INTRODUCTION:

Available information today in the web are stored in huge databases in a text format, including large amount of documents such as news articles, magazines, e-books, research papers ...etc. These collections are stored in a text databases from many sources in an unstructured text format that can't simply be used for processing by computers as it is in order to extract useful information from it as we cannot simply extract meaningful keywords that can act as terms used by web users to write a query that could return what the searcher exactly or partially needs [6]. And to do so, some pre-processing must be done on these documents in order to get it more beneficial for us.

With the large amount of variant collections of documents uploaded daily to the web, there is an increasing necessity to organize and index these documents in order to make it easier to the searchers retrieving it, and this organizing and Indexing operation should be done automatically.

One of the most interests and demanding field of information retrieval nowadays is Text (online Document) indexing which refers in general to automatically extracting meaningful keywords that can be used to build an index for the documents suitably [9]. This necessity of developing and using such indexing paradigm has been risen since there are very large percent of documents haven't author-assigned Key-Terms or Key-phrases as a document's metadata description that may influence their contents [1].

Text indexing is heavily applied on English documents, but unfortunately that's not true for Arabic ones. Hence, one of the most challenging problems in this context is applying text indexing mechanisms on Arabic text documents. We intend in this paper to make a comprehensive study on the most popular proposed methodologies to cover this interesting researching area.

The remaining of this paper is organized as follows: in section 2, we introduced sixteen research papers that deal with our topic, while in section 3 we will produce some comparison metrics between these methodologies and finally in secion 4 we end our paper with a conclusion section.

Related work:

Although there have been many works on text indexing through the last decade, but unfortunately there have been little deals with Arabic text documents, some of these works used single-term(word) indexing technique while others deal with phrase (sequence of words) terms indexing technique, some researchers making comparisons between deferent techniques and gave us preferability to one against others based on some experimental results, on the other hand other researchers propose some new technique or made some enhancements on existing ones either depends on statistical or un-statistical methods . Final set of papers proposed tools as key-terms extractors to be used in text indexing.

In [1], authors proposed a learning method for extracting key-phrases of Arabic text documents that doesn't depends only on statistical term weighting methods and distance, but it also depends on using both linguistic knowledge -which provides improvements in efficiency- along with machine learning technique aimed at classifying candidate key-phrases being extracted as valid key-phrases or not. This linguistic knowledge includes extracting lexical features of documents from a specially designed annotated Arabic corpus, it also includes using syntactic rules based on part of speech tags and allowed word sequences to extract candidate key-phrases. They used two learning algorithms, C4.5 decision tree classifier and GenEx (Genitor and Extractor) algorithm. Three basic steps were discussed in their paper: the first one is Linguistic pre-processing, second one is defining the candidate key-phrase terms and what features of these terms are considered discriminative using the linguistic knowledge, while the third step is determining the final key-phrases list using learning algorithms after making feature vector calculation. In First step every document is pre-processed in order to correct any possible writing mistakes. Then, the input document is stemmed into words and phrases. In their work, Arabic terms are not reduced to its stems, instead the abstract form of the word is only used (they meant by the abstract form is the same as making light stemming in which only affixes are removed and not reducing the words to its root), then Each word is analyzed to extract its lexical features. In second step, following some syntactic rules driven from linguistic knowledge, all of one, two or three successive words appear in a document that represent possible phrases are generated as n-gram terms. Also extracting the abstract form of the candidate phrase words is done here in this step. Finally, each candidate phrase given from the previous step is assigned a number of features used to evaluate its importance. In their algorithm that responsible of determining that importance, three factors were taken in consecration that determine the selection of such features and their corresponding values which are: (1)The absolute importance of the phrase which determined independently from its original document, (2) Heuristics, where the feature values are computed through investigating human experts manually written key-phrases, and finally (3) feature extraction process and their associated values are based on the abstract form of the phrases only. A learning model built using the LDA (Linear Discriminant Analysis) is used here as Training phase through training documents having author-assigned key-phrases. A set of 30 documents from different sources and domains with focus on the computer area has been collected, pre-processed, analyzed, and manually key-phrase-assigned. Some of these documents have author-assigned key-phrases; other documents haven't, so they have carefully been read by many field specialists in order to assign the proper document key-phrases. Experiments were held on three data sets contain 50 documents to measure the level of acceptance of the extracted key-phrases. Key-phrases extracted by the presented system were compared to those extracted by two Arabic key-phrase extraction systems: KP-Miner and Sakhr Keyword Extractor. Experiment showed that their tool has a significantly better performance than the existing Arabic extractor systems, where precision and recall values reach double their corresponding values in the other systems especially for lengthy and non-scientific articles. Their future works includes the use of simpler compact corpus with light stemmer and studding the effect of word class ambiguity on the system performance.

