{
    "paper_id": "O00-1006",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:59:04.639131Z"
    },
    "title": "Clustering Similar Query Sessions Toward Interactive Web Search",
    "authors": [
        {
            "first": "Chien-Kang",
            "middle": [],
            "last": "Huang",
            "suffix": "",
            "affiliation": {},
            "email": "ckhuang@mars.csie.ntu.edu.tw"
        },
        {
            "first": "Lee-Feng",
            "middle": [],
            "last": "Chien",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Academic Sinica",
                "location": {
                    "country": "Taiwan"
                }
            },
            "email": "*lfchien@iis.sinica.edu.tw"
        },
        {
            "first": "Yen-Jen",
            "middle": [],
            "last": "Oyang",
            "suffix": "",
            "affiliation": {},
            "email": "yjoyang@csie.ntu.edu.tw"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "A new effective log-based approach for interactive Web search is presented in this paper. The most important feature of the proposed approach is that the suggested terms corresponding to the user's query are extracted from similar query sessions, rather than from the contents of the retrieved documents. The experiment results demonstrate that this approach has a great potential in developing more effective web search utilities and may inspire more studies on advanced log mining mechanisms.",
    "pdf_parse": {
        "paper_id": "O00-1006",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "A new effective log-based approach for interactive Web search is presented in this paper. The most important feature of the proposed approach is that the suggested terms corresponding to the user's query are extracted from similar query sessions, rather than from the contents of the retrieved documents. The experiment results demonstrate that this approach has a great potential in developing more effective web search utilities and may inspire more studies on advanced log mining mechanisms.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Users' queries for Web search are usually short. For example, the average length of TREC topic description for conventional text retrieval is 15 tokens [11, 12] , while analyses of web search engine logs reveal that the average query length for Web search is about 2.3 tokens [6, 9] . Short queries means that the information about the user's intention provided to the search engine is very limited. To deal with the short query problem, interactive search techniques [2, 7] which attempt to identify the user's intentions and suggest more precise query terms are therefore commonly incorporated in Web search engine design.",
                "cite_spans": [
                    {
                        "start": 152,
                        "end": 156,
                        "text": "[11,",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 157,
                        "end": 160,
                        "text": "12]",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 276,
                        "end": 279,
                        "text": "[6,",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 280,
                        "end": 282,
                        "text": "9]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 468,
                        "end": 471,
                        "text": "[2,",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 472,
                        "end": 474,
                        "text": "7]",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "To determine more relevant query terms for each given query, the conventional interactive search processes often rely on the key terms in the retrieved documents [2, 7, 10] . The key term set is extracted either statically from the documents during preprocessing or dynamically on-the-fly. Since the precision rates of the retrieved documents are usually not high enough, the extracted key terms are often found not relevant and not very helpful in practical Web search services.",
                "cite_spans": [
                    {
                        "start": 162,
                        "end": 165,
                        "text": "[2,",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 166,
                        "end": 168,
                        "text": "7,",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 169,
                        "end": 172,
                        "text": "10]",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "In fact, extraction of relevant terms can be carried out by analyzing users' logs. In recent years, mining search engine logs has been obtaining more attention. Silverstein et al. [9] performed a second-order analysis on a log with a huge number of Web query terms. The results are then used to facilitate phrase recognition and query expansion [3] .",
                "cite_spans": [
                    {
                        "start": 180,
                        "end": 183,
                        "text": "[9]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 345,
                        "end": 348,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "In this paper, we propose a new approach based on log analysis for developing more effective interactive Web search engines. The most important feature of the proposed approach is that the suggested terms are extracted from similar query sessions, rather than from the contents of the retrieved documents. A query session is defined as a sequence of search requests issued by a user for a certain information need. The basis of the proposed approach is that two users with the same information need will issue common or related query terms. For example, in search for a subject regarding \"search engine technology\", a user may submit query terms such as \"search engine\", \"Web search\", \"Google\", \"Web search and multimedia\", while another user may submit \"Web search\", \"Lycos\". Therefore, if similar query sessions could be identified, query terms for the same information need can be extracted and applied to improve the effectiveness of search engines.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "The remainder of the paper will be organized as follows. Section 2 is a brief introduction to the idea of interactive search based on similar query sessions. The method proposed for segmenting query sessions from proxy logs will be described in Section 3. Then, how query sessions are clustered is addressed in Section 4.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "Section 5 will present some experiment results and a conclusion is given in Section 6. Fig.1 is an abstract diagram showing our idea for interactive search. Before introducing the basic idea of the proposed approach, the concept of query session is presented and defined below:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 87,
