| { | |
| "paper_id": "2019", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T14:49:04.148155Z" | |
| }, | |
| "title": "NSURL-2019 Task 7: Named Entity Recognition (NER) in Farsi", | |
| "authors": [ | |
| { | |
| "first": "Nasrin", | |
| "middle": [], | |
| "last": "Taghizadeh", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Tehran", | |
| "location": { | |
| "settlement": "Tehran", | |
| "country": "Iran" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Zeinab", | |
| "middle": [], | |
| "last": "Borhanifard", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Tehran", | |
| "location": { | |
| "settlement": "Tehran", | |
| "country": "Iran" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Melika", | |
| "middle": [], | |
| "last": "Golestani Pour", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Tehran", | |
| "location": { | |
| "settlement": "Tehran", | |
| "country": "Iran" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Mojgan", | |
| "middle": [], | |
| "last": "Farhoodi", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Maryam", | |
| "middle": [], | |
| "last": "Mahmoudi", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Masoumeh", | |
| "middle": [], | |
| "last": "Azimzadeh", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Heshaam", | |
| "middle": [], | |
| "last": "Faili", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Tehran", | |
| "location": { | |
| "settlement": "Tehran", | |
| "country": "Iran" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "NSURL-2019 Task 7 focuses on Named Entity Recognition (NER) in Farsi. This task was chosen to compare different approaches to find phrases that specify Named Entities in Farsi texts, and to establish a standard testbed for future researches on this task in Farsi. This paper describes the process of making training and test data, the list of participating teams (6 teams), and evaluation results of their systems. The best system obtained 85.4% of F 1 score based on phrase-level evaluation on seven classes of NEs including person, organization, location, date, time, money, and percent.", | |
| "pdf_parse": { | |
| "paper_id": "2019", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "NSURL-2019 Task 7 focuses on Named Entity Recognition (NER) in Farsi. This task was chosen to compare different approaches to find phrases that specify Named Entities in Farsi texts, and to establish a standard testbed for future researches on this task in Farsi. This paper describes the process of making training and test data, the list of participating teams (6 teams), and evaluation results of their systems. The best system obtained 85.4% of F 1 score based on phrase-level evaluation on seven classes of NEs including person, organization, location, date, time, money, and percent.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Named Entity Recognition (NER) is defined as the task of identifying relevant nouns such as persons, products, and genes which are mentioned in a text. NER is an important task as it is usually employed as a primary step in the other tasks such as event detection from news, customer support for on-line shopping, knowledge graph construction, and biological analysis (Bokharaeian et al., 2017) .", | |
| "cite_spans": [ | |
| { | |
| "start": 368, | |
| "end": 394, | |
| "text": "(Bokharaeian et al., 2017)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "NER is a famous and well-studied task in English (Yadav and Bethard, 2018) and some other languages like Arabic (Shaalan, 2014; Helwe and Elbassuoni, 2019; Taghizadeh et al., 2018) and German (Riedl and Pad\u00f3, 2018 ). However, this task is not highly examined in Farsi because there is no standard benchmark for it. Although there are some Farsi NER corpora such as PEYMA (Shahshahani et al., 2018) , ArmanPersoNER (Poostchi et al., 2016) , A'laam (Hosseinnejad et al., 2017) , and Persian-NER 1 ; none of them is known as standard data set to the research community. Moreover, the type of named entities and annotation guidelines are different in each corpus. Because of the diversity of annotation types and data sets which were used for training and test, the result of current researches on Farsi NER cannot be directly compared.", | |
| "cite_spans": [ | |
| { | |
| "start": 49, | |
| "end": 74, | |
| "text": "(Yadav and Bethard, 2018)", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 112, | |
| "end": 127, | |
| "text": "(Shaalan, 2014;", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 128, | |
| "end": 155, | |
| "text": "Helwe and Elbassuoni, 2019;", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 156, | |
| "end": 180, | |
| "text": "Taghizadeh et al., 2018)", | |
| "ref_id": "BIBREF15" | |
| }, | |
| { | |
| "start": 192, | |
| "end": 213, | |
| "text": "(Riedl and Pad\u00f3, 2018", | |
| "ref_id": "BIBREF11" | |
| }, | |
| { | |
| "start": 371, | |
| "end": 397, | |
| "text": "(Shahshahani et al., 2018)", | |
| "ref_id": "BIBREF14" | |
| }, | |
| { | |
| "start": 414, | |
| "end": 437, | |
| "text": "(Poostchi et al., 2016)", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 447, | |
| "end": 474, | |
| "text": "(Hosseinnejad et al., 2017)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The goal of this competition was to bring Farsi NER researchers together. We introduce a large scale corpus containing about 900K tokens as the training data for this task. To evaluate the participating teams, a test set was prepared which contains 150K tokens. The training and test set follow the same annotation schema. These data sets are publicly available for further researches 2 . The domain of all data is the news sentences because they are the most entity-rich.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Participants were allowed to use any public data and resources such as Farsi Wikipedia 3 and Farsi Knowledge Graph 4 (Sajadi et al., 2018) in addition to the official training data of the shared task in the process of making their system. In this case, they must thoroughly describe those resources and the way they used them.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "To the best of our knowledge, this is the first shared task in Farsi. Since Farsi belongs to the group of low-resource languages (Taghizadeh and Faili, 2016; Fadaei and Faili, 2019) , the availability of annotated corpora and resources will be very useful for future investigation in this language.", | |
