Citation Context

...ates each word by applying one of the four stemmers implemented in the IRIS Nice Stemmer module, 2 which consists of a simple plural remover (Frakes & Baeza-Yates, 1992, chap. 8), the Porter stemmer (=-=Porter, 1980-=-), the modified Krovetz inflectional stemmer, and the Combo stemmer. The modified Krovetz inflectional stemmer implements a modified version of Krovetz's inflectional stemmer algorithm (Krovetz, 1993)...

Citation Context

...rd relevance feedback model called the “passage feedback model”.sThe formula for feedback vector creation in the passage feedback model looks almost identical to the “Ide regular” formula (Ide, 1971; =-=Salton & Buckley, 1990-=-), except where the document vector d is replaced by p, the passage vector. ∑∑ −+= nonrelrel ppqq oldnew (4) Since the normalization factor of the Lnu weight is based on document length, an inverse do...

Citation Context

...ter the initial processing step described above, IRIS conflates each word by applying one of the four stemmers implemented in the IRIS Nice Stemmer module,2 which consists of a simple plural remover (=-=Frakes & Baeza-Yates, 1992-=-, chap. 8), the Porter stemmer (Porter, 1980), the modified Krovetz inflectional stemmer, and the Combo stemmer.sThe modified Krovetz inflectional stemmer implements a modified version of Krovetz’s in...

Citation Context

...ve linear model (Wong & Yao, 1990; Wong, Yao, Salton, & Buckley, 1991). The basic approach of the adaptive linear model, which is based on the concept of the preference relation from decision theory (=-=Fishburn, 1970-=-), is to find a solution vector that, given any two documents in the collection, will rank a morepreferred document before a less-preferred one (Wong et al., 1988). The goal of the adaptive linear mod...

Citation Context

...C., Salton, G., Allan, J., & Singhal, A., 1995) for query terms. Lnu weightssattempt to match the probability of retrieval given a document length with the probability of relevance given that length (=-=Singhal, Buckley, & Mitra, 1996-=-). Our implementation of Lnu weights was the same as that of Buckley et al. (1996, 1997) except for the value of the slope in the formula, which is an adjustable parameter whose optimal value may depe...

Citation Context

... (Porter, 1980), the modified Krovetz inflectional stemmer, and the Combo stemmer. The modified Krovetz inflectional stemmer implements a modified version of Krovetz's inflectional stemmer algorithm (=-=Krovetz, 1993) and restores the root form of plural ("-s,&-=-quot; "-es," "-ies"), past tense ("-ed"), and present participle ("-ing") words, provided this root form is in our online dictionary. Though this 1 A prior vers...

Citation Context

...from the pre-test results. 3.1. System Component Tests 3.1.1 Experiment Design Prior experiments, both in and outside of TREC, have shown the use of syntactic phrases to be only marginally effective (=-=Salton, 1968-=-; Lewis, Croft & Bhandaru, 1989). However, most of the findings were based on the performance of initial retrieval only and did not investigate the effect of automatically expanding the feedback query...

Citation Context

... dq∑ = = 1 Tdq , (1) where qk is the weight of term k in the query, dik is the weight of term k in document i, and t is the number of terms in the index.sWe used SMART Lnu weights for document terms (=-=Buckley, Singhal, Mitra, & Salton, 1996-=-; Buckley, Singhal, & Mitra, 1997), and SMART ltc weights (Buckley, C., Salton, G., Allan, J., & Singhal, A., 1995) for query terms. Lnu weightssattempt to match the probability of retrieval given a d...

Citation Context

...levant' or "nonrelevant," but they are not quite sure what a "marginally relevant" document should be. The question of what makes a document relevant is a fertile ground for resear=-=ch (Schamber, 1991; Barry, 1994-=-). In a prior research, we investigated the relationship between the proportionality and the degree of relevance and found that the number of relevant passages in a document corresponded directly with...

Citation Context

...sed.sIn previous research on this “collection fusion” problem, various strategies were employed to compensate for the potential incomparability of query-document similarity scores across collections (=-=Voorhees, Gupta, & Johnson-Laird, 1995-=-; Savoy, Calve, & Vrajitoru, 1997). Though the “raw score” merging method can be problematic when collection-dependent term weights (i.e. idf weight) cause the retrieval scores of similar documents to...

Citation Context

...f expanding the initial query byspseudo-relevance feedback (using terms from the top n documents) have been used by top performing participants in past TREC ad-hoc experiments (Buckley et. al., 1995; =-=Voorhees & Harman, 1997-=-). In addition to query expansion by phrases and automatic feedback, we also tested query expansion methods by using passages6 with matching initial query terms.sThree variations of query expansion by...

Citation Context

... IRIS a third relevance feedback model called the "passage feedback model". The formula for feedback vector creation in the passage feedback model looks almost identical to the "Ide reg=-=ular" formula (Ide, 1971-=-; Salton & Buckley, 1990), except where the document vector d is replaced by p, the passage vector. - + = nonrel rel p p q q old new (4) Since the normalization factor of the Lnu weight is based on do...

Citation Context

...precision, and the total number of relevant documents retrieved in the top 1000 documents. Optimum F is the highest F value in all retrieval iterations, where F is computed from recall and precision (=-=Shaw, 1986-=-) by the formula, P R F 1 1 2 + = . (5) 3.1.2 Results The analysis of retrieval results by all evaluation measures used showed a consistent pattern of improved retrieval performance with the larger fe...