In [2], the authors proposed system to create automatic indexing for Arabic books depending on text summarization and abstraction processes to collect the main topic and statement. The system contains two phases, the first phase make pre-processing for the documents by removing stop words, punctuation marks, non-letters, diacritics, and any English word or letter. After that compute the term frequency in document and record them in decreasing way, then remove the highest and lowest frequency and matches between the page number and the term in which they occur in the document. Finally add the generated index at the end of document. The second phase of the system is the same process of phase one but here the system makes check whether the document is already exist in the database or not, if it does not exist, then the system makes the same process in the first phase. The authors in this paper held their experiments on 50 documents from 10 different categories collected from several Arabic website, each document contains a number of words between [1500, 4000]. They used Precision and Recall as an evaluation metrics for their proposed methodology and by applying many experiments they came up with average precision for 5 documents with 0.998 while recall was 1.00 results.

In [3], authors proposed project to evaluate IR (information retrieval) efficiency and improve it by encoding the index terms using Huffman code and the results of their study showed a reduction of index size with reduce the term size by 40%, and consequently improve index transfer time. Their results minimizes the number of comparisons in each query process by 36% using binary search and consequently improves the retrieval efficiency and reduces the processing CPU-time for each query. The experiments were held on three different collections of documents, first document (bible.txt) contains 748,368 words with 12,596 distinct index terms, the second document which is (world192.txt) that contains 280,298 words with 18,862 distinct index terms and final one is (Encarta Articles) with 1,000 articles from Microsoft encyclopedia that contains 414,671 words with 28,285 distinct index terms.

In [4], the authors provide a new approach called Arabic Noun Phrase Extractor (ANPE), that present a certain methodology to extract the noun-phrases (NPs) from Arabic documents depend on three units like tokenization, part-of-speech tagging, and noun phrase identification. The authors used this approach to create information retrieval (IR) system using some of information retrieval techniques like Noun Phrase indexing, this system aimed to help the user within search to get the most likely relevant documents from the collection of documents. The authors depend in the following steps to implement the IR system: Step 1: the pre-processing by removing special characters, stop words, punctuation marks, non-Arabic letters, diacritics, and any English word for each input collection documents. Step 2: create the inverted file for each collection documents. Step 3: chosen the target query from source query list. Step 4: start the search with chosen query. Finally, Step 5: determined the similarity between chosen query and retrieved documents form collection document using the cosine similarity formula. The cosine similarity formula has archived most likely relevant documents that user need it by depending in some of equitation’s like TF-IDF. The authors in experiment process depend on two different data-set: “LCD” that collected by Agency France Press (AFP), and “collection of newspaper” that collected from the European Language Resources Distribution Agency. The authors on evaluation process depend on two different metrics like precision and recall measures, and apply two measures between proposed system using noun phrase indexing and other system used the single term indexing, and get more relevant documents by using noun phrase terms rather than using single terms.