| "cite_spans": [ | |
| { | |
| "start": 129, | |
| "end": 157, | |
| "text": "(Taghizadeh and Faili, 2016;", | |
| "ref_id": "BIBREF16" | |
| }, | |
| { | |
| "start": 158, | |
| "end": 181, | |
| "text": "Fadaei and Faili, 2019)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "So far, some researchers have been conducted on Farsi NER. Poostchi et al. (Poostchi et al., 2018) presented a BiLSTM-CRF model, which is a recurrent neural network obtained by a combination of a long short-term memory (LSTM) and a conditional random field (CRF). They presented a public data set for Farsi NER, called ArmanPersoNER, which includes six types of NEs: person, organization, location, facility, product, and event. Their model showed 77.45% of F 1 on ArmanPer-soNER. Shahshahani et al. (Shahshahani et al., 2018 ) presented a hybrid system consisting of a rule-based and a statistical system. The rule-based system composed of a large list of NEs in Farsi in addition to the regular expressions for detecting them. The statistical system is a CRF model trained by the PEYMA corpus. Their system reached 84% of F 1 for seven classes of person, organization, location, date, time, money, and percent, based on 5fold validation on the training data.", | |
| "cite_spans": [ | |
| { | |
| "start": 59, | |
| "end": 98, | |
| "text": "Poostchi et al. (Poostchi et al., 2018)", | |
| "ref_id": "BIBREF9" | |
| }, | |
| { | |
| "start": 481, | |
| "end": 525, | |
| "text": "Shahshahani et al. (Shahshahani et al., 2018", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Farsi NER", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Hossinnejad et al. (Hosseinnejad et al., 2017) presented a corpus named A'laam consisting of 13 classes of named entities. They split this corpus into two parts of 90% and 10% for the training and test, respectively, and trained a CRF model using the training part. They obtained 92.9% and 78.5% of precision and recall, respectively. Zafarian et al. (Zafarian et al., 2015) proposed a semi-supervised method for Farsi NER. They used an un-labeled bilingual data in addition to a small labeled data to train their system. They presented a bootstrap method that iteratively trains a CRF model using the labeled data as well as those unlabeled data that the current model predicts them with high confidence. Their data contains three classes of person, organization, and location. They reached 67.5% of F 1 . Current researches on Farsi NER use different data for the training and test. Most of these data are not public or annotated with diverse annotation schema. The evaluation methods of them are not similar and so their results cannot be directly compared.", | |
| "cite_spans": [ | |
| { | |
| "start": 19, | |
| "end": 46, | |
| "text": "(Hosseinnejad et al., 2017)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 351, | |
| "end": 374, | |
| "text": "(Zafarian et al., 2015)", | |
| "ref_id": "BIBREF19" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Farsi NER", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Participating systems have to predict NE tags for a set of tokenized documents. We defined two subtasks:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Task", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "\u2022 3-classes including person, organization, and location;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Task", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "\u2022 7-classes including date, time, money, and percent in addition to the three above classes.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Task", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "NEs that belong to four classes of date, time, money, and percent sometimes can be recognized using the rule-based or hybrid methods (Ahmadi and Moradi, 2015; Riaz, 2010) ; while NEs of the classes of person, organization, or location are often recognized based on the gazetteer lists and they are more subject to ambiguity. Therefore, we have separated these two subtasks and participants could submit different systems for them.", | |
| "cite_spans": [ | |
| { | |
| "start": 133, | |
| "end": 158, | |
| "text": "(Ahmadi and Moradi, 2015;", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 159, | |
| "end": 170, | |
| "text": "Riaz, 2010)", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Task", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "CoNLL 2003 defined the baseline of NER task a system which only selects complete unambiguous named entities that appear in the training data. We adapted this baseline as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline Method", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "\u2022 In case of overlap between two candidate named entities, the longer is selected. For example, consider three NEs of the training data: 1) \u202b\"\u202a/Iran\u202c\u0627\ufbfe\ufeae\u0627\u0646\"\u202c which is a location, 2) \u202b\u0627\ufeb3\ufefc\ufee3\ufbfd\"\u202c \u202b\ufeb7\ufeee\u0631\u0627\u06cc\u202c \u202b/\ufee3\ufea0\ufee0\ufeb2\u202cIslamic Consultative Assembly\" which is an organization, and 3) \u202b\u0627\ufbfe\ufeae\u0627\u0646\"\u202c \u202b\u0627\ufeb3\ufefc\ufee3\ufbfd\u202c \u202b\ufeb7\ufeee\u0631\u0627\u06cc\u202c \u202b/\ufee3\ufea0\ufee0\ufeb2\u202cIslamic Consultative Assembly of Iran\" which is an organization as well. To extract named entities from phrase \u202b\ufeb7\ufeee\u0631\u0627\u06cc\"\u202c \u202b\ufee3\ufea0\ufee0\ufeb2\u202c \u202b\u0627\ufbfe\ufeae\u0627\u0646\u202c \u202b/\u0627\ufeb3\ufefc\ufee3\ufbfd\u202cIslamic Consultative Assembly of Iran\", the baseline system selects whole phrase as an organization instead of separately tagging \u202b\u0627\ufeb3\ufefc\ufee3\ufbfd\"\u202c \u202b\ufeb7\ufeee\u0631\u0627\u06cc\u202c \u202b/\ufee3\ufea0\ufee0\ufeb2\u202cIslamic Consultative Assembly\" as an organization and \u202b\"\u202a/Iran\u202c\u0627\ufbfe\ufeae\u0627\u0646\"\u202c as a location.