ࡱ> Ut` Rbjbjss .ffffV!!!"$V\*'$(()))ZZD9[$‘đđđY]$hi8hY@Z8h8hff))06vsvsvs8hfn))‘vs8h‘vsvsn}~)'  2!k2b~F| <\r~|(n ~~X ][)_vsa, d-][][][s^][][][\8h8h8h8hVE:_d6R"9V:_"VVVffffff  Adapting Document Ranking to Users Preferences using Click-through Data Min Zhao, Hang Li, Adwait Ratnaparkhi, Hsiao-Wuen Hon, Jue Tang Feb. 15, 2006 MSR-TR-2006-15 Microsoft Research Microsoft Corporation One Microsoft Way Redmond, WA 98052 Adapting Document Ranking to Users Preferences using Click-through Data Min Zhao Institute of Automation Chinese Academy of Sciences Beijing, China m.zhao@mail.ia.ac.cn Hsiao-Wuen Hon Microsoft Corporation Redmond, WA, USA hon@microsoft.com Hang Li Microsoft Research Asia Beijing, China hangli@microsoft.com Jue Wang Institute of Automation Chinese Academy of Sciences Beijing, China jue.wang@mail.ia.ac.cn Adwait Ratnaparkhi Microsoft Corporation Redmond, WA, USA adwaitr@microsoft.com ABSTRACT This paper proposes a new approach for using click-through data to re-rank the documents retrieved by a search engine. The goal is to make the final ranked list of documents accurately represent users preferences reflected in the click-through data. Our approach combines the ranking result of a traditional IR algorithm (BM25) with that given by a machine learning algorithm (Nave Bayes). The machine learning algorithm is trained on click-through data (queries and their associated documents), while the IR algorithm runs over the document collection. We consider several alternative strategies for combining the use of click-through data and that of document data. Experimental results confirm that any method of using click-through data greatly improves the preference ranking, over the method of using BM25 alone. We found that a linear combination of scores of Nave Bayes and scores of BM25 performs the best for the task. At the same time, we found that the preference ranking methods can preserve relevance ranking, i.e., the preference ranking methods can perform as well as BM25 for relevance ranking. Categories and Subject Descriptors H.3.3 [Information Search and Retrieval]: Retrieval models General Terms Algorithms, Experimentation, Measurement, Performance Keywords Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CIKM04, November 813, 2004, Washington D.C., U.S.A. Copyright 2004 ACM 1-58113-000-0/00/0004$5.00. Users preferences, re-ranking, adaptation, click-through data, linear combination INTRODUCTION This paper is concerned with the problem of improving document ranking in information retrieval by using click-through data. Suppose that we have a search engine. Given a query, it can return a list of documents ranked based on their relevance to the query. During the use of the search engine we have collected users click-through data. The click-through data can represent the general trends of users preferences on the documents with respect to the queries, although some of the clicks might be noisy. Our goal here is to re-rank the documents in future search using the previously registered click-through data so that the most likely to be clicked documents are re-ranked on the top from among the relevant documents. That is to say, we are to re-rank documents on the basis of users preferences among the relevant documents. Usually users only look at the top ranked documents in searching results (cf., [19] [20] [21]), and thus if we can rank users preferred documents on the top, then we will be able to improve users search experiences, for example, save users time in search. To the best of our knowledge, no investigation has been conducted on the problem, although there is some related work as described in Section 2. In this paper, we employ a machine learning approach to address the above problem and investigate a number of simple, but in our view, fundamental methods. We employ BM25 for the relevance ranking in the initial search engine (In our experiments the search engines used for collecting click-through data were also based on BM25). We use the click-through data to train a classifier which can categorize queries into documents with scores. Specifically, queries are viewed as instances, documents are viewed as categories, and queries invoking clicks of a document are regarded as instances classified into the category. As classifier, we make use of Nave Bayes. Given a new query, BM25 and Nave Bayes can respectively construct a ranked list of documents. We consider a number of methods to combine the two lists. Linear combination is the main method among them, and the weights in the linear combination are also trained with click-through data. Three questions may arise here. First, do the preference ranking methods preserve relevance? In other words, are the documents ranked top by the methods still relevant? Second, are the methods indeed effective for preference ranking when compared with the baseline BM25? Which method is more effective in performing the task? Third, what kind of characteristics does click-through data have and how does it affect preference ranking? We conducted experiments in order to answer the above questions, using data from two search engines: Encarta and Microsoft Word Help. First, experimental results indicate that our preference ranking methods perform as well as or even better than BM25 for relevance ranking. (We evaluated relevance in terms of top k precision.) Second, experimental results indicate that the preference ranking methods using click-through data significantly perform better than BM25 for preference ranking. Furthermore, among the methods, the linear combination of scores from BM25 and Nave Bayes performs the best. (We evaluated preference in terms of click distribution). Third, we observed that the behavior of preference ranking based on click through data can vary according to time period. In conclusion, our methods can significantly improve preference ranking and at the same time preserve relevance ranking. RELATED WORK Use of Click-through Data Document retrieval can be described as the following problem: a search engine receives a query from a user, and then returns a ranked list of retrieved documents based on their relevance to the query. The list contains the links to the documents, as well as the titles and snippets of the documents. The user clicks the links of the documents which he considers relevant (and interesting), after reading the titles and snippets. Click-through data can be recorded in operation of the search engine, as described in [11]. Here, click-through data refers to pairs recorded during document retrieval. The clicks are the set of documents clicked by the user after sending the query. Some work has been done on using query logs in information retrieval and web data mining (e.g., [1]), because query logs contain a large amount of useful information. Here, by query logs we mean all the information which can be obtained during search of documents. Query log data, thus, contains click-through data. (Note that some researchers do not distinguish the two terms.) Methods of using click-through data have been proposed for meta-search function training, term expansion, and query clustering. For example, Joachims [11] proposes a method of utilizing click-through data in learning of a meta-search function. Specifically, he introduces a method of training a new model of meta-search function with click-through data, which he calls Ranking SVM. His method is unique in that it takes the relative positions of the clicks in a ranked list as training data. He has applied the method to the adaptation of a meta-search engine to a particular group of users. Oztekin et al [16] propose several methods for meta-search. They have compared the methods by using a new metric called average position of clicks calculated based on click-through data. The authors claim that the implicit evaluation can offer more statistically reliable results than explicit evaluation based on user relevance judgments. Cui et al [6] try to use click-through data to solve the problem of mismatch between query terms and document terms. With click-through data, the probabilistic correlations between query terms and documents terms can be extracted and high quality term expansion can be conducted. Furthermore, Beeferman etc. [4] propose methods for harvesting clusters of queries and web pages by using click-through data. Ling et al [13] investigate a similar problem, but focus on the issue of finding generalized query patterns and using them in the cache of a search engine. Meta Search and Relevance Feedback Meta-search attempts to improve ranking by combining ranking results returned by several lower level search engines, while relevance feedback manages to improve ranking by seeking feedbacks on relevance from users. Both of them consider performing ranking based on relevance, not based on users preference. Meta-search and related problems such as result fusion have been intensively studied [2] [3] [8] [9] [11] [12] [14] [16] [22]. One of the key issues in meta-search is to construct a meta-ranking function which combines the scores or ranks returned from lower level ranking functions. Many methods for meta-ranking functions have been proposed (cf., [2] [16]). The proposed meta-ranking functions are based on averaging, sum, linear combination, voting, interleaves, logistic regression, etc. Relevance feedback is an interactive and iterative process in which given a query the user is asked to mark retrieved documents as relevant or irrelevant, the query is then expanded with terms extracted from the relevant documents, next, documents are retrieved with the new query, and the process is repeated [18]. Since relevance judgments can become burdens to users, automatic ways of getting feedbacks, such as pseudo-relevance feedback [5], have been proposed. PROBLEM AND EVALUATION Preference Ranking Using Click-through Data The problem we address in this paper is whether and how we can use click-through data to improve preference ranking. More specifically, we consider a re-ranking (or adaptation) problem as follows: a search engine first receives a query from a user, and then presents a ranked list of retrieved documents based on their relevance to the query. Next, a re-ranking engine re-orders the ranked list of documents based on users preferences, using click-through data. The top ranked documents should be relevant, and preferred by users. In an extreme case, the re-ranking engine totally ignores the initial ranking and creates a new ranking. In another extreme case, the re-ranking engine only uses the initial relevance ranking. Preference and relevance are different notions, although closely related