Base_JSON/prefixY/json/Y11/Y11-1049.json

{
    "paper_id": "Y11-1049",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T13:39:11.145640Z"
    },
    "title": "An English-Chinese Cross-lingual Word Semantic Similarity Measure Exploring Attributes and Relations",
    "authors": [
        {
            "first": "Lin",
            "middle": [],
            "last": "Dai",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Beijing Institute of Technology",
                "location": {
                    "postCode": "100081",
                    "settlement": "Beijing",
                    "country": "China"
                }
            },
            "email": "dailiu@bit.edu.cn"
        },
        {
            "first": "Heyan",
            "middle": [],
            "last": "Huang",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Beijing Institute of Technology",
                "location": {
                    "postCode": "100081",
                    "settlement": "Beijing",
                    "country": "China"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Word semantic similarity measuring is a fundamental issue to many NLP applications and the globalization has made an urgent request for cross-lingual word similarity measure. This paper proposed a word semantic similarity measure which is able to work in cross-lingual scenarios. Basically, a concept can be defined by a set of attributes. The basic idea of this work is to compute the similarity between words by exploring their attributes and relations. For a given word pair, we first compute similarities between their attributes by combining distance, depth and relation information. Then word similarity are computed through a combination scheme. The algorithm is implemented based on an English-Chinese bilingual ontology HowNet. Experiments show that the proposed algorithm results in high correlation against human judgments, which encourages its broad application in cross-lingual applications.",
    "pdf_parse": {
        "paper_id": "Y11-1049",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Word semantic similarity measuring is a fundamental issue to many NLP applications and the globalization has made an urgent request for cross-lingual word similarity measure. This paper proposed a word semantic similarity measure which is able to work in cross-lingual scenarios. Basically, a concept can be defined by a set of attributes. The basic idea of this work is to compute the similarity between words by exploring their attributes and relations. For a given word pair, we first compute similarities between their attributes by combining distance, depth and relation information. Then word similarity are computed through a combination scheme. The algorithm is implemented based on an English-Chinese bilingual ontology HowNet. Experiments show that the proposed algorithm results in high correlation against human judgments, which encourages its broad application in cross-lingual applications.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Word semantic similarity measure plays a fundamental role in many natural language processing (NLP) applications such as information retrieval (IR), machine translation (MT) and word sense disambiguation (WSD). Many research efforts have been made in the past decades to improve the effectiveness of word semantic similarity measures. With the progress of the globalization, information emerged from a variety of languages to the Internet. Under this circumstance, crosslingual applications such as machine translation, cross-lingual information retrieval, cross-lingual text categorization and clustering etc., become more and more attractive. Consequently, crosslingual word semantic similarity measure becomes a meaningful research topic.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Although there are some statistical and hybrid measures which have good overall performance (Li et al., 2003; Xia et al., 2011; Hassan and Mihalcea, 2009) , the knowledge-based measures, to us, is still an important method. Even in a hybrid measure, the knowledge-based part is still a key part. Furthermore, if a large scale ontological knowledge base is available, the knowledge-based measures can be further improved if the knowledge base is comprehensively investigated and the similarity measures are elaborately designed.",
                "cite_spans": [
                    {
                        "start": 92,
                        "end": 109,
                        "text": "(Li et al., 2003;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 110,
                        "end": 127,
                        "text": "Xia et al., 2011;",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 128,
                        "end": 154,
                        "text": "Hassan and Mihalcea, 2009)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this work, we \"re-examine\" merely knowledge based word similarity measure and a crosslingual semantic similarity measure is proposed. As we know, concepts are language-independent and it is natural that each concept can be defined by several aspects, i.e., attributes. For example, the concept of the animal dog can be roughly defined by a kind-of attribute with value livestock and a modifier attribute with value domesticated. If all the words in this world are defined in this way, it is possible to compute the semantic similarity of any pair of words in spite of their language. This is the basic idea of this work. In this work, this idea is implemented by utilizing a bi-lingual ontology, HowNet (Dong and Donget al., 2006) , to locate attributes and concepts.",
                "cite_spans": [
                    {
                        "start": 706,
                        "end": 733,
                        "text": "(Dong and Donget al., 2006)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Although the proposed method can deal with English and Chinese solo-lingual word pairs, this paper focuses on cross-lingual ones. To evaluate our method, a series of experiments were conducted on a cross-lingual benchmark data sets with human ratings. A conclusion could be safely drawn from experimental results that the proposed measure is promising when compared with previous solo-lingual and cross-lingual word similarity measure.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The rest of this paper is organized as follows. Section 2 discusses the related works. The methodology of our method is presented in Section 3. Brief introduction of HowNet and its attribute network are introduced in Section 4. The implementation of the method is presented in Section 5. We evaluate the method in Section 6 before concluding this paper in Section 7.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Previous researches of word semantic similarity measure mainly focus on monolingual settings. Early research efforts have been devoted to design the knowledge-based measures, in which word synonym set in thesaurus plays an elementary role to word similarity calculation. In particular, WordNet (Fellbaum, 1998) has been widely adopted in word similarity measures for English due largely to the hierarchically organized synsets for English words. Researchers investigate taxonomic structure or ontological framework, and attempt to calculate the word similarity by counting conceptual distance. For example, (Yang et al., 2005; Alvarez et al., 2007) .",
