Effective Early Termination Techniques for Text Similarity Join Operator

Effective Early Termination Techniques for

Text Similarity Join Operator*

Selma Ayse ?zalp1, and ?zgür Ulusoy2

1Department of Industrial Engineering, Uludag University, 16059 Gorukle Bursa, Turkey

ayseozalp@https://www.sodocs.net/doc/bc5793553.html,.tr

https://www.sodocs.net/doc/bc5793553.html,.tr/~ayseozalp

2Department of Computer Engineering, Bilkent University, 06800 Bilkent Ankara, Turkey

oulusoy@https://www.sodocs.net/doc/bc5793553.html,.tr

Abstract. Text similarity join operator joins two relations if their join attributes are

textually similar to each other, and it has a variety of application domains including

integration and querying of data from heterogeneous resources; cleansing of data;

and mining of data. Although, the text similarity join operator is widely used, its

processing is expensive due to the huge number of similarity computations

performed. In this paper, we incorporate some short cut evaluation techniques from

the Information Retrieval domain, namely Harman, quit, continue, and maximal

similarity filter heuristics, into the previously proposed text similarity join

algorithms to reduce the amount of similarity computations needed during the join

operation. We experimentally evaluate the original and the heuristic based similarity

join algorithms using real data obtained from the DBLP Bibliography database, and

observe performance improvements with continue and maximal similarity filter

heuristics.

1 Introduction

The text similarity join operator, as its name implies, joins two relations if their join attributes, which consist of pure text, are highly similar to each other. The similarity between join attributes is determined by well-known techniques such as tf-idf weighting scheme [1] and cosine similarity measure from the Information Retrieval (IR) domain. The text similarity join operator has various application domains. Cohen [2], Gravano et al. [3], and Schallehn et al. [4] use this operator for the integration of data from distributed, heterogeneous databases that lack common formal object identifiers. For instance, in two Web databases listing research institutions, to determine whether the two names “AT&T Labs” and “AT&T Research Labs” denote the same institution or not, text similarity join operator may be employed.

Meng et al. [5] use the text similarity join operator to query a multidatabase system that contains local systems managing both structured data (e.g., relational database) and unstructured data (e.g., text). As an example let’s assume that we have two global relations: applicants containing information about job applicants and their resumes, and positions including the description of each job; then the text similarity join operator is *This research is supported by a joint grant from TüBITAK (grant no. 100U024) of Turkey and the National Science Foundation (grant INT-9912229) of the USA.

used to answer queries like “for each position, find k applicants whose resumes are most similar to the position’s description”. Jin et al. [6] employ similarity join operator for solving the problem of record-linkage in the context of data cleansing. In [7, 8], we describe similarity join operator to facilitate metadata based web querying [9].

We classify the text similarity join operators as top-k, threshold, and d irectional similarity join operators [8] such that the top-k similarity join takes two relations R and S, and an integer k as input, then joins tuple pairs from R and S according to the similarity of their textual join attributes, and returns k joined tuples having the highest similarity values. The threshold similarity join also takes two relations R and S, and a real threshold value in the range [0..1] as input, and joins tuples from R and S if the similarity of their textual join attributes is greater than or equal to the threshold value. The last similarity join operator, called directional similarity join, joins each tuple from relation R with k most similar tuples from relation S, and returns at most |R|*k joined tuples where |R| is the number of tuples in relation R. In this study, we focus on the directional similarity join operator, and we try to reduce the amount of similarity comparisons executed by employing some early termination heuristics (e.g., Harman, quit, continue, and maximal similarity filter) from the IR domain. These heuristics improve the performance of the join operation by considering only the tuple pairs that have high similarity to each other and ignoring the ones having small or no similarity. We also show through experimental evaluation that early termination heuristics improve the performance of the similarity join operator considerably in terms of the number of disk accesses made and the amount of similarity computations performed.

The rest of the paper is organized as follows. In the next section, we describe the related work. A brief summary of the previously proposed directional join algorithms and the early termination heuristics are presented in sections 3 and 4. In Section 5, we experimentally evaluate and compare all the algorithms in terms of the CPU time required for processing, the number of tuple comparisons and the number of disk accesses made. Finally, we conclude our discussion in Section 6.

2 Related Work

Recently, similarity join operator for both low and high dimensional data has become a popular research topic as it is used in variety of applications such as data integration, data cleansing, data mining, and querying. Different techniques have been used for the similarity join of low dimensional (e.g., text) and high dimensional data (e.g., multimedia, biological data). Among the text similarity join proposals, the works presented in [3, 4, 6] describe processing techniques for the threshold similarity join operator. In [2, 10, 11, 12, 13, 14], algorithms for the top-k similarity join operator are described.

Although numerous proposals exist for the threshold and the top-k similarity join operators, only Meng et al. [5] study the directional similarity join operator. They propose and experimentally evaluate three join algorithms namely, Horizontal-Horizontal Nested Loop (HHNL), Horizontal-Vertical Nested Loop (HVNL), and Vertical-Vertical Merge (VVM), which use the well-known similarity measure, tf-idf weighting scheme and cosine similarity measure for similarity comparisons. As the names of the algorithms imply, HHNL and HVNL algorithms are nested loops based join algorithms such that HHNL algorithm compares each document (tuple) pairs in the collections (relations), and HVNL

algorithm, on the other hand, uses the documents in one collection and the inverted file for the other collection to compute the similarities. Algorithm VVM, which is not nested loops based, uses inverted files on both collections to compute the similarities. The details of these similarity join algorithms and the early termination heuristics applied to these algorithms are given in the subsequent sections.

The similarity measure employed in [5] and also in this study is the cosine similarity measure with tf-idf weighting scheme [1] in which, each document (join attribute in the similarity join operator) is represented as a vector consisting of n components, n being the number of distinct terms (i.e., stemmed words) in the document collection, such that each component of a vector for a document gives the weight of the term i for that document. Weight of a term for a particular document is computed according to tf-idf value, where tf (term frequency) is the number of occurrences of term i within the document; and idf (inverse document frequency) gives more weight to scarce terms in the collection. The similarity measure is the cosine of the angle between the two document vectors such that the larger the cosine, the greater the similarity. Other measures such as Hamming distance, and longest commo n subsequence (LCS) for determining the similarity between short strings have also been developed. In [2, 3, 5] tf-idf weighting scheme and cosine similarity measure are preferred as the vector space model gives quite good matches even for short strings. Also, the vector space model allows the use of inverted indices, which makes possible for us to integrate some early termination heuristics from the IR domain during the similarity comparisons of tuples.

3 Directional Text Similarity Join Algorithms

The only study that has appeared in the literature for the directional similarity join operator were developed by Meng et al. [5] who presented three algorithms namely HHNL, HVNL, and VVM for the join operator. The HHNL (Horizontal-Horizontal Nested Loops) algorithm is a blind nested loops join algorithm, in which each tuple r in relation R is compared with every tuple in relation S, and k most similar tuples from S are joined with tuple r. In [5], the input relations R and S are read from disk. After reading X tuples from R into the main memory, the tuples in S are scanned; and while a tuple in S is in the memory, the similarity between this tuple and every tuple in R that is currently in the memory is computed. For each tuple r in R, the algorithm keeps track of only those tuples in S, which have been processed against r and have the k highest similarities with r. In the HHNL algorithm, and also in all other algorithms described in [5], a heap structure is used to find the smallest of the k-largest similarities.

The HVNL (Horizontal-Vertical Nested Loops) [5] algorithm is an adaptation of the ranked query evaluation techniques in the IR domain to the join operation. In an IR system, the aim is to find the k documents in the system which are most similar to the user query. For that purpose, most of the IR systems employ inverted files. In these systems, for each term t in the user query, the term is searched from the inverted index and the ids of documents containing term t are found. Then, the similarity calculations are performed only for those documents that have at least one common term with the user query. Algorithm HVNL is a straightforward extension of this method such that for each tuple r in R, the algorithm calculates the similarity of r to all tuples i n S having at least one common term with r, and selects the k most similar tuples from S. The advantage of

HVNL algorithm is that, it does not perform similarity calculations for all tuples in S as in the case of the HHNL algorithm. In the HVNL algorithm, the inverted file consists of (i) an inverted index which includes the index term (t), the number of tuples in S containing the index term (f t), and a pointer to its corresponding inverted list entry, and (ii) an inverted list which stores tuple id having the index term t, and the frequency of the term in that tuple (f s,t). In the HVNL algorithm the inverted index is stored in the memory, the inverted list entries, and the relations R and S are read from disk.

