{ "paper_id": "N13-1031", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T14:40:39.591819Z" }, "title": "Development of a Persian Syntactic Dependency Treebank", "authors": [ { "first": "Mohammad", "middle": [ "Sadegh" ], "last": "Rasooli", "suffix": "", "affiliation": { "laboratory": "", "institution": "Columbia University", "location": { "settlement": "New York", "region": "NY" } }, "email": "rasooli@cs.columbia.edu" }, { "first": "Manouchehr", "middle": [], "last": "Kouhestani", "suffix": "", "affiliation": { "laboratory": "", "institution": "Modares University Tehran", "location": { "country": "Iran" } }, "email": "m.kouhestani@modares.ac.ir" }, { "first": "Amirsaeid", "middle": [], "last": "Moloodi", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Tehran Tehran", "location": { "country": "Iran" } }, "email": "a.moloodi@ut.ac.ir" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This paper describes the annotation process and linguistic properties of the Persian syntactic dependency treebank. The treebank consists of approximately 30,000 sentences annotated with syntactic roles in addition to morpho-syntactic features. One of the unique features of this treebank is that there are almost 4800 distinct verb lemmas in its sentences making it a valuable resource for educational goals. The treebank is constructed with a bootstrapping approach by means of available tagging and parsing tools and manually correcting the annotations. The data is splitted into standard train, development and test set in the CoNLL dependency format and is freely available to researchers.", "pdf_parse": { "paper_id": "N13-1031", "_pdf_hash": "", "abstract": [ { "text": "This paper describes the annotation process and linguistic properties of the Persian syntactic dependency treebank. The treebank consists of approximately 30,000 sentences annotated with syntactic roles in addition to morpho-syntactic features. One of the unique features of this treebank is that there are almost 4800 distinct verb lemmas in its sentences making it a valuable resource for educational goals. The treebank is constructed with a bootstrapping approach by means of available tagging and parsing tools and manually correcting the annotations. The data is splitted into standard train, development and test set in the CoNLL dependency format and is freely available to researchers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The process of manually annotating linguistic data from a huge amount of naturally occuring texts is a very expensive and time consuming task. Due to the recent success of machine learning methods and the rapid growth of available electronic texts, language processing tasks have been facilitated greatly. Considering the value of annotated data, a great deal of budget has been allotted to creating such data.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction 1", "sec_num": "1" }, { "text": "Among all linguistic datasets, treebanks play an important role in the natural language processing tasks especially in parsing because of its applica-tions in tasks such as machine translation. Dependency treebanks are collections of sentences with their corresponding dependency trees. In the last decade, many dependency treebanks have been developed for a large number of languages. There are at least 29 languages for which at least one dependency treebank is available (Zeman et al., 2012) . Dependency trees are much more similar to the human understanding of language and can easily represent the free word-order nature of syntactic roles in sentences (K\u00fcbler et al., 2009) .", "cite_spans": [ { "start": 474, "end": 494, "text": "(Zeman et al., 2012)", "ref_id": "BIBREF25" }, { "start": 659, "end": 680, "text": "(K\u00fcbler et al., 2009)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction 1", "sec_num": "1" }, { "text": "Persian is a language with about 110 million speakers all over the world (Windfuhr, 2009 ), yet in terms of the availability of teaching materials and annotated data for text processing, it is undoubtedly a low-resourced language. The need for more language teaching materials together with an everincreasing need for Persian-language data processing has been the incentive for the inception of our project which has defined the development of the syntactic treebank of Persian as its ultimate aim. In this paper, we review the process of creating the Persian syntactic treebank based on dependency grammar. In this treebank, approximately 30,000 sentences from contemporary Persian-language texts are manually tokenized and annotated at morphological and syntactic levels. One valuable aspect of the treebank is its containment of near 5000 distinct verb lemmas in its sentences making it a good resource for educational goals. The dataset is developed after the creation of the syntactic valency lexicon of Persian verbs (Rasooli et al., 2011c) . This treebank is developed with a bootstrapping approach by automatically building dependency trees based on the (b) A simple non-projective dependency tree for a Persian sentence: \"It is based on that\".", "cite_spans": [ { "start": 73, "end": 88, "text": "(Windfuhr, 2009", "ref_id": null }, { "start": 1023, "end": 1046, "text": "(Rasooli et al., 2011c)", "ref_id": "BIBREF22" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction 1", "sec_num": "1" }, { "text": "Figure 1: Examples of Persian sentences with the dependency-based syntactic trees. 1(a) and 1(b) are examples of a projective and a non-projective dependency tree, respectively. The first lines show the original words in Persian. The pronunciation and their meanings are shown in the second line and the third line respectively. In the fourth line, the part of speech (POS) tags of the words are presented. Note that the words are written from right to left (the direction of Perso-Arabic script). The dependency relations are described in Table 2 . The relation is shown with an arc pointing from the head to the dependent. previous annotated trees. In the next step, automatic annotation is corrected manually.", "cite_spans": [], "ref_spans": [ { "start": 540, "end": 547, "text": "Table 2", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Introduction 1", "sec_num": "1" }, { "text": "The remainder of this paper is as follows. In Section 2, we briefly review the challenges in Persian language processing. In Sections 3 and 4, the details about the annotation process, linguistic and statistical information about the data and the annotator agreement are reported. In Section 5, the conclusion and suggestions for future research are presented.