                        "end": 92,
                        "text": "Fig.1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "The proposed approach is assumed that the query space of users is formed by clusters of users' query sessions, and a set of query sessions grouped in the same clusters contain similar information needs. For each input query session with a sequence of i query terms, the interactive search process is then designed to retrieve the most similar cluster of query sessions from the query space, and then extract relevant terms in the cluster as suggested terms for next search. Once the i+1th query term is selected, it forms a new query session with i+1 terms and the interactive process will perform again.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query session = (ID, R 1 ,\u2026, R m ) where ID means the identifier of a user submitting a sequence of requests to a search engine in a certain period of time. Each request R i = (t i , q i ) means user ID sends a query term q to the search engine at time t",
                "sec_num": null
            },
            {
                "text": "Terms",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Suggested",
                "sec_num": null
            },
            {
                "text": "R i-1 , R i , R i+1",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Suggested",
                "sec_num": null
            },
            {
                "text": "Popular Visited Pages with the Subject Based on the above definition and idea, the problem to be dealt with is then formulated.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Input Query Session",
                "sec_num": null
            },
            {
                "text": "For a set of query sessions from a query session log, the considering problem is to cluster these query sessions into different groups based on estimated similarity between query sessions. Each cluster can be defined as {S i | f(S i , S j ) > threshold}, in which f() is the similarity estimation function between query sessions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Query Session Clustering Problem",
                "sec_num": null
            },
            {
                "text": "The proposed approach, as shown in Fig. 2 , is composed of three processing modules: query session segmentation module, query session clustering module and relevant term extraction module. In the stage of query session segmentation, each query session will be segmented and extracted from a proxy log, according to the time gap between successive search requests. All of the extracted query sessions will form as a query session log. In the session clustering stage, the sessions with similar queries will be clustered and the cluster names extracted from composed high frequency terms. In the relevant term extraction stage, the relevance between the recorded query terms will be calculated and sets of relevant terms will be extracted for term suggestion applications in a search engine. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 35,
                        "end": 41,
                        "text": "Fig. 2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Overview of the Proposed Approach",
                "sec_num": null
            },
            {
                "text": "A common proxy server might easily have thousands of clients accessing the web through it. Not only the general HTTP requests could pass through the proxy server, all of search HTTP requests are same. Compared with common search engine logs, a proxy server's log records more rigid information for users' information access and, more importantly, the recorded search requests are not limited to certain search engines. However, a proxy log might record too much information and only some of them are useful in terms of search engine applications [13] . In our application it is sufficient to only use the following fields of logging information:",
                "cite_spans": [
                    {
                        "start": 546,
                        "end": 550,
                        "text": "[13]",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "A timestamp that indicates when a search request was submitted.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "A client address that indicates the IP address of the requesting instance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "A URL string that contains the request content.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "Since the experiments are just performing, the testing log is from NTU proxy servers and is still small. Some statistics of the testing proxy log are listed in Table 1 that queries for a single information need come clustered in time, and then there is a gap before the user returns to the search engine.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 160,
                        "end": 167,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "The method for query session segmentation is then proposed as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "The Method for Query Session Segmentation:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "For a proxy log, it will segment the whole log L = {T i | where",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "T i = (ID k , t i , q i )} into a set of query sessions {S i | S i = (ID k , R 1 ,\u2026, R m )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": ", where R i = (t i , q i ), and t it i-1 < threshold}, where t i is the timestamp when the query q i issued.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Segmentation",
                "sec_num": "3."
            },
            {
                "text": "To realize the performance of the above method, several experiments have been performed. Table 2 : Percentages of the extracted singleton and non-singleton query sessions, when the time threshold is set as 5 minutes.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 89,
                        "end": 96,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Analysis of Segmented Query Sessions",
                "sec_num": null
            },
            {
                "text": "As the definition of the session clustering problem in Section 2, the similarity estimation function is necessary and formulated below: ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Clustering",
                "sec_num": "4."
            },
            {
                "text": "Similarity",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Clustering",
                "sec_num": "4."
            },
            {
                "text": "( ) ( ) \u2211 < < < < = n j m i j i mn R R sim S S sim 1 , 1 2 1 2 1 / , ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Clustering",
                "sec_num": "4."