In [5], authors provide a new System of Indexing and Classifying of Arabic Texts (SICAT) based on a heuristic lemmatization without depending on any linguistic rule, and the proposed system work in specific domains like politics, sports, economics, medicine, and agriculture. The authors in his research depend on a lot of natural language processing (NLP) applications such as automatic abstract and the research of information applications. The core idea of proposed system is studying and extracting the pertinent terms from some of collected documents related to the available system domains by applying a set of formatting techniques on input documents, calculates average frequencies for each input documents, and finally classify these documents with suitable domain. The authors create his system depend on the indexing and the classification stages. The indexing stages contain three phase: the pre-learning, the lemmatization, and the frequencies calculation. The pre-learning, in this stage, the SICAT system make some of pre-pressing with the input documents by removing special characters like (#, %, @, *, /, +,?and!), stop words, punctuation marks, non-Arabic letters, diacritics, and any English word, that obtain the better accuracy of result. The lemmatization, in this stage, the SICAT system given a unique form for all term that have the same root to simplify the treatment and allow to calculate the frequency with an effective way, by analyses the input text and make comparison with two words with looking in the first, second word one and vice-versa. The frequencies calculation, in this stage, the SICAT system using TF-IDF approach to calculate the weight frequency for all terms in each domain, that allows to determine the pertinent terms. The classification stage, contain two phase: the extraction of keywords and classify the new document. The extraction of keywords, in this stage, the SICAT system extract the suitable pertinent terms within a list for each domain. Finally, the text documents classifications, in this stage, the SICAT system classify the input text document in to suitable domain. The authors in experiment process depend on special corpus composed around of(12500) words with (25) Arabic text document in specific domains like politics, sports, economics, medicine, and agriculture collected by authors from newspapers, Internet, and magazines with (5) Arabic text document for each domain. The authors on evaluation process depend on two different metrics like precision and recall measures, and apply two measures on different chosen domains, and the evaluated result they showed 90% precision and 85% recall.

In [6], Al-Shammari propose a similarity measure to stem Noisy Arabic document since there is a necessity to handle such texts (noisy one) which cannot be stemmed by applying common rules usually used to be applied on modern Arabic language. The data set used in her paper is collected from the Holy Book of Islam "Qur'an", she explained the reason behind using "Qur'an” by saying that the Qur'an orthography differs from the modern Arabic orthography. The evaluation metrics used to evaluate her stemmer performance was: 1) under-stemming index 2) over stemming index 3) the stemming weight. The experiment results when comparing this stemming technique with khoja stemmer and larky's light10 stemmer, showed promising future for this approach which encourage further research.

In [7], the authors provide new system called KP-Miner, that used in different area of information retrieval and the authors proved its effectiveness and efficiency by applying different comparisons and experiments with different widely used systems in the same area. KP-Miner system presents a new approach to extract the key-phrase with high accuracy from unstructured text of Arabic and English documents without relying on the training samples as it is in the machine learning extractors, by applying the following three steps: candidate key-phrase selection, candidate key-phrase weight calculation and key-phrase refinement. In the first step, KP-Miner system determines the location of key-phrases by using the punctuation location and stop words that given in the text. In the second one, KP-Miner calculate the key-phrase weight depend in TF-IDF algorithms while in the final step, KP-Miner allow the user to determine the number of key-phrase (n) that system return within a list, and the system return the top (n) of key-phrase depending on the key-phrase weight, by eliminating the lower weight and return the (n) key-phrase higher weight within a list to the user. KP-Miner system can be applied in any other language with simple adjustments like the stop word list, and Stemmer approach. The authors in evaluation process depends on using different metrics like precision, recall, the average number of keywords extracted correctly per document (Avg. Keys) and the average F-measure. In English language evaluation process the authors depends on 7 different data-sets: “NASA” with 141 documents, “FIPS” with 35 documents, “Journals” with 75 documents, “Aliweb” with90 documents, “CSTR” with 80 short abstracts of document collected by the KEA-3.0 system, “Journal 2”with 60 documents collected by authors from ACM digital library and http://cogprints.orgto make the initial test for KP-Miner system, and finally “Journal 3” dataset with 21 documents collected by authors from the ACM digital library and http://cogprints.orgto train KEA-3.0 system. Results of evaluation process when comparing between both the KP-Miner and KEA system when extracting seven keywords, the KP-Miner system returns roughly 22% more correct key-phrases rather than the KEA system and extract 18% and 15% more than when the KEA system extracting 15 and 20 keywords. In Arabic language evaluation process the authors depend on the following two data-set: 100 documents collected by authors form Arabic Wikipedia with average number of word is ±[804-934] and average number of key-phrase is ±[3.2-8.1], and the second data-set is 18 agricultural extension documents provided from Egyptian Ministry of agriculture were concatenated into one big document with 53,696 words. Result of evaluation process the authors present two experiments, the first experiment is comparison the KP-Miner system with other key-phrase extractors system (Sakhr) and found the KP-Miner system better than Sakhr system with 57.5%. The second experiment is extracting key-phrases from a domain-specific corpus and found 66% extracted key-phrases with 33 matched AGROVOC from 50 extracted key-terms.