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction 1", "sec_num": "1" }, { "text": "Persian is an Indo-European language that is written in Arabic script. There are lots of problems in its orthography such as encoding problems, hidden diacritics and writing standards (Kashefi et al., 2010) . A number of challenges such as the free or- der of words, the existence of colloquial texts, the pro-drop nature of the Persian language and its complex inflections (Shamsfard, 2011) in addition to the lack of efficient annotated linguistic data have made the processing of Persian texts very difficult; e.g. there are more than 100 conjugates and 2800 declensions for some word forms in Persian (Rasooli et al., 2011b), some words in the Persian language do not have a clear word category (i.e. the lexical category \"mismatch\") (Karimi-Doostan, 2011a) and many compound verbs (complex predicates) can be separable (i.e. the non-verbal element may be separated from the verbal element by one or more other words) (Karimi-Doostan, 2011b).", "cite_spans": [ { "start": 184, "end": 206, "text": "(Kashefi et al., 2010)", "ref_id": "BIBREF14" }, { "start": 374, "end": 391, "text": "(Shamsfard, 2011)", "ref_id": "BIBREF24" } ], "ref_spans": [], "eq_spans": [], "section": "Persian Language Processing Challenges", "sec_num": "2" }, { "text": "After the development of the Bijankhan corpus (Bijankhan, 2004) with the annotation of word categories, other kinds of datasets have been created to address the need for Persian language processing. Among them, a Persian parser based on link grammar (Dehdari and Lonsdale, 2008) , a computational grammar based on GPSG , syntactic treebank based on HPSG (Ghayoomi, 2012) and Uppsala dependency treebank (Seraji et al., 2012) (c) A simple dependency tree for a Persian sentence with a an inflected form of a prefixed verb \"I will not return to this house.\". The word is the prefix, the word is the auxiliary for the future and the word is the main verb. Notice that the prefix is attached to the auxiliary without any space and the remaining part of the verb is separated by a space. 3 Persian Dependency Treebank", "cite_spans": [ { "start": 46, "end": 63, "text": "(Bijankhan, 2004)", "ref_id": "BIBREF2" }, { "start": 250, "end": 278, "text": "(Dehdari and Lonsdale, 2008)", "ref_id": "BIBREF8" }, { "start": 403, "end": 424, "text": "(Seraji et al., 2012)", "ref_id": "BIBREF23" } ], "ref_spans": [], "eq_spans": [], "section": "Persian Language Processing Challenges", "sec_num": "2" }, { "text": "With the creation of the Virastyar spell checker software (Kashefi et al., 2010) , many open-source libraries were released for Persian word processing such as POS tagging, encoding refinement, tokenization, etc. Regarding the need for syntactic analysis of Persian texts, we decided to prepare a valuable linguistic data infrastructure for Persian syntax. In the first step, there was a need for choosing from the existing theories of grammar that best suits Persian. Among grammatical theories, we decided to choose the dependency grammar. In dependency grammar, syntactic relations are shown with dependencies between the words. In computational dependency grammar, each word has one head and the head of the sentence is the dependent of an artificial root word (K\u00fcbler et al., 2009) . A sample dependency tree is shown in Figure 1 (a) for a Persian sentence. Note that Persian sentences are written from right to left.", "cite_spans": [ { "start": 58, "end": 80, "text": "(Kashefi et al., 2010)", "ref_id": "BIBREF14" }, { "start": 765, "end": 786, "text": "(K\u00fcbler et al., 2009)", "ref_id": "BIBREF15" } ], "ref_spans": [ { "start": 826, "end": 834, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Motivation", "sec_num": "3.1" }, { "text": "There are several reasons for the preference of dependency grammar to grammars such as phrasebased structure grammars. Although in both of the representations, one can show the syntactic analysis of a sentence, dependency representation has the power to account for the free word order of many languages such as Turkish (Oflazer et al., 2003) and Czech (Hajic, 1998) and also Persian. As an example, a sample non-projective dependency tree for the Persian language is shown in Figure 1(b) . The recent advances in very fast dependency parsing models (e.g. (Nivre, 2009; Bohnet and Nivre, 2012) ), has made the syntactic processing task very popular in the recent decade.", "cite_spans": [ { "start": 320, "end": 342, "text": "(Oflazer et al., 2003)", "ref_id": "BIBREF19" }, { "start": 347, "end": 366, "text": "Czech (Hajic, 1998)", "ref_id": null }, { "start": 556, "end": 569, "text": "(Nivre, 2009;", "ref_id": "BIBREF18" }, { "start": 570, "end": 593, "text": "Bohnet and Nivre, 2012)", "ref_id": "BIBREF3" } ], "ref_spans": [ { "start": 477, "end": 488, "text": "Figure 1(b)", "ref_id": null } ], "eq_spans": [], "section": "Motivation", "sec_num": "3.1" }, { "text": "In the Persian language, in addition to the abundance of crossings of the arcs, another problem occurs with compound verbs and verbs in the progressive aspect: compound and progressive verbs are multi-word expressions that may be separated depending on the context. Persian compound verbs consist of a light verb and a non-verbal element and the non-verbal element can be a noun, an adjective (in rare cases) or a sequence of a preposition and a noun (Dabir-Moghaddam, 1997) . In addition, the nonverbal elements can also be inflected. The distance between the nonverbal element and the light verb on the one hand and the possibility of the nonverbal element being inflected on the other hand have made the task of compound verb identification very difficult. For example, in Bijankhan (Peykare) corpus (Bijankhan et al., 2011) , approximately 9% of nonverbal elements of compound verbs are placed away from the light verb for the compound verbs with the light verb /kaerdaen/ (to do) (Rasooli et al., 2011a). A group of Persian progressive verbs are composed of two words, the first being the simple past or the simple present form derived from the infinitive /dAStaen/ (to have) and the second being the past continuous or the present continuous form of the main verb. The first verb (an auxiliary) agrees with the second in number and person. The problem is that