text new page (beta)
English (pdf)
Article in xml format
Article references
Automatic translation
Send this article by e-mail
Cited by SciELO 
Similars in
SciELO 
Permalink
1 Introduction
A WordNet is a large lexical database where nouns, verbs, adjectives and adverbs are grouped into sets of cognitive synonyms, the so-called synsets [17]. Each synset represents a distinct concept and consists of a unique synsetID, synset members, and a gloss consisting of a brief definition and one or more examples showing the use of members in the synsets. Synsets are connected to others by means of semantic relations such as hypernymy or generalization, hyponymy or particularization, and meronymy or part-whole relation. Currently, the biggest WordNet is the Princeton WordNetfn (PWN) constructed manually since 1990. The PWN version 3.0 has 117,659 synsets including 82,415 noun synsets, 13,767 verb synsets, 18,156 adjective synsets and 3,621 adverb synsets.
In this paper, we discuss the feasibility of creating a Vietnamese WordNet (VWN) having the same structure as the PWN by bootstrapping from freely available resources. The remainder of this paper is organized as follows. In Section 2, we discuss related work. Section 3 describes the proposed approaches to build the VWN from existing resources. Results of our experiments and discussion are presented in Section 4. Section 5 concludes the paper.
2 Related Work
The research presented in this paper discusses an efficient method to generate a VWN with the same structure as the PWN. Therefore, this section highlights prior work on constructing WordNets based on the PWN. According to Vossen [25], the two common approaches to build a new WordNet in a target language T are the expand approach and the merge approach. Using the expand method, a new WordNet is created by simply translating the PWN to T, whereas using the merge method, an independent WordNet in T is firstly built and then aligned to the PWN. There have been a large number of efforts in various languages with the goal of constructing WordNets. We present a few prominent ones in this section.
2.1 WordNets Created Using the Merge Approach
A French WordNet was constructed from multilingual resources by Sagot and Fiser [20]. The authors performed word alignment and extracted bilingual lexicons from a multilingual corpus; then, every lexical entry was assigned a synsetID obtained from the Balkan WordNet [23]. They also translated the English WordNet to French using dictionaries and thesauri. The French WordNet was finally generated by merging synsets collected from the two methods. Their WordNet contains 32,351 non-empty synsets, and its accuracy based on manual evaluation is 80%.
Gunawan and Saputra [7] generated a prototype version of synsets for an Indonesian WordNet from a monolingual dictionary of Bahasa Indonesia and an Indonesian thesaurus. They first extracted synonym concepts from the thesaurus, combined them with entries in the monolingual dictionary and removed duplicate entries. Finally, a hierarchical clustering technique was applied to merge synsets. Their Bahasa WordNet consists of 60,673 synsets. No evaluation was performed.
A Hindi WordNetfn has been constructed manually by ‘looking up the various list meanings of words in different dictionaries’ [4]. The current version has 105,352 unique words and 40,457 synsets. The Hindi WordNet is the first WordNet for Indian languages and has been used to construct WordNets for other Indian languages (e.g., Marathi, Sanskirt and Gujarati) in the IndoWordNet project.
2.2 WordNets Created Using the Expand Approach
Oliver and Climent [18] compared the accuracies of WordNets created by several methods. The first WordNet was created using the Google translation machine to translate a sense-tagged corpus in English to Spanish. The generated WordNet had about 8,000 synsets with accuracy of 80%. In the second method, given a parallel corpus, an analyzer was used to tag senses of words with the English WordNet. Then, constructing a WordNet for Spanish became a word alignment problem. The accuracy of the second approach was lower than that of the first approach, and it depended on the size of the corpus. A bigger corpus increased the accuracy of the created WordNet. They also concluded that sense tagging introduced more errors than statistical machine translation.
Kaji and Watanable [9] constructed a Japanese WordNet by translating the PWN synsets to Japanese, by using a correlation matrix to deal with translation ambiguity. Later, Bond et al. [3] and Isahara et al. [8] constructed another Japanese WordNet by extracting synsets from the PWN and translating them to Japanese using bilingual dictionaries. They enriched the Japanese WordNet using the most common words obtained from different resources. This Japanese WordNet contained 57,238 synsets with 93,834 words.