to each other. If users think that some documents are more worth-reading than the others among the relevant documents, then  SHAPE \* MERGEFORMAT Figure 1. Comparison of retrieval problemswe say that they prefer those documents. In this paper, preferences are assumed to be from a group of users, not a single user. Preference is more subjective and dynamic, while relevance is more objective and static. When a query is fixed, its relevant documents are almost fixed, while users preferences may change depending on user groups and time periods. In this paper, we assume that click-through data can reflect the general trends of users preferences. (Note that users preferences can be measured by other factors like dwell time, cf., [7]). It is likely that a user clicks a document which he find uninteresting or even irrelevant later. The percentage of such noisy clicks over total clicks should be low, however. (Otherwise, the search engine will not be used by users.) Therefore, click-through data can and should be used in preference ranking. Figure 1 shows the differences between the current problem and other information retrieval problems: conventional search, meta-search, relevance feedback. The current problem focuses on preference ranking, while the other problems focus on relevance ranking. In practice, usually users only look at the top ranked documents in searching results (cf., [19] [20] [21]), and thus putting users preferred documents on the top is necessary and important for document retrieval. Click Distribution An ideal way of evaluating the performance of preference ranking might be to select a number of users and a set of queries, and let every user to explicitly judge preference on each of the documents to each of the queries. Furthermore, for each query, counting the number of preference judgments on each document, ranking the documents in descending order of their preference counts, and taking the order as the correct answer. This evaluation can give quite accurate results. However, it is costly, and thus is not practical. We consider here the use of click distribution for the evaluation. Click distribution is the distribution of document positions clicked by users in the ranked lists. If a user clicks k documents, a method that places the k documents in higher ranks (ideally the first k ranks) is better than a method that does not. The advantage of using the measure is that a large amount of click-through data can be used. The disadvantage is that click-through data may contain noises. However, as is explained above, the percentage of noisy data should be low. Thus, click distribution is a good enough measure for evaluation of general performance of preference ranking. METHODS We propose a machine learning approach to address the problem of preference ranking using click-through data. It turns out to be a problem of using click-through data to re-rank documents so that the most likely to be clicked documents are re-ranked on the top from among the relevant documents. We first use click-through data to train a classifier which categorizes queries into documents with scores. It turns out to be a problem of text classification. Specifically, we view queries as instances, documents as categories, and queries invoking clicks of a document as instances classified into the category. We next combine the results given by the classifier and the results given by the initial search engine. It is actually a problem of result fusion. Specifically, we combine the information from two data sources: the classifier and the search engine, in order to achieve better performances than using information from either of the sources alone. As the initial search engine, we employ the ranking method used in the Okapi system [17], which is called BM25 (BM hereafter). (This is because in the experiments described below click-through data were collected from search engines based on BM25.) For click-through data classification, we employ Nave Bayes (hereafter NB). For combination of the two methods, we consider a number of strategies: (1) data combination (DC), (2) ranking with NB first and then BM (NB+BM), and (3) linear combination of BM and NB scores (LC-S) and linear combination of BM and NB ranks (LC-R). BM and NB themselves are the combination methods in two extreme cases: the former does not make use of click-through data and the latter makes only use of click-through data. DC first combines click-through data with document data, and then uses BM. In practice, some search engines only return a ranked list of documents without scores. NB, NB+BM, and LC-R can still work under such circumstances. Notation D = {d} is a document collection. Q = {q} is a set of queries, and a query q consists of words. T = {t} is a set of words. R(q) = is a ranked list of retrieved documents with respect to query q, returned by a method. Some methods only return a ranked list, while other methods associate each di (1 d" i d" m) with a score S(q, d), satisfying S(q, d1) e" S(q, d2) e" & e" S(q, dm). C = {<q, c>} is a set of click-through data, where c = is a set of documents clicked by a user. Qd = {q} is the subset of Q related to document d in click-through data. BM Initial Ranking Function In BM, given query q, we first calculate the following score (using default values of parameters in the formula of BM25) of each document d with respect to it  EMBED Equation.3 , where tft,d is term frequency of word t in document d, dft is document frequency of word t in collection D, N is number of documents in D, dld is length of document d, and dlavg is average length of documents in D. We next rank documents according to their scores SBM(q,d), and then obtain a ranked list of documents RBM(q). NB Click-through Classification We can view each document in the given collection as a category, and the queries associated with each of the documents as instances classified into the corresponding category. The association can be found in click-through data (cf., Figure 2). Thus, we can construct a classifier on the basis of the click-through data and formalize the problem of document retrieval as that of text classification.  SHAPE \* MERGEFORMAT  Figure 2. Relationship between queries and documents In NB, we employ Nave Bayes as classifier [15]. Given query q (a sequence of words t1, t2, , tl) and category d, the conditional probability of p(d|q) is calculated as follows.  EMBED Equation.3 . We calculate the probabilities using click-through data.  EMBED Equation.3 . Similarly we have  EMBED Equation.3 . Here, Qd is the subset of queries in Q associated with document d, tft,Qd is term frequency of word t in Qd, tft is term frequency of t in Q, tfQd is total term frequency in Qd, tfQ is total term frequency in Q, and |VQ| is number of unique words in Q. Furthermore, p(d) and  EMBED Equation.3  are prior probabilities of category d and its converse category  EMBED Equation.3 , and we assume that they are equal for all documents. The converse category  EMBED Equation.3 consists of all documents except d. We use the log of the ratio of the two probabilities as the score of document d with respect to query q.  EMBED Equation.3  In NB, we rank documents by SNB(q,d) and then obtain RNB(q). DC Data Combination The above two methods rank documents based upon information from only one data source: either document or click-through. In the sequel, we consider a number of methods which combine the information from both data sources. The simplest method of combination might be to combine the queries to their associated documents, view them as pseudo-documents, and run BM with the pseudo-documents. More precisely, for each document d and its associated query set Qd, we create a pseudo-document d by combing d and Qd. Given a new query q, apply BM to the collection D={d} and get SBM(q,d) for each document. Let SDC(q,d) = SBM(q,d), we can rank the documents with the scores and obtain RDC(q). NB+BM NB First and BM Next In NB+BM, given the initial ranked list by BM, we pick up from the list those documents whose NB scores are larger than a predetermined threshold, rank them by NB, and put them on the top of the list. The assumption is that click-through data reflects users preferences and thus it is better to use the information first for preference ranking. Let RBM(q) be the ranked list of documents returned by BM. Some of them can be ranked by NB with the ranked list denoted as RNB(q). We put the documents in RNB(q) at the top, and the remaining documents behind according to RBM(q). We then obtain RNB+BM(q). For example, if RBM(q) = and RNB(q) = , then RNB+BM(q) = < d3, d2, d1, d4>. LC-S and LC-R Linear Combination We can employ a fusion strategy to combine the results of NB and BM. The most widely used method for such fusion is linear combination. We consider two sorts of linear combination methods. In Linear Combination of Scores (LC-S), we rank documents by linearly combining the normalized scores returned by BM and NB:  EMBED Equation.3 , where  EMBED Equation.3 and  EMBED Equation.3 are normalized to [0, 1],  EMBED Equation.3 is the weight. In Linear Combination of Ranks (LC-R) we rank documents by linearly combining the normalized ranks returned by BM and NB. Suppose that document d is ranked at ith position in RBM(q) and at jth position in RNB(q), and there are m documents in RBM(q) and n documents in RNB(q), then we have  EMBED Equation.3 , where  EMBED Equation.3 is weight. We determine the weight  in the linear combination by training on click-through data. As measure for training, we use Percentage of Clicks on Top N (PC-TopN for short). PC-TopN represents the percentage of clicked documents on top-N positions among all clicked documents. Actually, PC-TopN is the result at point N in a click distribution. Since we use click distribution for evaluation, it is natural to use the same measure for training (For training, we can use a single value instead of a distribution as the measure). Given a specific value of , we can create a linear combination model and calculate its corresponding PC-TopN value with the training data. From a number of possible values of , we select the one that maximizes PC-TopN with respect to the training data. That is to say, we choose the  that achieves the best performance in document preference ranking. Since there is only one parameter  in our linear combination model, there is no need to employ a complicated parameter optimization method. Other Methods We have also tried other methods including interleave (cf., [16]), BM first and NB next, and linear combination of NB score and BM rank. They do not work as well as the above methods. Due to