Citation Context

...he generated coauthorship weights express valuable information which should, and can, be taken into account for discovering the authorities of the authors within the community. The PageRank algorithm =-=[34]-=- can be applied to a undirectional coauthorship graph by transforming each undirectional edge into a set of two directional symmetrical98 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CY...

Citation Context

... instead of the organization to enhance expert finding. Since we use community context information to smooth the language model, our work is related to existing works in statistical language modeling =-=[24]-=-–[27], which is employed to discover documents related to a query in the documentbased model [12]. Ponte and Croft [24] were the first to apply the language modeling techniques in IR. From then on, ma...

Citation Context

...ead of the organization to enhance expert finding. Since we use community context information to smooth the language model, our work is related to existing works in statistical language modeling [24]–=-=[27]-=-, which is employed to discover documents related to a query in the documentbased model [12]. Ponte and Croft [24] were the first to apply the language modeling techniques in IR. From then on, many va...

Citation Context

... the language modeling techniques in IR. From then on, many variations on these traditional language models have been developed to improve the performance of IR, such as relevancebased language model =-=[28]-=-, title-based language model [29], and cluster-based language model [30]. Typically, a necessary and important step for the language model is to perform smoothing for the unseen query terms in the doc...

Citation Context

...k) and p(ai|Ck). At query time, our approach is performed as shown in Algorithm 1. The document-based model is approximately performed using the top k1 relevant documents obtained by top k algorithms =-=[35]-=-, and meanwhile, the community-sensitive AuthorRank is implemented using the top k2 relevant communities as well. In Section V-C4, we investigate and discuss the effect of these two parameters k1 and ...

Citation Context

... the evaluation of the task, three different metrics are employed to measure the performance of our proposed models, including precision at rank n (P@n), mean average precision (AP) (MAP), bpref [5], =-=[37]-=-, [38], and mean reciprocal rank (MRR). P@n measures the fraction of the top n retrieved results that are relevant experts for the given query, which is defined as # relevant experts in top n results ...

Citation Context

...ERTISE RETRIEVAL USING COMMUNITY-AWARE STRATEGIES 95 community-aware strategies are different. In this paper, we utilize the AuthorRank [21] to measure the authority based on the coauthorship network =-=[22]-=-, but it is query independent. We develop the query-sensitive AuthorRank as well as the adaptive ranking refinement strategy for the enhanced model. More recently, several expert finding approaches us...

Citation Context

...expertise for the given query. Traditionally, the expertise of a candidate is characterized based on the documents that have been associated with the candidate. One of the state-of-the-art approaches =-=[6]-=-, [7] is the document-based model to estimate the weighted sum of retrieval scores of all documents related to an expert candidate as a measure of candidate’s expertise. However, previous methods main...

Citation Context

...n these traditional language models have been developed to improve the performance of IR, such as relevancebased language model [28], title-based language model [29], and cluster-based language model =-=[30]-=-. Typically, a necessary and important step for the language model is to perform smoothing for the unseen query terms in the document, and several different smoothing methods have been proposed, such ...

Citation Context

...erform smoothing for the unseen query terms in the document, and several different smoothing methods have been proposed, such as Jelinek–Mercer smoothing and Bayesian smoothing using Dirichlet priors =-=[26]-=-, [27]. However, all these smoothing methods only consider the collection as a whole, while our proposed smoothing method uses the community context information to smooth the language model instead of...

Citation Context

...e from the categories described previously, there are various methods proposed to extend or enhance expertise retrieval in many ways. Macdonald and Ounis presented a voting model for expert search in =-=[15]-=-. Their algorithm investigated several data fusion techniques to aggregate document votes into a ranking of candidates without using community-aware strategies. In [16], the authors extended the exper...

Citation Context

... all documents associated with the candidate, and the ranking scores were estimated according to the candidate profile in response to a given query. On the other hand, document-based models [6], [7], =-=[13]-=- are also referred to as query-dependent method in [14]. They first rank documents in the corpus for a given query topic and then find the associated candidates according to the retrieved documents. I...

Citation Context

... compared by Balog et al. in [6]. The candidate-based approach was first proposed by Craswell et al. [10], which is also referred to as profile-based method or query-independent approach in [6], [11]–=-=[14]-=-. In these methods, a profile (“virtual document”) was built for each candidate based on all documents associated with the candidate, and the ranking scores were estimated according to the candidate p...