Sathapornrungkij and Pluempitiwiriyawej [21] proposed a semi-automatic method to construct a Thai WordNet from machine readable dictionaries. They designed a WordNet Builder system which extracted lexical, semantic, and translation relations from the English WordNet and a dictionary. The extracted data was then evaluated according to 13 criteria (e.g., monosemic one-to-one, polysemic one-to-one and polysemic many-to-one). The created Thai WordNet contained 19,582 synsets with a coverage of 80% at 76% accuracy. Later, Akaraputthiporn et al. [1] and Leenoi et al. [14, 15] constructed Thai WordNets from several bilingual dictionaries using a bi-directional translation method. They noted that using different input dictionaries created by different methods such as corpora-based methods or author’s expertise produced WordNets with different accuracies. In addition, cultural issues such as categorization, gender, and collective perception needed to be taken into account to maintain the structure of Thai data.
Saveski and Trajkovski [22] constructed a Macedonian WordNet using the expand approach. To remove irrelevant translations, the English synset gloss was translated into Macedonian, and then the Google similarity metric [5] was applied to compute the similarity scores showing the semantic relatedness between the translated gloss and the candidate words. The selected words were words with Google similarity distance with the translated gloss greater than a threshold. The Macedonian WordNet they created had 33,276 synsets.
Lam et al. [13] proposed several methods to create WordNets in many languages with limited resources. The authors generated WordNet synsets for a target language T by translating PWN synsets to T using the Microsoft Translator. The approach using direct translation (DR), the approach using intermediate WordNets (IW) and the approach using intermediate WordNets and a dictionary (IWND) were introduced to remove translation ambiguities. In the DR approach, synsets in the T WordNet were built by simply translating PWN synsets to T. The IW approach handled translation ambiguities by using different WordNets with the same structure as the PWN. For each synsetID in PWN, they extracted all synsets of intermediate WordNets and translated to T. The objects of their study included resource poor and endangered languages, which do not have many existing lexical resources. Hence, the IWND approach translated synsets having the same synsetID to English, and then translated them to T. The correct members of synsets were selected based on the occurrence counts of translation candidates. The authors claimed that the IW approach with 4 intermediate WordNets helped construct better WordNet synsets. They did not establish connections between synsets created.
WordNets created using the expand approach have the same structure as the PWN; however, their quality considering complex agglutinative morphology, presence of culture specific meanings and usages of words is not good compared to those of WordNets built using the merge approach. Generally, the expand approach is more widely used than the other.
3 Proposed Approaches
Generating a new WordNet for a language using the merge approach needs linguistic experts in the language. In addition, the VWN we want to create will have the same structure as the PWN. Therefore, the expand approach is the best choice to construct a VWN. Our work is based on the study of Lam et al. [13], and is divided into 3 parts: creating synsets, establishing connections among synsets and extracting glosses of synsets.
3.1 Creating Synsets
To create synsets for the VWN, we use the intermediate WordNets (IW) approach. Lam et al. [13] experimented using the IW approach with different numbers of intermediate WordNets, but they did not know how many intermediate WordNets are good enough to create a new WordNet of high quality. In addition to the WordNets used in their studies, we experiment with one more WordNet, the Thai WordNet. Table 1 presents information about WordNets used. All WordNets used are linked to the PWN version 3.0 and are obtained from the Open Multilingual WordNet [2].
First, we query synsetIDs of all synsets in the PWN. For each synsetID, we extract all members belonging to that particular synset in the PWN and other intermediate WordNets. Then, we translate all synset members in different languages to Vietnamese using a machine translator. As a result of this step, for every synsetID we have a list of translation candidates in Vietnamese. One drawback of the IW approach is that the coverage percentage of synsets created using the IW approach is lower than using the DR and IWND approaches.