the progressive auxiliary can be away from the main verb. The sample trees with compound verbs and progressive auxiliary verbs are shown in Figures 3(a) and 3(b) respectively.", "cite_spans": [ { "start": 451, "end": 474, "text": "(Dabir-Moghaddam, 1997)", "ref_id": "BIBREF6" }, { "start": 803, "end": 827, "text": "(Bijankhan et al., 2011)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Motivation", "sec_num": "3.1" }, { "text": "In this treebank, we followed the format of the CoNLL tab-separated format for dependency parsing (Buchholz and Marsi, 2006) . In addition to the lemma, we annotated part of speech tags (both coarse and fine grained) and person, number and tense-mood-aspect (only for verbs) of words in sentences. The details of the part of speech tags and other morphosyntactic features and dependency relations are shown in Tables 1 and 2, respectively. The part of speech tag set in this treebank is not the same as that of Bijankhan (Peykare) corpus (Bijankhan et al., 2011) and it is essential to convert the tagset in Peykare corpus to the tagset in this treebank, in order to use both datasets 2 . We also tried to use the writing standard of the Academy of Persian Language and Literature except for the cases where for a word there were several standards all of which were used in Persian written texts (e.g.", "cite_spans": [ { "start": 98, "end": 124, "text": "(Buchholz and Marsi, 2006)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Representation and Dependency Relation", "sec_num": "3.2" }, { "text": "/PAntSe/ (whatever)). We also prepared two representations for objects accompanied by the accusative case marker. In the first representation (done manually), we assume the accusative case marker /rA/ as the head of the two-2 It is important to note that the conversion between the coarse-grained POS tags is straightforward and does not need any special effort. word sequence object plus rA. The second representation, that is the automatic conversion of the first, is the reverse order of the first one in which the accusative case marker is the dependent of the direct object and the direct object is considered as the head of the aforementioned sequence. In the first representation, objects are much easier to find by the parser (because of the uniqueness of the accusative case marker in Persian and less distance of it from the verb as its head) but it may increase the number of non-projective arcs to the syntactic tree. We prepared both of the representations in two separate data packs. A sample comparison between the two structures is shown in Figure 4 .", "cite_spans": [], "ref_spans": [ { "start": 1057, "end": 1065, "text": "Figure 4", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "and", "sec_num": null }, { "text": "In the treebank, all words are single word forms (without spaces). There is only one exception for simple verb inflections where even multi-word tokens of simple verbs are shown as only one unit. The reason is that for many cases such as the case of inflections for prefixed verbs it is more straightforward to analyze the whole part instead of analyzing each There is also another feature for representing the typographical connectedness of words that are separated into two or more tokens with the values ISO (isolated word), NXT (attached to the next token) and PRV (attached to the previous token).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "and", "sec_num": null }, { "text": "part separately 3 . In Table 3 , possible types of the Persian verb inflections are shown. As seen in Table 3 , 6 forms of 14 inflection types of Persian verbs are multi-word tokens and for passive verbs they may be composed of more words than their active counterparts (since for passive verbs an auxiliary form derived from the infinitive /Sodaen/ is used). In Figure 3 (c), a sample tree with a multi-token pre-fixed verb is shown. As shown in the case of colored tokens, it seems more beneficial to put all morphemes of the word together before parsing. Furthermore, with the available Persian verb analyzer it is very easy to first preprocess the verbs 4 . ", "cite_spans": [], "ref_spans": [ { "start": 23, "end": 30, "text": "Table 3", "ref_id": "TABREF8" }, { "start": 102, "end": 110, "text": "Table 3", "ref_id": "TABREF8" }, { "start": 364, "end": 372, "text": "Figure 3", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "and", "sec_num": null }, { "text": "The annotation process consists of several consecutive steps. In Figure 2 , a summary of the bootstrapping approach in the annotation process is shown. At first, a collection of independent sentences have been collected randomly from the web. For the first 5000 sentences, we crawled Persian news texts and randomly sampled the sentences. For the remaining sentences, we first listed the absent verb lemmas in the 5000 sentences based on the verb list ex-tracted from the valency lexicon of Persian verbs (Rasooli et al., 2011c) and collected random sentences that included the absent verb lemmas in their words. We listed all possible inflections and per each verb lemma, sampled at most 8 sentences from the web. These sentences had to contain at least one present tense, one past tense, one passive voice and one future tense inflection unless we could not find them and were obliged to reduce the number. The sentences were not shortened and were kept with their original length and words. Finally, we manually removed sentences containing colloquial words. However, we did not remove loan words or cases of code-switching between latin-script words and Persian words in the sentences. The raw sentences were fed to the encoding and spell checking module. After spell correction, all sentences were tokenized and tagged with part of speech tags. All of the word processing steps were done using Virastyar library (Kashefi et al., 2010) . After tokenization and POS tagging, the tokenized sentences were fed to the Persian verb analyzing tool (Rasooli et al., 2011a) . In the next step, the preprocessed sentences were given to the dependency parser. We used MST parser (Mc-Donald et al., 2005) for parsing the sentences. In the final step, annotators corrected the errors of tokenization, POS tagging and parsing. In about every one to two weeks, the parser model was updated by training on the new version of the treebank. This process lasted 9 months and the number of annotators increased by time to speed up the process. In the first 6 