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline Method", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "\u2022 When two NEs are next to each other and have the same tag, they are merged. For example, there are two NEs in the training data: \u202b\ufe91\ufeec\ufee4\ufee6\"\u202c \u0662\u0662/22th of Bahman\" and \"\u0661\u0663\u06f5\u0667/1357\" which are date. In the test phase, the baseline system visits phrase \"\u0661\u0663\u06f5\u0667 \u202b\ufe91\ufeec\ufee4\ufee6\u202c \u0662\u0662/22th of Bahman 1357\", and separately selects these two phrases as date. Then, they are merged to be one mention. Our analysis showed that this heuristic is often true. However, in few cases it may be wrong. For example, consider following sentence:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline Method", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "\u202b\u0631\u0648\ufee3\ufe8e\ufee7\ufbff\ufe8e\ufbfe\ufbfd\u202c \u202b\u062f\ufbfe\ufb59\ufee0\ufee4\ufe8e\u062a\u202c \u202b\ufed3\ufeae\u0648\ufe97\ufe8e\u202c \u202b\ufb90\ufeee\u0631\ufee7\ufede\u202c \u202b\u0622\ufee3\ufe8e\ufee7\ufeee\u202c \u202b\ufe9f\ufe8e\ufee7\ufeb8\ufbff\ufee8\ufbfd\u202c \u202b\u0627\u0648\u0644\u202c \u202b\ufb94\ufeb0\ufbfe\ufee8\ufeea\u202c \u202b\u0627\ufeb3\ufe96.\u202c \"The first option for Amano's successor is Cornel Fruta, a Romanian diplomat.\"", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline Method", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "There are two adjacent mentions with the same person type: \"\u202b/\u0622\ufee3\ufe8e\ufee7\ufeee\u202cAmano\" and \u202b\ufed3\ufeae\u0648\ufe97\ufe8e\"\u202c \u202b/\ufb90\ufeee\u0631\ufee7\ufede\u202cCornel Feruta\", and merging them into one NE is not correct.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline Method", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "These examples reveal some challenges of Farsi NER. One challenge is that a unique named entity may appear in the text with different names. For example, \u202b\ufeb7\ufeee\u0631\u0627\u06cc\"\u202c \u202b\ufee3\ufea0\ufee0\ufeb2\u202c \u202b\u0627\ufbfe\ufeae\u0627\u0646\u202c \u202b/\u0627\ufeb3\ufefc\ufee3\ufbfd\u202cIslamic Consultative Assembly of Iran\", \u202b\u0627\ufeb3\ufefc\ufee3\ufbfd\"\u202c \u202b\ufeb7\ufeee\u0631\u0627\u06cc\u202c \u202b/\ufee3\ufea0\ufee0\ufeb2\u202cIslamic Consultative Assembly\" and \"\u202b/\ufee3\ufea0\ufee0\ufeb2\u202cassembly\" are different names of the same entity. While \"\u202b/\ufee3\ufea0\ufee0\ufeb2\u202cassembly\" is a common noun, it names an organization. It means that gazetteers are not sufficient for detecting boundaries of entity mentions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline Method", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "Another challenge is that two or more entity mentions may be adjacent in the sentence, in the sense that there is no word between them. They may have different or similar types. In case of similar types, it may be possible or not to merge them into a unique mention. For example adjacent entity mentions of date, mostly can be merged, such as \u202b\ufeb3\ufe8e\u0644\"\u202c \u202b\ufee3\ufe8e\u0647\u202c \u202b\ufee3\ufeec\ufeae\u202c \u202b\ufeeb\ufed4\ufe98\ufee2\u202c \u202b\u0648\u202c \u202b\ufe91\ufbff\ufeb4\ufe96\u202c 1398\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline Method", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "We presented a training data set which has two parts: the first part is PEYMA corpus (Shahshahani et al., 2018) containing 300K tokens; the second part has 600K tokens. The same annotation schema was used for annotating two parts. This annotation schema was prepared based on two standard guidelines: 1) MUC 5 and 2) CoNLL 6 ; then it was adapted for Farsi linguistic structures (Shahshahani et al., 2018) . In these data sets there are seven classes of named entities: person, organization, location, money, date, time, and percent.", | |
| "cite_spans": [ | |
| { | |
| "start": 379, | |
| "end": 405, | |
| "text": "(Shahshahani et al., 2018)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set Creation", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "Steps of creating data set include news collection, pre-processing, and named entity tagging. The test data has two parts: in-domain and out-of-domain. The former was sampled from the same news websites in the same period of time that the training data were collected. The latter was selected from different news websites at different times. Specifically, documents of the training data mostly were sampled from a few Farsi news websites between 2016 and 2017; while out-of-domain documents were sampled from multiple Farsi news websites from different countries of the world mainly in 2019. Therefore, in-domain documents are more similar in word distribution to the training data than the out-ofdomain documents.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set Creation", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "Therefore, in-domain documents are more similar in word distribution to the training data than the out-of-domain documents. Preprocessing on news documents was performed using Persianp toolkit (Mohseni et al., 2016) , which includes tokenization, sentence split, and normalization. Two annotators performed the annotation task, and the agreement between them is 95% which shows the quality of the annotations. The data format is similar to the CoNLL 2003, in which each line contains one word and empty lines represent sentence boundaries. Annotation format is IOB that encodes the beginning and inside of the entity mentions and type of them. ", | |
| "cite_spans": [ | |
| { | |
| "start": 193, | |
| "end": 215, | |
| "text": "(Mohseni et al., 2016)", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set Creation", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "Six teams have participated in both subtasks. Most of them opted for use of CRF models and deep learning methods specifically Bi-LSTM. Because these two models deal with sequence tagging problems. Word embeddings, n-grams, and POS tags were used as features by the systems. Morphological and orthographic features of Farsi phrases were used by some of the participants. Table 5 briefly shows the models and features used by the participants. ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 370, | |
| "end": 377, | |
| "text": "Table 5", | |