To increase the coverage percentage of synsets in the VWN, we improve the method to select translation candidates. The ranking method based on occurrence count is still applied to calculate the ranking value of translation candidates. The rank of a candidate w is calculated as below:
where:
—
—
—
—
The rank value of each translation candidate is in the range from 0.000 to 1.000. The greater the rank value of the candidate, the higher the possibility that it will become a synset member. Lam et al. [13] select translation candidates based on 3 scenarios: (i) All candidates with the rank values of 1.000 are accepted as correct translations. (ii) If there is no candidate with rank values of 1.000, the candidates with the highest rank value are selected as correct translations. (iii) For each synsetID, if all candidates have the same rank value, they skip all these candidates.
Their approaches to select candidates for each synsetID significantly reduce translation ambiguities; however, an issue is that they discard many correct translations. For instance, members of the synsetID 110399491, with a gloss ‘a father or mother; one who begets or one who gives birth to or nurtures and raises a child; a relative who plays the role of guardian’, obtained from PWN and JWN are {parent} and {
}.
Translations of these members are {cha mµ} and {phö huynh}, respectively. The criteria for selecting candidates by Lam et al. discard these two candidates which are both correct translations. So, we change the selection method: if all translation candidates of a synset have the same rank value, we compute the Google distance between each translation candidate pair to find the semantic relation among candidates using the NGD formula [6]:
where:
—
—
—
A pair of candidates is accepted as correct translations if the Google distance is smaller than a threshold
3.2 Establishing Connections Among Synsets
Synsets in PWN are linked to others by semantic relations, which are of 28 types in the PWN version 3.0. There are 285,348 relations among synsets. Lam et al. [13] did not establish connections among the synsets created. We establish connection among synsets in the VWN based on relations among synsets in the PWN using Algorithm 1. First, each Vietnamese synset created
Table 2 shows an example of establishing connections between synsetID 110399491 in the VWN with 2 synset members {cha mµ, phö huynh}. We note that we do not translate semantic relations to Vietnamese. Currently, the VWN constructed is managed based on the WNSQL projectfn.
Table 2 Example of synsets having connections to the synsetID 110399491 in the VWN
| Synset ID | Synset member | Gloss | Semantic relation | |
| PWN | VWN | |||
| 107970406 | family, family unit | gia đình, hộ gia đình | primary social group; parents and children | member meronym |
| 109772448 | adopter, adoptive parent | cha mẹ nuôi | a person who adopts a child of other parents as his or her own child | hyponym |
| 110332385 | female parent, mother | mẹ | a woman who has given birth to a child (also used as a term of address to your mother | hyponym |
| 110126708 | genitor | cha mẹ ruột | a natural father or mother | hypernym |
| 110654932 | stepparent | cha dượng | the spouse of your parent by a subsequent marriage | hyponym |
| 109918248 | kid, child | đứatrẻ | a human offspring (son or daughter) of any age | antonym |
3.3 Extracting Glosses of Synsets From the Viet WNMS
The project called Viet WNMSfn has constructed a Vietnamese WordNet for nouns, verbs and adjectives. This Viet WNMS project has been developed using the WNMS tool of the Asian WordNet project (AWN) [19] which provides a platform for building and sharing WordNets in Asian languages based on the PWN. The target of the Viet WNMS project is to build a Vietnamese WordNet consisting of 30,000 synsets and 50,000 words, including the 30,000 most common words in Vietnamese. The Viet WNMS project is divided into 2 partsfn:
— Translating the core of the PWN to Vietnamese. According to authors, the core of the PWN are words with high occurrence counts obtained from the BNC corpusfn.
— Manually adding concepts that exist only in Vietnamese. Currently, the Viet WNMS has 40,788 synsets and 67,344 words.
The approach to create the VWN, discussed in this paper based on the IW approach in [13], takes advantages of lexicons in several WordNets having the same structure as the PWN. As a result, our VWN has a better synset coverage percentage and includes common words not only in English but also in several other languages such as French, Finnish, Japanese and Thai. Moreover, our VWN has 4 POSes, including adverbs, whereas the Viet WNMS has 3 POSes. To the best of our knowledge, there is no published paper on this Viet WNMS project. We do not know anything about the structure of this WordNet. However, by manually checking several synsetIDs, we understand that these synsetIDs or synsetOffsets in the Viet WNMS are not the same as in the PWN. Hence, the Viet WNMS is likely to have a different structure compared to the PWN and our VWN.