months, we used 8 annotators and for the next 5 months, we hired 6 more annotators to speed up the process. The annotators and linguistic experts consisted of 1 PhD graduate (linguistics), 4 PhD candidates (linguistics), and 9 MA graduates or graduate students (7 linguistics, 1 Persian language and literature and 1 computational linguistics). All of the annotators were native Persian speakers.", "cite_spans": [ { "start": 1417, "end": 1439, "text": "(Kashefi et al., 2010)", "ref_id": "BIBREF14" }, { "start": 1546, "end": 1569, "text": "(Rasooli et al., 2011a)", "ref_id": "BIBREF20" }, { "start": 1673, "end": 1697, "text": "(Mc-Donald et al., 2005)", "ref_id": null } ], "ref_spans": [ { "start": 65, "end": 73, "text": "Figure 2", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Annotation Process", "sec_num": "3.3" }, { "text": "After finalizing the annotation of all raw sentences, we applied a rule-based potential error finder to find the potentially erroneous sentences. The rules were gradually collected in the process of the annotation by the annotators. All the potentially erroneous sentences were given to the annotators to be checked for potential errors. In Section 4.1, the statistics about the changes after the correction is reported. One of the main reasons for the double checking phase in the process is that based on our manual investigations of the annotations, we found some inevitable mistakes by annotators that could be solved by manual rules. Mistakes such as scrolling the drop-down list unintentionally and changing the part of speech tag or dependency relation and mistakes caused by tiredness and lack of concentration in addition to some of the changes of the linguistic conventions in the annotation. Since all cases of dependency relations in this treebank may be usually either a left-branching relation or a right-branching one and most of the relations are restricted to certain types of parts of speech, it is easy to capture the potential errors in the annotations based on the rules mentioned and to keep track of the changes in the linguistic conventions by searching the cues for those conventions (most of the changed conventions were made to very rare relations in the syntactic structure).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Annotation Process", "sec_num": "3.3" }, { "text": "In (Dligach and Palmer, 2011) , it is concluded that although doubly annotated corpora are more reliable, annotating more sentences only once is more beneficial; i.e. annotating each sentence only once is less time-consuming and more cost-effective. We annotated all the sentences only once (with an additional checking phase) except for the 5% of the sentences in order to estimate the quality of our linguistic conventions and agreement among the annotators. The statistics about the annotators agreement is reported in Section 4.1.", "cite_spans": [ { "start": 3, "end": 29, "text": "(Dligach and Palmer, 2011)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Annotation Process", "sec_num": "3.3" }, { "text": "Finally, 29,982 sentences were manually annotated. The details about the statistics is shown in Table 4. It is worth mentioning that 39.24% of the words in the treebank are tagged as noun, 12.62% as verb, 11.64% as preposition and 7.39% as adjective. The most frequent dependency relations are post-dependent (15.08%) and Ezafeh construction (10.17%). As shown in Table 5 , the number of nonprojective arcs in the second representation is a little bit less than the first. As mentioned earlier, the main reason is the dependencies between the direct object and words after the accusative case marker such as the example in Figure 4 . The change percentage after the correction of the potential errors is shown in Table 6 . It seems that the rules for finding the poten- tial errors were useful for correcting the errors.", "cite_spans": [], "ref_spans": [ { "start": 364, "end": 371, "text": "Table 5", "ref_id": "TABREF11" }, { "start": 623, "end": 631, "text": "Figure 4", "ref_id": "FIGREF2" }, { "start": 713, "end": 720, "text": "Table 6", "ref_id": null } ], "eq_spans": [], "section": "Statistics of the Treebank", "sec_num": "4" }, { "text": "The statistics about the agreement among the annotators is shown in Table 7 . We can also use the Kappa (Cohen, 1960) to measure the quality of the annotation based on the agreement among the annotators (pr(a) in Eq. 1) and the expected agreement or probability of chance (pr(e) in Eq. 1). If we consider the accuracy of the parser on the raw text without gold POS tags (approximately 75% labeled and 80% unlabeled accuracy based on our experience during the bootstrapping) and the POS tagger that we used during the annotation process (approximately 94%) as the probability of chance, we see that for all of the tasks in Table 7 , the quality of the annotaion is more than 0.81 and is considered as almost perfect according to (Landis and Koch, 1977) . k = pr(a) \u2212 pr(e) 1 \u2212 pr(e) (1)", "cite_spans": [ { "start": 104, "end": 117, "text": "(Cohen, 1960)", "ref_id": "BIBREF5" }, { "start": 728, "end": 751, "text": "(Landis and Koch, 1977)", "ref_id": "BIBREF16" } ], "ref_spans": [ { "start": 68, "end": 75, "text": "Table 7", "ref_id": null }, { "start": 622, "end": 629, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Annotators Agreement", "sec_num": "4.1" }, { "text": "As mentioned earlier, Persian is a language with its own challenges. We tried to overcome some of those challenges by preparing valuable linguistic Changes to Unlabeled Relations 4.91% Changes to Labeled Relations 6.29% Changes to POS Tags 4.23% Table 6 : Statistics about changes in the treebank after the manual correction of the potential errors.", "cite_spans": [], "ref_spans": [ { "start": 246, "end": 253, "text": "Table 6", "ref_id": null } ], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "Unlabeled Relations 97.06% Labeled Relations 95.32% POS Tags 98.93% Table 7 : Statistics about agreements among the annotators.", "cite_spans": [], "ref_spans": [ { "start": 68, "end": 75, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "datasets 5 . In addition to the preparation of the treebank, we prepared some useful desktop and webbased tools for searching in the dataset, obtaining statistics and viewing syntactic structures graphically. We hope to report more details about the linguistic aspects and the findings of the project in addition to our detailed experiments on the parsing task in future publications. We believe that this treebank is just the very first step to satisfy the need for Persian language processing. Our future aim is to add a semantic level to the annotation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "This research is done while working in Dadegan Research Group, Supreme Council of Information and Communications Technology (SCICT), Tehran, Iran. The project is fully funded by SCICT.