0 0.2 0.4 0.6 0.8 1 0.4 0.45 0.5 0.55 0.6 Weighting Factor (λ) Accuracy (CPT) Internal Corpus(CPT 16k) External Corpus(Cwiki 1000k) 0 0.2 0.4 0.6 0.8 1 0.15 0.2 0.25 0.3 0.35 0.4 Weighting Factor (λ) Accuracy (CPC) Internal Corpus(CPC 4k) External Corpus(Cwiki 1000k) Figure 5: The performance of KEWEh with varied weighting factor λ

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. 375

375 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