We notice that synsets in the Viet WNMS have glosses in Vietnamese, which we believe are constructed manually by experts. Therefore, we extract these glosses and add them to synsets in our VWN using Algorithm 2. We could not use synsetIDs or synsetOffsets to retrieve data from the Viet WNMS. Hence, for each word
(i) We query all synsets, including their glosses (each of which is called
(ii) We trace back to all synsets having
Then, we compute a cosine similarity score between each pair of
4 Experiments and Discussion
4.1 Experiments
The synsets and the semantic relations among them in the VWN are evaluated by 8 volunteers who use Vietnamese as mother tongue. We use the same set of 300 synsetIDs, randomly chosen from the synsets we create, and connections among them. Each volunteer is requested to evaluate using a 5-point scale: 5: excellent, 4: good, 3: average, 2: fair and 1: bad.
The VWN is built by translating the PWN and several intermediate WordNets to Vietnamese. The quality of translations and quantity of synsets are highly dependent on machine translators used. Lam et al. [13] used the Microsoft Translator API for translation. When we performed experiments in 2017 for this paper, the Microsoft Translator API was not available for free, and therefore we use the Yandex Translate APIfn.
We experimented by constructing VWNs using both our approaches, denoted by IW-NGD, and the IW approach [13] with 4 intermediate WordNets (PWN, FWN, WWN and JWN) and 5 intermediate WordNets (PWN, FWN, WWN, JWN and TWN) using the Yandex Translate API. Table 3 presents the number of synsets, their coverage percentages and average scores of the VWNs built. The VWNs generated using 5 intermediate WordNets have greater numbers of synsets and average scores.
Table 3 VWNs created using different approaches
| Approach | Number of intermediate WordNets | Synsets | Average score | % coverage |
| IW | 4 | 55,048 | 3.21 | 46.79% |
| IW | 5 | 61,808 | 3.61 | 52.53% |
| IW-NGD | 4 | 61,348 | 3.23 | 52.14% |
| IW-NGD | 5 | 78,285 | 3.73 | 66.54% |
Moreover, the IW-NGD approach creates VWNs of better quality in terms of the numbers of synsets and coverage percentages than the IW approach. The IW-NGD approach with 5 intermediate WordNets creates the best VWN in our experiment. So, we establish links among synsets in the best VWN created. There exist 80,413 semantic relations among 78,285 synsets created in the VWN. The average evaluation score of relations is 3.60.
The Viet WNMS has been published on a website but has limited web service capability. In addition, words in our VWN are not the same as words in the Viet WNMS. In particular, our VWN has many words which do not exist in the Viet WNMS; and contrarily, the Viet WNMS consists of many words that do not exist in our VWN. Currently, we have queried 2,094 words from the Viet WNMS, and then extracted synsets’ glosses for these words.
We carefully evaluate the glosses extracted and find that a value of 0.30 or higher for threshold β finds very good mapped glosses, with an average evaluation score of 4.60. Hence, such synset glosses (the ones extracted from the Viet WNMS) are accepted as the correct glosses and are aligned to the corresponding synsets in our VWN. We have extracted 4,555 glosses for synsets in our VWN. We believe that cooperation between the two Vietnamese WordNets is likely to produce a more extensive WordNet.
Table 4 presents some glosses extracted and aligned to the corresponding synsets in our VWN. In this table, Member means the synset member of the SynsetID in our VWN, Gloss in the PWN: the gloss of the SynsetID extracted from the PWN, GlossTrans: the translation of the Gloss in the PWN generated by a machine translator, CosineSim: the cosine similarity score between the GlossTrans and the Gloss extracted from the Viet WNMS.