                "cite_spans": [
                    {
                        "start": 294,
                        "end": 310,
                        "text": "(Fellbaum, 1998)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 607,
                        "end": 626,
                        "text": "(Yang et al., 2005;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 627,
                        "end": 648,
                        "text": "Alvarez et al., 2007)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "RELATED WORKS",
                "sec_num": "2"
            },
            {
                "text": "Some research efforts have been made to develop the corpus-based measures (jarmasz, 2003; Terra et al., 2003; Bollegala et al., 2007; Alvarez et al., 2007) or hybrid measures (Li et al., 2003; Strube et al., 2006) . The basic idea is to measure dependence between words by using statistics extracted from corpora. As a result, the corpus-based measures are found less effective than the knowledge-based ones in most word similarity measuring evaluations due to coverage problem.",
                "cite_spans": [
                    {
                        "start": 74,
                        "end": 89,
                        "text": "(jarmasz, 2003;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 90,
                        "end": 109,
                        "text": "Terra et al., 2003;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 110,
                        "end": 133,
                        "text": "Bollegala et al., 2007;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 134,
                        "end": 155,
                        "text": "Alvarez et al., 2007)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 175,
                        "end": 192,
                        "text": "(Li et al., 2003;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 193,
                        "end": 213,
                        "text": "Strube et al., 2006)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "RELATED WORKS",
                "sec_num": "2"
            },
            {
                "text": "In recent years, HowNet has been investigated by more and more researchers in several applications including word semantic similarity measuring. For example, (Liu and Li, 2005; Lin et al., 2009; Fan and Chen, 2009) . These works focused on mono-lingual context and were not evaluated on data sets with human ratings. Newly works began to apply HowNet on English data set (Dai et al., 2008) . Xia et al. (2011) combine HowNet and parallel corpus to mearsure cross-lingual word similarity. But their algorithms ignore the relations between sememes, which is proven to be very informative in this paper.",
                "cite_spans": [
                    {
                        "start": 158,
                        "end": 176,
                        "text": "(Liu and Li, 2005;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 177,
                        "end": 194,
                        "text": "Lin et al., 2009;",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 195,
                        "end": 214,
                        "text": "Fan and Chen, 2009)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 371,
                        "end": 389,
                        "text": "(Dai et al., 2008)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 392,
                        "end": 409,
                        "text": "Xia et al. (2011)",
                        "ref_id": "BIBREF17"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "RELATED WORKS",
                "sec_num": "2"
            },
            {
                "text": "Human beings judge similarity and dissimilarity between objects by their attributes. For example, when concept automobile and bicycle are concerned, people think they are similar because both of them have the is-a attributes with value vehicle. At the same time, people think they are different because an automobile is automatic while a bicycle is manual.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "There are numerous concepts in this world which are represented by different languages. To define concepts in an efficient and standard manner and make the definitions easy to be handled by computers, we can use a set of fundamental and language independent concepts to define other concepts. We call them meta-concepts. Furthermore, the real world is complex and there are abstract concepts as well as concrete concepts. To define all the concepts properly, meta-concepts must be organized into hierarchial trees.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "Because concepts are defined by meta-concepts, meta-concept similarity is extremely important to concept similarity. For a given concept pair {c 1 , c 2 }, we propose that the similarity between them s(c 1 , c 2 ) is a function of their feature sets as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s(c 1 , c 2 ) = f (f s 1 , f s 2 )",
                        "eq_num": "(1)"
                    }
                ],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "where f s 1 and f s 2 is the feature set of concept c 1 and c 2 respectively. Meta-concept similarity is the basis of similarity between feature (i.e., attribute) sets. Distance and depth information of meta-concepts in meta-concept hierarchies are fundamental evidences for computing meta-concept similarities. Given two meta-concepts in a hierarchy, there must exist the shortest path between them, which connects their hypernyms. We call their common hypernym on the shortest path the subsumer of them and the length of the shortest path the distance between them. Intuitively, meta-concept similarity is a monotonically decreasing function of the distance a monotonically increasing function with respect to the depth of subsumers.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "Besides, when computing meta-concept similarity, the impact of various relations should be considered. Actually, besides hypernym-hyponym relation which forms tree hierarchies, there are many other kinds of relations between concepts. Now, we propose that the similarity between meta-concept mc 1 and mc 2 is a function of the distance l, depth h, and relatedness r, as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s(mc 1 , mc 2 ) = f (f l (l), f h (h), f r (r))",
                        "eq_num": "(2)"
                    }