            },
            {
                "text": "The similarity between two composed query terms will be further described below.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Clustering",
                "sec_num": "4."
            },
            {
                "text": "Development of an effective relevance estimation function is important. Since our research is just in the beginning, only two kinds of relevance estimation functions were developed and tested. In the first method, the relevance between two query terms is simply calculated by the co-occurrence frequency value of the query terms in the segmented query sessions. In the second method, the relevance is calculated by the cosine value of the query terms' feature vectors.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Query Session Clustering",
                "sec_num": "4."
            },
            {
                "text": "In the first method, we define the relevance estimation function below.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Method I for Similarity Estimation of Relevant Terms",
                "sec_num": null
            },
            {
                "text": "Before calculating the relevance between query terms, a set of query sessions has been segmented and extracted from the testing proxy log. After preprocessing the query session log, we calculate the co-occurrence frequency between each unique query term and its associated terms occurring together in the same query sessions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "f(x, y i ) = co-occurrence(x, y i )",
                "sec_num": null
            },
            {
                "text": "We explain the calculation process with a simple example below. After segmenting the proxy log, it is assumed that we got five query sessions S1-5 and each contains several query terms from A to F, e.g., In this case, f(B, C) will be 3, because B and C occur together in three sessions, i.e., S1, S2 and S3. Although the above method looks straightforward, its obtained performance is really out of our expectation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "f(x, y i ) = co-occurrence(x, y i )",
                "sec_num": null
            },
            {
                "text": "In the first method, the relevance of two query terms needs a strong support of their co-existence in a certain number of query sessions. Using a VSM-like technique it can release such a constraint. The second method is based on vector space model, and it can be formalized as below.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Method II for Similarity Estimation of Relevant Terms",
                "sec_num": null
            },
            {
                "text": "Assuming there are two terms T 1 and T 2 :",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "f(x, y i ) = cos(FV(x), FV(y I )), FV(x) means feature vector of term x, FV(T i ) = {S j :N ij | Sj and Ti are coexist in query sessions, Nij is the count of their co-occurrence}",
                "sec_num": null
            },
            {
                "text": "T 1 {S 1 :N 11 , S 2 :N 12 , S 4 :N 14 , S 5 :N 15 } T 2 {S 1 :N 21 , S 2 :N 22 , S 3 :N 23 , S 7 :N 27 }",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "f(x, y i ) = cos(FV(x), FV(y I )), FV(x) means feature vector of term x, FV(T i ) = {S j :N ij | Sj and Ti are coexist in query sessions, Nij is the count of their co-occurrence}",
                "sec_num": null
            },
            {
                "text": "The relevance value of T1 and T2 is the obtained cosine or say the inner product value of these two vectors.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "f(x, y i ) = cos(FV(x), FV(y I )), FV(x) means feature vector of term x, FV(T i ) = {S j :N ij | Sj and Ti are coexist in query sessions, Nij is the count of their co-occurrence}",
                "sec_num": null
            },
            {
                "text": "( ) ( ) ( ) ( ) ( ) \u2211 \u2211 \u2211 \u2022 \u2022 = \u2022 = k k j j i i i N N N N T FV T FV T T f 2 1 2 1 2 1 2 1 2 1 cos ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "f(x, y i ) = cos(FV(x), FV(y I )), FV(x) means feature vector of term x, FV(T i ) = {S j :N ij | Sj and Ti are coexist in query sessions, Nij is the count of their co-occurrence}",
                "sec_num": null
            },
            {
                "text": "A issues to be dealt with in the clustering process, that is, what each cluster means and how to name these clusters. In order to find out the representative meaning of each cluster and avoid the difficulty in classifying short sessions, the clustering process is being developed as shown in Fig. 4 , which is designed as an incremental adaptive procedure.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 292,
                        "end": 298,
                        "text": "Fig. 4",
                        "ref_id": "FIGREF5"
                    }
                ],
                "eq_spans": [],
                "section": "The Clustering Process",
                "sec_num": null
            },
            {
                "text": "This procedure consists of 4 processing steps:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Clustering Process",
                "sec_num": null
            },
            {
                "text": "1. For each incoming query session, check whether there are certain common query terms between the session and existing clusters. If the common query terms exist, assign the session to these clusters.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Clustering Process",
                "sec_num": null
            },
            {
                "text": "clusters, calculate the similarity between the session and existing clusters. If the estimated similarity is higher than a predefined threshold, the session will be assigned to the cluster.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "If the incoming session doesn't have sufficient common query term with existing",
                "sec_num": "2."
