sbunlp/Farsnet

FarsNet is the largest and the most reliable Persian wordnet developed by Natural Language Processing Laboratory of Shahid Behehsti University(SBU) and is partially supported and granted by Iran Telecommunication Research Center (ITRC). The latest version of Farsnet (version 3) has more than 100,000 lexical entries (words or phrases) that contain about 40,000 sets of synonyms. At least one meaning is defined for each entry and each meaning participates in one and only one synonym set(synset). All synsets either participate in the hierarchy or are introduced as headings. In addition, each synonym set or at least one of its members has participated in at least one non-hierarchical relationship. Also, each synonym set is mapped to a synonym set, such as in Princeton WordPress 3.1, if possible.

• The structure of FarsNet

FarsNet like other wordnets in the world is built upon synsets; the synonym sets. It is mainly a graph of synsets and their relations. Each synset is made of synonym senses of words. In other words, FarsNet has a database of lexical entries. Each lexical entry which may be a word or a phrase, has one or more senses. Each word sense participates in exactly one synset. Synsets are interconnected to each other by synset relations and are mapped to PWN 3.1 synsets by equivalence or near-equivalence relations. Moreover word senses may participate in sense relations to be connected to other word senses. In this section the structure and features of lexical entries, synsets and relations and mappings will be discussed in more details. Some parts of what you will see is similar to PWN and some parts are different. -- Lexical entries: words and phrases and their morpho-syntactic features For each lexical entry we save the following information - Orthographic form: each entry has at least one orthographic (written) form. But some entries has more than one written form. We do not count various forms of writing a word as various entries but add them to “other written forms” feature of the entry. The main form is the most popular or the one suggested by Academy of Persian language and literature (APLL). - Pronunciation or phonetic form: each entry has 1 or more phonetic forms. The main one shows how the word is pronounced in formal Persian spoken in Tehran, the capital of Iran. Phonetic forms are saved in our own standard which can be easily converted to IPA. - Morpho-syntactic information: for each sense of each lexical entry besides its POS which is common in all wordnets we keep more detailed information such as choosing among formal/informal, abbreviation, idiom, derived/compound/simple/phrase, for all POS categories. Besides we choose among other features specific to each POS category too. For example for nouns we choose among proper-noun/mass-noun /number/gerund, name of person/animal/ time/location and singular/plural. For adjectives, specific features to choose among are comparative/superlative, and prenominal/post nominal. For verbs, the options are transitive/intransitive/ditransitive, active/passive, copula/simple/auxiliary and causative and for adverbs we determine we they are adverb of adjective/verb/adverb/sentence. It should be noticed that although the above information is present in FarsNet 2.5 and higher, they are not added to the uploaded LMF for multilingual wordnet as the required metadata were not available. -- Senses, Synsets and Relations Current version of FarsNet includes inter-POS as well as inner-POS relations. “Derivational form”, “antonymy”, “verbal part of”, “non-verbal part of”, “refers-to” are relations between word senses. “Verbal (non-verbal)-part-of” is a new relation between a compound verb and its verbal (non-verbal) component. “Refers-to” is another new relation between word senses which connects either an informal form of a word to its formal from or connects an irregular plural to its singular form. From the synset relations, in addition to hypernym (as between peach and fruit), hyponym (as between food and hamburger), various types of meronym (as between apple and apple juice) and holonym (as between car door and car) entailment (as between snore and sleep) and cause (as between kill and die) which were all present at FarsNet 1.0 as well as at Princeton WordNet, FarsNet 2.0 and later include the following relations(Shamsfard and Ghazanfari, 2016): - Has-attribute / is-attribute-of: the relation between a quantitative adjective and the attribute whose value is the adjective. For example the relation between heavy and weight or between warm and temperature. This relation is available in PWN too. - Domain / is-domain-of: the relation between a domain specific term and its corresponding domain. For example between Carbid and chemistry or between arthritis and medicine. This relation is available in PWN too. - Agent/ Is-agent-of: the relation between a predicate (verb) and the potential agent of it. For example between author and writing or chef and cooking. This relation and all coming after it are new to FarsNet from version 2.0 and higher and also do not exist in PWN 3.1. - Patient/ Is-patient-of: the relation between a predicate (verb) and the potential patient or theme of it. For example between eat and edible thing or write and letter. This relation - Instrument/ Is-instrument-of: the relation between a predicate (verb) and the potential instrument of it. For example between eat and spoon or write and pen. - Location/ Is-Location-of: the relation between a verb or noun and its location (a noun). For example the relation between cooking and kitchen or to cook and kitchen. - Salient defining feature: The relation between an adjective and the noun it often/ mainly describes. For example the relation between Stale and bread. Although PWN does not have the same relation, it has some relations between co-occurrences which can cover the relation “salient defining feature” somehow. - “Related to”- the relation between any two synsets which has a semantic relation other than the previous named relations. For example the relation between author and book or between school and teaching. As it was described some of the above relations just exist in FarsNet 2.0 and higher. Their creation is motivated by various NLP tasks. For example the relations between a predicate and its arguments such as agent, patient and instrument help semantic role labelers, word sense disambiguation (WSD) modules and information/ knowledge extraction systems to better find the corresponding relations and do their jobs. “Related-to” relation is used to relate any two synsets which has a sort of relation not included in the above named relations. Although the relation between some of the related concepts could be extracted by traversing the links in Princeton WordNet or FarsNet 1.0, the new relation specifies the important ones explicitly. It is mostly used in information retrieval and also in finding similarity between text components for example in text summarization. In addition to new relations added to FarsNet 2.0 and higher, we have added the argument structures of verbs too. FarsNet 2.0 is actually a combination of Persian WordNet and Persian VerbNet. It includes the verb frames (argument structure) of about 200 Persian simple verbs along with the selectional restrictions of their arguments. By argument structures we mean syntactic, grammatical and thematic arguments and their correspondences. These structures are discussed in (Shamsfard and Ghazanfari, 2016) in more details. All synsets in Persian are utilized by gloss and example, all of them have at least one hierarchical relation or semantic category or labeled as top-level-concept. All of them have at least one non-hierarchical relation and all of them are either mapped to PWN or labeled as no-mapping. -- Mapping FarsNet to Princeton English WordNet (PWN)