                ],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "where",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "f l (\u2022), f h (\u2022) and f r (\u2022)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "is the contribution of distance, depth and relation respectively. f (\u2022) is the framework that combine the three parts into final measure. It worth noting that values of both the distance and depth may cover a large range, while the interval of similarity should be finite with extremes of exactly the same or not similar at all. If we assign exactly the same with a value 1 and no similarity as 0, then the interval of similarity is [0, 1].",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BASIC IDEA",
                "sec_num": "3"
            },
            {
                "text": "To fulfill the task of this paper, a well structured knowledge base in which concepts are defined by feature(s) is demanded. In this section we briefly introduce the attribute and relation system of HowNet, which meets our request.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "As a knowledge base for natural language processing, HowNet provides plenty of taxonomically semantic knowledge as well as real-world knowledge. As a knowledge system that describes relations between concepts, HowNet attempts to construct a net structure of its knowledge base from the inter-concept relations and inter-attribute relations. This is the fundamental distinction between HowNet and other tree-structured lexical databases. The philosophy of HowNet entails its unique structure.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "The meta-concept in HowNet is called sememe, which is the smallest semantic unit that can't be reduced further and used to define concepts. Through combining sememes selected from a finite sememe set, HowNet can describe infinite concepts. There are 2219 basic sememes in the current version of HowNet which are organized into five taxonomies and four subsidiary taxonomies.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "As the sememes are concerned, in addition to being fundamental description units, there are complicated relations among them. The relations include hypernym-hyponym, antonym, converse, whole-part, material-product, etc. These relations are given in two ways. One way is the tree structure which represents hypernym-hyponym in taxonomies. The other way is the definition of sememes themselves. Although sememes are the meta-concepts used to define other concepts, the sememes are again made up by sememes. For example 1 , the frame of sememe bird is: {animal:materialOf={edible}, {eat:patient={\u02dc}}, {fly:agent={\u02dc}}}, and the frame of sememe food is: {edible:{cook:PatientProduct ={\u02dc}}}. Both of these two frames have a sememe edible. By this way, HowNet gives a strong relation between sememe bird and food.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "From the relations among sememes we can see that the actual structure of HowNet is indeed a net rather than merely a tree. Although there exist tree hierarchies in it, HowNet is totaly different to a traditional thesaurus. In a traditional thesaurus such as WordNet, concept is the minimum unit for word sense description and each concept is a node in a hierarchy. But in HowNet, sememes are nodes in hierarchies and each concept is defined by sememes.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "HowNet uses Knowledge Database Markup Language (KDML, in short) to construct the semantic expression of concepts and sememes. KDML consists of a set of grammar rules and keywords. The detailed introduction of KDML is beyond this work. We give several definition examples in Table 1 From the examples we can conclude several rules that are necessary for similarity computing.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 274,
                        "end": 281,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "-The first sememe in any definition expression gives the most essential sense, i.e., is-a attribute of corresponding concept.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "-A concept has one or more features. Each feature is defined by an attribute-value pair. The attribute symbols are in the front of the equal sign and values go after the equal sigh.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "-The sign\u02dcis a special and frequently used meta-concept which means the defined concept itself.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "The current version of HowNet includes 100,168 English words and 96,370 Chinese words. Its large coverage makes it competent to act as a knowledge base in many NLP researches.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "ATTRIBUTE AND RELATION SYSTEM IN HOWNET",
                "sec_num": "4"
            },
            {
                "text": "From section 4 we know that the nature of HowNet matches the methodology proposed in section 3 very well. This section gives the detailed implementation of the methodology.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "IMPLEMENTATION",
                "sec_num": "5"
            },
            {
                "text": "As already discussed, the sememe is the basic unit of word definition and sememe similarity is extremely important to word similarity. We will use function (2) to compute sememe similarity. In (2), there are three information resources, i.e., distance, depth and relation. The performance of different strategies for formula (2) will be experimentally investigated in section 6.3.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sememe Similarity",
                "sec_num": "5.1"
            },
            {
                "text": "Because concepts are defined by frames, concept similarity can be calculated from their definitions. From section 4 we know that a frame is constructed with sub-frames in a nested manner. The first sub-frame of a frame, which is always a simple sememe without attribute label, points out the essential feature of the corresponding concept. It specifies the is-a attribute of each concept. Meanwhile, the other sub-frames describe other necessary attributes.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "To simplify the computation and make full use of the semantic information provided by frames, we break a frame into a set of features. Sememes in all levels are reserved in this feature set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "Each feature is a {attribute-label, value} pair. We name the feature from the first sub-frame as primary feature and the corresponding sememe as primary sememe. And other features are called secondary feature and corresponding sememes are called secondary sememe. The attribute-label of a secondary feature reserves the attribute-labels of upper features. For example, the feature set of grin is {{CausePartMove},{PatientPart,part},{PatientPart-PartPosition,mouth},{PatientPartwhole,AnimalHuman}}.