            },
            {
                "text": "3. If the incoming session isn't assigned to any cluster, it will be sent to the delay queue for further processing. In this step, the incoming session will compare with other sessions in delay queue to check whether there are common query terms in the sessions that could be combined. 4. A standalone module will dynamically merge or split the clusters according to the new requirements or the new incoming sessions. When the similarity of two clusters are higher than another pre-defined threshold, merge will happen; when the cluster grows larger, split will happen. Merging and splitting are strategies for maintaining the similarity of query sessions in a cluster. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "If the incoming session doesn't have sufficient common query term with existing",
                "sec_num": "2."
            },
            {
                "text": "The above clustering process was just implemented. The sessions grouped by Step 1 are set that should contain at least two common query terms, and each obtained cluster is then named by the pair of common query terms with the highest frequency.",
                "cite_spans": [
                    {
                        "start": 75,
                        "end": 81,
                        "text": "Step 1",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Performance of Query Session Clustering",
                "sec_num": null
            },
            {
                "text": "Currently, there are about 700 initial clusters have been obtained from the query session log shown in Section 2. Table 3 Table 3 . An example of the obtained session clusters.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 114,
                        "end": 121,
                        "text": "Table 3",
                        "ref_id": null
                    },
                    {
                        "start": 122,
                        "end": 129,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Performance of Query Session Clustering",
                "sec_num": null
            },
            {
                "text": "It is worthy to note that clusters with similar names (that with shared query terms as the names of the clusters) usually contain similar information needs. Table 4 . An example which contains a number of obtained clusters with information needs related to \u5716\u7247(picture)",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 157,
                        "end": 164,
                        "text": "Table 4",
                        "ref_id": "TABREF6"
                    }
                ],
                "eq_spans": [],
                "section": "Performance of Query Session Clustering",
                "sec_num": null
            },
            {
                "text": "In fact, the proposed approach is also useful in relevant term extraction. We evaluate the proposed estimation methods with a testing set of query terms that were randomly selected from the testing proxy log. For Method I, the relevant terms are whose co-occurrence frequency large than 1, and for Method II the relevant terms are whose cosine value large than 0.25. The obtained preliminary result is shown in Table 5 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 411,
                        "end": 418,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Performance of Relevant Term Extraction",
                "sec_num": null
            },
            {
                "text": "The first column \"rank\" in Table 5 is the order of the testing terms in the extracted term set, which is sorted by their occurrences. The real query terms are listed in the \"term\" column, and their English translations are listed in the next column. The data in the \"freq\" column represents the occurrence of each query term. The \"total\" column indicates the numbers of all different co-occurred terms, and the \"related\" column the numbers of relevant terms among co-occurred terms that were checked manually. The next nine columns are the obtained statistics of the proposed methods.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 27,
                        "end": 34,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Performance of Relevant Term Extraction",
                "sec_num": null
            },
            {
                "text": "Each method consists of three columns, the first is the number of extracted relevant terms, the second is the number of correct relevant terms, and the third is the obtained accuracy. Note that the third method is the result merged with the proposed two methods. -\u66ff\u4ee3\u5f79 20",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Performance of Relevant Term Extraction",
                "sec_num": null
            },
            {
                "text": "\u2022 \u570b\u9632\u90e8:5 \u793e\u6703\u5f79:3 \u5175\u5f79:3 \u5357\u6295\u7e23\u653f\u5e9c:2 \u5167\u653f\u90e8:2 \u570b\u9632\u5f79:2",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Performance of Relevant Term Extraction",
                "sec_num": null
            },
            {
                "text": "In this paper, a new approach based on log analysis is proposed for implementing",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6."
            },
            {
                "text": "interactive Web search. The most important feature of the proposed approach is that the suggested terms corresponding to a user query are extracted from similar query sessions, rather than from the contents of the retrieved documents. Furthermore, the estimation of term relevance is also based on co-occurrence analysis of the query terms in query sessions. The experiment results presented in this paper are based on analysis of the proxy server logs. The results obtained so far demonstrate that the proposed approach is quite promising in respect to improving the effectiveness of interactive web search engines.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6."
