APPLYING GENETIC ALGORITHM IN QUERY IMPROVEMENT PROBLEM. Abdelmgeid A. Aly

Transcription

1 International Journal "Information Technologies an Knowlege" Vol. / [Project MINERVAEUROPE] Project MINERVAEUROPE: Ministerial Network for Valorising Activities in igitalisation - < [Rainie, 2006] Rainie, L. Life Online: Teens an technology an the worl to come (Annual conference of Public Library Association, Boston, March 23, 2006) - < an technology.pf> Authors' Information Detelin Luchev - Ethnographic Institute with Museum, BAS; Bulgaria, 000 Sofia, Moskovska str. 6A; luchev_etelin@abv.bg APPLYING GENETIC ALGORITHM IN QUERY IMPROVEMENT PROBLEM Abelmgei A. Aly Abstract: This paper presents an aaptive metho using genetic algorithm to moify user s queries, base on relevance jugments. This algorithm was aapte for the three well-known ocuments collections (CISI, NLP an CACM). The metho is shown to be applicable to large text collections, where more relevant ocuments are presente to users in the genetic moification. The algorithm shows the effects of applying GA to improve the effectiveness of queries in IR systems. Further stuies are planne to ajust the system parameters to improve its effectiveness. The goal is to retrieve most relevant ocuments with less number of non-relevant ocuments with respect to user's query in information retrieval system using genetic algorithm.. Introuction Several researchers have use the GA in IR an their results seem to inicate that this algorithm coul be efficient. In this vein, the main irections concern moifying the ocument inexing [] an [2] an the clustering problem [3]. It is not surprising therefore that there have recently appeare many applications of GAs in information retrieval. Most of them use the vector space moel, which also seems to be one of the most wiely, use moels in general [4]. These applications implement learning of the terms an/ or weights of the queries. Information Retrieval systems are use to retrieve ocuments that epen on or relevant to the user input query. The growth in the number of ocuments mae it necessary to use the best knowlege or methos in retrieving the most relevant ocuments to the user query. Information Retrieval systems eal with ata bases which are compose of information items ocuments that may consist of textual, pictorial or vocal information. Such systems process user queries trying to allow the user to access the relevant information in an appropriate time interval. The art of searching will be in the atabases or hypertext networke atabases such as internet or intranet for text, soun, images or ata, [4]. Thus an information system has its heart a collection of ata about reality [5]. Most of the information retrieval systems are base on the Boolean queries where the query terms are joine by the logical operators AND an OR. The similarity between a query an ocuments is measure by ifferent retrieval strategies that are base on the more frequent terms foun in both the ocument an the query. The more relevant ocument is eeme to be the query request. The most frequently use measures of retrieval effectiveness are precision, the percentage of the retrieval ocuments that are relevant an recall, the percentage of the relevant ocuments that are retrieve. Information retrieval is concerne with collection an organization of texts, responing to the requests of users for the information seeking texts, retrieving the most relevant ocuments from a collection of ocuments; an with retrieving some of non-relevant as possible. Information retrieval is involve in: - Representation, - Storage, - Searching,