The algorithm VVM (Vertical-Vertical Merge) em ploys sorted inverted indices with respect to the index terms on both of the input relations R and S [5]. The VVM algorithm scans both inverted files on the input relations at the same time. During the scan of the inverted indices, if both index entries correspond to the same index term, then similarities are accumulated between all tuples in the inverted lists of the indices. The VVM algorithm assumes that, both inverted files as well as relations R and S are read from disk. In order to store intermediate similarities between every pair of tuples in the two relations, the algorithm needs |R|*|S| accumulators1 that are stored in main memory. The strength of the algorithm is that it scans the inverted files only once to compute similarities between every pair of tuples. However, the memory requirement for the accumulator is so large that it cannot be run for relations having large number of tuples. As an example, let’s assume that both relations R and S consist of 100,000 tuples, and each similarity value requires 4 bytes (size of float), so the memory allocated for the accumulator should be at least 100,000*100,000*4 bytes = 40Gb. In this study, we do not consider the algorithm VVM due to its huge memory requirement.

4 Heuristic Based Directional Similarity Join Algorithms

In the subsequent sections, we first describe early termination heuristics [15] from the IR domain that we use to improve the performance of directional similarity join operation, and then we briefly introduce directional similarity join algorithms employing these heuristics.

4.1 Harman Heuristic

Harman et al. [16] proposed a heuristic to decrease the number of similarity computations performed during the search of similar documents to a user query. We apply this heuristic to the HVNL algorithm as it employs an inverted index over the relation S. The HVNL algorithm extended with Harman heuristic is called HVNL-Harman in which, for each tuple r in relation R, weights of the terms in r are examined, and the inverted index is accessed only for these terms having a weight greater than the 1/3 of the highest weighted term in r. This heuristic is implemented by modifying the original HVNL algorithm as presented in Figure 1. The HVNL-Harman algorithm considers S tuples which have high weighted terms in tuple r, and does not perform similarity computations for other S tuples that do not contain high weighted terms.

1Accumulator is a set of real numbers (A

) each stores an accumulated similarity between tuples r and s.

rs

Fig. 1. HVNL algorithm with Harman heuristic (HVNL-Harman)

4.2 Quit and Continue Heuristics

Moffat et al. [17] also suggested to sort the terms in the user query with respect to their weights in descending order, and to access the inverted index with respect to this order. They place an a priori bound (i.e., accumulator bound) on the number of candidate documents that can be considered for the similarity calculation. New documents are compared until this bound is reached. The idea behind this heuristic is that, terms of high weight are permitted to contribute to the similarity computation, but terms of low weight are not. When the bound is reached; i n the quit approach, the cosine contribution of all unprocessed terms are ignored, and the accumulator contains only partial similarity values for documents. In the continue strategy, documents that do not have an accumulator are ignored, but documents for which accumulators have already been created continue to have their cosine contributions accumulated. When the processing ends, the computation of full cosine values for a subset of the documents becomes completed.

As the quit heuristic allows only the partial similarity computation, it is not suitable for the directional similarity join operator. To find top-k similar tuples for a given tuple r, we need to have full cosine values and thus, we use the continue heuristic w ith the HVNL algorithm (HVNL-Continue). In this variation of the HVNL algorithm, for each tuple r of R, only s tuples from S which have high weighted terms in r are considered for similarity computations until the accumulator bound on the number of tuples that can be considered for similarity computations is reached. When the accumulator bound is reached, the full cosine similarities between tuple r and s tuples become computed and the k-most similar tuples to r are selected. In the HVNL-Continue algorithm, we need document vectors (term weights) for tuples in relation S to compute the full cosine similarity values. Term weights for each s tuple can be computed prior to the join operation by just passing over the relation only once as a one time cost. For the implementation of the HVNL-Continue algorithm, we modify the 3rd line of the algorithm in Figure 1 as “for each term t in (sorted terms list of) r”. Also, in the 6th and 12th lines of the algorithm, for each tuple r in relation R, we increase the value of a counter variable by 1 each time a new s tuple is considered

for similarity computation, and when the value of the counter becomes equal to the predetermined accumulator bound, the for loop in line 3 is exited, and the counter is reset.

4.3 Maximal Similarity Filter Heuristic

“Maximal similarity filter” [7] is another technique that may be used to reduce the number of tuple comparisons made during the directional text similarity join operation. Let u s = be the term vector corresponding to the join attribute of tuple s of S, where u i represents the weight of the term i in the join attribute. Assume that the filter vector f R = is created such that each value w i is the maximum weight of the corresponding term i among all vectors of R. Then, if cos_sim (u s, f R) < V t then s can not be similar to any tuple r in R with similarity above V t. The value c os_sim (u s, f R) is called the maximal similarity of a record s in S to any other record r in R.

In the HVNL algorithm with maximal similarity filter (HVNL-Max-Filter), the inverted list entries are accessed with respect to descending order of maximal similarity values of s tuples. For each term t in tuple r of R, the inverted index is entered and the similarity comparisons are stopped at the point when the maximal similarity value (cos_sim (u s, f R)) for the tuple s is less than the smallest of the k-largest similarities computed so far for tuple r, since it is not possible for s to be in the top-k similar tuples list. The maximum weight of a term for a given relation is determined while creating the vectors for the tuples, and the filter vector for each relation may be formed as a one-time cost. To apply this heuristic, we need to sort the inverted list entries with respect to maximal similarity values of tuples just once during the preprocessing step. The HVNL-Max-Filter algorithm is also very similar to the HVNL-Harman algorithm (Figure 1). One difference is, the 3rd line of HVNL-Harman is changed as “for each term t in r”. Also, in lines 6 and 12, similarity computations for s tuples having term t are performed if the maximal similarity value for the tuple s is greater than the smallest of the k-largest similarities computed so far for tuple r, otherwise the for loop in line 3 is exited.

We also apply the maximal similarity filter heuristic to the HHNL algorithm (i.e., HHNL-Max-Filter), in which we sort the tuples in relation S in descending order of their maximal similarity filter values as a preprocessing step, and we terminate the inner loop when the maximal similarity filter for the s tuple that is being processed is less than the smallest of the k-largest similarities computed so far for the tuple r.

5 Experimental Results

We compared the performance of HHNL, HVNL, HVNL-Harman, HVNL-Continue, HVNL-Max-Filter, and HHNL-Max-Filter in terms of the number of tuple comparisons made, the number of disk accesses required, and the CPU time needed. For the experimentation, we implemented these algorithms in C programming language under MS WindowsXP operating system. We did not include VVM since it requires huge amount of memory to keep intermediate similarities between tuple pairs. In the implementation, the relations R and S are stored on disk and each block read from the relations contain 10000 tuples. For the HVNL and its variations, the inverted index is in-memory, however the inverted list entries are stored on disk and up to 5000 inverted list entries are kept in the cache. An inverted list entry that is not in the cache is retrieved fro m disk by making

random disk access, and when the cache is full, the entry for the term having the least term frequency is replaced with the new entry.

In the experiments, we used a real dataset that consists of the bibliographic information of journal and conference papers obtained from the DBLP Bibliography database [18]. In the implementation of the directional text similarity join, the relations R and S do not contain any common tuple, and the relation R consists of bibliographic information of approximately 91,000 journal papers, and the relation S contains bibliographic information of 132,000 conference papers. The paper title attribute is chosen as the join attribute, and for each journal paper r in relation R, we try to find k conference papers from relation S having the most similar titles to the title of r. We created the vectors and the maximal similarity filters for the join attribute of each tuple in the relations R and S, and the inverted index on relation S in advance as the preprocessing step. We assumed that we have enough main memory to store the inverted index and the accumulators used for similarity calculations. The experiments were performed on a PC having Pentium III 450 MHz CPU and 320 MB of main memory.