| "ref_id": "TABREF4" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Participating Systems and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "There are different methods for the evaluation of NER systems. Two main methods are phrase-level and word-level evaluation. In the phrase-level evaluation, a phrase is counted as true-positive for class c, if both boundaries of the phrase and its predicted tag are correct. In contrast, in word-level evaluation, each word is considered separately. Therefore, the phraselevel evaluation is tougher than the word-level evaluation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation Metrics", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "We used evaluation script of conlleval 7 . This script computes three measures including precision, recall, and F 1 based on the standard definition. Evaluation of the 3-classes subtask has been performed based on the macroaveraging method. Accordingly, precision and recall are obtained by averaging of the precision and recall of the three classes of person, organization, and location.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation Metrics", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Evaluation of 7-classes subtask has been conducted using the micro-averaging method due to class imbalance problem, in the sense that frequencies of NE phrases belonging to four classes of date, time, money, and percent are very fewer than the three classes of person, organization and location, according to the Tables 3 and 4 . So, in this case, the microaveraging better evaluates the quality of systems.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 313, | |
| "end": 327, | |
| "text": "Tables 3 and 4", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Evaluation Metrics", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Participating teams mainly used sequence tagging methods including CRF and Bi-LSTM networks. The feature sets used by them include lexical, morphological, and structural features. Tables 6 and 7 show the evaluation results of 3-classes and 7-classes subtasks, respectively. Generally, results of the wordlevel evaluation are higher than the phraselevel evaluation. Moreover, the results of the evaluation by the in-domain data are higher than the out-of-domain data in terms of the F 1 score. All teams outperformed the baseline and ranking of the teams are the same based on all kinds of the evaluations. The best F 1 scores are 85.9% and 88.5% based on the phrase-level and word-level evaluation, respectively, which are obtained by the MorphBERT system (Mohseni and Tebbifakhr, 2019) .", | |
| "cite_spans": [ | |
| { | |
| "start": 756, | |
| "end": 786, | |
| "text": "(Mohseni and Tebbifakhr, 2019)", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 180, | |
| "end": 194, | |
| "text": "Tables 6 and 7", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Result", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "The second best system, Beheshti-NER-1 (Taher et al., 2019) , got near F 1 scores: 84.0% and 87.9% based on the phrase-level and word-level evaluation, respectively. These two systems used BERT model (Devlin et al., 2018) for training high accurate representation of Farsi tokens. BERT is a deep bi-directional language model that presented state-of-the-art results in a wide variety of NLP tasks. Both systems used the BERT to process a huge amount of un-labeled Farsi texts to obtain pre-trained word embeddings which then was fine-tuned for the NER task.", | |
| "cite_spans": [ | |
| { | |
| "start": 39, | |
| "end": 59, | |
| "text": "(Taher et al., 2019)", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 200, | |
| "end": 221, | |
| "text": "(Devlin et al., 2018)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Result", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "MorphoBERT used a morphological analyzer as a prior step before the BERT network. Farsi is rather rich-morphology and analyzing tokens to find their parts reveals the grammatical and semantic information. So, instead of embedding tokens of sentences into the network, MorphoBERT firstly decomposes tokens into constituents and then fed these constituents into the BERT network. Then, the representation of the sentence which was obtained from the BERT is given to a Bi-LSTM network. Additionally, a vector representing word cluster features is given to the Bi-LSTM. Finally, the Softmax layer produces a probability distribution over all classes (Mohseni and Tebbifakhr, 2019) .", | |
| "cite_spans": [ | |
| { | |
| "start": 646, | |
| "end": 676, | |
| "text": "(Mohseni and Tebbifakhr, 2019)", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Result", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "Beheshti-NER-1 system utilizes a CRF model on top of the BERT network. The motivation of using CRF is that an encoder like BERT tries to maximize the likelihood by selecting best-hidden representations, and CRF tries to maximize the likelihood by selecting best output tags (Taher et al., 2019) .", | |
| "cite_spans": [ | |
| { | |
| "start": 274, | |
| "end": 294, | |
| "text": "(Taher et al., 2019)", | |
| "ref_id": "BIBREF17" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Result", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "To better understand the details of the scores, we presented the F 1 scores of each 7 classes based on the phrase-level evaluation in Table 8 . Generally, the most F 1 scores were obtained by percent and money classes. Because there are specific keywords representing them and so there are high-precision patterns that specify entity mentions of these classes. Specifically, percent often comes with the keywords like \"\u202b/\u062f\u0631\ufebb\ufeaa\u202cpercent\"; while money appears with words and phrases denoting money like \"\u202b/\u062f\ufefb\u0631\u202cDollar\", \"\u202b/\u0631\ufbfe\ufe8e\u0644\u202cRial\", or \"\u202b/\ufbfe\ufeee\u0631\u0648\u202cEuro\". On the other hand, the least F 1 scores were obtained by the time class. Perhaps because the number of phrases in the training data having time tag is very few in comparison to the other classes.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 134, | |
| "end": 141, | |
| "text": "Table 8", | |