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "In(Seraji et al., 2012), multi-token verbs are considered as separate words.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "If it is needed to respect the exact format of CoNLL, spaces between the verb tokes should be replaced by a character such as underscore. Regarding the special fine-grained morphological tags for the verb such as tense-mood-aspect, it is also straightforward to separate all of the multi-word verbs and add new dependency relations between their words.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "A comprehensive description of the syntactic relations and morphosyntactic features is reported in the treebank official report (Dadegan Research Group, 2012) in the treebank package both in Persian and English.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "The project is funded by Iran Supreme Council of Information and Communication Technology (SCICT). We really appreciate the linguists who helped us in annotating: Farzaneh Bakhtiary, Parinaz Dadras, Maryam Faal-Hamedanchi, Saeedeh Ghadrdoost-Nakhchi, Mostafa Mahdavi, Azadeh Mirzaei, Sahar Oulapoor, Neda Poormorteza-Khameneh, Morteza Rezaei-Sharifabadi, Sude Resalatpoo, Akram Shafie, and Salimeh Zamani; and the programmers who helped us in the process of the development of the treebank: Seyed Mahdi Hoseini, Alireza Noorian, Yasser Souri, and Mohsen Hossein-Alizadeh; and also our colleagues who helped us find linguistic sources from the web: Azadeh Abbasi Abyaneh, Shima Zamanpoor, Narmin Ghaderi, Houra Nouri and Seyedeh Maneli Hashemian; and other colleagues especially Mahdi Behniafar. We thank Nizar Habash for his support of this paper and Weiwei Guo and three anonymous reviewers for their useful comments on the paper.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "A computational grammar for Persian based on GPSG. Language Resources and Evaluation", "authors": [ { "first": "Mohammad", "middle": [], "last": "Bahrani", "suffix": "" }, { "first": "Hossein", "middle": [], "last": "Sameti", "suffix": "" }, { "first": "Mehdi Hafezi", "middle": [], "last": "Manshadi", "suffix": "" } ], "year": 2011, "venue": "", "volume": "45", "issue": "", "pages": "387--408", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mohammad Bahrani, Hossein Sameti, and Mehdi Hafezi Manshadi. 2011. A computational grammar for Per- sian based on GPSG. Language Resources and Eval- uation, 45(4):387-408.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Lessons from building a persian written corpus: Peykare. Language resources and evaluation", "authors": [ { "first": "Mahmood", "middle": [], "last": "Bijankhan", "suffix": "" }, { "first": "Javad", "middle": [], "last": "Sheykhzadegan", "suffix": "" }, { "first": "Mohammad", "middle": [], "last": "Bahrani", "suffix": "" }, { "first": "Masood", "middle": [], "last": "Ghayoomi", "suffix": "" } ], "year": 2011, "venue": "", "volume": "45", "issue": "", "pages": "143--164", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mahmood Bijankhan, Javad Sheykhzadegan, Moham- mad Bahrani, and Masood Ghayoomi. 2011. Lessons from building a persian written corpus: Peykare. Lan- guage resources and evaluation, 45(2):143-164.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "The role of the corpus in writing a grammar: An introduction to a software", "authors": [ { "first": "Mahmood", "middle": [], "last": "Bijankhan", "suffix": "" } ], "year": 2004, "venue": "Iranian Journal of Linguistics", "volume": "19", "issue": "2", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mahmood Bijankhan. 2004. The role of the corpus in writing a grammar: An introduction to a software. Ira- nian Journal of Linguistics, 19(2).", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "A transitionbased system for joint part-of-speech tagging and labeled non-projective dependency parsing", "authors": [ { "first": "Bernd", "middle": [], "last": "Bohnet", "suffix": "" }, { "first": "Joakim", "middle": [], "last": "Nivre", "suffix": "" } ], "year": 2012, "venue": "EMNLP-CoNLL", "volume": "", "issue": "", "pages": "1455--1465", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bernd Bohnet and Joakim Nivre. 2012. A transition- based system for joint part-of-speech tagging and la- beled non-projective dependency parsing. In EMNLP- CoNLL, pages 1455-1465.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "CoNLL-X shared task on multilingual dependency parsing", "authors": [ { "first": "Sabine", "middle": [], "last": "Buchholz", "suffix": "" }, { "first": "Erwin", "middle": [], "last": "Marsi", "suffix": "" } ], "year": 2006, "venue": "Proceeding of the Tenth Conforence on Computational Natural Language Learning (CoNLL)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sabine Buchholz and Erwin Marsi. 2006. CoNLL-X shared task on multilingual dependency parsing. In Proceeding of the Tenth Conforence on Computational Natural Language Learning (CoNLL).", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "A coefficient of agreement for nominal scales. Educational and psychological measurement", "authors": [ { "first": "Jacob", "middle": [], "last": "Cohen", "suffix": "" } ], "year": 1960, "venue": "", "volume": "20", "issue": "", "pages": "37--46", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jacob Cohen. 1960. A coefficient of agreement for nom- inal scales. Educational and psychological measure- ment, 20(1):37-46.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Compound verbs in Persian", "authors": [ { "first": "Mohammad", "middle": [], "last": "Dabir-Moghaddam", "suffix": "" } ], "year": 1997, "venue": "Studies in the Linguistic Sciences", "volume": "27", "issue": "2", "pages": "25--59", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mohammad Dabir-Moghaddam. 1997. Compound verbs in Persian. Studies in the Linguistic Sciences, 27(2):25-59.