2 30 International Journal "Information Technologies an Knowlege" Vol. / Fining ocuments or texts which are relevant to some requirements for the information esire by a user. The input of GAs [6] is a population of iniviuals calle chromosomes, which represent the possible solutions to the problem. These are either generate at ranom or, if one has some knowlege can be use to create part of the initial set of potential solutions [7]. These iniviuals change (evolve) over successive iterations calle generations, via processes of selection, crossover, an mutation. The iterations en when the system no longer improves or when a pre-set maximum number of generations has been reache. The output of the GA will be the best iniviual of the en population, or a combination of the best chromosomes. To solve each problem, one has to provie a fitness function f. Its choice is crucial for the proper functioning of the algorithm. Given a chromosome, the fitness function must return a numeric value that represents its utility. This score will be use in the process of selection of the parents, so that the fittest iniviuals will have a greater chance of being selecte. This paper presents a GA base on Cosine fitness function an Classical IR in query optimization problems, an some applications of GA in information retrieval system. This algorithm has been applie on three well-known test collections (CISI, CACM an NPL). The goal is to retrieve most relevant ocuments with less number of nonrelevant ocuments with respect to user query in information retrieval system using genetic algorithm. 2. Anteceents 2.. Information retrieval moels Several retrieval moels have been stuie an evelope in the IR area, we analyze some of these moels, which are: Boolean moel. In the Boolean retrieval moel, the inexer moule performs a binary inexing in the sense that a term in a ocument representation is either significant (appears at least once in it) or not. User queries in this moel are expresse using a query language that is base on these terms an allows combinations of simple user requirements with the logical operators AND, OR an NOT. The result obtaine from the processing of a query is a set of ocuments that totally match with it, i.e., only two Possibilities are consiere for each ocument: to be or not to be relevant for the user s nees, represente by the user query [8][9]. Vector space moel. In this moel, a ocument is viewe as a vector in an n-imensional ocument space (where n is the number of istinguishing terms use to escribe contents of the ocuments in a collection) an each term represents one imension in the ocument space. A query is also treate in the same way an constructe from the terms an weights provie in the user request. Document retrieval is base on the measurement of the similarity between the query an the ocuments. This means that ocuments with a higher similarity to the query are juge to be more relevant to it an shoul be retrieve by the IRS in a higher position in the list of retrieve ocuments. In This metho, the retrieve ocuments can be orerly presente to the user with respect to their relevance to the query [8]. Probabilistic moel. This moel tries to use the probability theory to buil the search function an its operation moe. The information use to compose the search function is obtaine from the istribution of the inex terms throughout the collection of ocuments or a subset of it. This information is use to set the values of some parameters of the search function, which is compose of a set of weights associate to the inex terms [0][] Evaluation of information retrieval systems There are several ways to measure the quality of an IRS, such as the system of efficiency an effectiveness, an several subjective aspects relate to the user satisfaction. Traitionally, the retrieval effectiveness (usually base on the ocument relevance with respect to the user s nees) is the most consiere. There are ifferent criteria to measure this aspect, with the precision an the recall being the most use. Precision ( P ) is the rate between the relevant ocuments retrieve by the IRS in response to a query an the total number of ocuments retrieve, whilst Recall ( R ) is the rate between the number of relevant ocuments retrieve an the total number of relevant ocuments to the query existing in the ata base [9]. The mathematical expression of each of them is showe as follows: r f P =, f R = r f r ()