| "ref_id": "TABREF7" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Result", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "We have described the NSURL-2019 task 7: NER in Farsi. Six systems have processed the Farsi NE data. The best performance was obtained by the MorphoBERT system that is 85.4% of F 1 score based on the phrase-level evaluation of the 7-classes subtask. This system uses morphological features of Farsi words together with the BERT model and Bi-LSTM. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "https://github.com/Text-Mining/Persian-NER 2 https://github.com/nasrin-taghizadeh/ NSURL-Persian-NER", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://fa.wikipedia.org/wiki/ 4 http://farsbase.net/search/html/index.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://www-nlpir.nist.gov/related_ projects/muc/proceedings/ne_task.html 6 https://www.clips.uantwerpen.be/conll2003/ ner/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "A hybrid method for Persian Named Entity Recognition", | |
| "authors": [ | |
| { | |
| "first": "Farid", | |
| "middle": [], | |
| "last": "Ahmadi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hamed", | |
| "middle": [], | |
| "last": "Moradi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "2015 7th Conference on Information and Knowledge Technology (IKT)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--7", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Farid Ahmadi and Hamed Moradi. 2015. A hybrid method for Persian Named En- tity Recognition. In 2015 7th Conference on Information and Knowledge Technology (IKT), pages 1-7. IEEE.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "SNPPhenA: a corpus for extracting ranked associations of single-nucleotide polymorphisms and phenotypes from literature", | |
| "authors": [ | |
| { | |
| "first": "Behrouz", | |
| "middle": [], | |
| "last": "Bokharaeian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alberto", | |
| "middle": [], | |
| "last": "Diaz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nasrin", | |
| "middle": [], | |
| "last": "Taghizadeh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hamidreza", | |
| "middle": [], | |
| "last": "Chitsaz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ramyar", | |
| "middle": [], | |
| "last": "Chavoshinejad", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Journal of biomedical semantics", | |
| "volume": "8", | |
| "issue": "1", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Behrouz Bokharaeian, Alberto Diaz, Nasrin Taghizadeh, Hamidreza Chitsaz, and Ram- yar Chavoshinejad. 2017. SNPPhenA: a corpus for extracting ranked associations of single-nucleotide polymorphisms and phe- notypes from literature. Journal of biomed- ical semantics, 8(1):14.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "BERT: Pretraining of deep bidirectional transformers for language understanding", | |
| "authors": [ | |
| { | |
| "first": "Jacob", | |
| "middle": [], | |
| "last": "Devlin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ming-Wei", | |
| "middle": [], | |
| "last": "Chang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kenton", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kristina", | |
| "middle": [], | |
| "last": "Toutanova", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1810.04805" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre- training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Using syntax for improving phrase-based SMT in low-resource languages", | |
| "authors": [ | |
| { | |
| "first": "Hakimeh", | |
| "middle": [], | |
| "last": "Fadaei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Heshaam", | |
| "middle": [], | |
| "last": "Faili", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hakimeh Fadaei and Heshaam Faili. 2019. Us- ing syntax for improving phrase-based SMT in low-resource languages. Digital Scholar- ship in the Humanities.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Arabic Named Entity Recognition via deep co-learning", | |
| "authors": [ | |
| { | |
| "first": "Chadi", | |
| "middle": [], | |
| "last": "Helwe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shady", | |
| "middle": [], | |
| "last": "Elbassuoni", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Artificial Intelligence Review", | |
| "volume": "52", | |
| "issue": "1", | |
| "pages": "197--215", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Chadi Helwe and Shady Elbassuoni. 2019. Arabic Named Entity Recognition via deep co-learning. Artificial Intelligence Review, 52(1):197-215.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "A'laam Corpus: A Standard Corpus of Named Entity for Persian Language", | |
| "authors": [ | |
| { | |
| "first": "Shadi", | |
| "middle": [], | |
| "last": "Hosseinnejad", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yasser", | |
| "middle": [], | |
| "last": "Shekofteh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tahereh", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Emami", | |
| "middle": [], | |
| "last": "Azadi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Signal and Data Processing", | |
| "volume": "14", | |
| "issue": "3", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Shadi Hosseinnejad, Yasser Shekofteh, and Tahereh and Emami Azadi. 2017. A'laam Corpus: A Standard Corpus of Named En- tity for Persian Language. Signal and Data Processing, 14(3).", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Persianp: a Persian text processing toolbox", | |
| "authors": [ | |
| { | |
| "first": "Mahdi", | |
| "middle": [], | |
| "last": "Mohseni", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Javad", | |
| "middle": [], | |
| "last": "Ghofrani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Heshaam", | |
| "middle": [], | |
| "last": "Faili", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Computational Linguistics and Intelligent Text Processing -17th International Conference, CICLing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "75--87", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1007/978-3-319-75477-2_4" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mahdi Mohseni, Javad Ghofrani, and Hes- haam Faili. 2016. Persianp: a Persian text processing toolbox. In Computational Linguistics and Intelligent Text Processing -17th International Conference, CICLing 2016, Konya, Turkey, April 3-9, 2016, Re- vised Selected Papers, Part I, pages 75-87.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "MorphoBERT: a Persian NER system with BERT and morphological analysis", | |