the limitation of space, we only consider the above methods in this paper. EXPERIMENTS AND RESULTS We conducted three experiments on two data sets. The first experiment was to investigate whether our preference ranking methods can preserve relevance ranking. The second was to examine whether our methods outperform the baseline BM25 for preference ranking, and among them which method works best. The third is to investigate the characteristics of click-through data. Data Sets The two data sets used in our experiments are Encarta and Word. Encarta is an electronic encyclopedia on the web (http://encarta.msn.com), which contains 44,723 documents. Its search engine (based on BM) has recorded click-through data. We obtained data collected over two months. Word is from Microsoft Word Help, which contains 1,340 documents. In an experimental setting, its search engine (also based on BM) collected click-through data during four months. Details of the two data sets are shown in Table 1. Table 1. Details of Word and Encarta click-through data WordEncartaTime span4 months2 monthsTotal number of queries~34,000~4,600,000Total number of unique queries~13,000~780,000Average query length~1.9~1.8Average clicks per query~1.4~1.2 Table 2 shows randomly selected queries in the click-through data of each data set. Table 2. Example queries in Word and Encarta Wordxml, table contents, watermark, signature, speech, password, template, shortcut bar, fax, office assistant, equation editor, office shortcut bar, highlight, language, bookmark Encartaelectoral college, China, Egypt, Encarta, sex, George Washington, Christmas, Vietnam war, dogs, world war II, solar system, abortion, Europe The titles and snippets in both Word and Encarta are created by humans, and thus are very accurate. Thus, for the two data sets it is possible for users to make preference judgments based on title and snippet information. Training and Testing Data The click-through data in Encarta and Word were recorded in chronological order. We separated them into two parts according to time periods, and used the first parts for training and the second parts for testing. For Encarta, we had 4,500,000 click-through pairs for training, and 10,000 pairs for testing. For Word, we had 30,000 click-through pairs for training and 2,000 pairs for testing. For NB and NB+BM, we made use of the training data (click-through data) in the construction of classifiers. For LC-R and LC-S, we made use of the training data in the construction of classifiers and a randomly selected subset of training data (10,000 instances in Encarta and 2,000 instances in Word) in the tuning of linear combination weights. For DC, we utilized the training data and the document data in the construction of models. When training the weights in LC-R and LC-S, we tried different values of N for PC-TopN: PC-Top1, PC-Top2, PC-Top5, PC-Top10, and PC-Top20. It seems that N only slightly affects the weights and the final ranking results have little differences with different values of N used in training. Thus, in the experiments reported in this paper, N was fixed at 5. In order to conduct precise analysis, we further created three data sets on the basis of the same click-through data for both Encarta and Word. They are referred to as one_click, first_click, and all_clicks, respectively. In one_click, we used the pairs only having single clicks as instances. In first_click, we extracted the first clicks in all click-through pairs to create instances. In all_clicks, we used all click-through pairs as instances. For example, given two click-through instances and , one_click = {}, first_click = {, }, and all_clicks = {, }. Due to limitation of space, we only report the results with first_click in this paper. The results with one_click and all_clicks have the same tendencies. (a) Word (b) EncartaFigure 3. Click distributions with 20% Word training data and 10% Encarta training data    (a) Word (b) Encarta Figure 4. Click distributions with 100% Word and Encarta training data Experiment 1 We randomly selected 40 queries from the testing (click-through) data of Encarta and that of Word respectively. More precisely, for each data set, we sorted all the queries in terms of frequency, then equally divided the sorted set into four subsets, and randomly picked up 10 queries from each of the subsets. In this way, we were able to select queries from different frequency ranges. For each query, the six ranking methods described in Section 4 were employed to produce a ranked list, and the top-5 documents returned by the methods were merged together. Next, humans made judgments on the relevance of the documents. We then used the measure of top 1, 2, and 5 precisions to evaluate the performances of the six methods for relevance ranking. We report here the results when 10% of Encarta and 20% of Word training data were respectively used for training of the ranking methods, in accordance with the results reported for Experiment 2 in Section 5.4. The results in Table 3 indicate that all the preference ranking methods using click-through data improve upon or at least work as well as the initial ranking method BM. (Some of the improvements are statistically significant according to our statistical testing). Thus, it is safe to say that the preference ranking methods can preserve relevance ranking. It should be noted that NB, the method which ranks documents by using click-through data alone, can still achieve better results than BM. The result indicates that preference is strongly related to relevance. (a1) Word reverse order (b1) Encarta reverse order part1(c1) Encarta reverse order part2(a2) Word time order (b2) Encarta time order part1(c2) Encarta time order part2Figure 5. Percentages of clicks on top 5 with different sizes of Word and Encarta training data: in the top part, click-through data is added by reverse time order; in the bottom part, click-through data is added by time order. Table 3. Performance of six methods with respect to Word and Encarta data in terms of top 1, 2 and 5 precisions (Pre@1, Pre@2, Pre@5)Collection Method Pre@1 Pre@2 Pre@5 Word LC-S 0.850 0.700 0.455 NB+BM 0.700 0.625 0.450 NB 0.700 0.613 0.445 DC 0.700 0.600 0.435 LC-R 0.675 0.575 0.390 BM 0.650 0.537 0.390 Encarta LC-S 0.825 0.563 0.300 NB+BM 0.750 0.525 0.280 NB 0.750 0.525 0.280 DC 0.750 0.525 0.270 LC-R 0.725 0.500 0.260 BM 0.450 0.388 0.250 Experiment 2 We applied the methods of BM, NB, DC, NB+BM, LC-R, and LC-S to both Encarta and Word data for preference ranking. All the results with Encarta and Word show almost the same tendencies. We used different proportions of training data in creating NB, DC, NB+BM, LC-R, and LC-R. We only show the results here when 20% and 100% of Word training data, and 10% and 100% of Encarta training data are available. This is because they are representative of the general trends in the entire results. We found that the use of click-through data is effective for enhancing preference ranking. Nearly in all cases, the methods using click-through (i.e., LC-S, LC-R, NB, NB+BM, DC) perform better than the method without using it (i.e., BM). We also found that when the size of training (click-through) data is large, LC-S, NB+BM, and NB perform nearly equally best, and when the size of training (click-through) data is small, LC-S performs best. Therefore, it is better to always employ LC-S. The results were obtained on the basis of the measure: click distribution as described in Section 3.2. Again, click distribution represents the percentage of clicked documents on top N positions among clicked documents. The higher percentage of clicks on top, the better a method is. For each method, we evaluated the percentages of clicks on top 1, top 2, top 10, top 20, and all. We also evaluated the percentages of clicks in intervals between rank 1-1, rank 2-2, rank 3-5, rank 6-10, rank 11-20, and rank 21-. Figure 3 shows the click distributions when 10% of Encarta and 20% of Word training (click-through) data are respectively available. Among the methods, LC-S performs the best. Figure 4 shows the click distributions when 100% training data is available. Among the methods, LC-S, NB+BM, and BM perform best. (a) Word (b) Encarta part1(c) Encarta part2Figure 6. Percentages of unseen queries in testing data versus sizes of Word and Encarta training data We conducted sign test [10] to see the significance of differences between methods on percentage of clicks on top5 (significance level was set to 0.005). We use {A, B} to represent methods A and B do not have significant difference in performance, and use A >> B to represent method A is significantly better than method B. With 20% training data for Word and 10% training data for Encarta, we have: (1) Word: {LC-S, NB+BM} >> NB >> DC >> LC-R >> BM (2) Encarta: LC-S >> NB+BM >> NB >> {DC, LC-R} >> BM With 100% training data, we get the same results for Word and Encarta: (3) {LC-S, NB+BM, NB} >> {DC, LC-R} >> BM. Experiment 3 Click-through data can change according to user groups and time periods. We investigated how the recording time of click-through data affect the preference ranking methods. Figure 5 shows the results both when click-through data are added in reverse time order (a1), (b1), and (c1) and when click-through data are added according to time order (a2), (b2), and (c2). The measure used is percentage of top 5 clicks (one point at click distribution). For better plotting, we separate the results of Encarta: Figure 5-(b1)(b2) show the results when from 1% to 10% of training data available, and Figure 5-(c1)(c2) show the results when from 10% to 100% of training data available. The results indicate that click-through data (and thus users preferences) may change from time to time, and also indicate that to maintain good preference ranking results, it is only necessary to use the click-through data recorded most recently (i.e., reverse time order is better than time order). From the results, we also see that the performances of the methods will converge when enough click-through data is available. Discussions We discuss why the use of click-through data can help improve preference ranking and why LC-S can work best when different sizes of click through data are available. We examined how the proportion of unseen queries in testing data changes when training (click-through) data gets accumulated. Figure 6 shows the results. We see that the number of previously unseen queries in the testing data decreases when training data increases no matter whether it is in time order or reverse order. For Word, eventually only 34% of testing queries is previously unseen, and for Encarta, only 15% of