In [8], authors proposed an application of indexing approach and retrieval for Arabic language that Depends on N-grams (a set of combined sequence of characters). N-gram used for retrieving text. Text retrieval consists of three Steps, first step is document indexing: extract terms from text document, Second step is giving weight for index term: give each extract terms a weight, Third step is documents ranking: documents are ranked depending on similarity). The authors split their work into two sections: first one is segmentation and indexing of text, the second section is assigning n-grams as a vector using vector space model in which documents and queries are represented as vectors in order to find the similarity. The authors used 3 calculation metrics of similarity, the first one is TF-IDF weighting scheme: Term Frequency-Inverse Document Frequency, the Second is Dice's coefficient based on lexical similarities on word, the third is Cosine Coefficient based on calculate the cosine between document vector and query vector). The authors display Spelling problem in Arabic language such as "changing of the letter YEH (ي) to ALEF MAKSURA (ى) at the end of a word and combining HAMZA or MADDA with ALEF (e.g. أ, إ )" and they gave a solution for it as a normalization technique Change all "ALEFs" in the word to "ALEF", after that they display the benefits of using N-grams such as it can be applied on any language , it is insensitive to change spelling, words is bigger than vocabulary for n-gram. The core of the author's work is deleting of stop words from documents and normalize it, after that they choose "3-gram, 4-gram or 5-gram" of characters and represent it as vector of terms. Authors applied N-gram method on a dataset collected from " Lebanese government’s official journals" on 2676 Arabic documents by using 3 calculation type of similarity and display The result as table for each n-gram ( Fiftieth results for all table). For 3 grams, When using TF-IDF weighting scheme, they found a best results for F1 measure with 36.44, and for Dice's coefficient F1 equals 24.53 when a threshold equal 0.6. This means using trigrams (N=3) to index Arabic documents is the best choice in Arabic information retrieval.

Nashat Mansor and his colleagues in [9], proposed an auto indexing method for Arabic texts based on morphological analysis and using weight calculation to words based on not only it's occurrence frequency, but also on how the occurrence is spread across the hole document. Their method separated into 3 phases: pre-processing step, stem word extraction and finally weight assignment and index selection. From their experimental results they showed a result of 46% for average recall and an average precision of about 64% for 24 texts.

In [10], the authors proposed an enhancing on khoja stemmer to stem Arabic words that uses two dictionaries one for radicals and the other is for roots, and the reason behind using these two dictionaries is that sometimes stemmer may return many terms (Arabic words) that have deferent meanings to the original term (word root), this leads to limit all these word's semantic into only one, for instance the words "الكاتب", "الكتاب", "المكتب", "المكتبه", which means in English: the writer, the book, the desk, the library respectively, and hence reduce them to the root "كتب" which means wrote. In addition to this, they use three lists: list of infixes, prefixes and postfixes. from their experimental results, they showed a comparison between this approach and two others, one based on only root-dictionary, the other based on radicals-dictionary, their results gives great recall and precision on many Arabic texts from the corpus of press "ArabiCorpus".