In Figure 2, the results in terms of the number of tuple comparisons (i.e., similarity computations) performed by the implemented algorithms for different k values are presented. As displayed in the figure, the HHNL algorithm needs to make around 12 billion comparisons for each different k values to join R and S, while all versions of the HVNL algorithm do less than 900 million tuple comparisons for the same join operation. HVNL, and all variations of the HVNL algorithm perform much better than the HHNL algorithm, because of the fact that these algorithms employ inverted index on the input relation S, and they compare similarity of tuples which are guaranteed to have a similarity value greater than 0. The HHNL algorithm, on the other hand, makes similarity computation for all tuple pairs regardless of whether the tuples contain any common term or not.

Fig. 2. Number of tuple comparisons for all algorithms vs. k values

The maximal similarity filter heuristic reduces the number of tuple comparison about 25% for both the HHNL and the HVNL algorithms. We use continue and Harman heuristics with the HVNL algorithm only, as these heuristics are applicable when an inverted index is employed. The continue heuristic, in which accumulator bound is set to 5000 tuples, provides more improvement on the performance of the HVNL algorithm by decreasing the number of tuple comparisons by 50%. The Harman heuristic, on the other

hand, does not improve the performance of the HVNL algorithm, because term weights for our input data are quite close to each other. Changing the value of k does not affect the number of tuple comparisons except for the maximal similarity filter heuristic. As the k value increases, maximum similarity filter heuristic needs to make more tuple comparisons to find top k similar tuples.

We also computed the number of disk accesses (Table 1) required by all algorithms when the relations R and S, and the inverted list entries on the join attribute of relation S are stored on disk. In the disk access computation, we ignored the number of di sk accesses made for writing the joined tuples to the disk. According to Table 1, the number of disk accesses performed by the HHNL algorithms, which is approximately 150 disk accesses, is quite less than those obtained with the HVNL algorithms since for each term t considered during the similarity comparisons, the HVNL based algorithms read inverted list entries of term t (i.e., I t) by making a disk access if it is not in the memory. According to the Table1, the continue heuristic reduces the number of dis k accesses of the HVNL algorithm by 50%. The Harman and maximal similarity filter heuristics, on the other hand, do not lead to any reduction on the number of disk accesses required. This result is due to the fact that, the term weights in our dataset are close to each other and the Harman heuristic considers almost all terms in a tuple r during the similarity computations. The maximal similarity filter heuristic on the other hand, needs to access all the inverted list entries for all terms in a tuple r to find the s tuples having high maximal similarity values. Therefore, the maximal similarity filter heuristic only reduces the number of tuple comparisons performed when the inverted list entries are sorted with respect to the maximal similarity value of tuples.

Table 1. Number of disk accesses performed by all the algorithms for all k values

k HHNL HHNL-

Max-Filter HVNL HVNL-

Harman

HVNL-

Continue

HVNL-

Max-Filter

5-25 150 150 26282 26198 14176 26282

Fig. 3. CPU time required by all algorithms for the directional similarity join

Although the number of disk accesses performed by the HHNL based algorithms is quite less, the number of tuple comparisons is considerably higher than the HVNL based algorithms. To determine which group of algorithms is more efficient, we measured the

CPU time required by all of the join algorithms for k=10 and reported the results in Figure 3. As presented in Figure 3, the CPU time required to execute the join operation is 11725 seconds for the HHNL algorithm, and 2810 seconds for the HVNL algorithm, which implies that similarity computations take much longer CPU time than making disk accesses for retrieving inverted list entries. The maximal similarity filter heuristic reduces the CPU time by 16% for the HHNL and 20% for the HVNL algorithms. The continue heuristic makes 35% reduction in the processing time when the accumulator bound is set to 5000 tuples. The Harman heuristic, on the other hand, does not provide any improvement since it also does not make any reduction in the number of tuple comparisons and disk accesses.

For the continue heuristic, the accumulator bound is an important factor on the performance of the join algorithm. To show the effect of the accumulator bound on the join operation, we run the HVNL-Continue algorithm with different accumulator bounds and present the results in Table 2. We observed that, as the accumulator bound is decreased, the number of tuple comparisons falls, due to the fact that, the accumulator bound is an upper bound on the number of tuples that can be considered for similarity comparisons. The number of tuple comparisons made remain the same for different k values.

Table 2. The effect of accumulator bound for the continue heuristic

Accumulator Bound # of Tuple

Comparisons

# of Disk

Accesses

CPU

Time (sec)

Accuracy

5,000 372,448,481 14,176 1854 65%

10,000 604,454,778 20,001 2595 84%

15,000 732,112,934 22,678 2801 91% We examined the accuracy of the output produced by the algorithms that employ early termination heuristics as follows: Accuracy= |B∩H| / |B|, where B denotes the actual output set generated by the HHNL or HVNL algorithm, H is the output generated by the algorithm that employ any one of the early termination heuristics, and | . | denotes the set cardinality. We observed that the Harman heuristic generates exactly the same output as the HHNL, and HVNL algorithms; the continue heuristic, on the other hand, could achieve 65% accuracy when the accumulator bound is set to 5000 tuples, and the accuracy can be improved up to 91% when the accumulator bound is increased to 15000 tuples. As the accumulator bound is an upper bound on the number of tuples that can be considered for the similarity comparisons, it highly affects the accuracy of the continue heuristic. For the m aximal similarity filter heuristic, we observed that the accuracy of this heuristic is 100%, as it calculates the similarity for s tuples having maximal similarity value greater than or equal to the smallest of the k largest similarities computed so far for tuple r. Therefore, the heuristic considers all s tuples that can be in the result set by eliminating the ones that are not possible to be in the result.

6 Conclusion

Similarity based text join is a very useful operator to be employed in a variety of applications. In this study, we incorporate some early termination heuristics from the

Information Retrieval domain to achieve performance improvement for the text similarity join algorithms. We have demonstrated through experimental evaluation that nested loops based similarity join algorithm performs the best in terms of the number of disk accesses required; however, it compares every tuple pairs from the relations to be joined and leads to a huge amount of expensive similarity computations. Inverted index based join algorithm, on the other hand, achieves very small number of similarity computations while requiring large number of disk accesses. When we compare the processing time of the algorithms, we have demonstrated that the index based algorithm is superior to the nested loops based one, and we have observed further performance improvement by applying the maximal similarity filter and the continue heuristics to the index based join algorithm.

References

1.Salton, G.: Automatic Text Processing. Addison-Wesley (1989).

2.Cohen, W.: Data Integration Using Similarity Joins and a Word-Based Information

Representation Language. ACM Trans. on Inf. Sys., Vol. 18, No. 3 (2000) 288-321.

3.Gravano, L., Ipeirotis, P. G., Koudas, N, Srivasta, D.: Text Joins in an RDBMS for Web Data

Integration. In Proc. of WWW2003 (2003).

4.Schallehn, E., Sattler, K. U., Saake, G.: Efficient Similarity-Based Operations for Data

Integration. Data & Knowledge Engineering, Vol. 48 (2004) 361–387.

5.Meng, W., Yu, C., Wang, W., Rishe, N.: Performance Analysis of Three Text-Join Algorithms.

IEEE Trans. on Knowledge and Data Eng., Vol. 10, No. 3 (1998) 477-492.

6.Jin, L., Li, C., Mehrotra, S.: Efficient Record Linkage in Large Data Sets. In Proc. of the 8th

Int. Conf. on Database Systems for Advanced Applications (DASFAA’03) (2003).

7.?zsoyoglu, G., Alting?vde, I. S., Al-Hamdani, A., ?zel, S. A., Ulusoy, ?., ?zsoyoglu, Z.M.:

Querying Web Metadata: Native Score Management and Text Support in Databases. ACM Trans. on Database Sys., Vol. 29, No. 4 (2004) 581-634.

8.?zel, S. A.: Metadata-Based and Personalized Web Querying. PhD Thesis, Dept. of Computer

Engineering, Bilkent University, Ankara (2004).