            },
            {
                "text": "The results presented in this paper is just a beginning of mining log data toward developing more effective web search engines. Since this approach already demonstrates quite promising results, further investigation on mining log data deserves more of our attention. Further study may result in more advanced mining mechanism that can give us more comprehensive information about term relevance and allow us to identify users' information need more effectively. For example, some sort of thesaurus information may be derived from mining log data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6."
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF1": {
                "ref_id": "b1",
                "title": "The Paraphrase Search Assistant: Terminology Feedback for Iterative Information Seeking",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "G"
                        ],
                        "last": "Anick",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Tipirneni",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Proceedings of 22nd International Conference on Research and Development in Information Retrieval (SIGIR-99)",
                "volume": "",
                "issue": "",
                "pages": "153--159",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P.G. Anick and S. Tipirneni, \"The Paraphrase Search Assistant: Terminology Feedback for Iterative Information Seeking,\" in Proceedings of 22nd International Conference on Research and Development in Information Retrieval (SIGIR-99), pages 153-159, 1999.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Phrase recognition and expansion for short precision-biased queries based on a query log",
                "authors": [
                    {
                        "first": "E",
                        "middle": [
                            "F"
                        ],
                        "last": "De Lima",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "O"
                        ],
                        "last": "Pedersen",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Proceedings of 22nd International Conference on Research and Development in Information Retrieval (SIGIR-99)",
                "volume": "",
                "issue": "",
                "pages": "145--152",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "E.F. de Lima and J.O. Pedersen, \"Phrase recognition and expansion for short precision-biased queries based on a query log,\" in Proceedings of 22nd International Conference on Research and Development in Information Retrieval (SIGIR-99), pages 145-152, 1999.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Real life information retrieval: A study of user queries on the web",
                "authors": [
                    {
                        "first": "B",
                        "middle": [
                            "J"
                        ],
                        "last": "Jansen",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Spink",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Bateman",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Saracevic",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "SIGIR FORUM",
                "volume": "32",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "B.J. Jansen, A. Spink, J. Bateman, and T. Saracevic, \"Real life information retrieval: A study of user queries on the web,\" SIGIR FORUM, 32(1), 1998.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Phrasier: a System for Interactive Document Retrieval Using Keyphrases",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [
                            "S"
                        ],
                        "last": "Staveley",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Proceedings of 22nd International Conference on Research and Development in Information Retrieval (SIGIR-99)",
                "volume": "",
                "issue": "",
                "pages": "160--167",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Jones and M.S. Staveley, \"Phrasier: a System for Interactive Document Retrieval Using Keyphrases,\" in Proceedings of 22nd International Conference on Research and Development in Information Retrieval (SIGIR-99), pages 160-167, 1999.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Analysis of a very large AltaVista query log",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Silverstein",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Henzinger",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Marais",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Morics",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. Silverstein, M. Henzinger, H. Marais, and M. Morics., \"Analysis of a very large AltaVista query log,\" Technical Report 1998-014, Digital Systems Research Center, 1998.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Fast and Effective Query Refinement",