In [11], author’s provides a new tool to extract multi-word term (MWT) from unstructured Arabic document depending on linguistic specification. The methodology of term extraction depend on two main step, the selection of MWT-like units and set the ranking by means of statistical techniques for MWT-like unit and classify the MWT-like unit to their part-of-speech using diab’s tagger after that calculate the term weight by used the log-likelihood ratio (LLR) and Okapi BM-25. The authors in this study depend on specialized corpus that collects from different website with general domain in Arabic language that contains 1.062 documents and 475.148 words.

Authors in [12], proposed retrieving “Hadith” automatically, the authors used in their study a vector space model, because vector space model speeding up the calculation and gives a good results. In the first part of their research, they presented the "Hadith" structure. "Hadith" is compound of two parts: the first part is "Isnad"; the second part is "Matn". After that they propose a pre-processing on Arabic texts that depends on automatic pre-processing by using the word affixes and extracted roots from the "matn", after this process the indexing on "hadith" can be done ,they uses the TF-IDF (Term Frequency, Inverse Document Frequency) to give weight for each term. The equation is TF-IDF (w, d) = TF *IDF. After that they calculate the weight for each term. They used a vector space model to calculate the similarity through calculate cosines between “Hadiths" vector and query vector. Finally the authors proposed sequence diagram for indexing "Hadiths" and retrieving system for it to illustrate Interdependence between all Components of this system such as " user Interface allows Interaction between the user and the system by formulation the query, searching and Indexing, relevance feedback". Authors apply they work on 60 "Hadiths" and a set of query from "Sahih Muslim", they organize the work in 2 tables, one for Index terms weight and the other table for similarity. After that they found the Precision and Recall for 12 "Hadiths". A result of 80% for average recall and an average precision of about 66% have been shown in their experiments.

In [13], authors proposed a procedure to refine the store of index and speed up the system by decreasing the proportion of recall and precision. This procedure is done by using part of speech tagging. The authors tagged two hundred and forty-two Arabic documents from "the proceedings of the Saudi Arabian national conferences". In the first part of their research, they displayed the type of classic model that is using in information retrieval (Boolean model, vector model, probabilistic model) and they display how it is working for each of them, they use a vector space model to test the work because it has a set of positives things such as weight (K, d) is non-binary and positive and the query index term is weighted. After that they displayed the type of classification on Arabic word (none, verbs, particles). The core of the work is using part of speech tagging to choose index term that Increases performance of retrieval Arabic language. They use tagging system constructed by "ghassan kanaan and riyad al-shalabi and majdi sawalha". The first step of their core work is tagging Two hundred and forty-two Arabic documents in which nouns, verbs and other part of speech have been extracted and Puts each of them in a separated index term. The second step, they apply this 3-type of tagging for Two hundred and forty-two Arabic documents with (Thirty-six queries Created by 3 graduated students) on the information retrieval system. The authors made a program to calculate the recall=|Ra|/|R| and precision=|Ra|/|A| for 3-type extracted term (noun, verb, other part of speech) and represent it as Curve .Finally, the idea behind their work is showing that using noun as index term is decreasing the average of precision and recall below than 1% and speed up the information retrieval system and decrease the overhead on system because the noun on documents represent 55% of the word.

