In Natural Language Processing (NLP) wide cov­erage lexical databases are used in different area such as information retrieval and cross-language information retrieval. WordNet is an example for a lexical database that groups words into sets of synonyms and categorizes them in four categories: noun, verb, adjective and adverb and records various relations between synonym sets. A broad overview of the different PWN applications such as "Machine Translation", "Information Retriev­al", "Document Classification", "Query Answer­ing" and "Conceptual Identification" have been presented in (Morato et al., 2004). PWN was created and maintained since 1990s. After this WordNet for other languages have

Heshaam Faili

been under development and new projects start every year. PWN database contains about 150000 words organized in over 115000 synsets.

Manual construction of WordNet is a time consuming task and requires linguistic know­ledge. A number of automatic methods were pro­posed for constructing WordNet for other lan­guages that use PWN and other existing lexical resources. In order to help the development of WordNets for other languages rather than English, especially for European one, a project named Eu­roWordNet was found (Vossen, 1999), in which a number of automatic methods for construction of such databases were proposed (Farreres et al., 1998).

There have been some other efforts to create a WordNet for Persian language (Famian, 2007; Rouhizadeh et al., 2008; Shamsfard, 2008) but there exists no Persian WordNet yet that covers all Persian words in dictionary and comparable with PWN. These projects have tried to construct Persian WordNet in the manually or semi auto­matic manner. In (Shamsfard, 2008) a semi au­tomatic method is proposed in which for each Persian word a number of PWN synsets are sug­gested by system in order to be supervised by a human annotator to select a relevant synset. Based on experiments mentioned by Shamsfard (2008), the proposed WordNet extracted automat­ically by the system, retrieved about 70% accura­cy.

In this paper a fully automatic method for con­structing a large-scale Persian WordNet from available resource such as PWN, MRDs and cor­pora has been proposed. Our approach uses dif­ferent word similarity metrics like mutual infor­mation and WordNet similarity to map Persian words to appropriate PWN synsets.

In the related field of automatic and semi automatic WordNet construction, several efforts have been made. In (Shamsfard, 2008) a semi automatic method has been used for developing a lexical ontology called FarsNet for Persian language. About 1500 verbs and 1500 nouns have been gathered manually to make WorldNet's core. Then some heuristics and Word Sense Disambiguation (WSD) methods have been used to find the most likely related Persian synsets.

According to the first heuristic, a Persian word has only one synset if it's be translated to a single English word. In this case no ambiguity exists for the Persian word whose one of synsets will be equivalent with that of English word. In other cases, second heuristic is used: if two translations of a Persian word have only one common synset then for the Persian word this common synset is selected. The existence of a single common synset in fact implies the existence of a single common sense between the two words and therefore their Persian translations shall be connected to this synset (Shamsfard, 2008). For words whose English translations have more than one synset and second heuristic cannot find the appropriate synset, WSD methods have been used to select correct synset. For each candidate synset, a score is calculated using the measure of semantic similarity and synset gloss words. Manual evaluation of the proposed automatic method in this research shows 70% correctness and covers about 6500 Entries on WordNet.

In (Sathapornrungkij and Pluempitiwiriyawej, 2005) a semi-automatic approach has been described to construct the Thai WordNet lexical database from WordNet and LEXiTRON machine readable dictionaries. Thai WordNet synsets have been derived from the PWN. The candidate links between Thai words and synsets have been derived from semantic links which are obtained from WordNet and the translation links which are obtained from LEXiTRON. In order to derive links between Thai words and PWN synsets, 13 criteria have been used which are categorized into three groups: monosemic, polysemic and structural criteria. Monosemic criteria focus on an English word which has only one meaning. Such English word has one synset in PWN. Polysemic criteria focus on an English word which has multiple meaning. Such English word has multiple synset in PWN. Structural criteria focus on the structural relations among synsets with respect to WordNet 1.7. In order to verify links that constructed using these 13 criteria, stratified sampling technique has been applied and for each criterion 400 links have been verified manually. The results of verification show that the best criterion has 92% correctness and the lowest correctness is equal 49.25%.