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "Similarity of feature pairs is measured by:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s f t (f t 1 , f t 2 ) = k * s(s 1 , s 2 )",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "where s(\u2022) is the similarity of the corresponding sememe pair, and k is a scaling factor decided by their attribute labels. If their attribute labels are the same, k is set to 1.0, or it is set as a decimal with value from 0.0 to 1.0, which reflects the strength of their relations. Now, frame similarity is converted to similarity between feature sets. When computing the similarity between two given feature sets, following principles are considered.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "-Primary features are more informative than secondary features.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "-Attribute-label modifier and null are equal to other attribute-labels.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "-For two features both have value of the sign\u02dc, the similarity between them is 1 if their attribute-label is the same. Otherwise the similarity is a small value, which is experientially set to 0.1 in this work. -Similarity between any feature set and empty set is 0.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "To satisfy the above principles, we use the following formula to compute concept similarity.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "SimC(C 1 , C 2 ) = s set (f s 1 , f s 2 ) if s f t (pf t 1 , pf t 2 ) = 0 s f t (pf t 1 , pf t 2 ) \u2022 s set (f s 1 , f s 2 ) if s f t (pf t 1 , pf t 2 ) > 0",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "where f s 1 , f s 2 is the feature set of concept C 1 and C 2 , pf t 1 and pf t 2 is the primary feature of concept C 1 and C 2 . At last we compute the similarity of two feature sets s set (\u2022). The measure of feature-set similarity must satisfies two basic requirements.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "(1) Similarity between two exactly same sets should be 1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "(2) Given two sets, both have n elements and there are z identical elements. Their similarity should be z/n.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "Based on the understanding of above principles, we use following steps to compute feature-set similarity.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "(1) Find the feature-pair with the highest similarly from two feature sets. Accumulate their similarity value to sim sum . (2) Delete the two features from their original feature sets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "(3) Repeat 1 and 2, until any one feature set is empty. (4) The final similarity is: 2 * sim sum /(m + n), where sim sum is the accumulated feature-pair similarly, m and n are the size of two feature sets respectively.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Concept Similarity",
                "sec_num": "5.2"
            },
            {
                "text": "Word semantic similarity is decided by concepts carried by them. For the case of polysemy, we select the maximum concept similarity as word similarity, which matches the psychology of human beings. Let C 1 = {C 1i }(i = 1, .., m) and C 2 = {C 2j }(j = 1, .., n) denote concept sets of word W 1 and W 2 , respectively. The similarity between word W 1 and W 2 is:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word Semantic Similarity",
                "sec_num": "5.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "SimW (W 1 , W 2 ) = arg max i,j {SimC(C 1i , C 2j )}",
                        "eq_num": "(5)"
                    }
                ],
                "section": "Word Semantic Similarity",
                "sec_num": "5.3"
            },
            {
                "text": "6 Evaluation and Discussion",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word Semantic Similarity",
                "sec_num": "5.3"
            },
            {
                "text": "To evaluate an algorithm for word semantic similarity, we should investigate its performance against human common sense. A commonly used word-pair set is constructed by Rubenstein et al. (1965) for evaluating word semantic similarity measures (denoted as RG65). Twenty five years later, Miller et al. (1991) selected 30 pairs(denoted as MC30) from RG65 data set to form another data set. Earlier researchers use the 28 word pairs (denoted as MC28) of the Miller-Charles set due to the absence of two word pairs from WordNet. The data sets mentioned above are monolingual which can't be directly used in this work. Up to now, there is a critical lack of a benchmark set for the evaluation of cross-lingual similarity measures. To evaluate the performance of our algorithm under cross-lingual circumstances, we constructed a English-Chinese bilingual data set (denoted as EC62) from RG65. For each word pair, we form 2 pairs by translating the first and second word into Chinese respectively. There are 2*62 word pairs because the word oracle is not presented in the current version of HowNet. To make EC62 reliable, Native speakers of Chinese who are also highly proficient in English, were asked to translate the words in the two data sets. They were asked to disambiguate the words and avoid multi-word expressions, slang or culturally-biased terms in their translations.",
                "cite_spans": [
                    {
                        "start": 169,
                        "end": 193,
                        "text": "Rubenstein et al. (1965)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 287,
                        "end": 307,
                        "text": "Miller et al. (1991)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data Set",
                "sec_num": "6.1"
            },
            {
                "text": "As discussed above, useful information for measuring word similarity includes the distance between sememes, depth of consumers, and relations between concepts. We believe that, to get a good similarity measure, all these information should be taken into account trough a reliable scheme. This section experimentally investigates the contribution of different information to deduce a good combination strategy of these information.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Tasks",
                "sec_num": "6.2"
            },
            {
                "text": "Researchers evaluated the performance of word similarity measures on benchmark data sets. The best way of evaluate similarity measure on a data set is to perform training-testing framework. But the size of RG65 is limited, to train and test algorithms on fixed subsets of it will causes overfitting and bias. In this work, a cross-validation scheme is used to get the performance. We randomly divide EC62 into training set and testing set. The size of training sets is 84 and the size of testing sets is 40. For each strategy, its parameters are tuned on training sets, and it is tested on testing sets. Similar to most works, we adopt the correlation coefficient 2 as evaluation criteria. Each strategy will be evaluated 10 times and the average correlation on testing set is recorded as its performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Tasks",