testing queries is previously unseen. Therefore, using click-through data (already observed click records) can help improve prediction of users future clicks, i.e., preference ranking. Table 4 shows the distributions of queries and clicked document in click-through data. For both data sets, query frequency is heavily distributed on a small number of common query terms, and click frequency is heavily distributed on a small number of common documents. The statistics indicate that prediction of users clicks based on click-through data is at least possible for common queries and common documents. Table4. Distribution of queries and clicked document ids with different frequencies in click-through dataFrequency 1 2-3 4-10 11-50 51- Word query Type 76.5% 14.1% 6.3% 2.6% 0.5% Token 29.9% 12.5% 14.4% 21.5% 21.7% Word clicked doc id Type 3.2% 5.7% 22.6% 49.5% 19.0% Token 0.1% 0.4% 4.6% 35.8% 59.1% Encarta query Type 68.3% 19.7% 7.4% 3.1% 1.5% Token 11.4% 7.4% 7.2% 11.9% 62.1% Encarta clicked doc id Type 8.0% 11.5% 20.4% 29.1% 31.0% Token 0.1% 0.2% 0.9% 4.8% 94.0%  From the results, we see that the way of using click-through data, as we propose, can effectively enhance the performance of preference ranking. It appears reasonable that NB, NB+BM, LC-S perform equally well when click through data size is large enough. In such situation, NB performs much better than BM, and thus both in the back-off model of NB+BM and the linear combination model of LC-S, NB does dominate (i.e., NB either works frequently or has a large weight). Interestingly, when click-through data is not sufficient, LC-S performs best, suggesting that combining information from two different sources effectively helps improve performance even if training data is not enough. It is easy to understand LC-S always performs better than LC-R, as the former uses scores and the latter uses ranks, and there is loss in information when using LC-R. CONCLUSIONS We have investigated the problem of preference ranking based on click-through data, and have proposed a machine learning approach which combines information from click-through data and document data. Our experimental results indicate that click-through data is useful for preference ranking; our preference ranking methods can perform as well as BM25 for relevance ranking; it is better to employ a linear combination of Nave Bayes and BM25; Future work still remains. In our study, we used Nave Bayes as the classifier. It will be interesting to see how other classifiers can work on the problem. In our experiments, we tested the cases in which the document collection was fixed. It is an open question whether we can make a similar conclusion in the cases in which the document collection dynamically changes during recording of click-through data. REFERENCES Anick, P. Using terminological feedback for web search refinement a log-based study. In Proceedings of SIGIR03, 2003, 88-95. Aslam, J.A., and Montague, M.H. Models for Metasearch. In Proceedings of SIGIR01, 2001, 275-284. Bartell, B.T., Cottrell, G.W., and Belew, R.K. Automatic combination of multiple ranked retrieval systems. In Proceedings of SIGIR94, 1994, 173-181. Beeferman, D., and Berger, A. Agglomerative clustering of a search engine query log. In Proceedings of SIGKDD00. 2000, 407-416. Buckley, C., Salton, G., Allan J., and Singhal A. Automatic Query Expansion Using SMART: TREC 3. TREC-1994, 69-80. Cui, H., Wen, J.R., Nie, J.Y., and Ma, W.Y. Probabilistic query expansion using query logs. In Proceedings of WWW02, 2002, 325-332. Dumais, S., Joachims, T., Bharat, K, and Weigend, A. SIGIR 2003 Workshop Report: Implicit Measures of User Interests and Preferences. 2003. Fox, E.A. and Shaw, J.A. Combination of multiple searches. In Proceedings of TREC-2, 1994, 243-249. Greengrass, E. Information Retrieval: a Survey. 2000. http://www.cs.umbc.edu/cadip/readings/IR.report.120600.book.pdf Hull, D.A. Using Statistical Testing in the Evaluation of Retrieval Experiments. In Proceedings of SIGIR93, 1993, 329-338. Joachims, T. Optimizing search engines using clickthrough data. In Proceedings of SIGKDD02, 2002, 133-142. Lee, J.H. Analyses of multiple evidence combination. In Proceedings of SIGIR97, 1997, 267-276. Ling, C.X., Gao, J.F., Zhang, H.J., Qian, W.N., and Zhang, H.J. Improving Encarta search engine performance by mining user logs. International Journal of Pattern Recognition and Artificial Intelligence. 16(8), 2002, 1101-1116. Manmatha, R., Rath, T., and Feng, F. Modeling score distributions for combining the outputs of search engines. In Proceedings of SIGIR01, 2001, 267-275. Mitchell, T.M. Machine Learning. McGraw Hill, 1997. Oztekin, B., Karypis, G., and Kumar, V. Expert agreement and content based reranking in a meta search environment using mearf. In Proceedings of WWW02, 2002, 333-344. Robertson, S.E. et al. Okapi at TREC-3. In Overview of the Third Text REtrieval Conference(TREC-3), 1995, 109-126. Salton, G., and Buckley, C. Improving retrieval performance by relevance feedback. Journal of American Society for Information Sciences, 1990, 41:288--297. Silverstein, C., Henzinger, M., Marais, H., and Moricz, M. Analysis of a Very Large AltaVista Query Log  ,-9:STV , - . 9 : U ] ^ u {s{m]mh0h*Y@H*nHo(tH h*Y@h*YnHtHh*YnHo(tHh*Y hs;5CJ hs;5 hs;CJh1CJnHo(tHhs;CJnHo(tHhs;5nHo(tHhVhs;OJQJnHtHhs;hs;5CJ nHo(tHhs;hs;5CJ nHtHhs;5CJ nHo(tH hs;5CJ hs;$   TUV~7kd$$Ifl u}4 lad  $ $If^gd?WB$ $ p]p^a$gds;!O T U ^ w gd9fDgd*Ygd*Ygd*Ygd*Ygd*Y<gd*Y & F $ p VD]p^ gds;$ $ p]p^a$gds;u w ( ) / 0 E F N O f h   % , V Z     h}nHtHh*nHo(tHh|,nHtHh nHo(tHh(hG hG5CJh9fDnHo(tHh*Yh6~Xh*Y@h0@nHo(tH h0@h&h*Y@h.@nHo(tH h*Y@ h*Y@o(3 0 E F O h  dxgdGgdGgdG$a$gdg_d$a$gd*Ygd*Ygd9fDgd*Ygd*Ygd*Y ' 2 9 ? m x ,>(=Mcdüå}hg_d6CJ]hg_d6CJ]nHo(tHhg_dhg_dCJ]nHo(tHhG6]nHtHhG5\nHo(tH hG5CJhGnHo(tHh+8Y h?h?h$zhwmhKh)?<h?h3{hGh*nHo(tHh*/fg=1H&#$+D/0$gdg_d=1H&#$+D/0$gdg_d=1+Dgdg_dx=1H&#$+D/0$gdg_dxgdGxgdG %'3?QSfgt()-34>HOkost %&ĽگīѣīīīīĽīěĈh*nHo(tH he6 h(6hMJhG6h)+nHtHh(phMJhG6nHo(tH hRhGhGhnHo(tHhGnHo(tHhGnHtHh(nHtH hg_d]hg_dCJnHo(tH hg_dCJ2gtoK< I hHxEƀZͅF.gdGxgdHxEƀZͅF.gdG &56:AQ^dehmno  ./6@JTJK]w ܻܻܶܲܲ릝hMJnHtHhGnHo(tHh`nHtHhrh ho(hRhGo(h nHo(tHh>hDh)+h( hRhGhf:h(phGh*nHo(tHhXnHtH:/06AF|~&'01<JTZb (ķįħğħěˏhA EnHtHhA Ehd;h:YhWhRhGo(h/hKhKnHtHhsnHo(tHhs hRhGhGhunHtHhGnHtHhvnHtHhGnHo(tHh`nHtHhMJnHtH5(*5Ddgsw{~QVfiju -?3EH`p; < !־ֶ־֪֚֮֞֞h+HYh[Nhvhf:6hf:hRhGo(h^nphG6h+8Yhu h hG hGo(hA EhGhQthA EnHtH hRhGhmnHo(tHhGnHo(tH:I c "#$%&<(T)n**]HEƀZͅF.gdGxgdG x`gdGHEƀZͅF.gdG !!>!e!|!!!!!!!"" """"$"4"i"j""""""#6#<#=#G#s#u#z##########$$*$b$$$$$$󻲪ꪡꪲh1nHtHhh^nHtHho>inHtHhnHo(tHhGnHtHhsnHo(tHhTnHtHhhnHo(tHh 'nHo(tHhynHtH hG6h|,hGnHo(tHhGho>ihG63$$$$%,%.%/%1%9%C%E%L%O%c%e%f%n%o%t%x%y%%%%%%%%%%%%%%%%h&o&&&&&&&&&&&&&&' '' '&''','8'A'D'I'J'O'W']'^'_'y'z'{''ݿппݿh\3nHtHh3InHtHh 'nHo(tHhGnHo(tHhGhGnHtHh~nHo(tHh1nHtHh1nHo(tHH''''''''''(("('(:(<(F(J(Q(((((&)R)S)T)a)p)t){))))))))))**M*O*]*_*m*~*********+:+=+f+h+t+u++++++++++,,,hY#whG6nHo(tHhY#whG6nHtHh 'nHo(tH hGo(hGh3InHtHhGnHo(tHhGnHtHJ*+-///12e3dYN x`gd x`gd0HEƀZͅF.gdGHxEƀZͅF.gdG x`gdG,,#,$,',(,,,-,1,2,6,7,;,<,>,B,D,a,b,,,,,,,- -"-$-(-)-+-,->-?-^-n------.Y.^.r.|............$/&/'/0/I/R/p/q/{/|//////żhGhnnHo(tHhnnHtHh\3nHtHhwunHtHhnHo(tHh: nHo(tHhGnHo(tHhGnHtHH//00/121j1|1111111112>2E2k2n222d3e3f3g3|3}3~3333333333333򿻯wjwjwjwjh5aJnHo(tHh5aJnHtHjNhUnHtHjhUmHnHtHuhnHo(tHhnHtHjhUnHtHhh#oh0nHtHh0h0nHo(tHh%*nHo(tHh_L h(hdnHo(tHhGhGnHo(tH)e333335}8- x`gdDkdh$$IfTlN t644 laT$xx$G$If`a$gd`l#Fkd$$IfTl]N t644 laT$x$If`a$gdOl#3333344/4L4Q4{4444444j5|5555555S6[6%70797m77778888$8W8l8p8q8r8v8w8{8ϹϭϥϥƜƘ{wh ho(hRhGo(hoUnHo(tH hRhGhh.MnHo(tHhGnHtHhhhvnHtHhKGnHo(tHh(hGnHo(tHhGh>h Bh|,h#ohnHo(tH&h}Qh5OJQJaJnHo(tH.57889;P<==>RHxEƀZͅF.gdkA x`gdGHEƀZͅF.gdG x`gdxgd {8|88888888889999:;;/;A;B;C;F;U;;;;;;;;;;<<"<&<O<P<<<<<<<<<y=}======> >>мм״״񨟛hWanHo(tHhWahkAnHo(tHhkAh-nHtHhGnHtHhZIhGnHo(tHhhG6 hZIhGh1InHtH hG6hGnHo(tHhA<hGh hRhG8>>>?? ?(?-?B?D?G?V?`?y?????????????@@)@0@Q@X@|@}@@@@@@@@ A0A;AAARAYA`AfAgAüأϛdždžh/&bnHo(tHhY hE^hfSh8nhzhkAnHo(tHhrhRheo(he hRheh~)h|,nHtHh.nHo(tHhzohnHtHhnHo(tHhhkAhWa h.hWa1gAmAnAoAqAAAAAAAAAABBB`BaBBBBBBBBBCC?CDCGCKCoCtCwC{C|CCCCCCC]DgDoDDDDDDDDDDDD EE$E)E*EoAhBBD>EGEiEEEG>HHHxEƀZͅF.gdkAxgd x`gd =E>EGEHELEMERESEiEjEnEoEtEvEyEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE.F0FFFFFFGGGGG G GGGGG.G0G2G4G6G:GGVGXGZG\G^GbGdGfGhGjGlGnGpGrGtGxGzGGGGGGGGGGGGGGGGGGGGGHHHH H H H HHHH6H7HH@HDHEHXHYHnHoHpHrHyHHHHHHHHHHhkA6nHtHhkAh/&bhkAnHo(tHh/&bnHo(tHhkAH*nHo(tHhkA6nHo(tHhkAnHo(tHhkAnHtHHHHEI_I7JJJWLtLVK `gdgHxEƀZͅF.gdkA $x`a$gdkA x`gdkAHxEƀZͅF.gdkAHHI0I1IEIFIGIXIYIZI[I\IeIgIjIIIIIIIIIIIIIIIIIIIIIIIJJJ JJJ4J5J6JgJhJiJѥѥѥѥѥѥѥh/&b6nHo(tHh/&bh/&b6nHo(tHhkA6H*nHo(tHhkA6nHo(tHjhkAEHUnHtH$j׳C hkACJUVnHo(tHhkAnHo(tHjhkAUnHtHhkA6nHtHh/&bnHo(tHhkAnHtH1iJkJlJoJpJvJzJJJJJJJJJJJJJAKJKKKULVLWLXLYLnLoLpLqLrLsLtL{L|LLL÷ΰפטט{ncnchkA5aJnHtHhkA5aJnHo(tHjhkAUnHtH!jhkAUmHnHo(tHujhkAUnHtHhu8nHtHhkA h 5hkAhkA6H*nHo(tHhkA6nHo(tHhkAnHo(tHhkAnHtHh/&b6nHo(tHh/&bnHo(tHh/&b6H*nHo(tH&tLL\MuMMMMMOgPPPPlHxEƀZͅF.gdkA x`gd $x`a$gd x`gdg $x`a$gdkA x`gdkA $`a$gdg LLLLLLLLLLLLLLLMMMMMM M MMMMMM M;MM?M@MBMDMEMOMPMZM\M]M^MoMpMqMrMsMMMMhkAH*nHtHjhkAH*UnHtHjlhkAEHUnHtH$jaC hkACJUVnHo(tHjhkAUnHtHhkAH*nHo(tHhkA6nHo(tHhkA6nHtHhHnHo(tHhkAnHo(tHhkAnHtH3MMMMMMMMMMMMMMMMMMMNNN#N2N3N5N7N;N>N?NVNWN[N]N_N`NbNcNyNzN~NNNNNNNNNɻ점|||||||||||hkA6nHo(tHhkAnHo(tHj[ hkAEHH*UnHtH$jaC hkACJUVnHo(tHhkAnHtHjhkAH*UnHtHj hkAEHH*UnHtH$jسC hkACJUVnHo(tHhkAH*nHtHhkAH*nHo(tH0NNNNNNNNNNNNNNNNNNNOOOO OOOOOOO3O6O:OBOCODOEOIOLO^O_O`OsOtOuOvOxOOOOOOOOOj=hkAEHUnHtH$j ʮC hkACJUVnHo(tHjhkAEHUnHtH$jq`C hkACJUVnHo(tHjhkAUnHtHhkAH*nHo(tHhkA6nHo(tHhkAnHtHhkAnHo(tH6OOOOOOOOOOOOOO P PP!P-P2PKPLPMPdPePgPhPiPzP{P|P}P~PPPPPPPPPPPPPPPPxhXXnHo(tHhkA6H*nHo(tHhkA6nHo(tHjGhkAEHUnHtH$jaC hkACJUVnHo(tHhkA6nHtHjBhkAEHUnHtH$j ʮC hkACJUVnHo(tHjhkAUnHtHhkAnHo(tHhkAnHtH/PPPPPPPPPQQ6R9R=RARfRhRiRzR{RRRRRRRRRRRRRRRRRRRRRRRSSSSSSSSSSSSSĪ㪖h6H*nHo(tHh6nHo(tHhkA6nHtHhnHo(tHhS ^nHtHh{nHo(tHhgnHo(tHhnHo(tHhXXnHo(tHhkAhkAnHtHhkA6nHo(tHhkAnHo(tH6PQXRSSTdVVW]R x`gdHxEƀZͅF.gdkA x`gdkAHxEƀZͅF.gdkAxgdSSS,S2S3S5S6S7S8S9S>S?SASBSCSDSESFSIS~SSSSSSSSSSST+T4TTTTUUUUU7UzU{U}U~UUUUUUUUUUUUUUмۼۼۼ伤ۼۼۼۼЖhkA6H*]nHo(tHh(nHtHhS ^nHtHhmnHtHhkAnHtHhkA6H*nHo(tHhkA6nHo(tHhkAnHo(tHhnHo(tHhnHtHh6nHo(tH;UUUUUUUUUUUVVVVVVVVVVV!V"V#V%V&V'V-V.V0V7V8V9V;V nHo(tHjP!hkAEHUnHtH$jC hkACJUVnHo(tHj#hkAEHUnHtH$j{C hkACJUVnHo(tHjhkAEHUnHtH$jVC hkACJUVnHo(tHj3hkAEHUnHtH$jٳC hkACJUVnHo(tHjhkAUnHtHhkAnHtHhkAnHo(tH!WWWSXtYYYZ]aab x`gdkAHxEƀZͅF.gdkA x`gd $x`a$gdg x`gdg QXSXVXqXsXwXXXXXXXXYYYYY YYYYY Y!Y#Y$Y%Y6Y7YEYFYHYIYJYPYQY_Y`YbYcYdYlYtYuYvYYYYYYYYYصصصصj;#hkAEHUnHtH$jٳC hkACJUVnHo(tHjhkAUnHtHhkA6H*nHo(tHhkAH*nHo(tHhkA6]nHo(tHhkA6nHo(tHhkA6]nHtHhkAnHtHhkAnHo(tH5YYYYYYYY(Z.ZTZrZZZZZ[[ ["[([*[4[@[B[H[J[P[T[h[r[[[[[[[[ \2\3\9\K\\\\] ]^ ^^ ^"^$^L^X^^^^믤븜딸hmnHtHh@u>nHtHhkA6nHo(tHhkA6nHtHh@u>nHo(tHhkAnHo(tHj&hkAEHUnHtH$j,C hkACJUVnHo(tHhkAnHtHjhkAUnHtH:^^^^^^0_2_V_`_r_t_x_z_________```a*a,a:a@aPalaaaaaaaab b!b"b&b'b;bjXjyt,HxEƀZͅF.gdkAgdkAgd5~kd-$$IfTl0K+ t0644 laTXj-kkmor!r#r$r&r(rpbkdH$$Ifl40E& t 6@a'44 lae4f4$$&@#$/Ifa$b$gdz x`gd0 x`gdkA 0k:k@kLkMkOkPkRkWkXk`kfkxkkkkkkkkkkkkkkkkkkllmmmmmmmmmmmn/n=n>nfninwnnnnnnnnnnnoooooooooooooh0nHtHh0nHo(tHhfv nHo(tHhfv nHtHhkA6nHtHhkA6nHo(tHhkAnHtHhkAnHo(tHHopppp pppp$p)p2p7p? h?2_6nHtHh?2_nHo(tHh?2_nHtHhkAnHtHhkAnHo(tHh&nHtHh -nHtHh :nHtHhB@nHtH- y y y y yyy*y+yOyPytyuyvywyxyyyzy{y}yyyyyyyyyyyyyzzBzEzwz}zzzvevevevevevev!h1CJOJQJaJnHo(tHh1CJOJQJaJnHtHj\h1UnHtHj/h1UnHtHjhh1UnHtHh15nHo(tHh1OJQJnHtH!h15CJOJQJaJnHtHjeh1UnHtHj\h1UnHtHh1nHo(tHjh1UnHtH'yy+yPyuy $Ifgdjawkkd $$IfTlF &  !  