                "authors": [
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Velez",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Weiss",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [
                            "A"
                        ],
                        "last": "Sheldon",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "K"
                        ],
                        "last": "Gifford",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Proceedings of 20th International Conference on Research and Development in Information Retrieval (SIGIR-97)",
                "volume": "",
                "issue": "",
                "pages": "6--15",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "B. Velez, R. Weiss, M.A. Sheldon and D.K. Gifford, \"Fast and Effective Query Refinement,\" in Proceedings of 20th International Conference on Research and Development in Information Retrieval (SIGIR-97), pages 6-15, 1997.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Overviw of the sixth text retrieval conference TREC-5",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Voorhees",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "K"
                        ],
                        "last": "Harman",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Proceedings of the Fifth Text REtrieval Conference (TREC-5)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "E. Voorhees and D.K. Harman, \"Overviw of the sixth text retrieval conference TREC-5,\" in Proceedings of the Fifth Text REtrieval Conference (TREC-5), 1997",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Overviw of the sixth text retrieval conference TREC-6",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Voorhees",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "K"
                        ],
                        "last": "Harman",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proceedings of the Sixth Text REtrieval Conference",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "E. Voorhees and D.K. Harman, \"Overviw of the sixth text retrieval conference TREC-6,\" in Proceedings of the Sixth Text REtrieval Conference (TREC-6),",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "SQUID Frequently Asked Questions. Section 6. Squid Log Files",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Wessels",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Wessels, SQUID Frequently Asked Questions. Section 6. Squid Log Files. http://www.squid-cache.org/Doc/FAQ/FAQ-6.html",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Yam",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yam. http://www.yam.com.tw.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "text": "An abstract diagram showing our idea for interactive search.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF1": {
                "text": "An overview of the proposed approach",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF2": {
                "text": "shows the relationship between the time thresholds and the numbers of segmented query sessions. The time thresholds determine the maximum time gap between two successive requests from the same client. The values of the time thresholds were tuned from 0 seconds to 360 seconds. In the research of Silverstein et al, 5 minutes as suggested is a proper threshold value. With the same threshold value, the number of segmented query sessions is shown in Table 2. The percentages of the segmented singleton and non-singleton query sessions are found similar to those reported by Silverstein et al.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF3": {
                "text": "The numbers of segmented query sessions (that with more than 1 queries), regarding to the change of increasing time thresholds.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF4": {
                "text": "S1: {A, B} S2: {C, D, B} S3: {A, B, C} S4: {A, E} S5: {B, C, E, F}",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF5": {
                "text": "The work flow of the session clustering process.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "TABREF1": {
                "type_str": "table",