Current FarsNet (version 2.5 and later) is mapped to PWN 3.1 as the latest version of Princeton WordNet. In the current version of FarsNet, we have about 29600 Persian synsets from which about 6250 have no-mapping labels, about 19800 are mapped by equivalence relation to PWN and ab out 3400 is mapped via near-equivalence relation to PWN. Totally, more than 23200 out of 29600 Persian synsets of FarsNet are mapped via either equivalence or near-equivalence relations to about 20135 English synsets of PWN. The rest are labeled as “no-mapping”. There are more than 24700 mapping relations defined in FarsNet. Various mapping options will be discussed in the following.

--- No-mapping Having no-mapping label means that one of the followings is happened: - There is a lexical gap between Persian and English and the concept we are talking about in the Persian synset is not lexicalized in English. For example the synset containing فطریه /fetriye/ (the money Muslims pay for feeding poor people at the end of Ramadan) has no equal or near-equal synset in PWN and the concept is not lexicalized in English. So there will be a no-mapping label for this synset. - The word has a translation in English but it is not included in PWN. For example the translation of the word مژده (moZde) which means good news is not available in PWN. The same situation occurs for many named entities and instances in Persian such as the names of cities, poets, touristic attractions, etc. - The English translation of the word is available in PWN but not in the sense equal to what the Persian word carries. For example although the word ‘stationary’ is a translation of ‘نوشت افزار’ /neveStafzAr and is available in PWN, it means something different with what neveStafzAr means. In other words although the translated word is available, it does not have the appropriate sense corresponding the Persian word.