Citation Context

...desirable to discover and infer community information, which contains a number of documents and authors for each community. Some existing studies have been conducted about how to discover communities =-=[8]-=-, [9], but this is not the focus of our work: Here, we focus on the problem of enhancing expertise retrieval with community information and suppose the community information already exists. As our app...

Citation Context

...tity retrieval [1], [2] and expertise retrieval [3]. Expertise retrieval has received increasing interest since the introduction of an expert finding task in the Text Retrieval Conference (TREC) 2005 =-=[4]-=-, [5]. The task of expertise retrieval is to identify a set of persons with relevant expertise for the given query. Traditionally, the expertise of a candidate is characterized based on the documents ...

Citation Context

...development of information retrieval (IR) techniques, many research efforts in this field have been made to address high-level IR and not just traditional document retrieval, such as entity retrieval =-=[1]-=-, [2] and expertise retrieval [3]. Expertise retrieval has received increasing interest since the introduction of an expert finding task in the Text Retrieval Conference (TREC) 2005 [4], [5]. The task...

Citation Context

...valuation of the task, three different metrics are employed to measure the performance of our proposed models, including precision at rank n (P@n), mean average precision (AP) (MAP), bpref [5], [37], =-=[38]-=-, and mean reciprocal rank (MRR). P@n measures the fraction of the top n retrieved results that are relevant experts for the given query, which is defined as # relevant experts in top n results P @n =...

Citation Context

...d and compared by Balog et al. in [6]. The candidate-based approach was first proposed by Craswell et al. [10], which is also referred to as profile-based method or query-independent approach in [6], =-=[11]-=-–[14]. In these methods, a profile (“virtual document”) was built for each candidate based on all documents associated with the candidate, and the ranking scores were estimated according to the candid...

Citation Context

...xpert finding in academia domain has not been addressed much in the past. Li et al. [31] built an academic expertise-oriented search service, and they proposed a relevancy propagation-based algorithm =-=[32]-=- using the coauthorship network for expert finding. In addition, the expert finding task in a real-world academic field based on the DBLP bibliography was explored in [7], [33]. In this paper, we also...

Citation Context

... in IR. From then on, many variations on these traditional language models have been developed to improve the performance of IR, such as relevancebased language model [28], title-based language model =-=[29]-=-, and cluster-based language model [30]. Typically, a necessary and important step for the language model is to perform smoothing for the unseen query terms in the document, and several different smoo...

Citation Context

...opment of information retrieval (IR) techniques, many research efforts in this field have been made to address high-level IR and not just traditional document retrieval, such as entity retrieval [1], =-=[2]-=- and expertise retrieval [3]. Expertise retrieval has received increasing interest since the introduction of an expert finding task in the Text Retrieval Conference (TREC) 2005 [4], [5]. The task of e...

Citation Context

... propagation-based algorithm [32] using the coauthorship network for expert finding. In addition, the expert finding task in a real-world academic field based on the DBLP bibliography was explored in =-=[7]-=-, [33]. In this paper, we also focus on expert finding in an academic environment based on DBLP bibliography data. III. MODELING EXPERTISE RETRIEVAL The task of expertise retrieval is to retrieve a li...

Citation Context

... with previous methods, our proposedDENG et al.: MODELS FOR EXPERTISE RETRIEVAL USING COMMUNITY-AWARE STRATEGIES 95 community-aware strategies are different. In this paper, we utilize the AuthorRank =-=[21]-=- to measure the authority based on the coauthorship network [22], but it is query independent. We develop the query-sensitive AuthorRank as well as the adaptive ranking refinement strategy for the enh...

Citation Context

...n to smooth the language model, our work is related to existing works in statistical language modeling [24]–[27], which is employed to discover documents related to a query in the documentbased model =-=[12]-=-. Ponte and Croft [24] were the first to apply the language modeling techniques in IR. From then on, many variations on these traditional language models have been developed to improve the performance...

Citation Context

...val (IR) techniques, many research efforts in this field have been made to address high-level IR and not just traditional document retrieval, such as entity retrieval [1], [2] and expertise retrieval =-=[3]-=-. Expertise retrieval has received increasing interest since the introduction of an expert finding task in the Text Retrieval Conference (TREC) 2005 [4], [5]. The task of expertise retrieval is to ide...