3 International Journal "Information Technologies an Knowlege" Vol. / with r {0,} being the relevance of ocument for the user an f {0,} being the retrieval of ocument in the processing of the current query. Notice that both measures are efine in [0,], being the optimal value. 3. Query Definition This is the most extene group of applications of GAs in IR. Every proposal in this group uses GAs either as a relevance feeback technique or as an Inuctive Query By Example (IQBE) algorithm. The basis of relevance feeback lies in the fact that either users normally formulate queries compose of terms, which o not match the terms (use to inex the relevant ocuments to their nees) or they o not provie the appropriate weights for the query terms. The operation moe involving an moifying the previous query (aing an removing terms or changing the weights of the existing query terms) which taking into account the relevance jugments of the ocuments retrieve by it, constitutes a goo way to solve the latter two problems an to improve the precision, an especially the recall, of the previous query [9]. IQBE was propose in [2] as a process in which searchers provie sample ocuments (examples) an the algorithms inuce (or learn) the key concepts in orer to fin other relevant ocuments. This metho is a process for assisting the users in the query formulation process performe by machine learning methos. It works by taking a set of relevant (an optionally, non-relevant ocuments) provie by a user an applying an off-line learning process to automatically generate a query escribing the user s nees. Smith an Smith [3] propose a GA for learning queries for Boolean IRSs. Although they introuce it as a relevance feeback algorithm, the experimentation is actually closer to the IQBE framework. The algorithm components are escribe as follows: - The Boolean queries are encoe in expression trees, whose terminal noes are query terms an whose inner noes are the Boolean operators AND, OR an NOT. - Each generation is base on selecting two parents, with the best fitte having a larger chance to be chosen, an generating two offspring from them. Both offspring are ae to the current population which increments its size in this way. - The usual GA crossover is consiere [4]. No mutation operator is applie. - The initial population is generate by ranomly selecting the terms inclue in the set of relevant ocuments provie by the user, having those present in more ocuments a higher probability of being selecte. - The fitness function gives a composite retrieval evaluation encompassing the two main retrieval parameters (precision an recall). Yang an Korfaghe [5] propose a similar GA to that of Robertson an Willet s [6]. They use a real coing with the two-point crossover an ranom mutation operators (besies, crossover an mutation probabilities are change throughout the GA run). The selection is base on a classic generational scheme where the chromosomes with a fitness value below the average of the population are eliminate, an the reprouction is performe by Baker s mechanism. 4. System Framework 4.. Builing an IR System The propose system is base on Vector Space Moel (VSM) in which both ocuments an queries are represente as vectors. Firstly, to etermine ocuments terms, the following proceure is use: - Extraction of all the wors from each ocument. - Elimination of the stop-wors from a stop-wor list generate with the frequency ictionary of Kucera an Francis [7]. - Stemming the remaining wors using the porter stemmer, which is the most commonly use stemmer in English [4][8]. After using this proceure, the final number of terms was 6385 for the CISI collection, 726 for CACM an 7772 for NPL. After etermining the terms that escribe all ocuments of the collection, the weights were assigne by using the formula propose by Salton an Buckley [9]:

4 32 International Journal "Information Technologies an Knowlege" Vol. / 2007 Where aij is the weight assigne to the term t j in ocument appears in ocument aij = D i, tf ij log max tf 2 tf ij log max tf D i, n j is the number of ocuments inexe by the term N n i (2) tf ij is the number of times that term t j t j an finally, N is the total number of ocuments in the atabase. Finally, the vectors are normalize by iviing them by their Eucliean norm. This is accoring to the stuy of Noreault et al.[7], of the best similarity measures which make angle comparisons between vectors. A similar proceure is carrie out with the collection of queries, thereby obtaining the normalize query vectors. Then, the following steps are applie: - For each collection, each query is compare with all the ocuments, using the cosine similarity measure. This yiels a list giving the similarities of each query with all ocuments of the collection. - This list is ranke in ecreasing orer of similarity egree. - Make a training ata consists of the top 5 ocument of the list with a corresponing query. - Automatically, the keywors (terms) are retrieve from the training ata an the terms which are use to form a binary query vector. - Aapt the query vector using the genetic approach. 4.2 The Genetic Approach Once significant keywors are extracte from training ata ( relevant an irrelevant ocuments) incluing weights are assigne to the keywors. The binary weights of the keywors are forme as a query vector, an the aapting of the query vector as a chromosome. Then, the GA is applie to get an optimal or near optimal query vector, an the result of GA approach is compare with the classical IRS without using GA. Ga's are characterize by 5 basic components as follows: - Chromosome representation for the feasible solutions to the optimization problem. - Initial population of the feasible solutions. - A fitness function that evaluates each solution. - Genetic operators that generate a new population from the existing population. - Control parameters such as population size, probability of genetic operators, number of generation Representation of the chromosomes These chromosomes use a binary representation, an are converte to a real representation by using a ranom function. We will have the same number of genes (components) as the query an the feeback ocuments have terms with non-zero weights. The set of terms containe in these ocuments an the query is calculate. The size of the chromosomes will be equal to the number of terms of that set, we get the query vector as a binary representation an applying the ranom function to moify the terms weights to real representation. Our GA approach receives an initial population chromosomes corresponing to the top 5 ocuments retrieve from classical IR with respect to that query Fitness Function Fitness function is a performance measure or rewar function, which evaluates how each solution, is goo. In our work, we use the cosine similarity as fitness function with equation: N n i 2