In PWN, there is a gloss for each synset that can be used in automatic WordNet construction. In (Kaji and Watanabe, 2006) this information has been used for automatic construction of Japanese WordNet. Given an English synset, it calculates the score for each of its Japanese translation candidates according to the gloss appended to the synset. A pair of words is called associated if mutual information between them be larger than a threshold. The score is defined as the sum of correlations between the translation candidate and the associated words appearing in the gloss. Whereas availability of bilingual corpora is limited, for calculating pair wise correlation between the Japanese translations of an English word and its associated words an iteratively approach has been proposed that calculate this correlation without using bilingual corpora.

In (Lee et al., 2000) a set of automatic WSD techniques have been described for linking Korean words collected from a bilingual MRD to PWN synsets. For a given synset, 6 individual heuristic scores are calculated and then a decision tree is used to combine these scores to classify the synset as linking or discarding. In order to make the decision tree, a set of synsets have been labeled manually as linking or discarding and corresponding heuristic scores have been calculated and then used for training data set. To evaluate the accuracy of proposed method the candidate synsets of 3260 senses of Korean words have been classified manually as linking or discarding. This test set has been used to calculate precision of each heuristic. The results of experiments show that the precision of all heuristics is better than random mapping and the best heuristic have 75.21% precision. The combination of heuristics using decision tree shows 93.59% precision.

Each Persian word can have several English translations and each English translation has also several PWN synsets. For a given Persian word, a bilingual dictionary is used to extract English equivalent words, and then a set of candidate synset is generated using PWN that contains all synsets of English translations of Persian word. As in (Shamsfard, 2008), if the English transla­tion of a given Persian word has only one synset in PWN, then the Persian word is linked to this PWN synset directly, or if for a candidate synset at least two English translations belong to it, then Persian word is linked to this PWN synset.

In other cases, a score is calculated for each remaining candidate synset and the synset with maximum score is selected as an appropriate syn-set of the Persian word. Note that after selecting a synset, all synsets that share English words are removed from candidate synsets. The following resources have been used in the process of score calculation:

• PWN: synset words, synset definition and hypernymy relations have been used.

• Bilingual dictionary (Persian - English)

• Raw Persian text corpus for extracting related words of a given Persian word

• Raw English text corpus for extracting mutual information between English words

Text corpora have been used to extract the re­lated words of any given word. To do this, Mu­tual Information (MI) metric between any words in corpus and given Persian word are calculated and n-best words with higher MI values are se­lected. Mutual Information of pair x and x' is defined as follows:

In formula 1, n(x, x') is co-occurrence fre­quency of x and x' in corpus. This frequency is calculated using a window with specific size. n(x) is the frequency of word x in corpus and N is the number of unique words in corpus. So, in order to select the most related words for a given Persian word, an additional step is consi­dered. For each Persian word w, other related Persian words with highest mutual information are selected and considered as a set R = (ri, r2, ..., rn}. Then for each Persian word ri a similar process is used and a set of words is extracted that is called R,-. If Ri contains the word w, then ri is selected as the related word for w and other­wise discarded.

After extracting the related words of the given Persian word, a Persian to English dictionary has been used to find equivalent translation of each related word. These words are referred as Related Translation Set (RTS). In scoring algorithm words that appear in gloss of each synset and words that appear in hypernym synset are called Gloss Words (GW). These words are considered as related words to the candidate synset and dis­tinguish each synset from other.

Now for each candidate synset of a given Per­sian word a score is calculated that is based on the idea that two related words in the two-side languages share the same words in the correlation set. That is, if Persian word w relates to English synset e, then other co-related Persian words rj, r2, rn which have gotten the best MI respect to w, should be related to the same synset e again.

Based on the above notion, the score of each candidate synset S can be estimated as follow:

WiERTS eiEGW

The score of synset S is defined as summation on product of semantic similarity between words in RTS and synset S, and mutual information be­tween words in RTS and words in GW. In (Pe-dersen et al., 2004) several methods for calculat­ing semantic similarity based on WordNet's structure have been presented. Some of these me­thods are based on path lengths between concepts and some of them are based on information con­tent. One of these methods is named path in which for each word w and synset s is defined as inverse of shortest path length between any syn-set of w and s. In our experiments the measure path has been used and calculated using formula 3.

mmSie synsets ofw(Pat^(si> ^))

In formula 2 the words from RTS which has less similarity to synset s has little effect on the amount of score in synset.