ta'    44 laTuyvyxyzy|y $$Ifa$gdjawkkd$$IfTlF &  !  ta'    44 laT|y}yyyy $Ifgdjawkkdl $$IfTlF &  !  ta'    44 laTyyzzz@;gd(}Ekd $$IfTl=&a' ta'44 laTx$IfXD2gdjawkkd $$IfTlF &  !  ta'    44 laTzzzzzzzzzzz{ {{%{.{/{0{F{G{Q{R{k{l{p{q{ǹǹt`L`t`'h(rhkACJOJQJaJnHo(tH'h1hkACJOJQJaJnHo(tH!hkACJOJQJaJnHo(tHhkAnHo(tHhph~#5nHo(tHh~#5nHtHh!XT5nHtHhkA5nHo(tHhphkA5nHo(tHhphkA5h1nHtHh15nHo(tHh1OJQJnHtH!h1CJOJQJaJnHo(tHzF{G{R{Y{_{e{k{$$If`a$gd%l#0K$Dkd$$IfTlG t644 laTxx$If`gdPl#k{l{q{v{|{F+++$$If`a$gd%l#K$kd$IfK$L$TlrD  t0644 laTq{v{{{{{{{{{{{|||||&|(|@|B|W|Y|n|p|||||||||}} }f}g}j}n}y}}}Ŵū~vmhenHo(tHhenHtHh$0FnHtHhanHtHhkA6nHtHh nHtHhkAnHtHhkAhkAnHo(tH!hkACJOJQJaJnHo(tH'h1hkACJOJQJaJnHo(tH'hG9fhkA5CJOJQJaJnHtH$h1hkACJOJQJaJnHtH*|{{{{{)kd$IfK$L$Tl4rD  t0644 laT$$If`a$gd%l#K${{{{{$$If`a$gd%l#K${{{{{D)))$$If`a$gd%l#K$kd$IfK$L$Tl4rD  t0644 laT{{{{{)kd&$IfK$L$Tl4rD  t0644 laT$$If`a$gd%l#K${{{{{$$If`a$gd%l#K${{{{{D)))$$If`a$gd%l#K$kd$IfK$L$Tl4rD  t0644 laT{{{{{)kd:$IfK$L$Tl4rD  t0644 laT$$If`a$gd%l#K${{{{|$$If`a$gd%l#K$|||||D)))$$If`a$gd%l#K$kd$IfK$L$Tl4rD  t0644 laT|||||+kdN$IfK$L$TlrD  t0644 laT$$If`a$gd%l#K$||| |&|$$If`a$gd%l#K$&|'|(|.|4|D)))$$If`a$gd%l#K$kd$IfK$L$Tl4rD  t0644 laT4|:|@|A|B|)kdZ$IfK$L$Tl4rD  t0644 laT$$If`a$gd%l#K$B|E|K|Q|W|$$If`a$gd%l#K$W|X|Y|\|b|D)))$$If`a$gd%l#K$kd$IfK$L$Tl4rD  t0644 laTb|h|n|o|p|)kdn$IfK$L$Tl4rD  t0644 laT$$If`a$gd%l#K$p|u|{|||$$If`a$gd%l#K$|||||D)))$$If`a$gd%l#K$kd$IfK$L$Tl4rD  t0644 laT||||)kd$IfK$L$Tl4rD  t0644 laT$$If`a$gd%l#K$||||g}~YNGGGGxgd5 x`gd5HEƀZͅF.gd`Dkd $$IfTlG t644 laTx$If`gd%l#}}}}}}}}~#~1~@~Y~~~~~89:[ՁׁÂʂς35<=ɿ𷮷zqzfh\;h`nHtHh`nHo(tHh`nHtHh5nHo(tHh fnHtHh fnHo(tHhy7nHo(tHhkA6nHtHhkAnHo(tHhkAnHtHh*aJnHtHh 5aJnHtHh$0FaJnHo(tHh$0FaJnHtHh*nHtHh$0FnHtHhanHtH%5ʃ݃oddd $Ifgdrnnkd8$$IfTl4F &  "  ta'    44 laf4T $$Ifa$gdrn$a$gdrnxgd`xgd5 Ƀʃ܃݃()/09WXYZqv̈́ypypyhhkAnHtHh`nHo(tHh`nHtHhP5nHo(tH!hPCJOJQJaJnHo(tHhPCJOJQJaJnHtHhPOJQJnHtH!hP5CJOJQJaJnHtHjJ-hPUnHtHj!hPUnHtHjhPUnHtHhPnHtHhPnHo(tH&XYZ<55xgd`Hkd:$$IfTl46&a' ta'44 laf4T x$IfgdrnnkdW9$$IfTl4F &  "  ta'    44 laf4TST^_eftu$Z͆نچ !#$16;v߇\ш҈ڈۈ02~ȉ"*2@u}hnHtHh(nHtHhtnHtHhkAhkAnHo(tHhkAmHnHsHtHhkAnHtHht,7nHtHK$Z͆چ*6܋=ݏ x`gd5xgd'HxEƀZͅF.gdkA<gdPxgd5)*+5679@APct͋ڋ܋cdwxȌɌˌ׌݌=D`|}hnHtHh3nHo(tHh5nHo(tHht,7nHtHh/nHtHh\BnHtHh0InHtHh1nHo(tHh6)nHtHhBnHo(tHhkAnHo(tHh?2_nHo(tHhkAnHtH9}GH܏ݏ.1BGHIShi:;BCWXfgÑđǑDOWYǷhgnHo(tHh\BnHtHhBCJOJQJaJhBCJOJQJaJnHtHhBnHo(tHhjhBnHo(tH hBo(hBh5nHo(tHhkAnHtHhkAnHo(tH9ݏHISUY^dh$$Ifa$gdOK$Bkdq:$$IfTlG t644 laT$x$IfgdOl#hity1!!$$Ifa$gdOK$kd:$IfK$L$Tlֈ 2 E t0644 laTy$$Ifa$gdOK$kd;$IfK$L$Tl4֞ 2 E t0644 laf4T$$Ifa$gdOK$kd<$IfK$L$Tl4֞ 2 E t0644 laf4TϐԐِސ$$Ifa$gdOK$$$IfXD a$gdOkd=$IfK$L$Tl4֞ 2 E t0644 laf4T $$Ifa$gdOK$kd>$IfK$L$Tl4֞ 2 E t0644 laf4T"'-38=B$$Ifa$gdOK$BCkd?$IfK$L$Tl4֞ 2 E t0644 laf4TCDJPUZ`f$$Ifa$gdOK$fgkdw@$IfK$L$Tl4֞ 2 E t0644 laf4Tgp$$Ifa$gdOK$$$IfXD a$gdOkdiA$IfK$L$Tl4֞ 2 E t0644 laf4TÑ$$Ifa$gdOK$Ñđkd[B$IfK$L$Tl4֞ 2 E t0644 laf4TđőƑǑYy ,xgd5xgdkA x`gd5`gdBBkdMC$$IfTlG t644 laTx$If`gdOl# Sw ~ȕܕO'+ϘИ &-DG[\^fmݙ "#Xÿ h#o( h#hGh#hGo( hGo(hG h<^hkAhOhMnHtHhMhkAh5nHo(tHhkAnHo(tHhkAnHtHht,7nHtHAPgK & FxEƀZͅF.gd5K & FxEƀZͅF.gd5 [FEƀZͅF[]gd#xgdG x`gd5K & FxEƀZͅF.gd5 os-FEƀZͅF[]gd#FEƀZͅF[]gd#FEƀZͅF[]gd#X]rsuvYkln~ I^oy~͜ΜԜ՜ۜܜ69MNPX_w~ɟ˟֟`rsh#h 'o( h 'o( hSehG h4hG h5whGhG h#hGh#hGo( hGo(N s-FEƀZͅF[]gd#FEƀZͅF[]gd#FEƀZͅF[]gd# oas-FEƀZͅF []gd#FEƀZͅF []gd#FEƀZͅF[]gd#a͝-s-FEƀZͅF []gd#FEƀZͅF []gd#FEƀZͅF []gd#ޟs-FEƀZͅF[]gd#FEƀZͅF[]gd#FEƀZͅF[]gd#suJKС:    !"$%'(*+-.45689?@ACĿhg_d hg_d0Jjhg_d0JUh*Y h*Y0Jjh*Y0JUhLjhLU hC#o(hC#hC#nHo(tHhC#nHtH h7o(U h4hGhG h#hG hGo(h#hGo(3Hs-FEƀZͅF[]gd#FEƀZͅF[]gd#FEƀZͅF[]gd#. Technical Report SRC 1998-014, Digital Systems Research Center, 1998. Spink, A., Jansen B.J., Wolfram, D., and Saracevic, T. From e-sex to e-commerce: web search changes. IEEE Computer, 35(3), 2002, 107-109. Spink, A., Wolfram, D., Jansen B.J., and Saracevic, T. Searching the web: the public and their queries. Journal of the American Society of Information Science and Technology, 52(3), 2001, 226-234. Vogt, C.C., and Cottrell, G.W. Predicting the performance of linearly combined IR systems. In Proceedings of SIGIR98, 1998, 190-196.     PAGE  PAGE  PAGE  Ranked Docs Based on Relevance New Query User Search Engine Ranked Docs Based on Relevance Query Feedback Algorithm Qdn={ q2, } } Qd2={ q1, } } Qd1={q2, qm, } , } dn d2 d1 11111111 qm Relevance Feedback q2 q1 User Click-through Data Re-ranked Docs Based on Preference Re-Rank Engine Query and Ranked Docs Based on Relevance Query Preference Ranking Using Click-through Data Search Engine User Query Search Engine Conventional Search Ranked Docs Based on Relevance User ... Meta Search Engine Search Engine 2 Search Engine n Meta Search Ranked Docs Based on Relevance Query Search Engine 1 HҸs-% & FgdC#FEƀZͅF[]gd#FEƀZͅF[]gd#FEƀZͅF[]gd# !#$&')*,-678ABCLMNO[dno 1$7$8$H$gd&`#$ & FgdC#CDJKLNO[\cdnosyzʺ̺κϺкѺҺտէտէo^!hkACJH*OJQJaJnHtH$hkA6CJOJQJ]aJnHtH'hkA6CJOJQJ]aJnHo(tH!hkA5CJOJQJaJnHtH.hB*CJOJQJ^JaJnHo(phtH+hB*CJOJQJ^JaJnHphtH#hB*CJOJQJ^JaJphhLht ht0Jjht0JU"oyz{ʺ˺̺κϺ $a$gdkA 1$7$8$H$gd $1$7$8$H$a$gd$1$7$8$H$a$gd 1$7$8$H$gdҺԺպֺ׺ٺںݺ޺ߺdztp\Itdztp\It$hkA6CJOJQJ]aJnHtH'hkA6CJOJQJ]aJnHo(tHht!hkACJH*OJQJaJnHtH+hkAB*CJOJQJ^JaJnHphtH.hkAB*CJOJQJ^JaJnHo(phtH&hkAB*CJH*OJQJ^JaJph)hkA6B*CJOJQJ]^JaJph#hkAB*CJOJQJ^JaJph!hkACJOJQJaJnHo(tHϺݺߺ      $a$gdkA $a$gdkA 1$7$8$H$gdkAԿԑyԑhdSd@/d!hkACJH*OJQJaJnHtH$hkA6CJOJQJ]aJnHtH!hkA5CJOJQJaJnHtHht!hkACJOJQJaJnHo(tH.hkAB*CJOJQJ^JaJnHo(phtH+hkAB*CJOJQJ^JaJnHphtH.hkAB*CJH*OJQJ^JaJnHphtH)hkA6B*CJOJQJ]^JaJph.hkAB*CJOJQJ^JaJnHo(phtH&hkAB*CJH*OJQJ^JaJph&'()+,-./Ų{jXF1)hkA6B*CJOJQJ]^JaJph#hkAB*CJOJQJ^JaJph#hkAB*CJOJQJ^JaJph!hkA5CJOJQJaJnHtHhkACJOJQJaJnHtH'h$hkACJH*OJQJaJnHtH$hkA6CJOJQJ]aJnHtH$hXXCJH*OJQJaJnHo(tH$hXX6CJOJQJ]aJnHtHht!hkACJH*OJQJaJnHtH$hkA6CJOJQJ]aJnHtH()+-.123456789:;<=>QRU 1$7$8$H$gd 1$7$8$H$gdkA $a$gdkA $a$gdkA $a$gdXX/01>QRSUVWYZ_`stuŻ̛̛̰o]Y]]D)h?hB*CJOJQJ^JaJphh#hB*CJOJQJ^JaJph.hB*CJOJQJ^JaJnHo(phtH&hkAB*CJH*OJQJ^JaJph)hkA6B*CJOJQJ]^JaJph7h0h5B*CJOJQJ^JaJnHo(phtHht.hkAB*CJH*OJQJ^JaJnHphtH.hkAB*CJH*OJQJ^JaJnHphtHUVYZ_`stuӻԻڻ 1$7$8$H$gd$1$7$8$H$a$gdgd 1$7$8$H$gd$x1$7$8$H$XD2a$gd 1$7$8$H$gdkAŻƻӻԻڻۻ#$234?̢̺̺o̺̺V?,h_ h5B*CJOJQJ^JaJph1h5B*CJOJQJ^JaJnHo(phtH.hB*CJOJQJ^JaJnHo(phtH4h?h5B*CJOJQJ^JaJnHphtH.h5B*CJOJQJ^JaJnHphtH#hB*CJOJQJ^JaJphht)h?hB*CJOJQJ^JaJph4h?hB*CJOJQJ^JaJnHo(phtHڻۻ#$+23@GH1$7$8$H$YD gd1$7$8$H$XD gd 1$7$8$H$gdx1$7$8$H$XD2gd$1$7$8$H$XD a$gdF1$7$8$H$WD^F`gd?@FGHSTYZhiopuv{|ɼʼѼҼܼݼֻֿ{{{{{_{7h0h5B*CJOJQJ^JaJnHo(phtH+hB*CJOJQJ^JaJnHphtH.hB*CJOJQJ^JaJnHo(phtH#hB*CJOJQJ^JaJphht,h0h5B*CJOJQJ^JaJph&h5B*CJOJQJ^JaJph)h5B*CJOJQJ^JaJo(ph%HZhijoptuv| 1$7$8$H$gd $1$7$8$H$a$gd $1$7$8$H$a$gd$1$7$8$H$a$gd 1$7$8$H$gdɼҼܼݼ & FgdC# $1$7$8$H$a$gd 1$7$8$H$gd 1$7$8$H$gdݼ hC#o(h/ hG+hB*CJOJQJ^JaJnHphtHht#hB*CJOJQJ^JaJph 5 0&P:pz/ =!8"8#$% 3 0&P/ =!8"8#$% P 5 0&P:p9fD/ =!8"8#$% 3 0&P/ =!8"8#$% P 5 0&P:p1/ =!8"8#$% 3 0&P/ =!8"8#$% P 5 0&P:pP/ =!8"8#$% 3 0&P/ =!8"8#$% P 5 0&P:pC#/ =!8"8#$% 3 0&P/ =!8"8#$% P L$$Ifd !vh5}#v}:V l5}4ad DdD  3 @@"?$$If!vh5#v:V#l] t06,5T$$If!vh5#v:V#l t06,5TDd YQ\  c $A? ?#" `2>e;[H2`!e;[H:` (V' xڥKHTQǿsΝ;w9K'ZLabBZ-Θ4K]dPM{Z6M-D \ZU:׽cÜ_V^֊i\XMV+HOiNi>P^bp= VT"=1_cg+KU䗌6#{P fBNe߭S鋍:f<<'kcDzUɷN1PSyN?Q';RnT)Lxݎ<^øI=4䒖le`$HT_.;x Wyl{\~笱\C/ N[2C᜗^W'AoӷݳFȲ.|U0Qo}/>b#*Uf 0%톆i+d ;Vv8np*tWs8+̸!gʯɿD`Av10',3a;b~NjD6QDd 1D  3 @@"?Dd \\  c $A? ?#" `2d~{Vpyi,`!d~{Vpyi,^ `1P{xڥkAnc4i6jCRڂ詭6@XâA<"=E, ` C!7K.;iE&yyC  ^=lĩQBdDhMr.tZ,G0mCqEu6:[#5I87SrJc?X"sԢg#,Z1MGX )d&lejlew^klX5cғ{t~v_ooKGmλSRѭ_Hvø ޱ?:KRͺX)܂_'Ǫ%nO ʫ^5b1bY~_Z~E۶kmy ߫=~k ɳL {P=?TKpޣz8TD軜7R{9>nߛT8F"oȓ5/DmrzळX-q΀Ř\jef]_fDd @ E=\  c $A? ?#" `25UUߴ= 9 `!5UUߴ= M(8XxڥMhA߼ݦvkEM`AZ@oU*jam ؓ''z-4փ<"5O'5]߼̼e$,t;Qa":rjl@\䯑4vdž`BCsU~yUQNNOcq> 0E\HSzTޡ,s3泵FB0<PNz8DSoX!͔s?]l&-YY>ۇ!⳪s_bD3_$?`JI%AkS̮wr o57fϒ(Zep(M]sџ˺?_:ꊧ+ۉ'4g?I૬bu/K %[kyƋx)MW>^$yK?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~Root EntryF Fq, 2Data CWordDocumentEObjectPoolH 2q, 2_1135859607F 2 2Ole CompObjfObjInfo  #$%&'*/478;>?BEHKNSVWZ^_`abcdfghik FMicrosoft Equation 3.0 DS Equation Equation.39q y S BM (q,d)=tf t,d "log(N"df t +0.5df t +0.5)0.5     3 !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~Equation Native _1136025983 F 2 2Ole  CompObj f+1.5dl d dl avg +tf t,d {} t"q " FMicrosoft Equation 3.0 DS Equation Equation.39qR8L p(d|q)=ObjInfo Equation Native n_1135859941F 2 2Ole p(d)"p(t 1 ,t 2 ,...,t l |d)p(t 1 ,t 2 ,...,t l )  "p(d)"p(t|d) t"d " FMicrosoft Equation 3.0 DS Equation Equation.39qCompObjfObjInfoEquation Native _1136025995;"F 2 2 Q p(d)"p(t|d) t"q " =p(d)"tf t,Qd +1tf Qd +|V Q |{} t"q " FMicrosoft Equation 3.0 DS EqOle CompObjfObjInfo!Equation Native "|uation Equation.39q`a p(2d|q)=p(2d)"p(t|2d) t"q " =p(2d)"tf t "tf t,Qd +1tf Q "tf Qd +|V Q |{} t"q "_1135173745F 2 2Ole (CompObj)fObjInfo+ FMicrosoft Equation 3.0 DS Equation Equation.39q  p(2d) FMicrosoft Equation 3.0 DS Equation Equation.39qEquation Native ,9_11355284611F 2 2Ole -CompObj .fObjInfo!0Equation Native 10_1136026005$F 2 2Ole 2l d FMicrosoft Equation 3.0 DS Equation Equation.39q0vb S NB (q,d)=logp(d|q)CompObj#%3fObjInfo&5Equation Native 6_1135860136)F 2 2p(2d|q){} FMicrosoft Equation 3.0 DS Equation Equation.39q ß4  S LC"S (q,d)=S BM (q,d)+"S NB (q,d)Ole 9CompObj(*:fObjInfo+<Equation Native =_1135603798.F 2 2Ole @CompObj-/AfObjInfo0C FMicrosoft Equation 3.0 DS Equation Equation.39q3K S BM (q,d) FMicrosoft Equation 3.0 DS EqEquation Native DO_1135603835,@3F 2 2Ole FCompObj24Gfuation Equation.39q3Й4Z S NB (q,d) FMicrosoft Equation 3.0 DS Equation Equation.39qObjInfo5IEquation Native JO_11356042408F 2 2Ole LCompObj79MfObjInfo:OEquation Native P)_1135860145' =F 2 2 @y  FMicrosoft Equation 3.0 DS Equation Equation.39q }0Q4 S LC"R (q,d)=m"im+"n"jnOle QCompObj<>RfObjInfo?TEquation Native U_11356042686BF 2 2Ole XCompObjACYfObjInfoD[ FMicrosoft Equation 3.0 DS Equation Equation.39q @y Oh+'0  , L X d p|Equation Native \)1TableSummaryInformation(G]DocumentSummaryInformation8e4yr.OS'sc_u]Q9{=xHs$q+]ALvIa;kvD<_)`¸xͳ0C20ikX2q'"4Bf 1ﴛ{_'ͯD~p#`{stS\+da=N4{:[fgcGl3q"[qX#(pY1) 9fy2 =_*C6]%Dd 0hlq\  c $A? ?