--- Near equivalence relation For equal concepts (synsets with the same meaning) in Persian and English wordnets we use an “equivalence” relation to map them together. But there are some synsets which do not have the exact equal in English and we use the near-equivalence relation to map them. Near-equivalence is used for one of the following cases: - The POS of the synsets (words) in the two languages are different. For example there may be an adjective like دانشجویی /dAneSjuyi/ in Persian which has no adjective equivalent in English. It is rather equal to the word ‘student’ as a noun coming before another noun (e.g. student card). Thus is this case we have to map an adjective to a noun with near-equivalence relation. - Although the POS of both synsets are the same, there are some tiny differences between the semantics or pragmatics of the two synsets and there is no closer mapping available. In these cases we use near–equivalence relation to map the synsets with major similarities and minor differences which are very near but not exactly the same. {کاهگلی} in Persian is a color near brown but not exactly the same. So we map it to {brown} with a near-equivalence relation. --- n to m mappings The point to consider is that the mappings are not necessarily one-to-one. We may map more than one Persian synsets into an English synset or map a Persian synset into more than one English synsets. The first case is caused by more granularity in FarsNet and the latter is caused by more granularity in PWN. As an example for the first case, three Persian synsets {کویر} (arid land with salty soil and without vegetation), {صحرا} (a large land without vegetation) and {بیابان} (a land with little rain and sparse vegetation) are mapped into the English synset {desert} (arid land with little or no vegetation). The three Persian synsets have some differences but they may be used interchangeably in Persian too. Although all of them refer to lands with little or no raining, the characteristics of soil and plants in them are different and so we use different lexemes to refer to those concepts which are all mapped to {desert] in PWN. As an example for the second case we may mention a 1-to-2 Persian-English mapping. There are two English synsets one including misbehavior ( improper or wicked or immoral behaviour) and the other including mistreatment ( the practice of treating (someone or something) badly) which are both mapped to a single Persian synset{بدرفتاری، بداخلاقی}. As another example we may mention the word cousin in English which is equal to 8 Persian words for son of brother of mother, son of brother of father, son of sister of mother, … --- Hierarchical gaps and Different views In some cases although the synsets can be mapped, we have problems in mapping the hierarchical structures. In other words the synsets in one Persian hierarchical path may be mapped to English synsets in different hierarchical paths. This means that although synsets are mapped, the hierarchical structures are not the same. For example consider the synset {لوازم التحریر، نوشت افزار} /neveStafzAr/ (tools for writing, drawing, studying ,…). It is equivalent to ‘stationary’ which is not available in this sense in PWN. The concept consists of what we can buy from stationary stores and many of them usually can be found in pencil cases of students like pen, pencil, eraser, ruler, pencil sharpener, pencil case, … The nearest synset to it in PWN is {writing implement} (an implement that is used to write) which just contains the tools we write with such as pen, pencil, marker, etc. Ruler is somewhere under measuring implement, pencil sharpener under sharpeners (under implement) and eraser is somewhere under implement and there is no relation between them; nothing more than e.g. the relation between a pencil eraser and an iron! But in Persian all of these are children of /neveStafzAr/ while they may have other parents such as measuring implement as well.

citations: M. Shamsfard, A. Hesabi, H. Fadaei, N. Mansoory, A. Famian, S. Bagherbeigi, E. Fekri, M. Monshizadeh, and S. M. Assi, “Semi automatic development of farsnet; the persian wordnet,” in Proceedings of 5th global WordNet conference, Mumbai, India, 2010, vol. 29. M. Shamsfard, “Developing FarsNet: A lexical ontology for Persian,” in 4th Global WordNet Conference, Szeged, Hungary, 2008. M. Shamsfard and Y. Ghazanfari, “Augmenting FarsNet with new relations and structures for verbs,” in 8th Global WordNet Conference (GWC 2016), 2016. Taheri and M. Shamsfard, “Mapping farsnet to suggested upper merged ontology,” in Asia Information Retrieval Symposium, 2011, pp. 604–613. M. Rouhizadeh, M. A. Yarmohammadi, and M. Shamsfard, “Developing the Persian WordNet of verbs: Issues of compound verbs and building the editor,” in Proceedings of 5th Global WordNet Conference, 2010.

For more information please visit: http://farsnet.nlp.sbu.ac.ir/