5 International Journal "Information Technologies an Knowlege" Vol. / t x i i= t 2 xi i= y i t 2 y i i= (3) Selection As the selection mechanism, the GA uses simple ranom sampling [20][6]. This consists of constructing a roulette with the same number of slots as there are iniviuals in the population, an in which the size of each slot is irectly relate to the iniviual s fitness value. Hence, the best chromosomes will on average achieve more copies, an the worst fewer copies. Also, uses the elitism strategy [2], as a complement to the selection mechanism. If after generating the new population, the best chromosome of the preceing generation is by chance absent, the worst iniviual of the new population is withrawn an replace by that chromosome Operators In our GA approaches, we use two GA operators to prouce offspring chromosomes, which are: Crossover is the genetic operator that mixes two chromosomes together to form new offspring. Crossover occurs only with crossover probability P c. Chromosomes are not subjecte to crossover remain unmoifie. The intuition behin crossover is exploration of a new solutions an exploitation of ol solutions. GAs construct a better solution by mixture goo characteristic of chromosome together. Higher fitness chromosome has an opportunity to be selecte more than lower ones, so goo solution always alive to the next generation. We use a single point crossover, exchanges the weights of sub-vector between two chromosomes, which are caniate for this process. Mutation is the secon operator uses in our GA systems. Mutation involves the moification of the gene values of a solution with some probability P m. In accorance with changing some bit values of chromosomes give the ifferent brees. Chromosome may be better or poorer than ol chromosome. If they are poorer than ol chromosome they are eliminate in selection step. The objective of mutation is restoring lost an exploring variety of ata. 5. Experiments an Results To perform the trial, it was first necessary to generate test atabases. We create these from three of the best known test collections: the CISI collection (460 ocuments on information science), the CACM collection (3204 ocuments on Communications), an finally the NPL collection (,429 ocuments on electronic engineering). One of the principal reasons for choosing more than one test collection is to emphasize an generalize our results in all alternative test ocuments collections. The experiments are applie on 00 queries chosen accoring to each query which oes not retrieve 5 relevant ocuments for our IR system. From our experimental observation, the best values for this test ocuments collections at crossover probability P c = 0.8 an mutation rate is P m = 0.7 for GA. The following are the results of applying GA for 00 generations. 5.. The CISI Documents Collection The results for the GA are shown in table () an figure, using non-interpolate average Recall Precision relationship. From this table an the corresponing figure we notice that GA gives a high improvement than Classical IR system with.9%. Also, the average number of terms of query vector before applying GA is terms; these terms are reuce after applying GA to terms as average number.

6 34 International Journal "Information Technologies an Knowlege" Vol. / 2007 Table (): Shows the experimental results of applying GA on CISI Collection Average nine-point Recall Precision for 00 query in CISI Collection Recall Precision Classic IR GA GA Improvement % Average Classic IR GA Figure. Represents the relationship between average recall-precision for 00 queries on CISI 5.2. The NPL Documents Collection The results for this experiment are shown in table (2) an figure 2, using non-interpolate average Recall- Precision relationship. From this table an the corresponing figure, we fin that the GA gives a higher improvement than classic IR system with.5% as average values. Also, the average number of terms of query vector before applying GA is 34.4 terms; these terms are reuce after applying GA to 6.8 terms. Table (2): Shows the experimental results of applying GA on NPL Collection Average nine-point Recall Precision for 00 query in NPL Collection Recall Precision Classic IR GA GA Improvement % Average