Citation Context

...rithms difficult to generalize to large-scale data collection. In [18], the authors proposed a graph-based reranking model and applied it to expert finding to refine ranking results. Serdyukov et al. =-=[19]-=- modeled the process of expert finding as multistep relevance propagation over the expertise graphs. Recently, Jiao et al. [20] proposed an ExpertRank algorithm for expert finding from online communit...

Citation Context

...ting model for expert search in [15]. Their algorithm investigated several data fusion techniques to aggregate document votes into a ranking of candidates without using community-aware strategies. In =-=[16]-=-, the authors extended the expert search by identifying some high-quality evidence. Bogers et al. [17] presented some relevant expert finding techniques that combine multiple sources of expertise evid...

Citation Context

...able to discover and infer community information, which contains a number of documents and authors for each community. Some existing studies have been conducted about how to discover communities [8], =-=[9]-=-, but this is not the focus of our work: Here, we focus on the problem of enhancing expertise retrieval with community information and suppose the community information already exists. As our approach...

Citation Context

...tional units to smooth the relevance scores. If an organization unit can be clustered as a community with similar research topics, the performance could be improved. Furthermore, Karimzadehgan et al. =-=[23]-=- leveraged the organizational hierarchy to enhance expert finding, in which they proposed a hierarchy-based algorithm to smooth the relevance scores using their neighbors’ scores. They demonstrated th...

Citation Context

...work has been concentrated on expertise retrieval in enterprise corpora [6] or intranet data set [3]. By contrast, expert finding in academia domain has not been addressed much in the past. Li et al. =-=[31]-=- built an academic expertise-oriented search service, and they proposed a relevancy propagation-based algorithm [32] using the coauthorship network for expert finding. In addition, the expert finding ...

Citation Context

... academic papers and social citation network. They only conducted experiments on a small-scale data collection, which makes their algorithms difficult to generalize to large-scale data collection. In =-=[18]-=-, the authors proposed a graph-based reranking model and applied it to expert finding to refine ranking results. Serdyukov et al. [19] modeled the process of expert finding as multistep relevance prop...

Citation Context

...f pd(ai,q) and pc(ai,q), respectively. Another one is a weighted 2 http://incubator.apache.org/lucene.net/ Fig. 5. Sample of the DBLP XML records. version of the reciprocal rank data fusion technique =-=[36]-=-, which can be defined as S ′ 1 1 (ai) =μ2 +(1− μ2) (23) Rd(ai) Rc(ai) where Rd(ai) and Rc(ai) are the ranks of author ai in Rd and Rc, respectively. The new results are ranked according to the aggreg...

Citation Context

... expertise: candidate-based model and document-based model. These two models have been proposed and compared by Balog et al. in [6]. The candidate-based approach was first proposed by Craswell et al. =-=[10]-=-, which is also referred to as profile-based method or query-independent approach in [6], [11]–[14]. In these methods, a profile (“virtual document”) was built for each candidate based on all document...

Citation Context

...retrieval [1], [2] and expertise retrieval [3]. Expertise retrieval has received increasing interest since the introduction of an expert finding task in the Text Retrieval Conference (TREC) 2005 [4], =-=[5]-=-. The task of expertise retrieval is to identify a set of persons with relevant expertise for the given query. Traditionally, the expertise of a candidate is characterized based on the documents that ...

Citation Context

...ggregate document votes into a ranking of candidates without using community-aware strategies. In [16], the authors extended the expert search by identifying some high-quality evidence. Bogers et al. =-=[17]-=- presented some relevant expert finding techniques that combine multiple sources of expertise evidence such as academic papers and social citation network. They only conducted experiments on a small-s...

Citation Context

...agation-based algorithm [32] using the coauthorship network for expert finding. In addition, the expert finding task in a real-world academic field based on the DBLP bibliography was explored in [7], =-=[33]-=-. In this paper, we also focus on expert finding in an academic environment based on DBLP bibliography data. III. MODELING EXPERTISE RETRIEVAL The task of expertise retrieval is to retrieve a list of ...

Citation Context

...applied it to expert finding to refine ranking results. Serdyukov et al. [19] modeled the process of expert finding as multistep relevance propagation over the expertise graphs. Recently, Jiao et al. =-=[20]-=- proposed an ExpertRank algorithm for expert finding from online communities, while they combined the authority and relevance scores empirically without an adaptive refining strategy. Compared with pr...