9.?zel, S. A., Alting?vde, I. S., Ulusoy, ?., ?zsoyoglu, G., ?zsoyoglu, Z. M.: Metadata-Based

Modeling of Information Resources on the Web. JASIST, Vol. 55, No. 2 (2004) 97-110.

10.Fagin, R., Lotem A, Naor, M.: Optimal Aggregation Algorithms for Middleware. In Proc. of

PODS 2001 (2001).

11.Bayardo, R. J., Miranker, D. P.: Processing Queries for First Few Answers. In Proc. of

Conference on Information and Knowledge Management (1996) 45-52.

12.Chang, K. C., Hwang, S.: Minimal Probing: Supporting Expensive Predicates for Top-k

Queries. In Proc. of SIGMOD 2002 (2002) 346-357.

13.Natsev, A, Chang, Y. C., Smith, J. R., Li, C. S., Vitter, J. S.: Supporting Incremental J oin

Queries on Ranked Inputs. In Proc. of VLDB 2001 (2001) 281-290.

14.Ilyas, I. F., Aref, W. G., Elmagarmid, A. K.: Supporting Top-k Join Queries in Relational

Databases. In Proc.of VLDB 2003 (2003).

15.Vo, A. N., Krester, O., Moffat, A.: Vector-Space Ranking with Effective Early Termination. In

Proc. of ACM SIGIR 2001 (2001) 35-42.

16.Harman, D. K., Candela, G.: Retrieving Records from a Gigabyte of Text on a Minicomputer

Using Statistical Ranking. JASIS, Vol. 41, No. 8 (1990) 581-589.

17.Moffat, A., Zobel, J.: Self Indexing Inverted Files for Fast Text Retrieval. ACM Trans. on Inf.

Sys., Vol. 14, No. 4 (1996) 349-379.

18.Ley, M.: DBLP Bibliography. At http://www.iformatik.uni-trier.de/~ley/db/ (2001).

员工离职倾向研究论文

*******大学 自考本科生毕业论文 题目：员工离职倾向研究 专业： 姓名： 准考证号： 联系电话：

2010年9月2日 摘要 现在的企业管理不仅仅是业务管理，还有人才的管理。无论是大、中小企业，他们都希望自己企业的规模越来越大，并想尽一切办法来到达目的。但在竞争日益激烈的当下，一个企业如果不能逐步的完善自己，则必定会被社会所淘汰。那么怎样完善自己来加强竞争优势成了一个很重要的问题。 大家都知道企事业竞争中很重要的一项是人才的竞争。任何一个企业它的正常运行都离不开他的员工，拥有了人才就等于拥有了创造力、拥有了更强更持久的生命力。一个公司越大，里面的各种类型的人才就越多，公司的发展前景就越有希望,因此很多企事业都想尽办法吸纳人才、留住人才。他们有的采用很多种方式来调动员工的积极性使员工努力的工作，但即便是这样，企业的人员流动还是很大，员工在自己的岗位上工作一段时间就会产生离职倾向的事时有发生。

一个普通的员工流失不可怕，但如果是一批呢？这其中损失的有形和无形的成本又有多大？所以怎样才能留住人才，怎样解决员工离职成了每个企事业必须关注的问题，在解决这些问题之前我们必须弄清楚员工的离职倾向。 员工离职研究是组织行为学中一个重要的研究领域，研究表明，离职倾向是离职行为的直接前因变量，通过考察影响员工离职倾向的因素，能够很好地预测员工的离职行为。本文对员工离职倾向的影响因素的研究进展进行了综合分析，在诸多的研究成果中，影响员工离职倾向主要因素有员工个人因素、企业因素、外部环境因素。

Abstract Now the enterprise management is not only the business management but also the personnel management. Both large ,medium-sized and small enterprises, they all hope their own’s enterprises’ scopes are bigger and bigger and try their best to reach the goals. But in the increasingly fierce competition, if an enterprise cannot gradually perfect itself,it must be eliminated by society. So how to improve themselves to strengthen competitive advantage has become a very important problem. Everyone knows talent competition is one of the important items of business competition. Any enterprise’s normal operation is inseparable from the staff.Getting talents is equal to having creativity, stronger and more enduring vitality.The bigger a company is, the greater the

后员工离职倾向研究

1 绪论 1.1研究背景 纵观现在的社会，集信息化、全球化、激烈的竞争于一体，面对这样的现状，90后们从小就接受综合型教育，一直坚持全面发展的目标，在这样的模式下，就造成了人才济济的局面，现在的本科生已经处处皆是，就连研究生、博士也已经随处可见。前几年的金融风暴，导致国内的很多企业倒闭、重组，人员频繁流动。也因为这些原因，促成了现今人才富足职位空缺的局面。这样的就业形势，无疑是严峻的。 从去年下半年开始，新一代的主力军——第一批90后大学生已经陆续步入了工作岗位。我也是一名90后，从去年的12月份起，我开始了自己的实习生活。我的实习单位是在位于榆林的神华陕西国华锦界能源有限责任公司。以下简称“国华电厂”。 神华陕西国华锦界能源有限责任公司是一个以发电为主，煤炭业为主的能源企业，属于国家单位。由于单位规模大、员工人数多，管理繁杂，因此，在2004年时，单位将其管理服务中心外包给了陕西四季假日饭店管理有限公司。管理服务中心包含了国华电厂的行政楼、生产楼、检修楼、板房的行政前台业务，职工餐厅、职工公寓、以及国华电厂的所有卫生打扫工作。我实习的岗位是行政楼的行政前台，主要职责是负责行政楼所有的会议服务、会议室调配、布置工作，以及外来人员的接待、文件打复印，总经理、部门经理办公室的卫生打扫及其资料整理工作。除此之外，就是听从经理的调遣，配合经理服务公司的员工，使其工作顺利进行。 实习了三个月，在我的身边，相仿年龄的同事们频繁离职，有的虽然在职，但流露着鲜明的离职倾向。实习三个月结束后，回学校，准备论文，跟同学交流期间，了解到，同学中很多在实习后再说以后的打算，有很多都是对现有的工作犹豫不决，言谈举止间透漏着离职重新再做打算的想法。现在的就业形势这么严峻，一份工作来之不易，为什么在得到来之不易的工作后，实习一段时间后大家会有这么强烈的离职倾向？这是由什么原因导致的？ 对于企业来说，花费人力、财力，为企业招进一名员工，然后花着代价培养这名员工，目的只有一个，将这个员工培养成企业所需要的人才，为企业赢得利益！而像这种，在花费了成本之后，留不住人才，这无疑对企业来说是一大损失。对刚刚步入职业生涯的90后来说，挣着一份工资，还能有一个让自己成长，将理

企业知识型员工离职倾向的实证研究

华东交通大学 硕士学位论文 企业知识型员工离职倾向的实证研究 姓名：张树 申请学位级别：硕士 专业：企业管理 指导教师：万华 20071107

企业知识型员工离职倾向的实证研究 摘要 进入21世纪以来，知识经济占据了主导地位，企业员工的离职，尤其是企业中的知识型员工的高离职率不仅增大了企业的人力资源投入成本，而且损害了企业形象，造成企业内部其他员工的士气低落。这些现象为管理理论研究和实践提出了新课题。本文的研究正是在这个大背景下对能够显著预测知识型员工实际离职行为的离职倾向这个变量所进行的实证研究，以期找出显著影响知识型员工离职倾向的因素，以便采取相应有效的人力资源管理措施，留住企业需要的优秀人才，降低企业知识型员工离职为企业带来的损失，保持企业知识型员工的合理流动。 本文在对国内外已有的离职研究成果和企业知识型员工离职特点进行分析的基础上，提出了自己的研究构想。通过SPSS软件对问卷调查数据进行分析以验证自己的构想，并找到了能够显著预测知识型员工的离职倾向、工作满意度和组织承诺度的变量。 研究表明，人口统计学因素中，不同年龄和学历的知识型员工在离职倾向的表现上有显著的差异；在非工作因素中，除知识型员工可选择的工作机会未与其工作满意度和组织承诺度表现出显著的相关性外，其他各因素均与工作满意度、组织承诺度和离职倾向表现出显著的相关性；工作因素中，企业发展前景、组织公平性、晋升机会、薪酬水平、社会支持、职业成长度均与工作满意度、组织承诺度和离职倾向表现出显著的相关性，但是工作压力只与组织承诺度表现出显著的相关性，工作环境未与工作满意度、组织承诺度和离职倾向表现出显著的相关性。知识型员工的工作满意度和组织承诺度可以有效的预测其离职倾向；知识型员工的工作满意度水平可由工作投入度、薪酬水平、社会支持、职业成长度、工作压力五个变量来有效预测；知识型员工的组织承诺水平可由工作投入度、组织公平性、晋升机会、职业成长度四个变量来有效的预测，并且工作满意度和组织承诺度作为离职倾向的中介变量也得到了验证。 根据对知识型员工的离职倾向影响因素的分析，本文有针对性地提出了知识型员工的保持策略，对企业留住核心员工，保持竞争优势等管理实践具有一定的指导意义。 关键词知识型员工离职倾向实证研究