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Persian Dependency Treebank, Annotation manual and user guide. Supreme Council of Information and Communication Technology (SCICT)", "authors": [ { "first": "Dadegan", "middle": [], "last": "Research Group", "suffix": "" } ], "year": 2012, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dadegan Research Group. 2012. Persian Depen- dency Treebank, Annotation manual and user guide. Supreme Council of Information and Communication Technology (SCICT), Tehran, Iran.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "A link grammar parser for Persian", "authors": [ { "first": "Jon", "middle": [], "last": "Dehdari", "suffix": "" }, { "first": "Deryle", "middle": [], "last": "Lonsdale", "suffix": "" } ], "year": 2008, "venue": "Aspects of Iranian Linguistics", "volume": "1", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jon Dehdari and Deryle Lonsdale. 2008. A link grammar parser for Persian. Aspects of Iranian Linguistics, 1.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Reducing the need for double annotation", "authors": [ { "first": "Dmitriy", "middle": [], "last": "Dligach", "suffix": "" }, { "first": "Martha", "middle": [], "last": "Palmer", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the 5th Linguistic Annotation Workshop", "volume": "", "issue": "", "pages": "65--69", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dmitriy Dligach and Martha Palmer. 2011. Reducing the need for double annotation. In Proceedings of the 5th Linguistic Annotation Workshop, pages 65-69.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Bootstrapping the development of an HPSG-based treebank for Persian. Linguistic Issues in Language Technology", "authors": [ { "first": "Masood", "middle": [], "last": "Ghayoomi", "suffix": "" } ], "year": 2012, "venue": "", "volume": "7", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Masood Ghayoomi. 2012. Bootstrapping the develop- ment of an HPSG-based treebank for Persian. Lin- guistic Issues in Language Technology, 7(1).", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Building a syntactically annotated corpus: The Prague dependency treebank. Issues of valency and meaning", "authors": [], "year": 1998, "venue": "", "volume": "", "issue": "", "pages": "106--132", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jan Hajic. 1998. Building a syntactically annotated cor- pus: The Prague dependency treebank. Issues of va- lency and meaning, pages 106-132.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Lexical categories in Persian", "authors": [ { "first": "Gholamhossein", "middle": [], "last": "Karimi-Doostan", "suffix": "" } ], "year": 2011, "venue": "Lingua", "volume": "121", "issue": "2", "pages": "207--220", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gholamhossein Karimi-Doostan. 2011a. Lexical cate- gories in Persian. Lingua, 121(2):207-220.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Separability of light verb constructions in Persian", "authors": [ { "first": "Gholamhossein", "middle": [], "last": "Karimi-Doostan", "suffix": "" } ], "year": 2011, "venue": "Studia Linguistica", "volume": "65", "issue": "1", "pages": "70--95", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gholamhossein Karimi-Doostan. 2011b. Separability of light verb constructions in Persian. Studia Linguistica, 65(1):70-95.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Automatic Spell Checking in Persian Language. Supreme Council of Information and Communication Technology (SCICT)", "authors": [ { "first": "Omid", "middle": [], "last": "Kashefi", "suffix": "" }, { "first": "Mitra", "middle": [], "last": "Nasri", "suffix": "" }, { "first": "Kamyar", "middle": [], "last": "Kanani", "suffix": "" } ], "year": 2010, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Omid Kashefi, Mitra Nasri, and Kamyar Kanani. 2010. Automatic Spell Checking in Persian Language. Supreme Council of Information and Communication Technology (SCICT), Tehran, Iran.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Dependency Parsing. Synthesis Lectures on Human Language Technologies", "authors": [ { "first": "Sandra", "middle": [], "last": "K\u00fcbler", "suffix": "" }, { "first": "Ryan", "middle": [], "last": "Mcdonald", "suffix": "" }, { "first": "Joakim", "middle": [], "last": "Nivre", "suffix": "" } ], "year": 2009, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sandra K\u00fcbler, Ryan McDonald, and Joakim Nivre. 2009. Dependency Parsing. Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "The measurement of observer agreement for categorical data", "authors": [ { "first": "Richard", "middle": [], "last": "Landis", "suffix": "" }, { "first": "Gary", "middle": [ "G" ], "last": "Koch", "suffix": "" } ], "year": 1977, "venue": "Biometrics", "volume": "", "issue": "", "pages": "159--174", "other_ids": {}, "num": null, "urls": [], "raw_text": "Richard Landis and Gary G. Koch. 1977. The mea- surement of observer agreement for categorical data. Biometrics, pages 159-174.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Online large-margin training of dependency parsers", "authors": [ { "first": "Ryan", "middle": [], "last": "Mcdonald", "suffix": "" }, { "first": "Koby", "middle": [], "last": "Crammer", "suffix": "" }, { "first": "Fernando", "middle": [], "last": "Pereira", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics (ACL)", "volume": "", "issue": "", "pages": "91--98", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ryan McDonald, Koby Crammer, and Fernando Pereira. 2005. Online large-margin training of dependency parsers. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics (ACL), pages 91-98, Sydney, Australia.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Non-projective dependency parsing in expected linear time", "authors": [ { "first": "Joakim", "middle": [], "last": "Nivre", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing", "volume": "", "issue": "", "pages": "351--359", "other_ids": {}, "num": null, "urls": [], "raw_text": "Joakim Nivre. 2009. Non-projective dependency parsing in expected linear time. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing, pages 351-359.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Building a Turkish treebank", "authors": [ { "first": "Kemal", "middle": [], "last": "Oflazer", "suffix": "" }, { "first": "Bilge", "middle": [], "last": "Say", "suffix": "" }, { "first": "G\u00f6khan", "middle": [], "last": "Dilek Zeynep Hakkani-T\u00fcr", "suffix": "" }, { "first": "", "middle": [], "last": "T\u00fcr", "suffix": "" } ], "year": 2003, "venue": "Treebanks", "volume": "", "issue": "", "pages": "261--277", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kemal Oflazer, Bilge Say, Dilek Zeynep Hakkani-T\u00fcr, and G\u00f6khan T\u00fcr. 2003. Building a Turkish treebank. Treebanks, pages 261-277.