Table 4 Examples of glosses extracted
| SynsetId | Member | Gloss extracted | GlossTrans | Gloss in the PWN | Cosine Sim |
| 100887081 | sư phạm | nghề của một giáo viên | nghề của một giáo viên | the profession of a teacher | 1.00 |
| 104161981 | ghế | đồ đặc, được thiết kế để ngồi | đồ nội thất, được thiết kế để ngồi | furniture that is designed for sitting on | 0.76 |
| 300230843 | điều chỉnh | sửa đổi để chức năng tốt hơn | sửa đổi cho tốt hơn | modified for the better | 0.68 |
| 113548105 | lọc | loại bỏ các tạp chất | quá trình loại bỏ các tạp chất (như dầu hoặc kim loại hoặc đường) | the process of removing impurities (as from oil or metals or sugar etc.) | 0.62 |
| 300128572 | chưa từng có | không có ví dụ, tiền lệ hoặc sự tương tự trước đây | không có tiền lệ | having no precedent; novel | 0.58 |
| 301711614 | đau đớn | vô cùng đau khê | thể hiện đau đớn hoặc đau đớn | expressing pain or agony | 0.30 |
4.2 Discussion
Lam et al. [13] and we create VWNs using the IW approach and the same 4 intermediate WordNets. The only different resource used in the prior published experiments and experiments reported in this paper is the machine translator. The previously reported VWN had 72,010 synsets (61.20% coverage percentage) with an average score of 4.26, which is higher than the VWN reported in this paper. The VWN created by Lam et al. [13] was evaluated by native Vietnamese speakers in the US whereas the VWN created in this paper has been evaluated by native Vietnamese speakers in Vietnam. We claim that the translation quality significantly affects the VWN created. Then, an initial important step to build a good WordNet is to use a very good machine translator or dictionaries for translation.
The VWN we created for this paper is managed using WNSQL with 18 tables. The main tables in our project are: linktypes, lexlinks, semlinks, senses, synsets and words. In addition, as mentioned earlier, the PWN has 28 types of semantic relations. We have established only 15 relation types among the synsets we created. One reason for limited connectivity is that many synsets do not exist in the VWN.
Constructing a VWN using the expand approach may lead to problematic issues regarding language gap as discussed below.
— The PWN has concepts which cannot be translated to Vietnamese. For instance, synsetID 107573347 with a gloss ‘a canned meat made largely from pork’ has one member {Spam} which does not translate well to Vietnamese, although it could possibly be translated to ‘một dạng th࿋t heo đóng h࿙p’ or ‘đồ hộp Mỹfn’.
— Many concepts in Vietnamese do not exist in English. For example, synsetID 107804323 with a gloss ‘grains used as food either unpolished or more often polished’ has one member {rice}, which should be translated to ‘gạo’ in Vietnamese. To the best of our knowledge, in English, ‘rice’ can be also used for ‘cooked rice’ or ‘boiled rice’ which are both translated to ‘cơm’. The PWN does not contain synsets pertaining to ‘cooked rice’ or ‘boiled rice’. In Vietnamese, ‘gạo’ is different from ‘cơm’. A similar issue is identified by Sathapornrungkij and Pluempitiwiriyawej [24] when building a Thai WordNet.
— Parts-of-speech (POS) of words in English and their translations in Vietnamese may not be similar. For instance, the word ‘sad’ in the PWN has only one POS of adjective. This word is translated to ‘buồn’ in Vietnamese. In addition to the POS of adjective, the word ‘buồn’ has a POS of verb, meaning ‘having strong need to do something’fn and the PWN does not have this concept. Some examples showing the uses of the word ‘buồn’ are ‘bu࿓n ngủ’ (sleepy or need to sleep) and ‘bu࿓n cưi’ (to feel like a laugh coming because of something funny (to need to laugh at something)).
5 Conclusion
The purpose of our work presented in this paper has been to study the feasibility of constructing a Vietnamese WordNet with as many synsets as possible by bootstrapping from free lexical resources. We have created synsets and established connections among them.
We intend to improve translation by changing the Yandex Translate API to another better freely available machine translator (if we can find one), and freely available dictionaries [11, 12].
We are contemplating several potential approaches to translate glosses of synsets in the PWN to Vietnamese or to extract glosses of synsets from a Vietnamese corpus. To improve translation quality between English and Vietnamese of glosses, we will use the approach proposed in [10].
In addition, finding a good method to mine or combine information from the Viet WNMS as we have done will definitely improve the quality of our VWN.