| "authors": [ | |
| { | |
| "first": "Mahdi", | |
| "middle": [], | |
| "last": "Mohseni", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Amirhossein", | |
| "middle": [], | |
| "last": "Tebbifakhr", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the first International Workshop on NLP Solutions for Under Resourced Languages, NSURL '19", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mahdi Mohseni and Amirhossein Tebbifakhr. 2019. MorphoBERT: a Persian NER system with BERT and morphological analysis. In Proceedings of the first International Work- shop on NLP Solutions for Under Resourced Languages, NSURL '19, Trento, Italy.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "PersoNER: Persian Named Entity Recognition", | |
| "authors": [ | |
| { | |
| "first": "Hanieh", | |
| "middle": [], | |
| "last": "Poostchi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Ehsan Zare", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mohammad", | |
| "middle": [], | |
| "last": "Borzeshi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Massimo", | |
| "middle": [], | |
| "last": "Abdous", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Piccardi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "COLING 2016-26th International Conference on Computational Linguistics, Proceedings of COLING 2016: Technical Papers", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hanieh Poostchi, Ehsan Zare Borzeshi, Mo- hammad Abdous, and Massimo Piccardi. 2016. PersoNER: Persian Named Entity Recognition. In COLING 2016-26th In- ternational Conference on Computational Linguistics, Proceedings of COLING 2016: Technical Papers.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "BiLSTM-CRF for Persian Named-Entity Recognition armanpersonercorpus: the first entity-annotated persian dataset", | |
| "authors": [ | |
| { | |
| "first": "Hanieh", | |
| "middle": [], | |
| "last": "Poostchi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Ehsan Zare", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Massimo", | |
| "middle": [], | |
| "last": "Borzeshi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Piccardi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hanieh Poostchi, Ehsan Zare Borzeshi, and Massimo Piccardi. 2018. BiLSTM-CRF for Persian Named-Entity Recognition arman- personercorpus: the first entity-annotated persian dataset. In Proceedings of the Eleventh International Conference on Lan- guage Resources and Evaluation (LREC- 2018).", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Rule-based named entity recognition in Urdu", | |
| "authors": [ | |
| { | |
| "first": "Kashif", | |
| "middle": [], | |
| "last": "Riaz", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proceedings of the 2010 named entities workshop", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "126--135", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kashif Riaz. 2010. Rule-based named entity recognition in Urdu. In Proceedings of the 2010 named entities workshop, pages 126- 135. Association for Computational Linguis- tics.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "A Named Entity Recognition shootout for German", | |
| "authors": [ | |
| { | |
| "first": "Martin", | |
| "middle": [], | |
| "last": "Riedl", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sebastian", | |
| "middle": [], | |
| "last": "Pad\u00f3", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", | |
| "volume": "2", | |
| "issue": "", | |
| "pages": "120--125", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Martin Riedl and Sebastian Pad\u00f3. 2018. A Named Entity Recognition shootout for German. In Proceedings of the 56th An- nual Meeting of the Association for Compu- tational Linguistics (Volume 2: Short Pa- pers), pages 120-125.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Farsbase: A crossdomain farsi knowledge graph", | |
| "authors": [ | |
| { | |
| "first": "Mohamad", | |
| "middle": [], | |
| "last": "Bagher Sajadi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Behrouz", | |
| "middle": [], | |
| "last": "Minaei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ali", | |
| "middle": [], | |
| "last": "Hadian", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "SEMAN-TICS Posters&Demos", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mohamad Bagher Sajadi, Behrouz Minaei, and Ali Hadian. 2018. Farsbase: A cross- domain farsi knowledge graph. In SEMAN- TICS Posters&Demos.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "A survey of Arabic Named Entity Recognition and classification", | |
| "authors": [ | |
| { | |
| "first": "Khaled", | |
| "middle": [], | |
| "last": "Shaalan", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Computational Linguistics", | |
| "volume": "40", | |
| "issue": "2", | |
| "pages": "469--510", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Khaled Shaalan. 2014. A survey of Arabic Named Entity Recognition and classifica- tion. Computational Linguistics, 40(2):469- 510.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "PEYMA: A Tagged Corpus for Persian Named Entities", | |
| "authors": [ | |
| { | |
| "first": "Mahdi", | |
| "middle": [], | |
| "last": "Mahsa Sadat Shahshahani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Mohseni", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "ArXiv", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mahsa Sadat Shahshahani, Mahdi Mohseni, Azadeh Shakery, and Heshaam Faili. 2018. PEYMA: A Tagged Corpus for Persian Named Entities. ArXiv, abs/1801.09936.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Cross-Language Learning for Arabic Relation Extraction", | |
| "authors": [ | |
| { | |
| "first": "Nasrin", | |
| "middle": [], | |
| "last": "Taghizadeh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Heshaam", | |
| "middle": [], | |
| "last": "Faili", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jalal", | |
| "middle": [], | |
| "last": "Maleki", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Procedia computer science", | |