In [14], authors focused on the value of multiple type of terms' interaction in their information retrieval research, specifically examining both words and N-gram as indexing terms. In addition to that, they aimed to study the effect of using alternative term weighting schemes and effects of using cosine with corpus Centroid subtraction as a measure of similarity rather than computing it relative to the origin. In order to achieve these objectives, Authors built HAIRCUT system "The Hopkins Automated Information Retriever for Combing Unstructured Text". HAIRCUT system was implemented in java and it implements a vector model, using cosine as its similarity measure which might be computed either relative to the origin or with corpus Centroid subtraction. To compute weights of terms, Okapi BM 25 which is a variant of TF/IDF weighting had been used. Normal TF and TF-IDF weighting schemes were also supported. They listed terms that appear in a high percentage of documents as effect terms used to index documents. HAIRCUT performs elementary pre-processing on queries in the same way as documents in the collection through parsing and removing stop structure i.e. affixes. Their experiment's pre-processing included removing punctuation, down-casing letters, and mapping numbers to a single digit. A ward was any remaining "blank-delimited" series of characters. After doing that, TREC-7 was tested against two types of terms: words and 5-grams. When comparing the use of non-stemmed words against 5-grams, a good performance obviously was achieved in the 5-grams case. When they studied the variability in performance caused by experimenting deferent weighting schemes on the TREC-7 Topics, they found both Okapi and TF/IDF term weightings outperforms the use of normal TF alone. They came up with Four Key observations: In term of improvement, Okapi and TF/IDF term weighting schemes are nearly equivalent, which are better than TF scheme. At lower recall Percentages, the use of Okapi and TF/IDF appears to give better results for word-based terms than for n-gram ones. Results were based on unofficial runs, i.e. "the top one hundred documents are not guaranteed to have been assessed". The average number of documents in the top one hundred that were assessed in the Okapi and TF/IDF runs ranges from 92% to 95%, but the TF runs averaged only 78% for n-grams and 66% for words. Finally, they have yet to see circumstances under which the use of Centroid subtraction is justified, which will be their future work.

Comparison:

In this section, we make a comparison between all papers studded here in the review of literature that we have already made in the previous section. To make this study, we take in consideration the following factors: (1) Index type used that could be word term, phrase term or both. (2) Proposed Method that had been used in each paper (usually referred to the method name or technique used). (3) Statistical computations that pares may use either to compute similarity or term weight vector. (4) Model, which represents the information retrieval model that were used to calculate some statistical computations such as similarity, i.e. vector model. (5) D.S which describes data set used by researchers to apply their experiments on it. (6) Evaluation measures that had been used to evaluate the proposed methodology or system, i.e. recall, precession. (7) Experimental results that had been shown to proof the proposed method effectiveness. (8) Statistical, un-statistical that classify papers into methods that using statistical, un-statistical or both technique.

Comparisons based on proposed methodology

In table1 we represent some of comparison critics about proposed methodology. Comparison factors are Index type, Model, Statistical, un-statistical, Proposed Methodology, Statistical computations.

Data set, evaluation measures and experimental results comparisons

In table2 we present a comparison between previously discussed papers based on used corpus, measures used to evaluate the proposed methodology and finally experimental results.

Table1. Methodology Comparison

Paper #	Index Type	Model	Statistical Un-statistical	Proposed Methodology	Statistical Computations
[1]	Phrase Terms	-	Un-statistical	Learning method for extracting Key-phrases using Linguistic knowledge and learning algorithms (C4.5 DT, GenEx)	Term weighting and Distance
[2]	Word term	-	Statistical	Extract index term from Arabic books based on lexical analysis	TF and Term occurrence
[3]	Word term	-	Un-statistical	Improve the efficiency of IR system by encoding the index terms using Huffman code	Encoding the Index term by Huffman code
[4]	Noun Phrase	-	Statistical	ANPE system to extract the Arabic noun phrase depend on tokenization, part-of-speech tagging, and noun phrase identification	cosine similarity formula and TF-IDF
[5]	Word term	-	Both	SICAT system to indexing and Classification of Arabic Texts based on a heuristic lemmatization	TF-IDF
[6]	Word term	-	Statistical	Proposed similarity measure to stem noisy documents, since there is a necessity to handle such texts which cannot be stemmed by applying common rules applied on modern Arabic language.	-
[7]	Phrase term	-	Both	New system called KP-Miner used to extract key-phrase from unstructured text in Arabic and English documents without relying on training samples as in machine learning extractors	TF-IDF
[8]	Word term	Vector model	Statistical	Using N-gram for Arabic documents indexing	TF-IDF Dice’s coefficient Cosine coefficient
[9]	Word term	-	Statistical	Auto indexing method using morphological analysis and weighting scheme depends on term occurrence frequency and occurrence spread across whole document.	TF- - Term occurrence spread
[10]	Word term	-	Statistical	Enhancing on khoja stemmer by using two dictionaries, radicals and roots in addition to three lists: list of infixes, prefixes and postfixes	-
[11]	Multi word term	-	Statistical	New system called (MWT-extractor) used to extract multi-word term from Arabic documents depend on linguistic specifications	log-likelihood ratio (LLR) and Okapi BM-25
[12]	Word term	Vector model	Statistical	Using vector model for automatic indexing and retrieval of "HADITHS"	TF-IDF
[13]	Word term	Vector model	Statistical	Using part of speech tagging for Arabic documents indexing	TF-IDF
[14]	Word and Phrase Terms	Vector model	Statistical	- Indexing using Both Words and N-gram In addition to using alternative term weighting schemes - Using cosine with corpus Centroid subtraction as a measure of similarity rather than computing it relative to the origin	Okapi BM-25 TF-IDF TF