关于离职倾向调研报告

关于离职倾向调研报告 人员流动是企业发展过程中的必然现象，适当的离职率对企业的发展和保持活力是有益的，但过低或过高都会影响企业的发展越来越多的企业意识到人才的重要性，不仅更重视对员工忠诚度的塑造，同时也通过种种措施来保证核心员工的留驻，离职倾向调研报告。然而，总会有人因对企业现状不满而选择离开，人员流动是企业发展过程中的必然现象，适当的离职率对企业的发展和保持活力是有益的，但过低或过高都会影响企业的发展。 一、调查目的、对象及方法 1、调查目的 为了更加清楚全面的的了解员工离职的原因，从而找到合适的解决方法提高公司员工的满意度和幸福度，使他们为公司创造更大的价值。改进公司管理模式使公司管理更加科学和人性化，实现公司可持续性的和-谐的发展。 2、调查对象 通过对本企业近4年的离职记录进行统计及进行发放问卷的方 式进行调查。本次调查共统计离职记录190份，发放问卷200份，收回问卷200份，问卷有效率100%。完全符合抽样调查的原则，该样本具有代表性。 3 调查方法 主要采用问卷调查法、数理统计法、文献资料法等研究方法。在

查阅国内外有关人力资源文献资料基础上，对问卷进行认真设计，对回收的调查问卷运用统计学软件进行分析，并根据调查统计结果提出相关对策。 对于企业员工来说，离职相对较小，从时间、成本、组织工作量来说很方便。再而普查的结果更为准确。 二、调查内容 1、员工离职原因 从调查结果看，员工离职原因主要可分为薪酬水平原因（60%）；薪酬的公平性或激励性原因（32.1%）、个人身体、志向或家庭原因（25%）、缺乏足够的职业发展机会（34%）、个人价值得不到体现（27%）、难以承受工作压力（25%）、个人性格与工作内容不相适应（10%）、团队氛围原因（20%）、办公环境原因（30%）、行业性质或政策原因（5%）。 在通货高度膨胀，CPI不断攀高的情况下，薪酬水平无疑成为员工离职的首要原因，占60%，薪酬的公平性或激励性原因占32%；同 时越来越多的人开始根据自身情况，对自己的职业生涯进行重新规划，其中选择缺乏足够的职业发展机会而离职的占34%，辞职报告《离职倾向调研报告》。 2、员工离职时间 调查结果显示，员工的离职高峰期集中在第一季度，占总离职人数的53%，春节后成为员工跳槽的高发期，其后的三个季度中，离职率逐季下降。这成为一种普遍规律，不同岗位间差异性不大。 3、离职员工工作年限

知识型员工离职倾向原因分析

摘要 现在的企业管理不仅仅是业务管理，还有人才的管理。无论是大、中小企业，他们都希望自己企业的规模越来越大，并想尽一切办法来到达目的。但在竞争日益激烈的当下，一个企业如果不能逐步的完善自己，则必定会被社会所淘汰。那么怎样完善自己来加强竞争优势成了一个很重要的问题。 大家都知道企事业竞争中很重要的一项是人才的竞争。任何一个企业它的正常运行都离不开他的员工，拥有了人才就等于拥有了创造力、拥有了更强更持久的生命力。一个公司越大，里面的各种类型的人才就越多，公司的发展前景就越有希望,因此很多企事业都想尽办法吸纳人才、留住人才。他们有的采用很多种方式来调动员工的积极性使员工努力的工作，但即便是这样，企业的人员流动还是很大，员工在自己的岗位上工作一段时间就会产生离职倾向的事时有发生。 一个普通的员工流失不可怕，但如果是一批呢？这其中损失的有形和无形的成本又有多大？所以怎样才能留住人才，怎样解决员工离职成了每个企事业必须关注的问题，在解决这些问题之前我们必须弄清楚员工的离职倾向。 员工离职研究是组织行为学中一个重要的研究领域，研究表明，离职倾向是离职行为的直接前因变量，通过考察影响员工离职倾向的因素，能够很好地预测员工的离职行为。本文对员工离职倾向的影响因素的研究进展进行了综合分析，在诸多的研究成果中，影响员工离职倾向主要因素有员工个人因素、企业因素、外部环境因素。

Abstract Now the enterprise management is not only the business management but also the personnel management. Both large ,medium-sized and small enterprises, they all hope their own’s enterprises’scopes are bigger and bigger and try their best to reach the goals. But in the increasingly fierce competition, if an enterprise cannot gradually perfect itself,it must be eliminated by society. So how to improve themselves to strengthen competitive advantage has become a very important problem. Everyone knows talent competition is one of the important items of business competition. Any enterprise’s normal operation is inseparable from the staff.Getting talents is equal to having creativity, stronger and more enduring vitality.The bigger a company is, the greater the various types of talents inside,the more hopeful the prospect of a company. So many enterprises are doing their best to recruit and retain talents. Some of them have adopted a variety of ways to arouse enthusiasm of employees so that they would work hard.Even if it is like that, enterprises’turnovers are still very big. The things that employees will have a tendency to quit their jobs when they work for a period of time in their positions often happen,. The loss of a common employee is not terrible, but if in batches ? How great the loss of tangible and intangible costs are? So how to retain talented persons and how to solve employees’ resignament are plroblems that enterprises must pay attention to.Before solving these problems,enterprises must make clear of the employees’resignation tendency. Employee turnover is an important research field in organizational behavior research. Research shows that resignation tendency is the directly antecedent influence factors of resignation behavior,Investigating the employee turnover tendency is able to predict the employee turnoverf behavior.The text has comprehensively analyse the factors affecting the employee turnover tendency.Among the various research achievements, the main affecting factors of employee turnover tendency have individual factors, the enterprises and the external environment factors.

离职倾向调研报告范文

离职倾向调研报告范文 人员流动是企业发展过程中的必然现象，适当的离职率对企业的发展和保持活力是有益的，但过低或过高都会影响企业的发展越来越多的企业意识到人才的重要性，不仅更重视对员工忠诚度的塑造，同时也通过种种措施来保证核心员工的留驻，离职倾向调研报告。然而，总会有人因对企业现状不满而选择离开，人员流动是企业发展过程中的必然现象，适当的离职率对企业的发展和保持活力是有益的，但过低或过高都会影响企业的发展。 一、调查目的、对象及方法 1、调查目的 为了更加清楚全面的的了解员工离职的原因，从而找到合适的解决方法提高公司员工的满意度和幸福度，使他们为公司创造更大的价值。改进公司管理模式使公司管理更加科学和人性化，实现公司可持续性的和谐的发展。 2、调查对象 通过对本企业近4年的离职记录进行统计及进行发放问卷的方式进行调查。本次调查共统计离职记录190份，发放问卷200份，收回问卷200份，问卷有效率100%。完全符合抽样调查的原则，该样本具有代表性。 3 调查方法 主要采用问卷调查法、数理统计法、文献资料法等研究