| "volume": "142", | |
| "issue": "", | |
| "pages": "190--197", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nasrin Taghizadeh, Heshaam Faili, and Jalal Maleki. 2018. Cross-Language Learning for Arabic Relation Extraction. Procedia com- puter science, 142:190-197.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Automatic Wordnet Development for Low-Resource Languages using Cross-lingual WSD", | |
| "authors": [ | |
| { | |
| "first": "Nasrin", | |
| "middle": [], | |
| "last": "Taghizadeh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hesham", | |
| "middle": [], | |
| "last": "Faili", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Journal of Artificial Intelligence Research", | |
| "volume": "56", | |
| "issue": "", | |
| "pages": "61--87", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nasrin Taghizadeh and Hesham Faili. 2016. Automatic Wordnet Development for Low- Resource Languages using Cross-lingual WSD. Journal of Artificial Intelligence Re- search, 56:61-87.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Beheshti-NER: Persian named entity recognition using BERT", | |
| "authors": [ | |
| { | |
| "first": "Ehsan", | |
| "middle": [], | |
| "last": "Taher", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Abbas", | |
| "middle": [], | |
| "last": "Seyed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mehrnoush", | |
| "middle": [], | |
| "last": "Hoseini", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Shamsfard", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the first International Workshop on NLP Solutions for Under Resourced Languages, NSURL '19", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ehsan Taher, Seyed Abbas Hoseini, and Mehrnoush Shamsfard. 2019. Beheshti- NER: Persian named entity recognition us- ing BERT. In Proceedings of the first In- ternational Workshop on NLP Solutions for Under Resourced Languages, NSURL '19, Trento, Italy.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "A survey on recent advances in named entity recognition from deep learning models", | |
| "authors": [ | |
| { | |
| "first": "Vikas", | |
| "middle": [], | |
| "last": "Yadav", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Steven", | |
| "middle": [], | |
| "last": "Bethard", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 27th International Conference on Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2145--2158", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vikas Yadav and Steven Bethard. 2018. A survey on recent advances in named entity recognition from deep learning models. In Proceedings of the 27th International Con- ference on Computational Linguistics, pages 2145-2158.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Semisupervised learning for named entity recognition using weakly labeled training data", | |
| "authors": [ | |
| { | |
| "first": "Atefeh", | |
| "middle": [], | |
| "last": "Zafarian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ali", | |
| "middle": [], | |
| "last": "Rokni", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shahram", | |
| "middle": [], | |
| "last": "Khadivi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sonia", | |
| "middle": [], | |
| "last": "Ghiasifard", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "2015 The International Symposium on Artificial Intelligence and Signal Processing (AISP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "129--135", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Atefeh Zafarian, Ali Rokni, Shahram Khadivi, and Sonia Ghiasifard. 2015. Semi- supervised learning for named entity recog- nition using weakly labeled training data. In 2015 The International Symposium on Artificial Intelligence and Signal Processing (AISP), pages 129-135. IEEE.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "type_str": "table", | |
| "content": "<table><tr><td/><td colspan=\"4\">Lang #Doc #Sent #Tokens</td></tr><tr><td>Training Data</td><td>Fa</td><td colspan=\"3\">1,456 27,130 885,296</td></tr><tr><td>Test Data</td><td>Fa</td><td>431</td><td>4,154</td><td>144,526</td></tr><tr><td>ArmanPersoNER</td><td>Fa</td><td>-</td><td>7,682</td><td>250,015</td></tr><tr><td>CoNLL-2003</td><td>En</td><td colspan=\"3\">1,393 22,137 301,418</td></tr><tr><td>CoNLL-2003</td><td>Gr</td><td>909</td><td colspan=\"2\">18,973 310,318</td></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "Data Statistics" | |
| }, | |
| "TABREF1": { | |
| "type_str": "table", | |
| "content": "<table><tr><td/><td colspan=\"3\">: Statistics of Test data</td></tr><tr><td>Test Data</td><td colspan=\"2\">#Doc #Sent</td><td>#Tokens</td></tr><tr><td>In-domain</td><td>196</td><td colspan=\"2\">1,571 68,063 (47%)</td></tr><tr><td>Out-of-domain</td><td>235</td><td colspan=\"2\">2,583 76,463 (53%)</td></tr><tr><td colspan=\"2\">4.1 Data Statistics</td><td/></tr><tr><td colspan=\"4\">Table 1 represents general statistics of our</td></tr><tr><td colspan=\"4\">Farsi data sets including the number of arti-</td></tr><tr><td colspan=\"4\">cles, sentences, and tokens in comparison with</td></tr><tr><td colspan=\"4\">English and German data sets of the CoNLL</td></tr><tr><td colspan=\"4\">2003. The comparison reveals that the Farsi</td></tr><tr><td colspan=\"4\">training data is a large scale data set that can</td></tr><tr><td colspan=\"4\">be used for further researches on Farsi NER.</td></tr><tr><td colspan=\"4\">Table 2 shows details of the in-domain and</td></tr><tr><td colspan=\"4\">out-of-domain parts of the test data. The two</td></tr><tr><td colspan=\"4\">parts have a nearly equal number of tokens.</td></tr><tr><td colspan=\"4\">Tables 3 and 4 represent the total number</td></tr><tr><td colspan=\"4\">of phrases and the number of unique phrases</td></tr><tr><td colspan=\"4\">tagged for each class of named entities in the</td></tr><tr><td colspan=\"4\">training and test data. Considering the size of</td></tr><tr><td colspan=\"4\">each corpus, the test set is denser in terms of</td></tr><tr><td>the entity tags.</td><td/><td/></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "" | |
| }, | |
| "TABREF2": { | |
| "type_str": "table", | |