                "sec_num": "6.2"
            },
            {
                "text": "From section 5.2 we know that sememe similarity is the key part of word similarity measure. In this section, we experimentally investigate the performance of sememe similarity strategies. The parameter k of feature similarity function (3) is fixed at 0.8 through out the experiments according to our previous experiences. Strategy 1. Sememe similarity measure is linear and exclusively decided by the shortest distance between sememes. A segmental linear function is investigated and sememe similarity is defined as",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "S 1 (s 1 , s 2 ) = max{0, 1 \u2212 l L } (6)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where L is a parameter which makes sememe similarity to be 0 when l is larger than L. This formula ensures the sememe similarity value falls into [0, 1]. The correlation coefficient with optimal L between this strategy and human judgments of Rubenstein-Goodenough's is 0.8594.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "This value proves that the shortest path length is an effective information source for word similarity measure. Strategy 2. Sememe similarity measure is linear and exclusively based on the depth of the subsumer of given sememe pair. A segmental linear functions is used in this strategy.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "S 2 (s 1 , s 2 ) = min{1, h H } (7)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where H is a parameter which makes sememe similarity being 1 when h is larger than H. This formula ensures that sememe similarity value falls into [0, 1] . This strategy is plausible because it reflects the truth that higher subsumer results in more abstract meaning the two sememes share and vice versa. The correlation coefficient with optimal H between the results of this strategy and human ratings is 0.8443. This value proves that the depth of subsumers is an effective information source for word similarity measure. Strategy 3. Sememe similarity is merely decided by the relationship between given sememe pair. We can simply model the contribution of relations between two sememes as a segmental linear function:",
                "cite_spans": [
                    {
                        "start": 147,
                        "end": 150,
                        "text": "[0,",
                        "ref_id": null
                    },
                    {
                        "start": 151,
                        "end": 153,
                        "text": "1]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "S 3 (s 1 , s 2 ) = min{1, D r R } (8)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where D r is the depth of the deepest relation-sememe, R is a parameter that makes similarity to be 1 when D r is larger than R. D r is set to zero if there is no relation-sememe. The correlation coefficient with optimal R between the results of this strategy and human ratings is 0.8343. This value proves that the relation between sememes should be considered when designing word similarity measure. Strategy 4. Strategy 1 is nonlinearly combined with Strategy 2 through multiplication. This strategy considers both the shortest path and depth of subsumer.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "S 4 (s 1 , s 2 ) = S 1 (s 1 , s 2 ) \u2022 S 2 (s 1 , s 2 )",
                        "eq_num": "(9)"
                    }
                ],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "The average correlation coefficient between the results of this strategy and human ratings is 0.8713, greater than that of Strategy 1, 2 and 3. This experiment illustrates that a simple combination of the shortest path length and the depth of subsumers using multiplication can significantly increase the accuracy of the similarity measure.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "Strategy 5. Similar to Strategy 4, sememe similarity measure is a function of the length of the shortest path and the depth of subsumer. But exponential functions are considered in this strategy. the sememe similarity is defined as",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "S 5 (s 1 , s 2 ) = e \u2212\u03b1l \u2022",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "e \u03b2h \u2212 e \u2212\u03b2h e \u03b2h + e \u2212\u03b2h (10)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where \u03b1 \u2265 0 and \u03b2 \u2265 0 are parameters scaling the contribution of the shortest path and the depth of subsumer, respectively. The average correlation coefficient between this strategy and human judgments of Rubenstein-Goodenough's is 0.8861, which is greater than that of Strategy 4. Strategy 6. To consider the shortest path, depth and relation at a same time, Strategy 3 is linearly combined with Strategy 4.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "S 6 (s 1 , s 2 ) = \u03bb \u2022 S 3 (s 1 , s 2 ) + (1 \u2212 \u03bb) \u2022 S 4 (s 1 , s 2 )",
                        "eq_num": "(11)"
                    }
                ],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where \u03bb is the parameter which leverages the weight of S 3 and S 4 . The average correlation coefficient with optimal \u03bb between this strategy and human judgments of Rubenstein-Goodenough's is 0.8372. This value is less than Strategy 4, so this combination of relation contribution with length and depth cannot produce a better similarity measure.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "Strategy 7. Strategy 3 is linearly combined with Strategy 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "S 7 (s 1 , s 2 ) = \u03bb \u2022 S 3 (s 1 , s 2 ) + (1 \u2212 \u03bb) \u2022 S 5 (s 1 , s 2 )",
                        "eq_num": "(12)"
                    }
                ],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where \u03bb is a parameter which leverages the contribution of S 3 and S 5 . The average correlation coefficient between this strategy and human judgments of Rubenstein-Goodenough's is 0.8499, which does not bring a better correlation than using S 5 alone. Strategy 8. Strategy 3 is used as a modifier of Strategy 4.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "S 8 (s 1 , s 2 ) = min{1, S 4 (s 1 , s 2 ) \u2022 (1 + \u03bb \u2022 S 3 )} (13)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where \u03bb is a parameter which leverages the impact of S 3 on S 4 . When \u03bb is 0, this strategy is equal to S 4 . The average correlation coefficient between this strategy and human judgments of Rubenstein-Goodenough's is 0.8945. This measure performs nearly at a level of human replication, where correlation between individuals is 0.9015 (Resnik, 1999) . The strongest correlation against human similarity judgments is obtained at \u03bb=2. Strategy 9. Strategy 3 is used as a modifier of Strategy 5.",