7 International Journal "Information Technologies an Knowlege" Vol. / Classic IR GA Figure 2. Represents the relationship between average recall-precision for 00 queries on NPL 5.3. The CACM Documents Collection The results for this experiment are shown in table (3) an figure 3, using non-interpolate average Recall Precision relationship. From this table an the corresponing figure, we notice that GA gives a higher improvement than that with classic IR system 5.3%, as average values. Also, the average number of terms of query vector before applying GA is 60.7 terms; these terms are reuce after applying GA to 6.83 terms. Table (3): Shows the experimental results of applying GA on CACM Collection Average nine-point Recall Precision for 00 query in CACM Collection Recall Precision Classic IR GA GA Improvement % Average Classic IR GA Figure 2. Represents the relationship between average recall-precision for 00 queries on CACM

8 36 International Journal "Information Technologies an Knowlege" Vol. / Conclusion From the previous results, it can be conclue that our GA approach gives more sophisticate results than classical IR system in our test collections. Also, the average number of terms is reuce after applying GA. The experiments evelope use three of the relative ocument collections (CACM, CISI an NPL), an compare the results of two variant systems (Classical IR an GA). The latter algorithm achieves the best performance an it obtains better precision than the first approach. Our GA approach use to aapt the weights of query terms is to get high precision results. References [] Hollan, J. H., Aaptation in Natural an Artificial Systems, University of Michigan Press, Ann Arbor (975). [2] K. A. DeJong, An Analysis of the Behavior of a Class of Genetic Aaptive Systems, Ph.D. Thesis, University of Michigan (975). [3] V. V. Raghavan an B. Agrwal. "Optimal etermination of user oriente clusters: An application for the reprouctive plan". In Proceeings of the secon conference on genetic algorithms an their applications, Hillsale, NJ (pp ), 987. [4] Baeza-Yates, R. an Ribeiro-Neto, B., Moern Information Retrieval, Aisson, 999. [5] R. Korfhage Robert. Information storage an Retrieval. John Wiley & Sons, Inc [6] Golberg, D. E., Genetic Algorithms in Search, Optimization, an Machine Learning, Aison-Wesley, Reaing, MA. (989). [7] Z. Michalewicz. Genetic algorithms + ata structures= evolution programs. Berlin: Springer- verlag, 995. [8] Salton, G. an McGill, M. H., Introuction to Moern Information Retrieval, McGraw-Hill, 983. [9] Van Rijsbergen, C. J., Information Retrieval, secon e., Butterworth, 979. [0] A. Bookstein. Outline of a general probabilistic retrieval moel, Journal of Documentation 39 (2) (983). [] N. Fuhr. Probabilistic moels in information retrieval, Computer Journal 35 (3) (992). [2] H. Chen et al., A machine learning approach to inuctive query by examples: an experiment using relevance feeback, ID3, genetic algorithms, an simulate annealing, Journal of the American Society for Information Science 49 (8) (998). [3] M.P. Smith, M. Smith. The use of genetic programming to buil Boolean queries for text retrieval through relevance feeback, Journal of Information Science 23 (6) (997). [4] J. Koza. Genetic Programming. On the Programming of Computers by means of Natural Selection, The MIT Press (992). [5] J. Yang an R. Korfhage. Query moifications using genetic algorithms in vector space moels, International Journal of Expert Systems 7 (2) 65 9 (994). [6] A.M. Robertson an P. Willet. Generation of equifrequent groups of wors using a genetic algorithm, Journal of Documentation 50 (3) (994). [7] T. Noreault, M. McGill an M. B. Koll. A performance evaluation of similarity measures, ocument term weighting schemes an representation in a Boolean environment. Information retrieval research. Lonon: Butterworths (98). [8] M. F. Porter. An algorithm for suffix stripping. Program, 4(3), (980). [9] G. Salton an C. Buckley. Improving retrieval performance by relevance feeback. Journal of the American Society for Information Science, 4(4), (990). [20] M. Goron. Probabilistic an genetic algorithms in ocument retrieval. Communications of the ACM, 3(0), (988). [2] D. C. Blair. "Language an Representation in information Retrieval". Amsteram: Elsevier, 990. Authors' Information A. A. Aly Dept. of Computer Science, Faculty of Science, Minia University, El-Minia, Egypt; abelmgei@yahoo.com