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Unsupervised identification of Persian compound verbs", "authors": [ { "first": "Mohammad", "middle": [], "last": "Sadegh Rasooli", "suffix": "" }, { "first": "Heshaam", "middle": [], "last": "Faili", "suffix": "" }, { "first": "Behrouz", "middle": [], "last": "Minaei-Bidgoli", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the Mexican international conference on artificial intelligence (MICAI)", "volume": "", "issue": "", "pages": "394--406", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mohammad Sadegh Rasooli, Heshaam Faili, and Behrouz Minaei-Bidgoli. 2011a. Unsupervised iden- tification of Persian compound verbs. In Proceedings of the Mexican international conference on artificial intelligence (MICAI), pages 394-406, Puebla, Mexico.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Effect of adaptive spell checking in Persian", "authors": [ { "first": "Mohammad", "middle": [], "last": "Sadegh Rasooli", "suffix": "" }, { "first": "Omid", "middle": [], "last": "Kashefi", "suffix": "" }, { "first": "Behrouz", "middle": [], "last": "Minaei-Bidgoli", "suffix": "" } ], "year": 2011, "venue": "7th International Conference on Natural Language Processing and Knowledge Engineering", "volume": "", "issue": "", "pages": "161--164", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mohammad Sadegh Rasooli, Omid Kashefi, and Behrouz Minaei-Bidgoli. 2011b. Effect of adaptive spell checking in Persian. In 7th International Confer- ence on Natural Language Processing and Knowledge Engineering (NLP-KE), pages 161-164, Tokushima, Japan.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "A syntactic valency lexicon for Persian verbs: The first steps towards Persian dependency treebank", "authors": [ { "first": "Mohammad", "middle": [], "last": "Sadegh Rasooli", "suffix": "" }, { "first": "Amirsaeid", "middle": [], "last": "Moloodi", "suffix": "" }, { "first": "Manouchehr", "middle": [], "last": "Kouhestani", "suffix": "" }, { "first": "Behrouz", "middle": [], "last": "Minaei-Bidgoli", "suffix": "" } ], "year": 2011, "venue": "5th Language & Technology Conference (LTC): Human Language Technologies as a Challenge for Computer Science and Linguistics", "volume": "", "issue": "", "pages": "227--231", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mohammad Sadegh Rasooli, Amirsaeid Moloodi, Manouchehr Kouhestani, and Behrouz Minaei- Bidgoli. 2011c. A syntactic valency lexicon for Persian verbs: The first steps towards Persian de- pendency treebank. In 5th Language & Technology Conference (LTC): Human Language Technologies as a Challenge for Computer Science and Linguistics, pages 227-231, Pozna\u0144, Poland.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Bootstrapping a Persian dependency treebank. Linguistic Issues in Language Technology", "authors": [ { "first": "Mojgan", "middle": [], "last": "Seraji", "suffix": "" }, { "first": "Be\u00e1ta", "middle": [], "last": "Magyesi", "suffix": "" }, { "first": "Joakim", "middle": [], "last": "Nivre", "suffix": "" } ], "year": 2012, "venue": "", "volume": "7", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mojgan Seraji, Be\u00e1ta Magyesi, and Joakim Nivre. 2012. Bootstrapping a Persian dependency treebank. Lin- guistic Issues in Language Technology, 7(1).", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Challenges and open problems in Persian text processing", "authors": [ { "first": "Mehrnoosh", "middle": [], "last": "Shamsfard", "suffix": "" } ], "year": 2009, "venue": "5th Language & Technology Conference (LTC): Human Language Technologies as a Challenge for Computer Science and Linguistics", "volume": "", "issue": "", "pages": "65--69", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mehrnoosh Shamsfard. 2011. Challenges and open problems in Persian text processing. In 5th Language & Technology Conference (LTC): Human Language Technologies as a Challenge for Computer Science and Linguistics, pages 65-69, Pozna\u0144, Poland. Gernot Windfuhr. 2009. The Iranian Languages. Rout- ledge.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Hamledt: To parse or not to parse", "authors": [ { "first": "Daniel", "middle": [], "last": "Zeman", "suffix": "" }, { "first": "David", "middle": [], "last": "Mare\u010dek", "suffix": "" }, { "first": "Martin", "middle": [], "last": "Popel", "suffix": "" }, { "first": "Loganathan", "middle": [], "last": "Ramasamy", "suffix": "" }, { "first": "Zedn\u00e9c\u01e9", "middle": [], "last": "Jan\u0161t\u011bp\u00e1nek", "suffix": "" }, { "first": "", "middle": [], "last": "Zabokrtsk\u1ef3", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the Eighth Conference on International Language Resources and Evaluation (LREC'12)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Daniel Zeman, David Mare\u010dek, Martin Popel, Lo- ganathan Ramasamy, Jan\u0160t\u011bp\u00e1nek, Zedn\u00e9c\u01e9 Zabokrtsk\u1ef3, and Jan Haji\u010d. 2012. Hamledt: To parse or not to parse. In Proceedings of the Eighth Conference on International Language Resources and Evaluation (LREC'12), Istanbul, Turkey.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "uris": null, "num": null, "text": "Diagram of bootstrapping approach in the development of the dependency treebank." }, "FIGREF1": { "type_str": "figure", "uris": null, "num": null, "text": "Examples of Persian sentences with the dependency-based syntactic trees. The format of the representation is the same asFigure 1." }, "FIGREF2": { "type_str": "figure", "uris": null, "num": null, "text": "A sample sentence with two kinds of representations of object-verb relation. The first one is done manually and the second automatically by converting the dependencies in the first representation." }, "TABREF2": { "type_str": "table", "text": "are the efforts to satisfy the need for syntactic processing in the Persian language. I am going from this house\". The PROG is a relation between a verb and its progressive auxiliary.", "num": null, "content": "
root
NVE
NPP
POSDEPNPOSTMOD
root
kaerdaem to bA zijAdi sohbaethAje
did (1st, sing) you with a lot speaking(s)
VPR PP ADJN
(a) A simple dependency tree with compound verb
for a Persian sentence: \"I spoke with you a lot\".