                "cite_spans": [
                    {
                        "start": 337,
                        "end": 351,
                        "text": "(Resnik, 1999)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "S 9 (s 1 , s 2 ) = min{1, S 5 (s 1 , s 2 ) \u2022 (1 + \u03bb \u2022 S 3 )} (14)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where \u03bb is a parameter which leverages the impact of S 3 on S 5 . When \u03bb is 0, this strategy is equal to S 5 . The average correlation coefficient between this strategy and human judgments of Rubenstein-Goodenough's is 0.8667, less than that of S 5 . Strategy 10. Relation between sememes is used to improve Strategy 4. An exponential nonlinear function is involved in this strategy.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "S 10 (s 1 , s 2 ) = min{1, S 4 (s 1 , s 2 ) \u2022 exp(\u03bb \u2022 D r )}",
                        "eq_num": "(15)"
                    }
                ],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "where D r is the depth of the deepest relation-sememe, \u03bb is a parameter which adjusts the contribution of relation. When \u03bb is set to 0, this strategy is equal to S 4 . The average correlation coefficient between this strategy and human ratings is 0.8469, , less than that of S 4 . Strategy 11. Relation between sememes is used to improve Strategy 5. The same exponential nonlinear function in Strategy 10 is involved.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "S 11 (s 1 , s 2 ) = min{1, S 5 (s 1 , s 2 ) \u2022 exp(\u03bb \u2022 D r )}",
                        "eq_num": "(16)"
                    }
                ],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "This strategy is equal to S 5 when \u03bb is set to 0. The average correlation coefficient between this strategy and human judgments of Rubenstein-Goodenough's is 0.9035, which is greater than Strategy 8 and a little bit greater than individual average. The strongest correlation against human similarity judgments is obtained at \u03bb=0.22.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Results",
                "sec_num": "6.3"
            },
            {
                "text": "Referring to results in Section 6.3, we can reach several observations. From the results of S 1 , S 2 and S 3 , we can see that the shortest path, depth of subsumer, and relation between concepts all contributes a lot to word similarity. When compared with previous research results, we can see that with any single factor, the performance is comparable to early work. It is particular worth noting that the results of Strategy 3 indicate the importance of relations in computing word similarity. Strategy 4 and Strategy 5 combines the shortest path length and the subsumer depth in different ways. They have better performance than Strategy 1 and 2 which uses single factor. This proves that combining effective information sources with a proper strategy will produce higher performance. Strategy 6 and Strategy 7 explores three kinds of information at the same time in different ways. But both of them produce lower performance than Strategy 4 and Strategy 5. This explains that relationship can not be linearly combined with distance and depth to improve algorithm's performance. From Strategy 8 to Strategy 11, relationship is used as a modifier of Strategy 4 or Strategy 5. When we use a linear function of relation to modify Strategy 4 in which distance and depth are used in linear manner (Strategy 8), or when we use a nonlinear function of relation to modify Strategy 5 in which distance and depth are used in nonlinear manner (Strategy 11), we can get higher performance. Otherwise, lower performance will be reached (Strategy 9 and Strategy 10). As Strategy 4 and Strategy 5 are concerned, we can see that Strategy 5 is better than Strategy4. It also can be seen that Strategy 11 is better than Strategy 8. This proves that nonlinear exponential function works better than linear function on this problem. As a conclusion, we propose an algorithm for word semantic similarity measuring using Strategy 11 to compute sememe similarity. According to the experimental results, optimal parameters are \u03b1=0.3, \u03b2=0.4, \u03bb=0.22. However, since our method combines three kinds of information in a intuitive and experimental way, we believe the performance could be further improved if better combination strategies are found.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "DISCUSSION",
                "sec_num": "6.4"
            },
            {
                "text": "The state-of-the-art monolingual algorithms (Alvarez et al., 2007) (Yang et al., 2005) get correlation values around 0.90 against Rubenstein-Goodenough's human ratings, which is very close to the human performance. On the English-Chinese data set EC62, our algorithm reaches an average correlation of 0.9035, which is comparable to monolingual measures of similarity, higher than (Xia et al., 2011) 's result. This grade of performance encourages the use of word similarity measuring algorithms in various cross-lingual applications.",
                "cite_spans": [
                    {
                        "start": 44,
                        "end": 66,
                        "text": "(Alvarez et al., 2007)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 67,
                        "end": 86,
                        "text": "(Yang et al., 2005)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 380,
                        "end": 398,
                        "text": "(Xia et al., 2011)",
                        "ref_id": "BIBREF17"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "DISCUSSION",
                "sec_num": "6.4"
            },
            {
                "text": "It is worth noting that the proposed method can be competent to English or Chinese sololingual word similarity measure, with fairly better performance than cross-lingual context. Due to the lacking of space, we won't give out the detailed result here.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "DISCUSSION",
                "sec_num": "6.4"
            },
            {
                "text": "This paper presents an attribute and relation based measure to calculate semantic similarity of English-Chinese word pairs. The key feature of this work which differs from the previous ones is that our algorithm explores attributes and semantic relations that connect sememes and concepts. Experiments on English-Chinese cross-lingual data sets show that the performance of the new measure is close (i.e. correlation coefficient value of 0.9035) to human judgments as well as comparable to monolingual word similarity measures. It can thus be concluded that attribute and semantic relation contribute a great deal to word similarity measuring and HowNet is applicable and effective in the task for English-Chinese cross-lingual scenarios.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CONCLUSION AND FUTURE WORK",