Persian WordNet constructor components are Word Translator, Related Word Extractor, Synset Extractor and Synset Selector. Persian words and their selected synsets are input and output of this system. Persian word is given as input to the Word Translator and Related Word Extractor components. In our experiment, 10 words with highest MI to the given Persian word are ex­tracted using Related Word Extractor. For this purpose, 3000 documents of IRNA newspaper text corpus have been used. IRNA is a news agency published their news on different lan­guages, mainly on Persian. In order to count the number of co-occurrences of words x and x', a window with the size of 20 words was considered. Translations of related words and candidate syn-sets are given to Synset Selector and appropriate synsets for the given Persian word are selected. In this step PWN is used for semantic similarity calculation and an English text corpus (USENET corpus) is used to calculate mutual information. Table 1 shows the number of words and docu­ments in the Persian and English text corpora. About 30698 Persian words from AryanpourPersian to English dictionary has been used for constructing Persian WordNet.

As it was mentioned in the previous section, Persian words were linked to PWN synsets in the two different ways. Some links was selected di­rectly without calculating their score by using some heuristics. We call these links as unambi­guous links. Some of these links are shown in table 2. As it shown in the table, unambiguous links are wrong in some cases. For example in the case of '<barchasb>tag', a verb synset is selected while the Persian word is noun, so the selection is judged as incorrect. If the part of speech tag in­formation of word is used in this example the correct synset would be selected.

Another type of links are ambiguous links, in which a scoring method is used for selecting the appropriate synset. Examples of these links are shown in table 3. As it's shown in the table, the word '<karmozd>commission' has been linked to 6th sense of word 'commission' that is wrong. In constructed Persian WordNet also word '<farman>commission' has been linked to this sense of word 'commission' but the word '<karmozd>commission' and the word '<farman>commission' have less similarity to­gether. In this example link between '<farman>commission' and 6th sense of word 'commission' is an unambiguous link. Therefore we can avoid of selecting this synset for '<karmozd>commission' using this information. In order to evaluate the quality of the selected links, 500 Persian words have been randomly selected and the accuracy of selected synsets has been evaluated manually. Table 4 summarizes the results of this evaluation. As it's shown in the table, the precision of unambiguous links is about 95.8% while this precision is 76.4% for the am­biguous links. The weighted average precision of the whole links in our automatically generated Persian WordNet is 82.6%, which outperforms the only comparable semi-automated Persian WordNet which was previously presented by (Shamsfard, 2008), about 12.5%. Also, by com­paring the PWN coverage rate of these Persian WordNets, it reveals that our result covered 29716 entries on PWN which it is about 4 times more than the previously generated Persian

WordNet.

The experimental results reveal that in PWN there is a short gloss for some synsets which makes the calculated score for those synsets to be lower than other candidate synsets of a given Per­sian word. This problem can be overcome by normalizing the scores of candidate synset of a given Persian word, i.e. by dividing the score of each synset by the number of words in GW. An­other solution of this problem is proposed by (Kaji and Watanabe, 2006). In (Kaji and Watanabe, 2006), the gloss is given as a query to text retrieval engine and the words that appear as the answer of this query are used instead of the words of gloss. In our experiments, the first solution is chosen which retrived the results shown in table 4.

This paper explored a method for automati­cally linking WordNet synsets to Persian words using pre-existing lexical resources such as Per­sian and English text corpora and PWN. The proposed method calculates a score for each candidate synset of a given Persian word and selects the synset with maximum score to be linked to the Persian word. This score is calcu­lated considering related words of Persian word and words that appear in gloss of synset. A pre­liminary experiment shows that this method can be used to construct Persian WordNet. In the proposed method for each Persian word synsets with maximum calculated score are selected without considering other Persian words. In fu­ture work we intend to adapt our method and contribute other Persian word in order to select a synset for a given Persian word.