#" `2u8Rpn?RbjQ\`!I8Rpn?Rbj@k|xcdd``~$d@9`,&FF(`TɁYRcgbR 63gaP5< %!  @_L ĺEB>2X@F&;35V ZZǰ@pT9LF]F\ 3jVf`jTHL`=&b *!-7 v 0y{@Ĥ\Y\ bPd+_> 1NDd T\   c $A? ?#" `2U$KD:Y1`!)$KD:Yf` !XJxcdd``$d@9`,&FF(`TI)QRcgbR Р@ UXRY`,@1[)Ar؅,L ! ~ Ay ۏ/? N쀘ˈHgJ%E\`W3M!I)$5$ŠV 5΃!v0m>NDd '\   c $A? ?#" ` 2<̾; M]; {B`!̾; M]; {B`\xڝKQgfwޝPBM 5`Dc}pXTB@C#r² be!R; !*Iܼy6g;s{!`\ $VuDZ[@!¯}ZӾk-Y'M%|n0dqRפ't6%'Kż%,e~#C˸]Rkz.;~\o6/%c: ->pՂr΂!ż}G'~ohJDd h{\   c $A? ?#" ` 2~]R b˺@̜w`!~]R b˺@̜B@@h $|xcdd``Vdb``baV d,FYzP1n:&&q! KA?H1Z30< UXRY7S?&,e`abM-VK-WMc~x V8pNHq50拁jV?E@|80oEUoŃP6TYP3Cԣ$0HjB8?K/FUÁ>U$ zz/z/JB~N0/89W yg, ܤr@('[bȈ0~ 盰fW~?]g,j~\Αlk `P121)Wĸ `g!, S,Dd TYR\   c $A ? ?#" ` 2|u&PZt{X;`!Pu&PZt{*@ 8 XJxcdd``ed``baV d,FYzP1n:&&6! KA?H1Z10 UXRY7S?&,e`abM-VK-WMc~x V8pNHq50VjV,  רD$${qd |?je#( |ypATRĤ\Y\ bPd+_> 1*p/Np)RnaP]W @/YݎYm2c0ҴO\wĮCeӨh3͖B -͹F:LU$elˬ'M\|V()_TeۮHuuks\7ަKAox :']MUT} ӝb8;HyZa2ӿ)I5~16cߺ!"4Z`;<v(k}vE;^,^Krѿϓ/t[!e"|?E)H9H*_bv/G*חP,@]ERe@1/1.H*_ѿ`{V5&E@<xʖ>H6ERe}nRIk~/BJ"|va懻5Z?N~^+I83wQwX@F(OS=0섏??iOlV7ys q'piHr9IÑh^UńnΖ;H yuU!qq )IGٺ08JupG%WԾT|_#^o>HHGs߹ps?Uh4˜h#^od$](G|n5ps"=tf&7zky{"(G‡G)#1aLS(q[蠕IuA!'\Ig3R^{FJq&#wQykgFs_9_7_7Fu#q eb?p[Lu39_|+K'XJ힉5!<φ{c$U xV]ty3}ٰJ|좝%ʲP'Fo _QWzaaWu=`_-/cұ|1ů{AmMt^OP}txfmi7O=}s.r,l{3rB(-:{ru^Yr)?;겏c_2z}5+o;9۫$QXW~1ƃB RiaXNg5~~.]uvžCVk‡`EϚlQC!1J|n1˃0͑:)}bd({19+wxp;k=sE3OU2,矟;%l{CyJ lZVf?Dbbgn˻n\G|*:1{/( l+ޛoOϊ;:_u=y%:\Fi3}sj6{[z~WЧxoz%yyz?\u4S]r[3d̻}{})Vk>W2}4'>%4dUB_J"bm&3㭋L3~줝LX#Ĵf`Hżd=u7OPY6uB=ѮP y'k6_{|},kƎVq`>'WhUg.jϩHDŽ+U Dd  )%6  3 A?", |bC^||6u6 ;@= |bC^||6u6STd8zb x\}lTw~>u`N u`>Z(>@JhadP EHD++U)UԔJPsBMۦqU*$@iwfn<^=?w;7o{vR鈣۪)>Tt٪J9sJ}Zq ?n*" ԓ@Q PB:>F&`qWduSeLaұxt%kK2en\eғToz2{zzxNtR=^y.cddnY_c_'Q~X˾2/eB"]zzNi} =j>@4<u16ZMgUA5:PjڪPenxv1 OIByw'w*v+6Wb<>ش< 皊ς5̻BXpخ Uͫ-m;wѕO_iA*绷_ToDVQ)n7wMt~mBjٕJˎm*QsoIahqf 21Zawr#3tmX'%ۛ6KRuot*+l3 f\[d>®Aöbu/Ka|E7GݖQl-XX6o'-y7ܕulzr45F&o#DҚx#h'!r.iHk fB$Ґ`=8=#iHk* ϵB~x$ iMǵa-433lLƕeۛOϩS=a0z M&V= v/z[9&a-%QuR=D=g85~'.̯ek`btEI7&t+pwx:r<ȱ_Sm8K v g5l迀Ƒ8BQWI=gy&)ǼX{mjDMS]#m9HxAÐ(9ˌȰ@+ϕ‘r$L…#nӐ&?|~q}8ku{]8Ra5*KhHri_+RG6QW|8c#FuU"'"0a:Qp868*vGw!_s!cY+pt< GG#iǑ+035\NWM9ӝgvuVi|39;k>W-׹=#e@Uf9UQ*;Vϼ*3o[-o|Uj}V{V{VuU>:6[`7X\>p[Hm9_zMG/E)'A>JYz\ה]KgTbd0?Wy_ӏY[g?Ce (e,GloV1XpN7^o9[5'Gݦ̧>}<^m@'m0e{6: Oa5׀vV_M$~űPV+aE hǛ`B eoH䅭"9)[0cc)@}0e1:㱢4e1&cn6}X^bGGM4X^wK|+:Ykșn ʈ ^v78/?܂OС8ȧB|Tv^O|rtwqA$ȣux t)J<Xˎl[tBM, im)7e=F3}>x.>uQ UTFN;0 3"yyja<˻lmR[[d컩I@A#t?VӔ}dϛ;8&1$%c B+R#연|ݽXz̃`&{?vN/Ujt=.S=;mOcIb "oߝED3 ;r[Ŀ1s`Ǫ,,o!c^}cz꽕ߕWjJ9 45ɥc)e Rulw1~[SͲn[ˬeVzdO'J<4XI]]RnA Wvb]:q,=1W`p͌}g3bp~5&W'kh,|dO 4#hQ6-‡ ah!h(NyGw4e,}VY̝K|`VNk> xF蟂3MF[B* 4e A ?r1|b?w }q7A\+j3|ѿ1/QEk 3s(>J I4z@  ;MF)˝E #4e~> |yf?E^'W ewp,K ǫw7,wBL7ASAyA]1? 121ytS^rb2\_pʫ4rQˣѷG׌g܄κ.x!U 0Ud Rsi~E<}xh۸YpMY'UX(^ Ud N+M.ˣG-̣GvQIVd pi=ryyzSI<:{7gp!MrHgo4s`?=JJ)p=$shv{Hm۬bTdm}6U>Y˅oY.rjHo??oZ7u eJ`ƎsX߇*J*K0wex淓ϰzsg;F4ܬ10k.~eleˮI ~%?UlcU%obL%jg`g O4oKi[ǘsPb}i}M6rpmrH~-&y' (%$ _(ylʹ?"ylr; d ȫI^K IRv݇|>YoWJ){I|߻X Q-DϤ5iS7BQjFh㽐vh\kKƅ&*@OVb|9U,^Sec4ǫ7~G;*zi\D}QT9j=8 -7)GYqX>?:=^y2<|i |y6'?DԞDTgs7x+&|G~ O{ġmPF16?H/[>GZ&ݤ;nZ^#XχvA {vd}5zN缀$D g"_\IepvU#_ژ{\p6BߑKk;Gig٥˱4-xE5#9+&q_WM Dd /*6  3 A?"H \!׷IM V$ U@= \!׷IM V`@V8a< x[}lTq M  MLȩYIL1֑8"񒢖H(X%K*jJJmJDƕh\*Fލws14~3s>G)<Q5~%ޝT_oVQ ~fu<PD>mTjP-{ǀXS@(܂lOd#)XoרƸ8g a\оq?.ɗwH}*y \64VB~M>ujڥFd۸?6OM[A7/tqUO<㼌muŕiSk&OǮ&$vպ+ۏ_I^g[#C~{kꕷ_$G%Ūov~>Mĕ䫿*i`c%99E~ݎ2ڱ'8ɱ2bpq,=1W``͌} S$I[ٷŸ=P&`g l%_P= Z8zy\Ǡ/8LLEu+ GorH8rYq*GُKG~uSÑ"\"4\^EG8M NSSQ5 XHmptRD](G}gG6eqeqTG:h $JÑhd$MqˋwGG8{vYiT(MrHmpv"M4s 烞h%sDHAi{? )4P6طB\&4l;B? )w4i^gk "K:4G~;ǼSc;4e,cVYɝKrhHI)(>+OCY"MF_feH'c>#@JHSQ6 &7qX78wr>ƸHS;E\+ j3r1XNH$kw6g!ўQSab.Éh$B ?(?\."M%!w G\X|q4xAPYQj&*$ÑU>U>U> oY+rj?d߶߶[׭u eo;c}n+ɗa:<#ψv>݄d\>G|zZi \bez>c5y"vom#o _ ??bn >_hey_M2%5_N^6iڤG~bx0GTʂe|R{#*) \SLp._&Hm"V>s@YN޺꜑hYϲ}!e.yrR)\OM5a?3V""rv u |-p0mR{B!8r/ẛ !zR % TQRRER:Ѧ)i-!?@Q;37{/;gog?n2tzY<߽Wf)XfR:^!;#WyR@uw; |+@xbBM@vMtm!tp2!].q^BV\5a5Ӂfq8tP a:yC-Krx&F8꧙0 \~)j<~zTZ^ת6yp8su^XiT;O{Uڦ> v{l#`g7 LO%{9qW-?SI c0clZED]kwqg:|ЮUݯ+]\(K%W{F}bם ;*rcxm,T+es mU%E{"ƓV|{->)1Eaw^#3tmX-fr|щrm:CICX?OUGc74K?g971?r<\ŸAݏZ:SVzJoϷgZ3wm|+|+ vzi>ZRϽ,\沘9mXLgher~߳"c{q a;tik$|g8A}|yH_iNZ~܌빹3|]vQvYod^/)׶R-Rn'k4eY'G~8j8::,p4X=25HVd Ba"M~[B1)_vJ(9-+GG'!p.(>q1]zg$c3r15 XH[CD]ȗGcSG8݉ib/e+6Xc,$ wUd nsvi>N|Yq&QbJ#\`8J7&b\ÉÑ,#V4ӐoY#{/5 ĵ ' Dü.>*CQ"Mb]|BDoBe"M~zHy]*l V"MYV!<&#oEEܼ¾W ?& H7G;ΗQyH7yጐ? 1g4v"W4s+[+&[#oѾB~ x$MFƳ g5 Å|7QH7h_'/AL&[#o0 d3WF@LAdk|LȣiuJع#=DlD?Mc cϯy9>;M[ymjƟL̋fb9ð绮sq'p61HSAäȻJQGk㵹X rHbNi}b]UeGo3z4x.p^HMX*r @y"M+eQyG)(tXWeX&cd 4ͻ//^̌k8Cਿ!o0\y,2gHȗGN04pg}GjFaUq,QI: 0grKvӓ}$LY?)su{Fʜj*r&'Q+=[ >֦ϼ3om+9p[?m?maahhϲΠ!/;N+>v'E= \N6Q~8ﳬ5|g E!g|g!/9>{ |MZM|a_sWmj "+=)I _w~”^Y qUkIЮ72)<J{ Zx(?0ꆴ|q1+*+FNVbȘ}2qPU 3q5ۇM4.ٛX+p6yj@~K ya@'o>@URI!OycdAHT҃6]zs*iM}?ql՟k8nv@KMz6$oQ o^ xa_%>/s#gy㘼X]J*K{k}7 ~:ɷs=oW=nn9{.='k ZV=:,Td|BM `A:Xcs 8 S[62B"K0D*h䃟&mnY&mln;žmWsYWx"Gp-CJu/z液>afAq':C#8-{|u7EHO 7ofoQqaV Dd  (,6  3 A?" FhU8.z p@= FhU8.zSTd8zb x\}lT| q8M[\H g@Z||0Y%S P  R"%T"RuEMӦW H??@Q;37{/;Cٝݝ];J) TtQ-5RLS*| Re}< ȏJH4P=𣐮r+RWU+nJ윸bu2鹠äcqty%gK2e^\e3PztNOih6P)М!ArY[ThvqݟZòp_\.CR[;Tw2L8)UclW9ghC0We+- CjJ 9q C~cVNI(ګWu?W1Lc/~>lZtJsUŦgYN,ylW*ƕgމ]s}(UVz}fwrVNq1k! u}?M_4l*bsoIah qf u>21NawV#SlmXfR|щr=KWq%sAҐć)e 2_2vsY7\wPKVVyU`ϱXSy6>`Lf,3}\ϻ-u綘9cLfher~?i{Y֮dnmjzǰ rZ45Fh[! 9G45F@sHCZo!&cxICZoW9=Z!iICZo0{Z{GjW"DV#n#ܟm$ҴOCg='nӓKAnh>q5O8xx$_}%!r.iZ/}Bm"M~z@ky. <'4kžx#|G{o#dwwh0"PnVD2S/wS:BlSBm|HO ;_G_"MFf;h;Dl}[ nN;HlGn:6g3Hlgj3}-L[ H7g< &[vaݽBmK=h('pr$MF7 yb"MnvFn|x$MF~ן縧:??6y>O9ËlS$"^"w7lNh"MZsㆣ#\fULhX|Y*RQXp$mptr!G/G ^#59QSDd z! (FN0f5!JT$" Xӈ0(8v;m#i/>a&40=`QwUk*NGtkHhx$M8yQMu 8wx#5S3`M#toٲG{U#iڧ!Ln~zϕi}>#w~H3P5YD5oʃμ53o*[ >|3_5gE>|UUU~*fOΠMsX-V}y R[fpa?eܗyE)'A>VEYz[\ה]Kbd?W菇0zM!/A _BY` jK)F?u;C/͵߰Ĕ&[իev)3bo{h·z燬u=z@_Qa>{:zMYizc\smn6ڱ,k=ty/D;?a~7e|<\:#;dvYjf+΍zM^;vxF3#Gv(]+"V^QV-qy1U1#@0e1:㱢4e1&cxS>Ro؇ճ_\fσN5׮/|__AgwsU~;yu:'T?VR'<-VgQ<3+VGt ͟U7ĒvئnrSfc4<uX6WA~(vaN'ofE"y:1|w+s;h+df&cNr\ Ǖ}d;8&94%c{jA/KXD^gk1sm1q0c˔Ҍo3aNf{ 7Of*bnwE[Q阛mީ裹n]w,r}tLJO(t FZC0NY0gL-ڲ18Oۘz2STIw֓o|g|Y{O? !&_?ۖ#~k><?͌{mrl~m׌[l߹#mO,{AaϋJ7k2m뮈WT"a:e!t?O֬#;w}@Ku\DZھ67U[}, ~@=$ Ű;o>}`@V8a< x[}l[خp8#;[{ACTr b fG*(֊@Bi$H^TQ"~QPj- QPi޻7w\~y3̼ͧg-K"k)J??R+UmJY;ԯ޴r(0j`nC`|?U t#G}-K #6rHrEf( WtÉDB-p`>0HZ7Wu1ёyȼe~aυ7b{zôǸ.6}c,[ͥ~v,)_ۋدy{`"; Ɗڍj7CwPUN5'os/~nw+JJA· =a;{Cī01O2AWkR /5bH_u5:'J}/QWw_^O ;vĻVXD KEjJq1gK,TO_+ўCl phƭ컍>1QR[aܑR ^ߟbXrfwg̓>36Q/S:9cC>X 1,&9y:8kIͣ[v  ω[sbA5Uh<6#YnãVMqˣGG?<{.Ί8 !I.4rir9?x>3GK/*#4;ľxt'>N4ľ"п1b_gqK4Nlx_$|& ]!h;޳Y.gW:pS{#`VNk>w ? 1 |G6F%%h5F.%!BL(}/~kk~O裏|#l]n|VH廝QCki>:+OCY&_lGcB}ZMN?qF说k~G hu8穙p> {sLL2ezp$] Ƃ .I??ܩ> 1}Mz#1o2>ycгbcN8,/A`x)4r QˣNwGZg2pnBg]<UUUe9S=,,d~ }L`qoV1\y!%[_uAn&6Q }Orv1yv㻓cy<~v.I7\SYO%mWan`j>yYkyּݟg{>yN(Kl~v%p=kw؏J!LZe;~]OCBmSBq<_2.u7iE>zR̩"'yMS{lU%xs7sPޯ(u8VfMJ>Bw!4z; ު;QQ:OZTQ(S`0X(ك9Bbz9lM{<ӟm^k~Oq Ajsw~h7 wz^ A"9ǧ}jϙ?RYwNm9T/m|>~;wڦg8;l5>8ghF^% ?lq7t H Dd /16  3 A?"B  I js2BTzX @= I js2BTzX`@V8a< x[}l[l`8 &N MH᠍ITr*@5M08RشVK$T,A+A+zREEBmSGQ{Rh,h*I[]{y;];oݛ7z0[JV{+BR)k;;qbT0톣EP 0 !\k>Xᘣ>UdË {1l9$cӳiۄT.'*_UoJ:[_Tiy"N[*0Wˍ=_bЦzJ[?:D͖mWA>IB܊L o߭<DD;r{c'7V~\|󆆆όMW<όPC1 !I.v| N@߰.U\^x;oƎ@zs }U }U 1HHU>*#J*g:.W߃~BF;A~)Or8]R g p͌}!hSb_ C+R/ngw }wAQEϊ9@s ?/WvzmMF' AQ6Oe&h5FKZ߁4AvAb~[a=2BQ6]>!huvYs:Ns;ѿ ,AQ6G' |=,]pZMFq {3JmǙGq;{>6 Zb=sp/0my7^Yxj +exs#;PJݪȼ8U[`Ӳ2Z0R?_x[7x;>\sUJ el8#{P?׮|[uLI ~goͫx38nnOMzn'{{o'N眐&GDM!g"\Neq9)*_ǛQ{\X6|ߑKk32+3첅ה[L0y%_;o͐pS|3+N}y$IfK$L$q!vh55#v#v:V l4 ta'55/  f4Ty$IfK$L$q!vh55#v#v:V l4 ta'55/ f4T_$IfK$L$q!vh5a'#va':V l4 ta'5a'/ f4Td$$If!vh5a'#va':V l4I t 6@a'5a'e4f4zDd j6  3 A?"w;zC4lF&@=w;zC4lF &#x]lǝ i\Tpdb@ąr8.BBk/pv~8zK=5JڞQQ YgZKnM)ΪJrNg֟Y ;3Hߝ|;Ƅ"*!.j˶{7Ofm/T&D&"}LXnoB4vL!Vڐ! ׹ qW)\Pb_tI|>|i D^U7я[?^-7˱WNb5%NLiZ|]>v:Lڱ{1Uض!=sHu6Qu8_*u_wީ`{?_Ԏ4(m~WN^'tՊD:YY)_kÇ+ƻO6weޱ)=nI.j}/_ժbynlo!SI[ruwt9jsuUo=xVY Ǚ+<=c~ֆq;sNjwx7x5~8 ~zPc*ͽN#ix clrї?&L_bVh(T?,zdE.-vQϲ|[i^k)Ox>j7<ٴ{Yn˗YHF- [:q|2s]׉-Ěۋ]%:$i˜uu.Wq,בN]E,җ9u7d( ܖ9u+HYNHϠW(^g̩+Q nt9^9uUPĥ_kt<s^v}TK#)vѡ[Mv}#vt֟\s`;uc@3H)PzZI=uz u+ TC=tzZO=ոz)R=ũ'SUuإq>̓6-Fy-?[R(u}f?6#׊^܉ E$+wvzjV4y;N]= G؈:ZGwu:uܾ`exFoqn + lhkC+aC5[V`C2׆@GUQj|Gvz9¾O~Pw`n>?mGQ`n{s>[ \}<}| u; l}õ@'ﭮNZQV&6W'۠k*I#Itr t:itu҈_ItI\4/C/F/=,]`7]k7iz4I`]A˝ߺ`7?08f_©;(N7~Էkzu 6d;um:kN\Agk)VY t u[ EMV\碯ז"ؖ0D} >C.’ēRS23²ij4h4 h$>(,H<(,HMN"mpu(axS)6{h''臔;֠ 5W9(ʨYl29M?ٷl#)c6}_1~m%N926d;V\rN92\#ўg;.B jUN92kxcF9׊;1g82[g}ReyMԵqely}"~uqd/Dd;_}O6CC+$"/ʵ (Kq5,= 8dxK|uQ%y&KS(bc,ʏ7%W~εhQ%g9g+YƸϿP(ʒg&ADQT~FB/ _Ff-Rﶚg97&$/e&B{vq?}X |K2dk25uOgʛmvQ/J6rc6X(oGQ i6z髝?p]r:'Ecl>fwʵj|JYN8.Gy5N>򍒏M:yI8ފ `+|˅Tr?qƐr%J|jFDi¹(5f)n{~5o" ufuXj]\]&[xNÛ,x Rw~~Aّ>^ux1cБjox؃V{,x;GQGI7)S'T[t$ۯ{*\x?JQQ:zA(o:RmQ7&OB?ǂ ((uVѷ8׶HKG\e*C?:m=oОՑjΤUx?E mE&XxH^}ŃV჆Y >HQѨqjt6ׂ8ՑjFLI3=&>bϓu=R7Hw#QT[t4{& ;31 >NQQc:zq??Ցjd.ZOE5|/6Ȇ?|ty/=Z=~ ,xs<; O\.a񋿝U[z n7~2:NYR?14txƒV_<wZSqߔ*~GcI8X8yT90 _zx?m |fT<4 ovz=xφDTՔ*]бj_ŃV!#QT񣭌휆U[Gz--GCmqz)UiX/t?Yq \|{Du/a}Ƅ`80y#;iؙOXI |9:V|}ߛ~"qv%Cyذ~K[ e(suP6(_2|&_5GQDJFTTHz( MjNԃ|RӚ|ԜF)Q!||Yԇ|!;K5P3Y7Rw^o5QA|QqMjQCET|H&(ߤ&_5(M #&_2,-g?Rؖb{c kG,fVF\]oї903g-$j$z[ȿ̓P_CukB`i\كPlbn>s<7 =q߃|f8yxHu_&y9$ ; y'!ه2O_JQGmKH&dANl=4t~ˉK~Of`_<|Svt!`J m a?xZ>4g 0S壨O,g y 7d^}TTroZk~^)z>T,Yzp Тb'c7+8|˲`^eb!5*>z[n߾9ؔR)8])>xoцGiKcO˿}: ,b;x> ~n> 641/*\aHOasYwK mO(@xUEv_c:]܃vaeF|dsgw'߹&bKߌ9[0_}_~M~ݛl߈}!ü{~2>Gnʝcu|:ek{GQ;s1fq7<WB\oHlt&ն>nF;kbqc|ſSR>>՜kpÎG9}cps..y &5&犙{Vbնo$xL1-s\,K! tIÎO91TX30?&{fc<^m[}dkWy\Jx-CO}yWlقW{_lU SK/}'sM'pewS+;j]GCn}:~?נǮ.~`\\/H5;Vr|x+*\DMR |0ggDgdl3g5lKf::~;+LsulV(ŏ!?|e}6#gj~qӫ>: L n``Yڌ5 ;3X[~A ۛo9x /W[Ͳ ,k؞}YR g3KfyLxSo}ؠ*Eϔ|uE?]r{F;vޜ']7oleg) ()>hbL1EGo\;R W`><ނ:gQG? ު(= T*u ~_L<ێ'1l422/}Uo7|3(,ғq[ۜ/}G$ ~ Zn j&n "mYMW)x+xG< ># >|*%WH[T7UZ+y̗Tv`gpZBsG -Bl uF^!] -,F(qDpars(7(k"@.dqYB,PEy9,er g7S^Ynb8yb(7BW6#ַ2/%M^&^JU_K1䥄xYA,8^&/R+2q<}_rFyq&7͏(/&H%f^i]2 xxreNrbŔε!~agC밖}[n\VB5t*gۈzjxـ>1y/U!2`x8^VL^ʈՔ2&/+;(oY.qЋ)FrvNyqvS/6f)M!8(2즐 8*G&G.\M+~i7{]ԛmσVVg'4ۄk;)]M䦖8n*U&7UݔWM䦒YOyqܔ+Ln*W]rr⦔2A#.`gC!O#~qjg7}hH?u9~6ͦl&Mq&|i)䧜)E8SǍj^ԥ . 5lN^ԥm҆.kKko.|CͣD4kqKn/Bw/ N^^e@s t里70ּK ~h.C_>x.#kI;ݽK eH/uF[r^-9 GF:x23boG W)buo a[k½~ 7co[A鍄m4^ԯ~g~m?w /`]C * ^N?nYj|rSjqm'`ׅet8B9ݔ^ԯWs= Zx u(F܋ h~FD.&؈{Q YQ0(ĒO~#EIOQ7H"Š+=uI=ޗ+֍Sx痰ZK Qs%qEYSr ]Tn/imQ"{h(<6D[8kÝe׌:Q$(Sxqp$bQe _ޠ!ߋcoO28z\%ʑhGmߢ8_i"Ga??| D9mh+c|4~"GmX[ 3HŊ.R }nwin"Gdp%ʑhG=[e~6lh#E3ú-GXqԏjY*v|QLiD83~1(G-V F-+a>!ߋ`'9_&ʑhG'ZV৔POjS Y`ϸ\;CH#{}m?c˟/?D@>vS` ?>G2}Ey#7U+N i Qy^c:j^I;4OVVm}{YNGIZ)?yW:kRT"leh픖ݧؽ\oQvb)odӕaQ߃$?cӒ=$g{x-{m {̛#d!yV{r{abk]nV_ƨŚ4kĕVe|0 zYdɉ\GD:٪V(gQSI||#6)q؈q*,a7R[ ibLNg̮lj%w9ةNsg}]/@>9J(w;Fc>%?Nc4m~Tlje ݄QOW (Q ~=[a/v{^&=S!Wa/+cyc 8QMmvy-ԍ1z\ uS axs;(bܽbn+X up"m HrҔ5]~,cLEt2+>v`lYѼ_Y| ޵5_(SEHcN:s-pK<2Dd j6  3 A?"1KXc=-WpYʼn @=KXc=-WpYʼn &#x]}l;18uBP}8ğ0 IhBKbHH98I1JjPI9&qrE(tX$TTA:ICdjUV7awg|ݙ۷7o1\Q|}>all4YŘd,l'@%0^XF |*  H3g |J(\~*k![nvx_L:~(7tZdpܿKص؟N[SKLfca u|\|Or+-~`\,Hop7skTl\Ei{7m>c b ll v)d~ 4j+mϟ Ei;שWQo#,42-\/wͺ#s8F|L!_dyl|}rڍ \c&tr9;at5sc@-rOuSLQ*;*7߫;s.kpE#ث{|P_+\f|_Iv<9fy2UI`Q d0~n}L:G$qi+4jQ7iMoQf|x= ߠK<1O(xB|d wܗ:i+*csX/~e$1XC<υֱJjgm 9F<(2Cx s.j2]1 YVB)6Yz)6h+ )VD% b"j+2u%!7h+zek?Xv=rJq-f[jL\iPIJ0&T/I#1/1(z7o7M(Y?0kM^/w/Q{S1M4'˩-N-Řu*L)р^jR7p`r@)\c\,z} uzzץpQ3#.