                "sec_num": "7"
            },
            {
                "text": "The inspiring results stimulate a few future works. For instance, we will explore the feasibility to integrate our measure in real NLP applications such as word sense disambiguation, cross-lingual information retrieval, and machine translation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CONCLUSION AND FUTURE WORK",
                "sec_num": "7"
            },
            {
                "text": "The current version of HowNet is bilingual, i.e., English and Chinese. Sememes are identified by English sememes and Chinese sememes at the same time. Both English version and Chinese version of sememe set is competent to define words of any language. We only give the English version of the sememes.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Let x and y the variables with covariance \u03c3xy and standard deviations \u03c3x and \u03c3y. The correlation coefficient between x and y is \u03c1xy = \u03c3xy \u03c3x\u03c3y .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "A Graph Modeling of Semantic Similarity between Words",
                "authors": [
                    {
                        "first": "M",
                        "middle": [
                            "A"
                        ],
                        "last": "Alvarez",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Lim",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "ICSC '07: Proceedings of the International Conference on Semantic Computing",
                "volume": "",
                "issue": "",
                "pages": "355--362",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alvarez, M.A., and S. Lim. 2007. A Graph Modeling of Semantic Similarity between Words. ICSC '07: Proceedings of the International Conference on Semantic Computing, pp.355-362. IEEE Computer Society.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Measuring semantic similarity between words using web search engines",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Bollegala",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Matsuo",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Ishizuka",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "WWW '07: Proceedings of the 16th international conference on World Wide Web",
                "volume": "",
                "issue": "",
                "pages": "757--766",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bollegala, D., Y. Matsuo, M. Ishizuka. 2007. Measuring semantic similarity between words using web search engines. WWW '07: Proceedings of the 16th international conference on World Wide Web, pp.757-766. ACM.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Measuring Semantic Similarity between Words Using HowNet. ICCSIT '08: Proc. of the International Conference on Computer Science and Information Technology",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Dai",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Xia",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "601--605",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dai, L., B. Liu, Y. Xia and S. Wu. 2008. Measuring Semantic Similarity between Words Us- ing HowNet. ICCSIT '08: Proc. of the International Conference on Computer Science and Information Technology, pp.601-605. IEEE.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Hownet And the Computation of Meaning",
                "authors": [
                    {
                        "first": "Z",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "Q",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dong, Z., and Q. Dong. 2006. Hownet And the Computation of Meaning. World Scientific Publishing Co., Inc..",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Similarity Computation of Low-frequency Chinese Words. FSKD '09",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Fan",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Sixth International Conference on Fuzzy Systems and Knowledge Discovery",
                "volume": "",
                "issue": "",
                "pages": "524--528",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fan, X., and J. Chen. 2009. Similarity Computation of Low-frequency Chinese Words. FSKD '09. Sixth International Conference on Fuzzy Systems and Knowledge Discovery, pp.524-528. IEEE.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "WordNet: An Electronic Lexical Database (Language, Speech, and Communication)",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Fellbaum",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fellbaum, C. 1998. WordNet: An Electronic Lexical Database (Language, Speech, and Commu- nication). The MIT Press.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Cross-lingual Semantic Relatedness Using Encyclopedic Knowledge",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Hassan",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Mihalcea",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proc. EMNLP'09",
                "volume": "",
                "issue": "",
                "pages": "1192--1201",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hassan, S., and R. Mihalcea 2009. Cross-lingual Semantic Relatedness Using Encyclopedic Knowledge. Proc. EMNLP'09, pp.1192-1201.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Roget's Thesaurus and Semantic Similarity",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Jarmasz",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Szpakowicz",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of Conference on Recent Advances in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "212--219",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jarmasz, M., S. Szpakowicz. 2003. Roget's Thesaurus and Semantic Similarity. Proceedings of Conference on Recent Advances in Natural Language Processing (RANLP 2003), pp.212-219.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "An Approach for Measuring Semantic Similarity between Words Using Multiple Information Sources",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Bandar",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Zuhair",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Mclean",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "IEEE Educational Activities Department",
                "volume": "",
                "issue": "",
                "pages": "871--882",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Li, Y., Bandar, A. Zuhair and D. McLean. 2003. An Approach for Measuring Semantic Similarity between Words Using Multiple Information Sources. IEEE Trans. on Knowl. and Data Eng., pp.871-882. IEEE Educational Activities Department.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Research on HowNet-based Chinese Word Lexical Semantic Similarity Measurement",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the Eighth International Conference on Machine Learning and Cybernetics",