方法。在查阅国内外有关人力文献资料基础上，对问卷进行认真设计，对回收的调查问卷运用统计学软件进行分析，并根据调查统计结果提出相关对策。 对于企业员工来说，离职相对较小，从时间、成本、组织工作量来说很方便。再而普查的结果更为准确。 二、调查内容 1、员工离职原因 从调查结果看，员工离职原因主要可分为薪酬水平原因（60%）；薪酬的公平性或激励性原因（%）、个人身体、志向或家庭原因（25%）、缺乏足够的职业发展机会（34%）、个人价值得不到体现（27%）、难以承受工作压力（25%）、个人性格与工作内容不相适应（10%）、团队氛围原因（20%）、办公环境原因（30%）、行业性质或政策原因（5%）。 在通货高度膨胀，CPI不断攀高的情况下，薪酬水平无疑成为员工离职的首要原因，占60%，薪酬的公平性或激励性原因占32%；同时越来越多的人开始根据自身情况，对自己的职业生涯进行重新规划，其中选择缺乏足够的职业发展机会而离职的占34%，辞职报告《离职倾向调研报告》。 2、员工离职时间 调查结果显示，员工的离职高峰期集中在第一季度，占总离职人数的53%，春节后成为员工跳槽的高发期，其后的三个季度中，离职率逐季下降。这成为一种普遍规律，不同

员工离职倾向影响因素研究

员工离职倾向影响因素研究 ——以“80后”离职员工为例 2012-12-17 10:50:27 来源：《技术经济与管理研究》2012年第3期摘要：研究采用问卷调查法和单变量方差分析法探讨了“80后”员工的离职倾向在个体和组织因素上的差异。研究发现“80后”员工的离职倾向在性别、年龄、目前组织工作时间、岗位性质和组织类型等因素上差异不显著；在受教育程度、工作年限和组织规模等因素上差异显著。由此得出，“80后”员工在离职倾向方面具有较强的一致性，他们的职业主体意识较强，具有职业化取代企业化的倾向。在管理实践中，由于本科比专科和研究生学历的“80后”员工有更强的离职倾向，组织需要更加关注本科学历“80后”员工的离职情况；由于小规模组织中的“80后”员工与较大规模组织中的员工相比，有着较高的离职倾向，小规模组织需要更为关注“80后”员工的离职问题。 关键词：离职倾向，教育程度，组织承诺，工作满意度 一、引言 “80后”员工是指出生于20世纪80年代(1980-1989年)的职场人员，他们正逐渐成为承载现代知识和技能的主力军，成为企业未来价值的主要创造者和获取竞争优势的核心资源。同时，“80后”员工工作环境的流动性、工作态度的独立性和职业多样性有日渐增强的趋势。频繁跳槽、缺乏忠诚是对“80后”员工职场表现的主要批评，“80后”员工的流失成为组织面临的严峻而紧迫的现实问题。如何让“80后”员工持续、稳定地为组织服务成为了组织管理者关注的新课题。

员工的离职倾向是指员工离开组织的心理倾向。一般认为离职倾向是员工一系列撤退认知的最后一个阶段，最能预测员工离职行为的发生(Telt &Mayer，1993)。探讨“80后”员工的离职倾向规律，并依此制定相应管理对策能使组织防止“80后”员工的过度流失，从而达到节约成本、发挥资源优势和创造更大价值的目的。 本研究旨在探讨“80后”员工的离职倾向在个体和组织因素上的差异，假设“80后”员工的离职倾向在个体和组织因素上差异不显著。 二、研究方法 1.被试和施测 研究采用问卷调查法。问卷以电子形式放于网络，问卷网址是 https://www.sodocs.net/doc/bc5793553.html,survey334995。在全国范围内联系“80后”在职员工，邀请他们在网上作答。共回收问卷273份，其中有效问卷259份，有效率为94.87%。 2.测量工具 采用自编自制的《离职倾向量表》共有3个问题。问题分别是：①经常有辞职的念头；②最近可能会试着去寻找新的合适的工作机会；③时常想去别的单位工作。此量表采用Likert5级量表的形式，从“非常不同意”到“非常同意”，分别用1～5分表示。

知识员工离职倾向关键性影响因素分析

知识员工离职倾向关键性影响因素分析* ———以科研事业单位为例 徐荣，曹安照 （安徽工程科技学院，安徽芜湖241000） 摘要：了解影响离职倾向关键性因素对稳定知识员工队伍具有积极的意义。本文以相关文献为基础，找出引起知识员工离职倾向的相关因素；从实证调查入手，通过GM (1，N)建模，对影响知识员工离职倾向的影响因素进行了灰色处理，得出了影响知识员工离职倾向的四个关键性影响因素：工作环境、可选择工作机会的主观感知、福利报酬、情感性承诺。 关键词：知识员工；离职倾向；灰色关联分析；工作满意度；组织承诺 中图分类号：F272.92文献标识码：A 0 引言 知道员工的离职倾向是有意义的，因为离职倾向通常被认为是预测员工离职行为最有效的指标(Carsten & Spector，1987)[1]，知道哪些因素最能影响员工离职倾向则更有意义。因为从管理的角度看，对某一项具体工作而言，把握其中4~5项关键指标是最有执行力的[2]。但目前的相关研究对知识员工离职倾向形成的影响因素作出关键性判断的较为鲜见，鉴于此，本研究拟从众多影响离职倾向的因素中提炼关键性因素，为相关组织针对知识员工离职行为进行管理提供依据。 1 影响因素选择 影响员工离职倾向的因素很多，但是各影响因素对于离职倾向的解释力是不同的，从文献回顾看，工作满意度、组织承诺、可选择工作机会的主观感知这三类因素对于员工离职倾向具有较强的解释力(Hill & Danier，1990；Frankas &Tetrick，1989；赵西萍，刘玲，张长征；2003)[3-5]。 工作满意度是员工对其工作或工作经历评估的一种态度的反映。Aronld & Feldman在1982年提出，工作满意度量表包括工作本身、上司、经济报酬、升迁、工作环境和工作团体等[6]。由于知识员工工作的特殊性，其工作满意度的结构和其它员工的工作满意度结构应该是有差异的，只是目前这方面的研究较少。本文参考Aronld等人量表基础上，对调查对象进行了访谈，了解知识自身的需求及现状，并且从管理者的角度了解管理中的困惑，以知识员工自身的特点出发来设计问卷，经过这样的过程，将量表修正为工作环境、晋升机会、*安徽省教育厅人才专项重点资助项目（项目编号：2009SQR080）

最新员工离职前的表现及预防

员工离职前的表现及 预防

员工离职前的表现及预防 一、主要用途 使用《员工离职前的表现》检测表，找到哪些员工可能有离职的倾向；利用《导致员工离职的主要因素》清单，分析造成员工有离职倾向的主要因素是什么；运用《预防员工离职的方法》清单，采取积极的应对措施，有效地做好员工离职的预防工作。 二、检测表及清单 员工离职前的表现

注：如果上述现象在该员工身上发生较多，您需要对此员工加以关注了。 三、导致员工离职的主要因素 企业优秀的人才产生离职倾向以至选择离开，一般可能有以下几种原因： 一、个人因素及工作本身的压力 ●因家庭原因造成的：如经济压力、家庭矛盾等，对员工的工作产生影响。 ●工作性质不适合个人的兴趣与能力，分权不当，大材小用或小材大用。 ●工作环境不利于工作顺利开展，或者直接影响个人的身体健康。 ●角色模糊，岗位职责不清，员工不清楚自己的定位的目标。 ●被要求去做很多的事（有些事不是自己的工作范畴），又得不到足够的时 间，使用员工产生过度负荷感。 ●不向员工清楚地沟通目标及决定，以致有时候当员工完成一项工作时，却遭 到企业的拒绝，严重打击士气。 二、企业管理因素 ●企业发展前景不清晰，不断重整部门、转调员工职务，以及改变企业走向， 让员工没有安全感。 ●企业没有清晰的薪资体系、绩效考核标准或薪资体系、绩效考核标准不合 理，造成分配不公。