| "content": "<table><tr><td>Data</td><td>PER</td><td>ORG</td><td colspan=\"4\">LOC MON DAT TIM PCT Total</td></tr><tr><td colspan=\"6\">Training 12,495 14,205 15,403 1,294 4,467 571</td><td>997 49,432</td></tr><tr><td>Test</td><td>2,738</td><td>3,160</td><td>4,081</td><td>357</td><td>1,147 165</td><td>156 11,804</td></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "Number of total phrases tagged per class" | |
| }, | |
| "TABREF3": { | |
| "type_str": "table", | |
| "content": "<table><tr><td>Data</td><td colspan=\"6\">PER ORG LOC MON DAT TIM PCT Total</td></tr><tr><td colspan=\"5\">Training 5,228 4,547 2,738 1,008 1,910 338</td><td colspan=\"2\">453 16,020</td></tr><tr><td>Test</td><td>1,470 1,326 1,015</td><td>288</td><td>628</td><td>114</td><td>97</td><td>4,917</td></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "Number of unique phrases tagged per class" | |
| }, | |
| "TABREF4": { | |
| "type_str": "table", | |
| "content": "<table><tr><td>Team</td><td>Model</td><td>Word Embeddings</td><td>Features</td></tr><tr><td>MorphoBERT</td><td>BERT + BiLSTM</td><td>BERT for token representation word2vec for word clustering</td><td>cluster number of words, morphology</td></tr><tr><td>Beheshti-NER-1</td><td>Transformer-CRF</td><td>BERT</td><td>-</td></tr><tr><td>Team-3</td><td>CRF</td><td>-</td><td>-</td></tr><tr><td>ICTRC-NLPGroup</td><td>CRF</td><td>-</td><td>n-gram, lemma, linguistics rules</td></tr><tr><td>UT-NLP-IR</td><td>CRF</td><td>-</td><td>POS, NP-chunk, word n-gram, char n-gram, stem, lemma</td></tr><tr><td>SpeechTrans</td><td>SVM</td><td>-</td><td>word unigram, char 5-grams, POS, stem, normalized surface</td></tr><tr><td>Baseline</td><td>heuristic</td><td>-</td><td>-</td></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "Description of Participating Systems" | |
| }, | |
| "TABREF5": { | |
| "type_str": "table", | |
| "content": "<table><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">Test Data</td><td/><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td colspan=\"5\">Phrase-level evaluation</td><td/><td/><td/><td/><td colspan=\"5\">Word-level evaluation</td><td/></tr><tr><td>Team</td><td colspan=\"3\">In-domain</td><td colspan=\"3\">Out-of-domain</td><td/><td>Total</td><td/><td colspan=\"3\">In-domain</td><td colspan=\"3\">Out-of-domain</td><td/><td>Total</td></tr><tr><td/><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td></tr><tr><td>1 MorphoBERT</td><td colspan=\"18\">88.7 8 87.9</td></tr><tr><td>3 Team-3</td><td colspan=\"18\">87.4 77.2 82.0 87.4 73.4 79.8 87.4 75.0 80.7 89.2 79.5 84.1 89.5 74.7 81.4 89.3 76.9 82.7</td></tr><tr><td colspan=\"19\">4 ICTRC-NLPGroup 87.5 76.0 81.3 86.2 69.6 77.0 86.8 72.3 78.9 90.1 78.2 83.7 88.7 70.2 78.4 89.4 73.5 80.7</td></tr><tr><td>5 UT-NLP-IR</td><td colspan=\"18\">75.3 68.9 72.0 72.3 60.7 66.0 73.6 64.1 68.5 87.3 71.9 78.9 86.4 61.1 71.6 86.9 65.7 74.8</td></tr><tr><td>6 SpeechTrans</td><td colspan=\"18\">41.5 39.5 40.5 43.1 38.7 40.8 42.4 39.0 40.6 66.8 38.3 48.7 66.2 35.2 46.0 66.6 36.4 47.0</td></tr><tr><td>7 Baseline</td><td colspan=\"18\">32.2 45.8 37.8 32.8 39.1 35.7 32.5 41.9 36.6 46.2 42.6 44.3 45.2 35.1 39.5 45.9 38.4 41.8</td></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "Evaluation of systems for subtask 3-classes 85.5 87.1 86.3 83.8 85.0 87.3 84.5 85.9 92.5 86.7 89.5 91.5 84.0 87.6 92.1 85.2 88.5 2 Beheshti-NER-1 85.3 84.4 84.8 84.4 82.6 83.5 84.8 83.3 84.0 90.5 87.2 88.8 89.7 85.0 87.3 90.1 85." | |
| }, | |
| "TABREF6": { | |
| "type_str": "table", | |
| "content": "<table><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">Test Data</td><td/><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td colspan=\"5\">Phrase-level evaluation</td><td/><td/><td/><td/><td colspan=\"5\">Word-level evaluation</td><td/></tr><tr><td>Team</td><td colspan=\"3\">In-domain</td><td colspan=\"3\">Out-of-domain</td><td/><td>Total</td><td/><td colspan=\"3\">In-domain</td><td colspan=\"3\">Out-of-domain</td><td/><td>Total</td></tr><tr><td/><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td><td>P</td><td>R</td><td>F 1</td></tr><tr><td>1 MorphoBERT</td><td colspan=\"18\">88.4 84.8 86.6 86.0 83.1 84.5 87.0 83.8 85.4 94.0 89.1 91.5 91.8 85.7 88.6 92.8 87.1 89.9</td></tr><tr><td>2 Beheshti-NER-1</td><td colspan=\"18\">84.8 83.6 84.2 83.9 82.0 83.0 84.3 82.7 83.5 91.4 87.3 89.3 89.7 85.7 87.7 90.4 86.5 88.4</td></tr><tr><td>3 Team-3</td><td colspan=\"18\">87.4 77.3 82.0 87.3 72.8 79.4 87.3 74.7 80.5 91.3 84.1 87.5 90.9 77.9 83.9 91.1 80.7 85.5</td></tr><tr><td colspan=\"19\">4 ICTRC-NLPGroup 87.0 76.1 81.2 86.2 70.2 77.4 86.5 72.7 79.0 89.2 83.1 86.1 89.8 76.5 82.6 89.7 79.4 84.2</td></tr><tr><td>5 UT-NLP-IR</td><td colspan=\"18\">77.3 70.2 73.6 74.1 61.9 67.5 75.5 65.4 70.1 92.7 79.3 85.4 91.1 68.4 78.1 91.9 73.1 81.4</td></tr><tr><td>6 SpeechTrans</td><td colspan=\"18\">38.0 34.5 36.2 38.9 33.6 36.0 38.5 34.0 36.1 76.1 32.9 45.9 74.9 30.3 43.2 75.7 31.5 44.5</td></tr><tr><td>7 Baseline</td><td colspan=\"18\">32.8 45.7 38.2 32.0 38.1 34.8 32.4 41.3 36.3 50.6 47.8 49.2 42.6 35.1 38.5 46.5 40.9 43.5</td></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "Evaluation of systems for subtask 7-classes" | |
| }, | |
| "TABREF7": { | |
| "type_str": "table", | |
| "content": "<table><tr><td>Team</td><td colspan=\"6\">Named Entity Classes PER ORG LOC DAT TIM MON PCT</td><td>F1</td></tr><tr><td>1 MorphoBERT</td><td>90.4</td><td>80.3</td><td colspan=\"2\">87.1 78.9 71.0</td><td>93.6</td><td colspan=\"2\">96.8 85.4</td></tr><tr><td colspan=\"2\">2 Beheshti-NER-1 81.8</td><td>80.8</td><td colspan=\"2\">88.0 77.8 75.8</td><td>85.1</td><td colspan=\"2\">91.6 83.5</td></tr><tr><td>3 Team-3</td><td>79.9</td><td>77.2</td><td colspan=\"2\">83.9 74.7 64.3</td><td>92.1</td><td colspan=\"2\">97.4 80.5</td></tr><tr><td colspan=\"5\">4 ICTRC-NLPGroup 76.2 75.93 82.8 76.0 67.1</td><td>91.3</td><td colspan=\"2\">93.6 79.0</td></tr><tr><td>5 UT-NLP-IR</td><td>63.4</td><td>58.8</td><td colspan=\"2\">78.2 76.1 69.1</td><td>84.5</td><td colspan=\"2\">93.5 70.1</td></tr><tr><td>6 SpeechTrans</td><td>24.3</td><td>23.5</td><td>63.1 12.0</td><td>4.1</td><td>0.3</td><td>0.7</td><td>36.1</td></tr><tr><td>7 Baseline</td><td>23.5</td><td>38.1</td><td colspan=\"2\">44.2 41.6 30.3</td><td>13.7</td><td colspan=\"2\">36.6 36.3</td></tr></table>", | |
| "num": null, | |
| "html": null, | |
| "text": "Details of phrase-level evaluation for subtask 7-classes (values are F 1 score)" | |
| } | |
| } | |
| } | |
| } |