                "content": "<table><tr><td>.</td></tr></table>",
                "html": null,
                "text": "Some statistics of the testing proxy log.",
                "num": null
            },
            "TABREF3": {
                "type_str": "table",
                "content": "<table><tr><td>Given two sessions, S 1 = (ID K , R 11 ,\u2026,R 1m ) and S 1 = (ID L , R 21 ,\u2026,R 2n ), in which R ij</td></tr><tr><td>is the j-th query term occurred in session S i which is issued by a client. The</td></tr><tr><td>similarity estimation function is defined as:</td></tr></table>",
                "html": null,
                "text": "",
                "num": null
            },
            "TABREF5": {
                "type_str": "table",
                "content": "<table><tr><td>Cluster Name</td><td colspan=\"2\">Sessions in Cluster</td><td/></tr><tr><td colspan=\"2\">\u53f0\u5927__\u53f0\u7063\u5927\u5b78 10: \u53f0\u5927</td><td>\u53f0\u7063\u5927\u5b78</td><td/></tr><tr><td/><td>1: +\u53f0\u5927</td><td colspan=\"2\">\u53f0\u5927 \u53f0\u7063\u5927\u5b78</td></tr><tr><td/><td>1: \u5716\u66f8\u9928</td><td colspan=\"3\">\u53f0\u5927\u5716\u66f8\u9928 \u53f0\u5927 \u53f0\u7063\u5927\u5b78</td><td>\u53f0\u7063\u5927\u5b78\u5716\u66f8\u9928</td></tr><tr><td/><td colspan=\"4\">1: \u53f0\u5927\u4e2d\u570b\u6587\u5b78\u7cfb \u53f0\u5927 \u53f0\u7063\u5927\u5b78</td></tr><tr><td/><td>1: \u53f0\u5927</td><td>\u53f0\u7063\u5927\u5b78</td><td>\u5175\u5b78</td></tr><tr><td/><td colspan=\"3\">1: \u81fa\u5927\u7269\u7406 \u53f0\u5927 \u53f0\u7063\u5927\u5b78</td></tr><tr><td/><td colspan=\"4\">1: \u53f0\u7063\u5927\u5b78 \u53f0\u5927 \u53f0\u7063\u5927\u5b78^\u62db\u6a19</td><td>\u53f0\u7063\u5927\u5b78^\u5de5\u7a0b</td><td>\u62db\u6a19^\u53f0\u5927</td></tr><tr><td/><td colspan=\"3\">1: \u53f0\u5927\u96fb\u6a5f \u53f0\u5927 \u53f0\u7063\u5927\u5b78</td></tr><tr><td/><td>1: \u53f0\u5927</td><td>\u53f0\u7063\u5927\u5b78</td><td colspan=\"2\">\u570b\u7acb\u53f0\u7063\u5927\u5b78</td></tr><tr><td/><td>1: \u793e\u6703\u5b78</td><td colspan=\"3\">\u53f0\u5927\u5716\u66f8\u9928 \u53f0\u5927 \u53f0\u7063\u5927\u5b78</td><td>\u5e02\u7acb\u5716\u66f8\u9928</td></tr><tr><td/><td>1: \u53f0\u5927</td><td colspan=\"3\">\u53f0\u5927\u91ab\u5b78\u9662 \u53f0\u5927\u91ab\u5b78\u7cfb \u53f0\u7063\u5927\u5b78</td></tr><tr><td/><td colspan=\"4\">1: \u6642\u5831\u80b2\u6a02 \u6642\u5831\u80b2\u6a02\u80a1\u4efd\u6709\u9650\u516c\u53f8</td><td>\u6642\u5831 \u53f0\u5927 \u53f0\u7063\u5927\u5b78</td><td>\u5927\u5b78\u806f\u62db\u653e\u699c</td><td>\u699c\u55ae</td></tr><tr><td/><td colspan=\"3\">1: \u53f0\u7063\u5927\u5b78 \u53f0\u5927 \u6210\u529f\u5927\u5b78</td></tr><tr><td/><td>1: \u5716\u66f8</td><td>\u53f0\u5927\u5716\u66f8</td><td colspan=\"2\">\u53f0\u5927 \u53f0\u7063\u5927\u5b78</td></tr><tr><td/><td>1: \u6930\u6797</td><td colspan=\"2\">\u53f0\u5927 \u53f0\u7063\u5927\u5b78</td><td>\u53f0\u7063\u5927\u5b78\u8a08\u7b97\u6a5f\u4e2d\u5fc3</td></tr><tr><td/><td>1: \u53f0\u5927</td><td>\u53f0\u7063\u5b78\u5927</td><td colspan=\"2\">\u53f0\u7063\u5927\u5b78</td><td>\u6bd4\u8cfd</td></tr></table>",
                "html": null,
                "text": "illustrates an example of the clusters. The numbers ahead in each row of Column 2 are frequency values of the corresponding sessions. Based on our initial observations, the relevance of the clustered query sessions is often high. It is obviously higher than that obtained with document-based approach in our experiences.",
                "num": null
            },
            "TABREF6": {
                "type_str": "table",
                "content": "<table><tr><td>Cluster Name</td><td>Translation</td></tr><tr><td>\u5361\u901a\u5716\u7247__kitty</td><td>cartoon picture __ kitty</td></tr><tr><td>\u53ef\u611b\u5716\u7247__\u5361\u901a</td><td>lovable picture __ cartoon</td></tr><tr><td>\u6bcd\u89aa\u7bc0\u5716\u7247__\u6bcd\u89aa</td><td>mother's day picture __ mother</td></tr><tr><td>\u6bcd\u89aa\u7bc0\u5716\u7247__\u6bcd\u89aa\u7bc0</td><td>mother's day picture __ mother's day</td></tr><tr><td colspan=\"2\">\u6bcd\u89aa\u7bc0\u5716\u7247__\u6bcd\u89aa\u7bc0\u5361\u7247 mother's day picture __ mother's day greeting card</td></tr><tr><td>\u6bcd\u89aa\u7bc0\u5716\u7247__\u8db4\u8db4\u718a</td><td>mother's day picture __ bear</td></tr><tr><td>\u6bcd\u89aa\u7bc0\u5716\u7247__\u5eb7\u4e43\u99a8</td><td>mother's day picture __ carnation</td></tr><tr><td>\u5716\u7247__kitty</td><td>picture __ kitty</td></tr><tr><td>\u5716\u7247__\u5361\u901a</td><td>picture __ cartoon</td></tr><tr><td>\u5716\u7247__\u5361\u901a\u5716\u7247</td><td>picture __ cartoon picture</td></tr><tr><td>\u5716\u7247__\u5e03\u4e01\u72d7</td><td>picture __ pudding dog</td></tr><tr><td>\u5716\u7247__\u6bcd\u89aa\u7bc0</td><td>picture __ mother's day</td></tr><tr><td>\u5716\u7247__\u6bcd\u89aa\u7bc0\u5361\u7247</td><td>picture __ mother's day greeting card</td></tr><tr><td>\u5716\u7247__\u6bcd\u89aa\u7bc0\u5716\u7247</td><td>picture __ mother's day picture</td></tr><tr><td>\u5716\u7247__\u76ae\u5361\u4e18</td><td>picture __ picachu</td></tr><tr><td>\u5716\u7247__\u6709\u8da3</td><td>picture __ funny</td></tr><tr><td>\u5716\u7247__\u72d7</td><td>picture __ dog</td></tr><tr><td>\u5716\u7247__\u8db4\u8db4\u718a</td><td>picture __ bear</td></tr><tr><td>\u5716\u7247__\u684c\u9762</td><td>picture __ theme</td></tr><tr><td>\u5716\u7247__\u684c\u9762\u738b</td><td>picture __ themeking</td></tr><tr><td>\u5716\u7247__\u795e\u5947\u5bf6\u8c9d</td><td>picture __ pokemon</td></tr><tr><td>\u5716\u7247__\u52d5\u7269</td><td>picture __ animal</td></tr><tr><td>\u5716\u7247__\u52d5\u756b</td><td>picture __ animation</td></tr><tr><td>\u5716\u7247__\u5eb7\u4e43\u99a8</td><td>picture __ carnation</td></tr><tr><td>\u5716\u7247__\u904a\u6232</td><td>picture __ game</td></tr><tr><td>\u5716\u7247__\u904a\u6232\u4e0b\u8f09</td><td>picture __ game download</td></tr><tr><td>\u5716\u7247__\u5716</td><td>picture __ graph</td></tr><tr><td>\u5716\u7247__\u5716\u7247\u4e0b\u8f09</td><td>picture __ picture download</td></tr><tr><td>\u5716\u7247__\u5716\u5eab</td><td>picture __ picture bank</td></tr><tr><td>\u5716\u7247__\u5716\u6a94</td><td>picture __ picture file</td></tr><tr><td>\u5716\u7247__\u6f2b\u756b</td><td>picture __ comic</td></tr></table>",
                "html": null,
                "text": "is an example which contains a number of clusters with information needs related to \u5716\u7247 (picture). In these clusters, \u5716\u7247(picture) will relate to several different kinds of search subjects, including characters in cartoon (e.g. kitty and pokemon), downloading, online picture banks, greeting cards for some festivals, and etc. These similar clusters could be further taken as sub-clusters of the information needs. The obtained information would be very useful in performing term suggestion in interactive search process.",
                "num": null
            },
            "TABREF8": {
                "type_str": "table",
                "content": "<table><tr><td>University includes \u53f0 \u5927 \u91ab \u5b78 \u9662 (medical school), \u53f0 \u7063 \u5927 \u5b78 \u8a08 \u7b97 \u6a5f \u4e2d \u5fc3</td></tr><tr><td>(computing center), \u53f0\u5927\u5716\u66f8\u9928 (library), \u53f0\u5927\u96fb\u6a5f (department of electrical</td></tr><tr><td>engineering),</td></tr></table>",
                "html": null,
                "text": "The performance obtained with the proposed methods.AnalyzingTable 5., we can find that Method I favors high frequency terms (e.g., term frequency > 50). It is really suited in applications that need not many but accurate relevant terms. However, for the query terms with not high frequency, we might rely on Method II. On the other hand, for those low frequency terms (term frequency < 10), Method II can not maintain a consistent performance. The effectiveness of this method is not reliable. In order to realize the effectiveness of the obtained result, we list an example of the extracted relevant query terms inTable 6.The query term is \u53f0\u7063\u5927\u5b78 (Taiwan University). The obtained relevant terms can be classified into 4 major categories. The first category is abbreviations including \u53f0 \u5927, +\u53f0\u5927 (\"+\" is the query syntax). The second is synonyms with different character forms like \u81fa\u7063\u5927\u5b78, with additional prefix like \u570b\u7acb\u53f0\u7063\u5927\u5b78, or nick name in semantics like \u6930\u6797 (Palm trees). The third is the sub divisions of Taiwan \u53f0\u5927\u4e2d\u570b\u6587\u5b78\u7cfb (department of Chinese literature), \u53f0\u5927\u91ab\u5b78\u7cfb (department of medicine). The final category is the events happened in Taiwan University, like \u62db\u6a19, \u5de5\u7a0b, \u5927\u5b78\u806f\u62db\u653e\u699c and \u699c\u55ae. An example of relevant terms extracted with the proposed methods.For more references, there are several examples that were not used in the testing are also illustrated below: \u5716\u7247:4 \u52d5\u756b:3 \u5716\u6a94:3 \u7db2\u9801\u88fd\u4f5c:2 \u5716:2 \u7db2\u9801\u5716\u5eab:2 \u904a\u6232\u4e0b\u8f09:2 \u4e16\u754c\u5730 \u5716:2 \u5730\u5716:2",
                "num": null
            }
        }
    }
}