●承诺员工的绩效和薪资福利待遇没有及时兑现。 ●当员工可以在其他企业找到薪水高二至三成的工作时，企业却宣布暂时不给 员工加薪或升迁。 ●领导处事不公，对某些员工偏心，给予他们较好的办公室，或者较舒适的出 差机会等。 ●领导管理不讲究方法，对于工作绩效不佳的员工不加以指导，却只在其犯错 误的时候加以指责或批评；或者批评员工时不就事论事，而涉及到员工的性格及其他问题，引起员工的反感情绪。 ●企业不赋予员工做决定的权力。 ●企业升迁没有能力的员工，真正的人才却没有受到重视。 三、企业文化问题 ●企业文化让员工感觉自己难以融入：如沟通、协调不畅；工作职责、流程不 清；规章制度不健全等。 ●企业虽然强调团队合作，但是却塑造部门或员工间必须激烈竞争的环境。 ●企业人际关系复杂，团队人心涣散、失和，合作不愉快。 四、心理因素 ●每天重复同样的工作，没有新鲜感，感觉原来的东西被掏空，学不到新的东 西，对工作提不起兴趣，没有成就感。 ●员工无法应付持续巨大的工作压力对个人能量和资源的过度要求，工作过度 劳累造成脑力、体力的透支，对人身心健康的冲击。 ●在公司受到挫折，或所做的工作不被认可，导致无法得到自我实现。

关于离职倾向调研报告

竭诚为您提供优质的服务，优质的文档，谢谢阅读/双击去除 关于离职倾向调研报告 人员流动是企业发展过程中的必然现象，适当的离职率对企业的发展和保持活力是有益的，但过低或过高都会影响企业的发展越来越多的企业意识到人才的重要性，不仅更重视对员工忠诚度的塑造，同时也通过种种措施来保证核心员工的留驻，离职倾向调研报告。然而，总会有人因对企业现状不满而选择离开，人员流动是企业发展过程中的必然现象，适当的离职率对企业的发展和保持活力是有益的，但过低或过高都会影响企业的发展。 一、调查目的、对象及方法 1、调查目的 为了更加清楚全面的的了解员工离职的原因，从而找到合适的解决方法提高公司员工的满意度和幸福度，使他们为

公司创造更大的价值。改进公司管理模式使公司管理更加科学和人性化，实现公司可持续性的和-谐的发展。 2、调查对象 通过对本企业近4年的离职记录进行统计及进行发放问卷的方式进行调查。本次调查共统计离职记录190份，发放问卷200份，收回问卷200份，问卷有效率100%。完全符合抽样调查的原则，该样本具有代表性。 3调查方法 主要采用问卷调查法、数理统计法、文献资料法等研究方法。在查阅国内外有关人力资源文献资料基础上，对问卷进行认真设计，对回收的调查问卷运用统计学软件进行分析，并根据调查统计结果提出相关对策。 对于企业员工来说，离职相对较小，从时间、成本、组织工作量来说很方便。再而普查的结果更为准确。 二、调查内容

1、员工离职原因 从调查结果看，员工离职原因主要可分为薪酬水平原因（60%）；薪酬的公平性或激励性原因（32.1%）、个人身体、志向或家庭原因（25%）、缺乏足够的职业发展机会（34%）、个人价值得不到体现（27%）、难以承受工作压力（25%）、个人性格与工作内容不相适应（10%）、团队氛围原因（20%）、办公环境原因（30%）、行业性质或政策原因（5%）。 在通货高度膨胀，cpI不断攀高的情况下，薪酬水平无疑成为员工离职的首要原因，占60%，薪酬的公平性或激励性原因占32%；同时越来越多的人开始根据自身情况，对自己的职业生涯进行重新规划，其中选择缺乏足够的职业发展机会而离职的占34%，辞职报告《离职倾向调研报告》。 2、员工离职时间 调查结果显示，员工的离职高峰期集中在第一季度，占总离职人数的53%，春节后成为员工跳槽的高发期，其后的三个季度中，离职率逐季下降。这成为一种普遍规律，不同岗位间差异性不大。

探讨企业员工满意度与离职倾向的关系

：员工工作满意度，不仅直接关系着企业的生产效率、顾客感知和利润水平，而且还能有效地反映组织绩效和企业管理中存在的问题，帮助企业进行自我诊断。本研究通过问卷调查，以S企业为例，探讨企业员工满意度与离职倾向的关系，结合实际提出提升员工满意度的建议。 关键词：工作满意度离职倾向建议 随着社会的进步，人才竞争变得日益激烈，为了吸引、稳住人才，各企业家都将目光重点投向自己的员工，更加关注员工的需要和感受。哈佛大学的一项调查研究表明：员工满意度每提高3个百分点，企业的顾客满意度提高5个百分点；员工满意度达到80％的公司，平均利润率增长要比同行业其它公司高出20％左右。通过员工满意度调查，可以明确企业在管理中存在的问题，根据发现的问题，提出针对性的解决方案，在实践中予以解决和改进，比如监控企业绩效管理的成效，及时预知企业人员的流动意向。 工作满意度的概念最早由霍波克（Hoppock）在其著作《Job Satisfaction》中提出，工作满意度是工作者心理和生理两方面对环境因素的满足感受，即工作者对工作情境的一种主观反应。回顾以往的研究，工作满意度的含义可分为三类：第一类，综合性定义；第二类，与期望差距的定义；第三类，参考架构性定义。本研究将结合综合性定义来测量整体工作满意度，即员工工作满意度是指员工对工作本身及工作相关各个方面的一种态度和看法，包括对企业、工作本身、工作回报、管理者、同事关系和工作环境等方面的满意程度。 Porter 和Steers提出离职倾向是当员工经历了不满意以后的一个退缩行为。离职倾向是预测离职行为的重要方法。相关研究发现，员工满意感对员工的缺勤、怠工、离职以及绩效等都有一定的影响和预测作用。了解员工的离职倾向，针对高离职倾向的员工，加强沟通与辅导，改善管理，减少不必要的员工离职，减少组织的损失。 一、问卷设计与调查 1、问卷设计 问卷的第一部分是个人的基本情况，包括性别、工龄、年龄、学历等内容。第二部分是员工工作满意度问卷，借鉴了国内外相关研究[3]，结合我国和S公司实际情况，提出调查的七个维度，分别是：①对企业的满意，包括对公司发展前景、管理制度等的满意度；②对领导的满意度，包括对不同层面管理工作的满意度等；③对工作回报的满意度，包括对绩效管理、报酬、福利、培训与晋升等的满意度；④对同事关系的满意度，包括人际沟通等的满意度；⑤对工作本身的满意度，包括工作胜任感、成就感等；⑥对工作环境的满意度，包括工作环境，工作气氛等；⑦对生活的满意度，包括生活娱乐设施等的满意度。第三部分，离职倾向。问卷采用5级评分法。 2、样本情况 本次调查的S制药企业，是国家“863”计划和国家“十一五”科技支撑计划等国家重点科技计划项目承担单位，公司设备先进，崇尚科学的管理方法。为确保问卷调查的可靠性，问卷由经过培训的第三方——心理学研究生，统一发放，当场填写收回。共发放调查问卷220份，收回196份，剔除无效问卷14份，共获得有效问卷182份，有效回收率为82.7%，其中男90人，女92人；生产部门120人，管理部门62人；29岁及以下104人，30岁至39岁51人，40岁及以上27人；学历：初中18人，高中或中专82人，大专及以上82人；工龄在5年以下有31人，6至10年95人，11年及以上56人。 3、方法 本研究选用的统计方法有相关分析、回归分析、方差分析等。首先，对自编问卷的信效度检验；其次，了解S企业员工工作满意度水平，讨论员工满意度各维度与离职倾向的关系；最后，比较不同人口统计特征的员工满意度差异。数据分析使用SPSS11.5软件。