The NVE is a relation between a light verb and its
nonverbal element. As shown in the tree, not only
the nonverbal element is not near the light verb, but
also it is inflected for plurality (i.e. speakings).
root
PROG
VPP
POSDEP
NPREMOD
root
miraevaemxAne Pin PaezdAraem
go (pres.cont., 1st sing.) house this from have (pres., 1st sing.)
VN PREM PPV
(b) A simple dependency tree for a Persian sentence with a pro-
root VPP gressive auxiliary: \"root POSDEP NPREMOD
barnaexAhaem gaeSt xAne Pin be
return (future, neg., 1st sing.) house this to
VN PREM PP
", "html": null }, "TABREF4": { "type_str": "table", "text": "Morphosyntactic features in the Persian dependency treebank", "num": null, "content": "
CPOSFPOSPersonNumberTMA
AJP (positive)
ADJ (adjective)AJCM (comparitive)
AJSUP (superlative)
ADR (address term)PRADR (pre-noun)
POSADR (post-noun)
ADV (adverb)SADV (genuine)
CONJ (coordinating conjunction)CONJ (conjunction)
IDEN (title)IDEN (title)
N (noun)ANM (animate) IANM (inanimate)SING (singular) PLUR (plural)
PART (particle)PART (particle)
POSNUM (post-noun modifier)POSNUM (post-noun modifier)
POSTP (postposition)POSTP (postposition)
SEPER (separate personal)
JOPER (enclitic personal)
DEMON (demonstrative)1SING (singular)
PR (pronoun)INTG (interogative)2PLUR (plural)
CREFX (common reflexive)3
UCREFX (noncommon reflexive)
RECPR (reciprocal)
EXAJ (exclamatory)
PREM (pre-modifier)QUAJ (interrogative) DEMAJ (demonstrative)
AMBAJ (ambiguous)
PRENUM (pre-noun numeral)PRENUM (pre-noun numeral)
PREP (preposition)PREP (preposition)
PSUS (pseudo-sentence)PSUS (pseudo-sentence)
PUNC (punctuation)PUNC (punctuation)
ACT (active)1SING (singular) See Table 3
V (verb)PAS (passive)2PLUR (plural)
MOD (modal)3
SUBR (subordinating clause)SUBR (subordinating clause)
", "html": null }, "TABREF5": { "type_str": "table", "text": "", "num": null, "content": "", "html": null }, "TABREF7": { "type_str": "table", "text": "Dependency relations in the Persian dependency treebank", "num": null, "content": "
Tense/Aspect/MoodAbbreviationExamplesxordaen: to eat, 1st, sing.
ImperativeHA/boxor/
Indicative FutureAY/xAhaem xord/
Indicative Imperfective PerfectGNES/mixordePaem/
Indicative Imperfective PluperfectGBES/mixorde budaem/
Indicative Imperfective PreteriteGES/mixordaem/
Indicative PerfectGN/xordePaem/
Indicative PluperfectGB/xorde budaem/
Indicative PresentH/mixoraem /
Indicative PreteriteGS/xordaem/
Subjunctive Imperfective PluperfectGBESE/mixorde bude bASaem/
Subjunctive Imperfective PreteriteGESEL/mixorde bASaem/
Subjunctive PluperfectGBEL/xorde bude bASaem/
Subjunctive PresentHEL/boxoraem/
Subjunctive PreteriteGEL/xorde bASaem/
", "html": null }, "TABREF8": { "type_str": "table", "text": "Tense/Mood/Aspect Types in Persian verbs", "num": null, "content": "", "html": null }, "TABREF10": { "type_str": "table", "text": "Statistics about the frequency of words in the Persian dependency treebank.", "num": null, "content": "
# Non-Projective1st Rep. 2nd Rep.
Number of Arcs122818512
Percent of Arcs2.471.71
Number of Sentences65744838
Percent of Sentences21.9316.14
", "html": null }, "TABREF11": { "type_str": "table", "text": "Statistics about non-projective relations in the Persian dependency treebank for both of the representations.", "num": null, "content": "", "html": null } } } }