ꨭ.pQgrQG\R[DGd.[rh)>c#a⣑ڗH=|#KGS[} 'zzצpRq59zM '5ט}X5ƽ}qjMP׬j{fUS:eͪ'UM*[߈0g(G1G5*BF`m $uI2dd8dd8dd8dd8dd8d6,p|yQR7Y,`2rChVS]#FnBVШ_kR.Fs9ęF|ǖR}RO>u6|;^F ˌm؉n(ƱET/1bNc[IEq7}=8A#ml =iu6>U/28v.Ջ8q6~8׈u/ژpillLPˠq]UJxQAg_k;"E]x6H6SˈqTBxQ͈/ǜ^K96隣K{~O݉k;G7S-~HeGmy-~ڼh.x觘^E[DܳE[b.m-*R;2HŊWesoߋz$=r$bQ+OTYoEgGޅ9.G/Vu UVs;SwC9٩lp2E.G/V 9r;|^(̟pl>HeYXqUY*kh^"Gq>x8zr$bQSx? > ޏ}/rP||<`ΦˑG_ .;|@B9Jf'rtQR!?<Ѱȃ.G/V-e#ECDyF6[vf=}s˳hD 8}_,MxߨѿQe+Qǭp,bV ޮ_QxM ]k ;):k.嚠[fPN ޭ_Q5x/ $E_/Urd[菺{X7/v%^)nx菹=q<fq܁>e+x菻qB&oh!G3X.BV9|P?(,r8$CQ8¡CIW9F,U 8}_-O9 >?(oe|9sG~ƾksi_+kkvRi}].}ݒ}]MN*]/Ҿݰod_ZQm7G틹JŨj=Tb/Ҿ>/~Q{ԾK`ߠd_Z6@%Q.;$2D3Toإ}`߈d0Pa煾L_Ro8ybi3V^\\gq2옂?ЕjG/A+_rFvvW7[_BK K]w!/_BS_+ }qZ/E.#wLC`~>e<ߍc+0ϋS\rп:YR:t :N[?RmFu|a_`_pu_I|m>KOlNyo?S?6ǁ×g'&|A-t!{=VVIV{$d= ́)z@LLէ^<s#6m$/~iWh1im6~)- wK:7wߤ,`eY4g <țm '[+iGo?No4)`Ul x !F/tDH|FLc2,뼺t  tox?YоcZ{]UHSIqaF#$1dg(Ksl^[NQ߯RٓM~9 Y=y;_te'%j/;]J-&_s o_쟫'g/:QҰM틝쳐x?j>&=_='4~ƞl d)Iʏ?P9z s E4]_}-.vODlu^ۯԽ% Sb ]=c+sH;锏DZ;@uҴfUż;LG,RWaMc@$$If!vh5 5 5! #v #v! :V l ta'5 5! /  / / /  T$$If!vh5 5 5! #v #v! :V l ta'5 5! / / / /  TDd j6  3 A?"=gp.~%1 3S@=gp.~%1 3S &#x]lE~p:!&?Q \|q%\~M 1$3kwPh XQ*,=\ɵ:KON;UQ3>7g?NjMvf>|33wcBǐJU qRSgQQn110dd~ !j'&6ɓ% \$P7'[>1vۘl3R#e^_5׸;Ck⎮SJYSbo73%q!qH0Y|%iE~ 񪙿 [y^b,Il6mu/Y n/={v~Bslvǜn.>S'"FT`LD.5Nl?:_2~>>µ(mxM2h+wU嫢Ze,'Si/.|r̒;Ŵ#;.\웙N)׺MT}/]=֪bfi -km:Em3\JuQo']>o}?Y1c>]WjmI;>>>Z0 [s3ۥܘss~ΔN)g_8ys'GS;?_>% y\K[~t_G*rB3n9u vu[]ǹG{q7_܇l%eSԭ?n+'K)'-v8(_-'wMSĩ/_% N]%|;ćU:3eܧĩns⟟sF.gς?gOk㬁\fpXCg+GY8[ ξt9[ ξqg =9u cK,%~|P UnvolNv'gΡYpfn&ooJ]99+B,; ^wGqxW ma`y%7.͐.kBZWVʬ"?k)+Cb 8\Uy.#G\W[Wwel_')9 y.ǻn^Uh<8]!v~{ޘB>cn;ڊ-y|=H>|m_Q#w& juAQ8zuy75iOP_7G֒:w?B]8K\t~RNSP}ܯԩ!^@+^1x탄~]HNZZ fm㬉lt962i[ ex]P|ՑU:uV[:4wh2 d$ޭ(,H<(Ox&MFҚӚ23²ijH篧gƃr _u_272zEŊd8ľqK|ۣ8q,Zsl܌cM(7cxSfc;5a*u]rrm潃FG݋tLnخ)G}h=省(/^Krm=H P^=;(8M|qq;~rIӼlhk,N96{ Ìu N3 ;"ic "c:ނ lgᔣhc=)W{;0S[zS{Dmk}f7q9-]|HEߣlQ~?Eі>7m(2͖~έKmI6[y_S?-25{BۣhKy-}(ڢb]HhgF¬w 8k95 (!|1g"0G'}vcO GѾG61n} :(gO6$k}Vh_1&}=]Hwʵj|룟e8u{{ˁ-wQoq}FIh_Sm O{Wr5j'E2T7k|9d>eCğ7y,j.4-rX׺DBU~?|oصf^jCVQ^gܕjaפf8*d m#=LPT[8a¥tقo(ra.axqB9RmHU|Gn)Ga[(G-^u ?eR9JqU18zkcB9Rm⨇LY=QbpczʑjG^; ?~܂(r֭S-SxqO˔U[E\Y ޣO>Gxq$Yy*|g&>h?CƤ~T(G-^ s)oL|؂P9_Y﯑=T[8e ӔUEC"Gz!sHkdN0NwG{: `_kO\*b9/cE 7[m-!ވrQ.ƨĨ8Q1#ƴLjAxc:woq?icd8Vm. ~Т;;{,Ve;j_CV {,xE*+ZR&oǪ-~ze[~܂ o(FpcY=~%-B["q1x84b|i#F7dp`|Ȣ?Q3Fr񧵌?Ǫ-~zY=~[L)gccB8Gcֈ^480Q ~Ѣ?(jl }dK[wR/gCK>C윻. 瀝sQ#}of_jԎQCEju J]H/Ҿ׫ٗBM/J=HҾf_jNt/ @i"j_:}oH/!/Ҿa7ٗAJHڗ i(ٗEEF/ mV3moK!Vv˹m_Ia%%;رc 4CBbQ< ?ԿCkw[_KMkCnl_vK,J7zJcD-AYe2%׹Ca>uMw^x^ЏQQB_ExC!w ˤzWX¢.WF}MC m7. Ƚk4%w0SS㗹25ރA %WP:\K7 iB2rafl8Ň/|M pxnt N]˲`^b;0v<05=;csgw>LB^ۣHVխ1f!˲ :^N~DmEIc^{N1y2R1sMb<ҿK_aF7r>X#eyte9;޿.3q-TsVw` uo/ϑc2v97zq<8㸌e|D5F'O-~)w3}Cm?s($<8l?|q})[O>Z6cBxn2x2og =b|vO3^"BmןU9ƻ3:^pz#!7[Lx[4em_ OynZӖl*k_o9}϶X:\uT˝/SERϯ2k.weԇIu-Dd j6  3 A?"_GM ~s@=w_GM ~ &#Ex]pT߻\b#T +QA@ $I%i CӌcutNFFIj2dD:ce R淏彷 yo~^1RLƾ`?26&b,N?XֈI }bW2 `(HFP)?Rpя++Cm)-Dqn (Ŧna0׳1_p]U|͔He6b vuYE~d;V/iߣ#K]8vt 6A%?zc^m''࢏߾o] jo:.9 9;'O\fcguS a~HbϾO,9:*T>r> 6dy5<],sł_dS,Z뾬f;s߯߱>eM~Wsss3OvkXΛgzz1T/XxsV U|L\&?QWOԯ}TfqRz}>VFKcm=~2pQsw-Ff4qu~;z%!QzE ]m1>=HuV&g|Nua]>ڂ.̞eP]ՎS)}cow pSawQlñɔz6(PyզmPMr%m+ƥN#n*RVNuV;/ю㕬%^fXBuދ,>.l;b!>Pl>-탢n6󓸘{[\SyFa3<=0nq1aG 'Y:ߦel!{A~$?,?i ?yxL5IT5/5zHe^aRARݪ$^,^V/T<ŰrT^~b*/H_`I//T#tJ#`ݖ&N~Buے|f6⢁꒸,.ꈋ(U'qQ[W[\TTWE)l]aqQA\ҤsV ]jJ霵J(K,>J\d}z6C?=v=3D W/1:⥑x$^a::bU'R W[T/UTWiه{%$8)⤜8)8Qp'8|[|[|H(TUCrQhGaKxp[CvNPl#҆QĚ6oRH,xm+X }OTH}MXOP9d웰lr" 7a>AD}²} +`4},$~8i-䰟ۨ{ rg6|پߟb*ұRQkV)UA \P#yN:PM."qKqM:P-t}gQy>Gcb3cfөOل狲u#~4k}E3E<*8(Qݤc3Uc1x 2{+v8Rc8NLw'q*Eُ:3x_;tgK)犲u`k}Z]? j6vxM>Úz l%~*p}O+.Q=t%`-G]?fҥqT8|Ԥ^uiM4#}{S~ԥsK]sJA҅X٤1ěo@w?{t9}\.54ǿCw?I!eix悗c 8=߁5g"N{!ѷʫ [-~ԯ:kFlي5@H]B啄y?׎8ߤ_;~ e"G<}O@3{4]TMcz<_׃uI㈱OoB\*/$lQ^ۤig` +y?usB.l_~Z7rp{=m'|~vMGƺ7{T9cSol[!ߏ3q`T9c z[wh~mQ6ptk0GSHŎc6x7l3݃p/]hcG͇8oʑG|ݽԦOj~ I#GoؗKr$bi𯷕g \wNGĥvrb=^ w}1:?ئ6 #!>tWbYnX;q\6.%ڶL[x/3/o)46zذv8겶Qmm\uƀG=&rc9 рo7oo)qYX_Jmڪ'\=|A~Χ~ؿq[<>]ۛO!c96mR̀w{K}P9q#'^S{{Z~p,̦?nOxWؿrWNx|O9u|&˩>Q/0:SN,[vK}ɇiݗ|ʝuzԯ*EJ5>կɣ~o_lR#&O/N5Rܧxԯ*PM+)TvAEvRvQcЯ[ѯjtROqwRѯjNR8Q_/՜iJ>cA:џ$%w-C|6}1W?Ib s.^# mI6 aF~'_km ?0<ȟk#?ϓ|A}6Gx/>?|ޣD5>$t{Vy.{]2r﶑[?[QDϡ+6r(wy{&)FL~tя p 8q+}%}>Xvֱ:VAG#l [Ol+F.v'aUnB-[+ֻ=́a]9Hlt2Gs05L9(ߋH;2^x/RB +c6ks͑k{vy Kdžh~p*[51O< "e${e\]Ϗ@/ffUfgw~Us]Hs1j|L]6.[^D;2/q[${|rĜius$.1H!:;C?_N||˧пjP)ABqg6sqt>G1!7qc;Q|k1\c~wt6I\gw//]e&/N0 t,L<_1wnX|YovA~9~-6>sxPw5P[(O1f~)o/keM5j=hBT>HAOJw/x8H}0bor]'XBZ dWߩ`,ɢ&j,S_'_, nYw|?>bpY8B` 85DDd j6  3 A?"l8UB'ab@=Zl8UB'a &#(x]}l;pB6p;;@RR !8$(uAͧ TD )UX$Rҩn$$ TZo~vw3H=4޾7yo׻>E19ش[S}iIٔ_Rc!YJb L3?]f$.u kt3?t|07tRBSM,bfF6(} y?(lU]l?!#C>| 3cW DZ2~)?;Vz|xXc}+U!˴N1%EY!6n|j {matl:_?Fi;9管F/X?x1 dAu|'e;S"4Xz|4P1Ժ\S5Ժ[yߖU'Z/LrY}ϝ^*v,ah7ϯηd֟&Z2h3ʒI|OD:dyq&%B ;/pt6<cq.B pvubΓLrvvߏ:99>F:KF|5Slm-yR 1?| \> _w+nXV RHg(I}'u|$\@FF"Wvç?grEy"/0s뛅1'7K~LV7Moj5~S5H&ߤ ޮ1)x\ >'s^Or[*Q_Jz-_Q@7[Z+y̗$fV Hv4Rn=+RS)!eԕ@r_rE!׎YYF"<[Huņ|Eh;cT+dbYQWDBSߨ+r)G' 7gpl)2Zu6״n=+P~<TWC䡒xXJuef;XCoԕQ9"q H,Q8,!VBu Rx(@_/0yX@&TKAuC0<7yO并~/-Da5jSx@?ך<˩"?T?,QATQ(bM1/".b<KLJ\C䡈dZ{P])'D>8~N4m9E XkFʭgDkR8i@19YCMu )ԣLNzO:{+.Z{K1{1f86)TWܬruTMrU,/UTQyD%8ɒ8DG{j⣁RI䣎B)>R>rM}]c/j\CI*k{ ZQCNpQ4rE G*GQ]D''5Rn%!#$$#vI풌 ǣb cz$H2K2K2KH2OH2d8>LII2zްna "N c-X'7RZjU#լ2W应)[:kMaF6lcET.m9Fً6،V %1L5-wjm'mAmXwlcb*/m)ዌm:;'ƾT.6־ƶГ6.eƾT.2֋]6ߓ6q^mGB^hhXl{gΣr>m7^Q\lx6Ke~]XHxQGgy({Fɯǜ^K16况}UҝxownO.aRJыxΖfɋZH'Xyіfm'lD~?l-1uta>:mq8-q-{x/ל˅-8 ۽hKe{2/2ߡwE[DQ3Rl)k9b{ofͳFk3BɌSE&š|&3n7kZo#b{ѾT A)T3DAJ ǟ2eLEl&`fx\&?qL) 8<}gfźnw)}aW(?'%[,d[U|-gli?fk8J X.G-VmPe7;| ^(gH2(r$bQ3Pe_Û5x {v< ~럽M#+*+TMwL@ 9V6{M#+#& 2y ~ȑUAs(f* ?0 G\>:/֎c]rߨ7kor7ypv`M͎u[rVZ2o-axG9nWI?YI8mX:Mwh#?Q}nݳ/']>|1>T#{ѾKa_d_jZL)Q]:$کJm=j_ԥ}d_jPRԣ\%bTM.J1Ҿ㰯WjztRG틻 HũfJ/ҾA7$ٗ!* RJxԾa}$S}^hD1ݶ~%-|׀c-x %:o;HME#/u2 4(Wo @j 5;./?ׅJMA, \鿲Ww}4_[g*y|9,q!j1w0npqȯebJM6U2 Lxo[nH@ύa<>X6Η%b`STsJx}ctKn"V =s1ۖ|7E/a<Ǘc5?U?[Jv,l|BVqOৡ?]Z/v9M-UE\'_*c堋u 8##a;9`!+W^ ~Z 4|Las~wBvؓo "L]R1aeV;k0Fk|#F>ݹugxbܽbn+Xu/qb+ ?Ƿ !J]S+%^^%n_c qՔS9D%s/ĊΏhb F@Qcylnq/ kbҖ-%5̮~2>11!>M0,ԝp/tk֯g|>~'JKL-e=}}:1 χݥNiǟ]zm/~U? G]KsqMu6R*X5b#T'W,U+ʫoS(/ |.Pk8CWp+\G!ܐ7 y!57oP@9` L&v;L9.YXOzlICG$\Æo^$E"`^$E"`^$JDNzCmȻ !6 9C>fX>^ޚm[b޿D-t_V7W۷c<!\sg^Z֯{\+DpOpdukaȕ:??z[~G㻓;v:=%ޤH=%=%z$O:TSt$1h?' d)Ǯңytx^&<}IQPJGhuH>\=Mmzs X-GV!V!V߁!#V ۝X5 VM&Ī RRӣXp;H&v"^dīī^ތxr<xg7)99$xUrQS2FG3bt0(葌LшZC0̜^>7\e^)wt9 %}#rak:GZg]r-$Ec3\Y8[H:>u9Nn=r yS|x p} j_EGx}7EyÃ|< $W{IYDGx0x\YwaP$}TSO.\DQp.H /ott}=,EGf{?~}WSsՂs?{Ogjb!N c% FG?u_3ݹ vG?uo_J3=QT("|i/KIAMs1]{ o+= 3(;ig"AEg9 ,(]w2GٿQp. <ɍ*]r#ɍ(/:1y/ mٶ|%s\?gns'JF|xW8dޞ΍ ClP}w}.Bد hlá5_/ e_~_b .s+(o>E__|Q'Nn]% ]>tm!Aq`H+|6^sbH to}%/}ۨϵ8mixJ9PF '?6v"vGD R]l bMO`[tԵ$I]dCa]m,Dݰ7`Q/-'[ܧw\Wr%3k>|I/y|N$+&fV>EWֆ0}eoPޗnyyKx-ǸĴ?ű,'^y Dd M /06   3 A?" N;U*N@ "@= N;U*N@>Y%K x[lUWR be 1vgnoA:W6PVS5tĄ8a0da6DtY"?fs~yzwWz;|~|9g>@!ÿ34yz яʈm#hr7 Rz"fdS{T nU/ON)b-z+fRVwglq:T= I]ĒyvO\}%ggg\ qkŀdgh%l(E /G)bri%<_ iEлaI?vce'hϺYw_ǂQ~ gaNו3ԁ95=F~Ji1(]T$њʢ{Eq*uA9\ra)r T@)ri'9})| z?ωN[PއoPoEP(A{5+(4kQE%EѧwY2lXh۴xO|Uq02^pK.4J\'7˰O_7prr|X,'Y=W=KwW8gפr6]cԀrU9IoLkSب)lY¶V3=)38 W|sFq3 0 no7 z/3nQ|ќnqtg$IǼa;'ަixlq6L>Oe.~:_1Uc^E=Uc^E=Uc^E=Uq7\ ~2C?e' 13_ȼamϵ/UL'??j'B7ے{;K^hEmGT|@F'!b_y>ڒfzF֏}P2Cܮ@ҶSN:C(9ZG@Qr/B; N r*3-b P__)([_>3s|z-bg{h=FQ1"F/ F;kȈQ5x)Ĩ ֌r :1jEw֑Ao:FQh1cԋ=ZW/b0 0%xv$mx"-x!IKʈe8!|恹Oj2}ܹ,OV-u1^KV[[׺X\lca@[(> \-$DU.Q̑m/8(=hCi^6r,9mFM7J<65#fq@8q,iZl<ɶ`le'j ,m,ZrF클Q;ltmu$/Ԗmo@փh:C{wpӶl<^Bփh:{7~{}6 ڧ]xwsۜclv0xòD;/bo7`SQ}A|ε/o]s_r3f~Pc$Juj_#Oz>;E (1%_S{o7Y }诒w栿ϗofq>'&_k|A/17U*=/_W8/ơ"7dGO yx(^(K<廐 qzwrV;}rmz}v{~A zw7|z<-^-ĝ\ #gwKo˵7x,;&qjN '``;!HkWq1P~pmb(!WE11(pYƠni֟/HAz}>,wE4E蛥;g g>3]*;'A8q=n}>cZ^*fB-Et譐YJKGW]O?0z=#o ;ڄRF.|:jOBVTG[n(c@g_̴UYh ^!)ۅbl4 xZn(2ƶ}/6~s\;˝z^y&lnR_QﮎWMUm{G+g|P>k50u%Ac2Wz9ͼ,9f|EW` HT?s@>NJ򻧰i_r3c&p1k]~H};>^[&Y;N8-dD<0N6q*1Zo\l|]5h!n}Yw~0BƳ7[g1g]b^:sgN>0m,"}O,u7e_r3^/63?Ʊ": }U Dd M /06 ! 3 A? " .g*̟o -@= .g*̟o>Y%K m x[olTw~>!ŪKRb &2 n{;\Hb(nI[0v+;9vbVhU44|i>؉PT4y~Eƻ3;ogfwv޻}b; ?ۄC♼.z-CyDq-4^z Rtf]&j[ XmoOb.[A)˖Xr͒Ĭ NLbccc+\ eGK'ڎ9 Yu6={q} %v&s)#Ky$_o wĽA$'lVًٮ(4Yr|*'%Db n%QU,feTWxUlAОC:$ڇGEyU>(bT'Lϟ(FE2t)OEWG>Mx/E\o;տz-ʕ(?ݎz&{8Ѻ~m~Y"h25tnt[WyZ&A7knpo; %xr>@j%&1p91\2ǥnf<ylNk]G{͹܋7|x>gⷈ[.GT\O> Author$a$CJOJQJkHPOP Paper-Title $xa$5CJ$OJQJkHFO"F Affiliations$a$ CJOJQJJ2J  Footnote Textp^`pCJOB Bulletp & Fp>Th^`p4 @R4 Footer  !,Ob, E-Mail<LOrL Abstract$ & Fx@& a$5CJ;@ List Number 3p & F8>Th.^8`POP Captions$H#$+Dp/0$a$5CJ>O> References$ & Fa$.)@. Page NumberHC@H Body Text Indenth`hRYR  Document Map-D M OJQJ^JB"@B Caption$a$5\^JaJtH TB@T Body Text#/H&#$+D@/0$CJ6U@6 Hyperlink >*B*ph4@4 Header ! !<"<  Balloon Text"CJaJt@3t }Q Table Grid7:V#0 #$Pa$FV@AF FollowedHyperlink >*B*ph`d B'QB Comment ReferenceCJaJ:b:  Comment Text&$a$@jab@ Comment Subject'5\>+> 9fD Endnote Text ($G$a$>*> 9fDEndnote ReferenceH* +2Aah|} &K[!'<M^klmnԙ                         !"#$%                 w v u t s &=>p o n ?CDEFG +2Aah|} &K[!'<M^klmn      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGUkmv6wԙ7lD{TUdt o K<Ic< T!n""#%'''*++++-/013P4556o9h::<>=G=i===>>???E@_@7AAAWCtCC\DuDDDDDFgGGGGHXIJJKdMMNNNSOtPPP QRTTU`WdYYYYYYYYYYYYYYZZ$Z%Z:Z?ZDZEZ^ZcZhZiZjZZZZZ[[[[:\;\]4] ^^`adddeeeeeeeuevexeze{e}eeeeeeeeergkkkkl,lQlRlTlVlXlYlqllllmm"n#n.n5n;nAnGnHnMnRnXn^ndnenfnlnrnxn~nnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnooo oooooo!o'o-o3o4o5o8o>oDoJoKoLoQoWo]ocodoeohonotozo{o|o}ooCp`r]svvvvvvvv4w5wwxxy6y}yyycz~$%/15:@DEPU[afkpqrx~ƃ̃̓΃ԃكރ  &,16<BCL[`ekqw}~{U,~É^iKc֍[K= bՓq o$%&'3<FGQRSXY02A 022@09@09@09@09@09@09A 022@0B;@0B;@0B;@0B;A 022@02=@02=@02=@02=@02=@02=@02=@02=@02=@02=@02=@02=@02=A 022@0ZC@0ZC@0ZCA 022@0NF@0NFA 022@0 I@0 I@0 I@0 I@0 I@0 I@0 I@0 I@0 I@0 IA 022 @ 0A 0|Q|Q@0S@0S@0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S @0S@0S@0S@0S@0S @0S @0S @0S@0S @0S @0S @0SA 0|Q|Q@0Y@0Y@0Y@0Y@0Y@0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j @0j A 0|Q|Q@0a@0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0t @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a `0a @0a @0a @0a @0a @0a @0a @0a A 0|Q|Q@0j@0j@0j@0z @0z @0z @0z @0z @0z @0z @0z @0z @0z @0j@0j@0j@0j@0j@0j@0jA 0|Q|Q@0t@0u{@0u{@0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ `0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{@0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ `0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ `0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{@0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ `0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{ @0u{@0|@0| @ 0@0 0 0@ 0@0 @ 0@ 0@ 0@ 0@ 0@ 0@ 0@ 0@ 0@ 0 @ 0 @ 0 @ 0 @ 0 @ 0@ 0@ 0@ 0@ 0@ 0@ 0 @0@0@0@0@0@0@0 00K00K00K000K000K00K000K00K000K00K00K000K00}0K00zK00z00@00@0@00@0@00@0@0@00000000000@00@00@00@0@00@0@00@00000000000000K060I0@00@00K0;0C0K0<0@K0<0@0K0=0<K0=0<K0=0<0K0>09K0>090K0?05K0?05K0?050K0@030K0A010K0B0.K0B0.0K0C0,0K0D0*0K0E0'K0E0'0K0F0%0K0G0#0K0H0 K0H0 0K0I0K0I00K0J0K0J0K0J00K0K00K0L00K0M00000K0Q0 K0Q0 K0Q0 0K0R0 0K0S0000000PTU^w 0EFOhd   f g t o K<Ic< T!n""#%''')*e+++++-/0013P4556o9h::<>=G=i===>>???E@_@7AAAWCtCC\DuDDDDDFgGGGGHXIJJKdMMNNNSOtPPP QRTTUUVW`WdYYYYYYYYYYYYYYZZ$Z%Z:Z?ZDZEZ^ZcZhZiZjZZZZZ[[[[:\;\<\]4] ^^`adddeeeeeeeuevexeze{e}eeeeeeeeeeerghikkkkl,lQlRlTlVlXlYlqllllmm"n#n.n5n;nAnGnHnMnRnXn^ndnenfnlnrnxn~nnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnooo oooooo!o'o-o3o4o5o8o>oDoJoKoLoQoWo]ocodoeohonotozo{o|o}ooCprq`r]s`uvvvvvvvvv4w5wwxxy6y}yyycz[|}~~~$%/15:@DEPU[afkpqrx~ƃ̃̓΃ԃكރ  &,16<BCL[`ekqw}~5{U,ՙ000000000000000000000000000000000000000  00\ 0\ 0\ 0\ 0\ 0\ 0\ 0\ 0\   0 0110>0>0>0>0>0>0>0> 0110{!0{!0{!  