Table2. Data set, evaluation measures and experimental results comparisons

Paper #	Data Set	Evaluation Measures	Experimental Results
[1]	- Specially designed annotated Arabic corpus to extract lexical features. - 30 doc from different sources focus on the computer area for learning phase. - Three Data Set contain 50 Doc for testing	Recall Precision	Precision and recall reach double value of KP-minor and Sakhr extractor systems specially for lengthy and non-scientific articles
[2]	50 documents with 10 different categories collected from several Arabic website and each document contains between [1500, 4000] words.	Precision and recall	With 5 documents the average precision 0.998 and average recall 1.00 results
[3]	Three data set: “bible.txt” with 748,368 words, “world192.txt” with 1443.283 words, and 1000 article from Microsoft Encarta	size of index and execution time	Reduction of index size by 40% and Reduces the number of comparisons process for each query by 36%
[4]	Two data set: "LCD" collected by Agency France Press and collection of newspaper collected by European Language Resources Distribution Agency	Precision and recall	Get more relevant documents by using noun phrase terms rather than using single terms
[5]	25 Arabic documents in different domain collected by authors from newspapers, Internet, and magazines with 12500 words	Precision and recall	90% average precision and 85% average recall
[6]	Holy Book of Islam "Qur'an"	- Under stemming. - Over stemming. - stemming weight	Promising results showed When comparing this technique with khoja stemmer and Larky's light stemmer
[7]	Five data set in English with 421 documents,100 documents in Arabic form Arabic Wikipedia with average #word is ±[804-934] and average #keyphrase is ±[3.2-8.1] and one big document with 53,696 words	Precision, recall and average F-measure	The KP-Miner system better than KEA system with 22% corrected key-phrases in English document, and it's better than Sakhr system in Arabic documents by 57.5%
[8]	2676 Arabic document from-the Lebanese government’s official journals	Avg. recall Avg. precision	41% recall 32% precision Similarity Result: For 3-grams: -VSM =36.44 -Dice’s coefficient =24.53
[9]	24 general-purpose texts with various lengths	Avg. recall Avg. precision	46% recall 64% precision
[10]	ArabiCorpus	Avg. recall Avg. precision	Their methods gives great recall and precision comparing with stemmers with one dictionary
[11]	Special corpus that contain 1.062 documents with 475.148 average words collected by authors from different website like ”Al-Khat Alakhdar” and ”Akhbar Albiae”	Precision and recall	The average precision in SWT is 26,1%, MWT_C is 29,7% with 3,6% enhancement, and MWT_D is 31,9% with 5,8% enhancement
[12]	60 "Hadiths" From "Sahih Muslim”	Avg. Recall Avg. precision	80% recall 66% precision
[13]	242 Arabic Doc from Saudi Arabian national conferences proceedings	Avg. Recall Avg. precision	Using noun as index term decreasing the average of pr and re below than 1%
[14]	TREC-7	Performance Recall Precision	- Better results of 5 grams comparing with Words terms - Okapi and TF-IDF outperform TF - For low recall level, Okapi and TF/IDF better for word terms than for n-gram