员工离职倾向影响因素研究下

员工离职倾向影响因素研究——以“80后”离职员工为例（下） 2012年12月17日10:39 来源：《技术经济与管理研究》2012年第3期作者：字号 打印纠错分享推荐浏览量 四、讨论 通过对“80后”员工离职倾向在个体和组织因素上的差异分析，我们认为： 1.“80后”员工在离职倾向方面具有较强的一致性。在本研究中，“80后”员工的离职倾向除受教育程度、工作年限和组织规模等少数因素显著影响以外，不受其他因素，包括性别、年龄、目前组织工作时间、岗位性质和组织类型等个人和组织因素的影响。因此，我们在员工离职研究方面，基本上可以对“80后”员工做整体研究，不需要进行分类研究。在管理实践当中，组织在对“80后”员工进行离职方面的管理时，也基本上可以面向组织中所有的“80后”员工进行统一管理。 2.“80后”员工的职业主体意识较强。从“80后”员工的离职倾向不受性别、年龄和目前组织工作时间等因素的影响来看，他们在职业选择方面具有较强的自主意识，性别、年龄等传统制约职业发展因素对他们的影响较小。因此，组织要充分尊重女性员工，为她们提供与男性员工同等的工作条件和工作机会。同时，组织对年龄偏大和在组织工作时间相对较长的“80后”员工在离职方面要给予足够的重视，他们的流失会对组织产生更大的威胁和损失。 3.“80后”员工具有职业化取代企业化倾向。从“80后”员工的离职倾向不受岗位性质、组织类型等因素的影响来看，他们更多地显现出对自身职业的忠诚，而非对企业的忠诚。相对于企业的发展，他们更注重自身的职业特长和喜好，更关注自身的发展，具有较清晰、明确的职业定位和职业发展规划。因此，组织对“80后”员工进行职业管理时要有双赢的理念，做到充分了解和尊重他们的职业发展目标，在组织目标和他们的职业发展目标之间寻找共同点、建立一致性，并为他们在组织中实现职业目标提供支持和帮助。 4.组织需要更加关注本科学历“80后”员工的离职问题。由于本科比专科和研究生学历的“80后”员工有更强的离职倾向，一方面组织在能力和学历之间，需要更加重视能力，为本科学历“80后”员工提供更大的上升空间；另一方面需要重视培训，为本科学历“80后”员工提供进修和培训机会以胜任更高职位。 5.组织需要更加关注参加工作1～3年的“80后”员工的离职问题。由于有1～3年工作经历的员工有着相对较高的离职倾向，组织需要与他们保持畅通、充分的沟通，努力加强他们对组织的情感承诺，并随着他们能力和绩效的增长，及时的给予加薪和晋升。 6.规模小的组织需要更为关注“80后”员工的离职问题。由于小规模组织中的“80后”员工的离职倾向较强，因此为留住员工，它们需要充分发挥小规模组织自身优势，如提高综合素质的岗位较多，能与员工做充分的面对面沟通，能为员工提供较多的晋升机会等。 五、结论 “80后”员工的个体因素中，性别、年龄、目前组织工作时间和岗位性质等个体因素方面对其离职倾

(完整版)90后员工离职倾向研究毕业设计

90后员工离职倾向研究 1 绪论 1.1研究背景 纵观现在的社会，集信息化、全球化、激烈的竞争于一体，面对这样的现状，90后们从小就接受综合型教育，一直坚持全面发展的目标，在这样的模式下，就造成了人才济济的局面，现在的本科生已经处处皆是，就连研究生、博士也已经随处可见。前几年的金融风暴，导致国内的很多企业倒闭、重组，人员频繁流动。也因为这些原因，促成了现今人才富足职位空缺的局面。这样的就业形势，无疑是严峻的。 从去年下半年开始，新一代的主力军——第一批90后大学生已经陆续步入了工作岗位。我也是一名90后，从去年的12月份起，我开始了自己的实习生活。我的实习单位是在位于榆林的神华陕西国华锦界能源有限责任公司。以下简称“国华电厂”。 神华陕西国华锦界能源有限责任公司是一个以发电为主，煤炭业为主的能源企业，属于国家单位。由于单位规模大、员工人数多，管理繁杂，因此，在2004年时，单位将其管理服务中心外包给了陕西四季假日饭店管理有限公司。管理服务中心包含了国华电厂的行政楼、生产楼、检修楼、板房的行政前台业务，职工餐厅、职工公寓、以及国华电厂的所有卫生打扫工作。我实习的岗位是行政楼的行政前台，主要职责是负责行政楼所有的会议服务、会议室调配、布置工作，以及外来人员的接待、文件打复印，总经理、部门经理办公室的卫生打扫及其资料整理工作。除此之外，就是听从经理的调遣，配合经理服务公司的员工，使其工作顺利进行。

实习了三个月，在我的身边，相仿年龄的同事们频繁离职，有的虽然在职，但流露着鲜明的离职倾向。实习三个月结束后，回学校，准备论文，跟同学交流期间，了解到，同学中很多在实习后再说以后的打算，有很多都是对现有的工作犹豫不决，言谈举止间透漏着离职重新再做打算的想法。现在的就业形势这么严峻，一份工作来之不易，为什么在得到来之不易的工作后，实习一段时间后大家会有这么强烈的离职倾向？这是由什么原因导致的？ 对于企业来说，花费人力、财力，为企业招进一名员工，然后花着代价培养这名员工，目的只有一个，将这个员工培养成企业所需要的人才，为企业赢得利益！而像这种，在花费了成本之后，留不住人才，这无疑对企业来说是一大损失。对刚刚步入职业生涯的90后来说，挣着一份工资，还能有一个让自己成长，将理论运用到实践中的机会不是很好吗？为何他们会产生离职倾向甚至做出离职抉择？ 因此，本文选择了90后大学生短期实习后离职倾向这一问题来进行研究。希望能够得到一些有价值的研究结果，给企业跟90后大学毕业生一些对策和建议，帮助企业留住人才，减少人员流失成本的同时，也能帮助90后大学生更好稳定的就业。 1.2研究目的与意义 员工离职一直被国内外学者重视跟关注。因为员工的流失不仅会给企业带来巨大的财力、人力损失，而且，高的离职率会使企业内部人心惶惶，工作业绩下降，同时会使企业在外的声誉受损。 目前，国内外研究离职时一般选取的群体都是以60-80为主，很少有人对90后这个新生群体的离职问题进行研究。对于90的研究，很多研究都是落脚于90后的性格特质，价值观，人生观。但是就目前来说，90后已经成了我国就业生命军中的主力，只有他们很好的就业，贡献自己的力

某企业青年员工离职倾向分析研究

12580青年员工离职倾向分析研究专题 摘要 12580呼叫中心青年员工居高不下的离职率，严重影响到了12580呼叫中心工作的正常运转、增加了呼叫中心的运营成本，为呼叫中心人力资源管理带来了难度，同时给公司运营带来了较大的负面影响。本研究在访谈和问卷调查的基础上，对12580青年员工的离职倾向进行了研究分析，结果发现该中心的青年员工离职倾向明显。在青年员工离职情况的五个项目调查中：“我可能在未来一年中离开单位”选项的平均值最大，为3.78。青年员工离职倾向主要是受个人因素和组织因素的双重影响。其中，从个体因素考虑：性别、年龄、职务是影响青年员工离职的重要因素。从组织因素考虑：影响因子由高到低排列为：工作人际关系、工作环境、薪酬待遇、管理体制、绩效管理。本研究根据上述调查的数据分析和结果，提出了：搭建公司人际交往平台、创造工作环境、提高绩效管理水平等改进完善公司人力资源管理实效和综合水平的意见建议。 关键词：12580；青年员工；离职倾向；影响因素

ABSTRACT The high turnover rate of young staff in Call Center has seriously affected the normal work of call center, increased the contact center operating costs, brought difficulty for human resources management of Call Center. At the same time, it also brought certain effect for companies operating. The present study is based on the interview and questionnaires, select a number of staff in 12580 to analysis turnover of young staff. The result as follow: Young employee turnover tendency obvious. In the case of the five young employee turnover in the research project, "I might leave units in the next year, the average" option for 3.78.The rate of turnover about young employee are affected by individual factors and organizational factors. Among them, from individual factors, gender, age, the position is an important factor of young staff turnover. From the organization, the impact factor according to consider from high to low arrangement for: work of interpersonal relationship, working environment, compensation, management system, performance management. Finally, according to the results of the study, it put forward some measures in case of turnover, such as pay more attention to interpersonal communication, create the working environment, improve the level of performance management. keywords：12580; young staff; leaving tendency; effect factor