0 0&&0&0&0&0& 0&0& 0&0&0&0&0& 0&&0/0/0/  0040404040404 0440K<0K<0K<0K<0K<0K< 0440>0>0>0> 0440@0@0@0@0@0@0@0@0@0@0@0@0@ 0440F0F0F 0440I0I 0440L0L0L0L0L0L0L0L0L0L 0440S  00T 0TT0V0V0V0V 0V 0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0V0      !"#$%&'()*+,-./012546789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrsuvwxyz{|}~V0V0V 0V 0Vt0V0V 0V 0V0V0V 0TT0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\0F\ 0TT0 e0 e0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e@0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e@0 e@0 e @0 e @0 e @0 e @0 e@0 e@0 e@0 e@0 e@0 e@0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e`0 e@0 e @0 e @0 e @0 e @0 e@0 e@0 eA 0TT0m0m0m0m0m0m@0m@0m@0m@0m@0m@0m@0m@0m@0m@0m@0m0m0m0m0m0m0m 0TT0x0x0x0x  00m|0m|0m|@0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m| 0m| 0m| 0m| 0m| 0m| 0m|0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m| 0m| 0m| 0m| 0m| 0m| 0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m| 0m| 0m| 0m| 0m| 0m| 0m|0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m| 0m| 0m| 0m| 0m| 0m| 0m|0m|0m|@0m|0m|0m|0m|0m| @ 0@0πP 0π{00 @0 TUV~ TU^w 0EFOhd   f g t o K<Ic< T!n""#%''')*e+++++-/0013P4556o9h::<>=G=i===>>???E@_@7AAAWCtCC\DuDDDDDFgGGGGHXIJJKdMMNNNSOtPPP QRTTUUVW`WdYYYYYYYYYYYYYYZZ$Z%Z:Z?ZDZEZ^ZcZhZiZjZZZZZ[[[[:\;\<\]4] ^^`adddeeeeeeeuevexeze{e}eeeeeeeeeeerghikkkkkkl,lQlRlTlVlXlYlqllllmmm"n#n.n5n;nAnGnHnMnRnXn^ndnenfnlnrnxn~nnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnooo oooooo!o'o-o3o4o5o8o>oDoJoKoLoQoWo]ocodoeohonotozo{o|o}ooCprq`r]s`uvvvvvvvvvvv4w5w6wwxxy6y}yyycz[|}~~~$%/15:@DEPU[afkpqrx~ƃ̃̓΃ԃكރ  &,16<BCL[`ekqw}~5{U,~É^iKc֍[K= bՓq o$%&'3<FGQRSXY`ght}Ɨȗɗٗݗߗ )*-.1278KLM\fqrzߘ  2@ABGHLMNT[bcst̙͙ΙϙЙљҙՙ0000000000000000 0 00000000000000000000000000000000000000000000000000 00g 0g 0g 0g 0g 0g 0g 0g 0g  0 0<<0I0I0I0I0I0I0I0I 0<<0n"0n"0n" 0 0''0'0'0'0' 0' 0' 0' 0'0'0'0' 0''000000 0050505050505 0550>=0>=0>=0>=0>=0>= 0550?0?0?0? 0550A0A0A0A0A0A0A0A0A0A0A0A0A 0550G0G0G 0550J0J 0550dM0dM0dM0dM0dM0dM0dM0dM0dM0dM 0550T 00U 0UU0VW0VW0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW 0VW0VW0VW0VW0VW VW 0VW 0VW0VW VW 0VW 0VW0VW 0UU0]0]0]0]0]0] 0] 0] 0] 0] 0] 0] 0] 0] ] 0] 0] 0] 0] 0] 0] 0] 0] 0] 0] ] 0] ] 0]  0UU0e0e0e0e0e0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e e 0e 0e0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e 0e  0UU0}o0}o0}o0}o0}o0}o0}o0}o0}o 0}o 0}o 0}o 0}o 0}o 0}o 0}o 0}o 0}o 0}o0}o0}o0}o0}o0}o0}o0}o 0UU0y0y0y0y 0UU0~0~0~0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ ~0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ ~0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~ 0~0~0~0~0~ 00 0 0 00 0 0^ 0^ 0^ 0^ 0^ 0^ 0^ 0^ 0^ 0 ^ 0 ^ 0 ^ 0 ^ 0 ^ 0^ 0^ 0^ 0^ 0^ 0^ 0^ 0^0>00>0@0h00@0h00@0h00@0h00@0@0@0@0@0@0@0@0@0h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000@00000000@000000000@00000PTU^w 0EFOhd   f g t o K<Ic< T!n""#%''')*e+++++-/0013P4556o9h::<>=G=i===>>???E@_@7AAAWCtCC\DuDDDDDFgGGGGHXIJJKdMMNNNSOtPPP QRTTUUVW`WdYYYYYYYYYYYYYYZZ$Z%Z:Z?ZDZEZ^ZcZhZiZjZZZZZ[[[[:\;\<\]4] ^^`adddeeeeeeeuevexeze{e}eeeeeeeeeeerghikkkkkkl,lQlRlTlVlXlYlqllllmmm"n#n.n5n;nAnGnHnMnRnXn^ndnenfnlnrnxn~nnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnooo oooooo!o'o-o3o4o5o8o>oDoJoKoLoQoWo]ocodoeohonotozo{o|o}ooCprq`r]s`uvvvvvvvvv4w5wwxxy6y}yyycz[|}~~~$%/15:@DEPU[afkpqrx~ƃ̃̓΃ԃكރ  &,16<BCL[`ekqw}~5{U,~É^iKc֍[K= bՓq oՙ000000000000000000000000000000000000000  00\ 0\ 0\ 0\ 0\ 0\ 0\ 0\ 0\   0 0110>0>0>0>0>0>0>0> 0110c!Proceedings Template - WORDEnd User Computing Services Normal.dot Yunbo CAO13Microsoft Office Word@%@H3 2@ji[B@H3 2q՜.+,0 hp|  ACMM Proceedings Template - WORD Title  FMicrosoft Office Word Document MSWordDocWord.Document.89q0c!0c!  0 0{&{&0&0&0&0& 0& 0& 0&0&0&0&0& 0{&{&0/0/0/  0040404040404 04403<03<03<03<03<03< 0440|>0|>0|>0|> 0440@0@0@0@0@0@0@0@0@0@0@0@0@ 0440F0F0F 0440I0I 0440xL0xL0xL0xL0xL0xL0xL0xL0xL0xL 0440S  00T 0TT0jV0jV0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV 0jV0jV0jV0jV0jV 0jV 0jV 0jV0jV 0jV 0jV 0jV0jV 0TT0.\0.\0.\0.\0.\0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\ 0.\0.\0.\0.\0.\0.\0.\0.\0.\0.\  0TT0d0d0d0d0d0d0d0d0d0d0d0d0d0d0d0d0d0d0d0d0dd0d0d@0d@0d@0d@0d@0d@0d@0d@0d @0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d@0d@0d@0d@0d@0d@0d @0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d `0d @0d @0d @0d @0d @0d @0d @0d A 0TT0n0n0n0n0n0n@0n@0n@0n@0n@0n@0n@0n@0n@0n@0n@0n0n0n0n0n0n0n 0TT0x0x0x0x 0TT0m|0m|0m|@0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|0m|0m|0m|0m|0m|0m|0m|0m| 0m|@0m|0m|0m|0m|0m| @ 0@0l@ 0l@ 0l@ 0l@0 @ 0@ 0@ 0@ 0@ 0@ 0@ 0@ 0@ 0@ 0@ 0 @ 0 @ 0 @ 0 @ 0 @ 0@ 0@ 0@ 0@ 0@ 0{00  """"""----------------------------------------0u  &(!$',/3{8>gA=EHtLPWbffgHghggh^iXj(r@rrrrsyuy|yyzk{|{{{{{{{{|||&|4|B|W|b|p||||ݏhyBCfgÑđ aHoϺUڻHRTVXZ^begknqw{Se+}++E@Y@[@WCoCrC\DpDrDDDDDDDFFF_FsFuFFFFgG{G}GNNNNNNNOO-OAOCOtPPPPPPԙ_:_::::::::::::: ")0!!!$  _2$H>G/AyA\Nb$Cy8VPz b$]`[q|Ub$G%ha< 42$H>G/AyA\N@/L 8  .(  d B+  j 3  s"*?` k  c $X99?B+  l  s * ?"6@ NNN?NB* z Z ' ,=UA m   )P n  <Cqr99qr99n "`:?,=U"/ C o  <Bqr99qr99o "`21 %4 B p  <Aqr99qr99p "`7f(A A`B q  c $DW5X(W5ZB r B S D7J(7 s  <=qr99qr99s "`' ,D/ = t  <<qr99qr99t "`>Y=Tl@ < u  <;qr99qr99u "`:?1=U5 ; v  <:qr99qr99v "`k(\233= : w  T9qr99qr99w  ?"6@`NNN?NlG7 M< 9B x 6D?"0@NNN?NT3.W?:B y  6D?"0@NNN?NU34>Y=B z 6D?"0@NNN?NT34W?:B { 6D?"0@NNN?NT37::?:B | <D?"0@NNN?NT3,W?4B } <D?"0@NNN?NT31W?4B ~  <D?"0@NNN?NT34W?7B   0D?"0@NNN?NU3:>@   <8qr99qr99 "` 2r< 8   <7K@K@ "`! L'  7   <6K@K@ "`_ .{  6   <5K@K@ "`'M "*i  5   <4K@K@ "``"7 %  4`B   c $DD ' ZB  S DD '    <3qr99qr99 "`M D i  3   <2K@K@ "`'* 2   61K@K@ "`4n)- 1   <0K@K@ "`-# 0`B   c $D'ZB  S D/#'   </K@K@ "`#))5 /   <.K@K@ "` .#w .ZB  B S D    <-K@K@ "`) E -bb   # , "`e% ,   B+qr99qr99 "` +`B   c $Df"f"eZB   S D f e   <(qr99qr99 "`%  (   <qr99qr99 "`&)    <qr99qr99 "`Tue    Bqr99qr99 "`IS  `B   c $Dg ZB  B S D(l )   <qr99qr99 "`~ " B   <D?"0@NNN?N"&   <qr99qr99 "`5c    Bqr99qr99 "`#]&@ r"   6ZGvH;JI`1!B(   <qr99qr99 "`#%& 6  &"-.  3  s"*?n   c $X99?"` &"-.   N*A8!A8!  ?"6@ NNN?N ! *   N)A8!A8!  ?"6@ NNN?N "$ )   NA8!A8!  ?"6@ NNN?N (+    TA8!A8!  ?"6@ NNN?N ]%`( B   s *D?"0@NNN?N] !Z$B  s *D?"0@NNN?N] !"$B  s *D?"0@NNN?N] !*   s * ?"6@ NNN?N :".   NOO  ?"6@ NNN?N!RR"    NOO  ?"6@ NNN?N!R"R%    NOO  ?"6@ NNN?N!)Rr,    TOO  ?"6@ NNN?N!%R) TB   C DO $!+B  TK@K@ D?"0@NNN?NO'!*   N OO  ?"6@ NNN?N2!R"     N OO  ?"6@ NNN?NR"2!%     N OO  ?"6@ NNN?N)2!r,     T OO  ?"6@ NNN?N%2!)  B S  ?  ~+pCԙj t t  _Ref62127459 _Ref62127429 _Ref62127355 _Ref62127395 _Ref62127533 _Ref62127298 _Ref62127302 _Ref62127303i֍bq ՙZaԓnՙEv/ v/# v/!v/4l"v/L)#v/#$v/O%v/l$ &v/ 'v/4w (v/ )v/< *v/D= +v/|,v/DL -v/ .v//v/ 0v/,£ 1v/l2v/D 3v/䒥 4v/V5v/ 6v/ 7v/ 8v/D!9v/ :v/#;v/ v/ ?v/" @v/t Av/ Bv/< Cv/d2 Dv/ Ev/!Fv/< Gv/"Hv/ԅ Iv/ Jv/ Kv/Dã Lv/'Mv/d Nv/$Ov/ Pv/<\ Qv/l& Rv/Sv/| Tv/Uv/}Vv/{ Wv/|\ Xv/$"Yv/4~Zv/ [v/$ \v/\ ]v/^v/T#_v/"`v/av/bv/$| U^^kw!!*FOO\hp n"===rrrrrrqvqvwvxxxyyy Nՙ      !#"$%'&()*+,-./012345768:9;<=>A?@BCD  ]guu~(//NXffow r"===rrrrrruvyvyvxxxyyyŏŏRՙ   !#"$%'&()*+,-./012345768:9;<=>@A?BCD >@*urn:schemas-microsoft-com:office:smarttags PostalCode:D*urn:schemas-microsoft-com:office:smarttagsStreet;E*urn:schemas-microsoft-com:office:smarttagsaddress9C*urn:schemas-microsoft-com:office:smarttagsplace8B*urn:schemas-microsoft-com:office:smarttagsCity=>*urn:schemas-microsoft-com:office:smarttags PlaceType=<*urn:schemas-microsoft-com:office:smarttags PlaceName9A*urn:schemas-microsoft-com:office:smarttagsStateB7*urn:schemas-microsoft-com:office:smarttagscountry-region>?*urn:schemas-microsoft-com:office:smarttags PersonName )EDCBA@?>C<<>BC7?BCA7?BC7?>C<<>BC7BCA7BCA7CBCABCABCABCACBACBABC<C<<>CCioa!j!{>}>>>>>>>>>>>>>??>?@?e@j@@@@@ AAlAoAD D=D@DDD5E;E[E]E`EcEEEEEEEGGIIIIJJ6J9JBJEJOOPPOQSQcQgQQQ+S/SSS``bbbbbbbbccqc{ccddd9dCdddddddy5yinnsǒΒ27IRӕܕ&'җԗՙ  ++===0>>>??_@d@hAlAtC}CDDEFGGXIIJJ2J6JLLzL~LLLLLMMNNPPPPSSdYlYZZy5y{{&'̗Η35ՙ333333333333333333333333333333333333U^g t <cn""''0155>=G=>>??E@_@AADDDDGGNNTTUUVW`WYYZjZZZ[[]4]eeeeeelRl#nouvvvyy~~%ЈÉމ^i K6cč֍2["K=}ސ [9by(qڔ ʖ&'EGPSWYfhǗɗޗ(*,.0268Kr  ?BFHLNacrt˙ՙ&'ՙ BrPcO"Z8xl$*i!jo@.@.@..@...@ ....@ .....@ ......@ .......@ ........hh^h`o(. H\H^H`\hH) \^`\hH. \^`\hH. 4\4^4`\hH)  \ ^ `\hH. | \| ^| `\hH.  \ ^ `\hH) \^`\hH.hh^h`OJPJQJ^Jo(nH\H^H`\OJQJo(hHn\^`\OJQJo(hHu\^`\OJQJo(hHl4\4^4`\OJQJo(hHn \ ^ `\OJQJo(hHu| \| ^| `\OJQJo(hHl \ ^ `\OJQJo(hHn\^`\OJQJo(hHuhh^h`o(. H\H^H`\hH) \^`\hH. \^`\hH. 4\4^4`\hH)  \ ^ `\hH. | \| ^| `\hH.  \ ^ `\hH) \^`\hH.hh^h`CJOJQJo([] !joxlPcO"!jo!jo b                         \yA<,eb|*Y 5-6 ) s;tv-/Y3{}?mBfo|!<)OcR z U N >? _L Y g Yq RR  x y  &* ;gX(rgZh^@(r *:0DMlm|)+<?gz "^Sb: $D|SPmOS1u?7S[`eB  : v fv }!gX"c"#S$o$"%X% '6)`)%*Z=+[+|, ---=7-_. .X.z//M/Q0]0 1 23,[34 5y(5~6j6t,7Q7y78u8{8<9S9V9 :(:.:R:f:9;WX;\;)?<|,>@u>B%?@N?@FAkA BSB?WB\B rB CCEC>LCSC'D9fDA ELEmME.NE$0F`5F8FsBF{F$GKG1&H=H0IZIIuIzIKZK CLZLokLqLM8M ;MzM M+NYN[N3lNpN1OO6OvOP-'PSP_PQ/Q:QkQCnQ}Qo5R+S=S"T?HT!XTRkTqUVVijllzlmnn5n8nrn#op(p)dp^npr)sggszsHtAttuwuZv:v; vLvavylvY#wjaw#*xSxyhgythyz$zZz {{_|C)}~~f~n~&t~&,R;B@1I U Bv!"CQy5G5Krw:''.^ ,ChBX`m5hD.O0C4=|Way#Idj+"&'Ge Fh !{{!S@d.lwSN fA 8dP\"k&rf g*7pm{"*/FW#(E5L>8=Od~)+2[xm!Z~X .,oITnx .d;C @FP@9$:ny7O^"Kr!ODDWyu`!&)BgCMJr 9h5eizToU=-(1'c<$\Qt> R\mUK5ti`DAED4J_BR"'^_ ~#+\3leY:;c  W4e}l 4O/]DmzLm"sG(}d $~3Iwm8 Y0]C[s -U/6~"Uv3R1prRTeZSc5![%)0a0>Ha os04cVd;/ C#W$*=kMd!(F -j4 zoGvDD7[7fS4'OH!.)>@6>'X3#7.M4og6gEXpbG,T mkX:#HX:fp5J]hs6TjxD L4FO2&0:Zl78VX[ f V ^< T!e+++++dYYYYYYYYYYYYYYZZ$Z%Z:Z?ZDZEZ^ZcZhZiZZZ[[[:\;\dddeeeeeeeuevexeze{e}eeeeeeeeeekkkkkl,lQlRlTlVlXlYlqllllmmm"n#n.n5n;nAnGnHnMnRnXn^ndnenfnlnrnxn~nnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnooo oooooo!o'o-o3o4o5o8o>oDoJoKoLoQoWo]ocodoeohonotozo{o|o}ovvvvvvvvv4w5w$%/15:@DEPU[afkpqrx~ƃ̃̓΃ԃكރ  &,16<BC[`ekqw}~ՙgn"@\\msrasia\AdobePS Print DriverNe01:AdobePS Acrobat DistillerAdobePS Acrobat Distiller\\msrasia\AdobePS Print DriverS 4dA4PRIV0''''\KhC\\msrasia\AdobePS Print DriverS 4dA4PRIV0''''\KhC#d>?PQRTؔԙP@PFP@PZP@P^P@Pp@UnknownHang Li ((MSR)) Gz Times New Roman5Symbol3& z Arial;& z Helvetica;[SOSimSun3z Times5& zaTahomaU MiriamTimes New Roman;Wingdings"h[ͅF,|f,|f ,qMqM4d2qHP?Bh2Proceedings Template - WORDEnd User Computing Services Yunbo CAO      CompObjjq