                "volume": "",
                "issue": "",
                "pages": "282--287",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lin, P., Y. Zhang and T. Zhang. 2009. Research on HowNet-based Chinese Word Lexical Seman- tic Similarity Measurement. Proceedings of the Eighth International Conference on Machine Learning and Cybernetics, pp.282-287.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Word Semantic Similarity Computation based on HowNe. The 3rd Chinese lexical and semantic proseminar",
                "authors": [
                    {
                        "first": "Q",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "17--30",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Liu, Q., and S. Li. 2005. Word Semantic Similarity Computation based on HowNe. The 3rd Chinese lexical and semantic proseminar, pp.17-30.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Contextual correlates of semantic similarity. Language and Cognitive Processes",
                "authors": [
                    {
                        "first": "G",
                        "middle": [
                            "A"
                        ],
                        "last": "Miller",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [
                            "G"
                        ],
                        "last": "Charles",
                        "suffix": ""
                    }
                ],
                "year": 1991,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "1--28",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Miller, G.A., and W.G. Charles, 1991. Contextual correlates of semantic similarity. Language and Cognitive Processes, pp.1-28. Psychology Press.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Semantic Similarity in a Taxonomy: An Information-Based Measure and its Application to Problems of Ambiguity in Natural Language",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Resnik",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Journal of Artificial Intelligence Research",
                "volume": "",
                "issue": "",
                "pages": "95--130",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Resnik, P. 1999. Semantic Similarity in a Taxonomy: An Information-Based Measure and its Application to Problems of Ambiguity in Natural Language. Journal of Artificial Intelligence Research, pp.95-130.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Contextual correlates of synonymy. Commun",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Rubenstein",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "B"
                        ],
                        "last": "Goodenough",
                        "suffix": ""
                    }
                ],
                "year": 1965,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "627--633",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rubenstein, H. and J.B. Goodenough. 1965. Contextual correlates of synonymy. Commun, pp.627-633. ACM Press.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "WikiRelate! computing semantic relatedness using wikipedia. AAAI'06: proceedings of the 21st national conference on Artificial intelligence",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Strube",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Ponzetto",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "1419--1424",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Strube, M., and S. Ponzetto. 2006. WikiRelate! computing semantic relatedness using wikipedia. AAAI'06: proceedings of the 21st national conference on Artificial intelligence, pp.1419-1424. AAAI Press.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Frequency estimates for statistical word similarity measures",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Terra",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Clarke",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "NAACL '03: Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology",
                "volume": "",
                "issue": "",
                "pages": "165--172",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Terra, E. and C. Clarke. 2003. Frequency estimates for statistical word similarity measures. NAACL '03: Proceedings of the 2003 Conference of the North American Chapter of the Asso- ciation for Computational Linguistics on Human Language Technology, pp.165-172. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Measuring semantic similarity in the taxonomy of WordNet. ACSC '05: Proceedings of the Twenty-eighth Australasian conference on Computer Science",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Powers",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "315--322",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yang, D., and D. Powers. 2005. Measuring semantic similarity in the taxonomy of WordNet. ACSC '05: Proceedings of the Twenty-eighth Australasian conference on Computer Science, pp.315-322. Australian Computer Society, Inc..",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Measuring Chinese-English Cross-Lingual Word Similarity with HowNet and Parallel Corpus",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Xia",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Yao",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Jin",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proc. CICLing",
                "volume": "II",
                "issue": "",
                "pages": "221--233",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xia, Y., T. Zhao, J. Yao, and P. Jin. 2011. Measuring Chinese-English Cross-Lingual Word Similarity with HowNet and Parallel Corpus. Proc. CICLing 2011, Part II, pp.221-233. LNCS.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF0": {
                "num": null,
                "text": ".",
                "type_str": "table",
                "content": "<table><tr><td>Word</td><td>Definition</td></tr><tr><td colspan=\"2\">cock f ruit asylum {InstitutePlace:domain={medical},{doctor:content={disease:CoEvent {bird:modifier={domesticated}{male}} {fruit} ={mad}},location={\u02dc}}}</td></tr><tr><td>grin jewel stove</td><td>{CausePartMove:PatientPart={part:PartPosition={mouth},whole= {AnimalHuman}}} {material:MaterialOf={treasure}} {tool:{WarmUp:instrument={\u02dc}},{burn:location={\u02dc}}}</td></tr></table>",
                "html": null
            },
            "TABREF1": {
                "num": null,
                "text": "The examples of definition expression of concepts in HowNet",
                "type_str": "table",
                "content": "<table/>",
                "html": null
            }
        }
    }
}