Citation Context

...ht of term k in the query, dik is the weight of term k in document i, and t is the number of terms in the index.sWe used SMART Lnu weights for document terms (Buckley, Singhal, Mitra, & Salton, 1996; =-=Buckley, Singhal, & Mitra, 1997-=-), and SMART ltc weights (Buckley, C., Salton, G., Allan, J., & Singhal, A., 1995) for query terms. Lnu weightssattempt to match the probability of retrieval given a document length with the probabili...

Citation Context

... dichotomous "relevant' or "nonrelevant," but they are not quite sure what a "marginally relevant" document should be. The question of what makes a document relevant is a fert=-=ile ground for research (Schamber, 1991-=-; Barry, 1994). In a prior research, we investigated the relationship between the proportionality and the degree of relevance and found that the number of relevant passages in a document corresponded ...

Citation Context

...ction fusion” problem, various strategies were employed to compensate for the potential incomparability of query-document similarity scores across collections (Voorhees, Gupta, & Johnson-Laird, 1995; =-=Savoy, Calve, & Vrajitoru, 1997-=-). Though the “raw score” merging method can be problematic when collection-dependent term weights (i.e. idf weight) cause the retrieval scores of similar documents to vary in different collections (D...

Citation Context

...reference relation from decision theory (Fishburn, 1970), is to find a solution vector that, given any two documents in the collection, will rank a morepreferred document before a less-preferred one (=-=Wong et al., 1988-=-). The goal of the adaptive linear model, in essence, is to construct a query vector that ranks the entire document collection according to the user’s preferences.sSince the user’s preferences are not...

Citation Context

...the interactive experiment to optimize performance with feedback. 2.4sFeedback Models 2.4.1sAdaptive Linear Model Currently, the default relevance feedback model of IRIS is the adaptive linear models(=-=Wong & Yao, 1990-=-; Wong, Yao, Salton, & Buckley, 1991).sThe basic approach of the adaptive linear model, which is based on the concept of the preference relation from decision theory (Fishburn, 1970), is to find a sol...

Citation Context

...nts. 2.2 Phrase Construction In our TREC-6 experiments, we constructed a statistically significant, two-word collocation index by extracting co-occurring word pairs within a window of 4 words (Haas & =-=Losee, 1994-=-; Losee, 1994) and selecting those that co-occur with statistically significant frequency (Berry-Rogghe, 1974). Though this collocation index worked very well in some cases, its overall effect on retr...

Citation Context

...o-word collocation index by extracting co-occurring word pairs within a window of 4 words (Haas & Losee, 1994; Losee, 1994) and selecting those that co-occur with statistically significant frequency (=-=Berry-Rogghe, 1974-=-). Though this collocation index worked very well in some cases, its overall effect on retrieval effectiveness did not appear to be significant (Sumner et. al., 1998). Furthermore, the computational c...

Citation Context

...ere a is a constant, and b is the difference vector resulting from subtracting a less-preferred document vector from a more preferred one. (For details about how this difference vector is chosen, see =-=Sumner et al., 1998-=-.) The choices for the constant a and the starting vector q (0) are very important since they can influence not only the composition of the solution vector but also the number of errorcorrection cycle...

Citation Context

...ractive track runs 1sIntroduction In our TREC-5 ad-hoc experiment, we tested two relevance feedback models, an adaptive linear model and a probabilistic model, using massive feedback query expansion (=-=Sumner & Shaw, 1997-=-).sFor our TREC-6 interactive experiment, we developed an interactive retrieval system called IRIS (Information Retrieval Interactive System1), which implemented modified versions of the feedback mode...

Citation Context

...periment to optimize performance with feedback. 2.4sFeedback Models 2.4.1sAdaptive Linear Model Currently, the default relevance feedback model of IRIS is the adaptive linear models(Wong & Yao, 1990; =-=Wong, Yao, Salton, & Buckley, 1991-=-).sThe basic approach of the adaptive linear model, which is based on the concept of the preference relation from decision theory (Fishburn, 1970), is to find a solution vector that, given any two doc...

Citation Context

...xperiments. 2.2sPhrase Construction In our TREC-6 experiments, we constructed a statistically significant, two-word collocation index by extracting co-occurring word pairs within a window of 4 words (=-=Haas & Losee, 1994-=-; Losee, 1994) and selecting those that co-occur with statistically significant frequency (Berry-Rogghe, 1974).sThough this collocation index worked very well in some cases, its overall effect on retr...

Citation Context

...ere α is a constant, and b is the difference vector resulting from subtracting a less-preferred document vector from a more preferred one.s(For details about how this difference vector is chosen, see =-=Sumner et al., 1998-=-.)sThe choices for the constant α and the starting vector q(0) are very important since they can influence not only the composition of the solution vector but also the number of errorcorrection cycles...

Citation Context

...ce or efficiency rather than content, or can contain subsections of various information content as in Congressional Record and Federal Register documents. The findings from passage feedback research (=-=Melucci, 1998-=-) as well as from comments made by IRIS users at large indicate that determination of relevance is sometimes based on certain portions of a document rather than the entirety of it. To test out this th...

Citation Context

...7). Though the "raw score" merging method can be problematic when collection-dependent term weights (i.e. idf weight) cause the retrieval scores of similar documents to vary in different col=-=lections (Dumais, 1993-=-; Voorhees et. al., 1995), we thought longer queries and massive retrieval window used for subcollection creation might mute its adverse effects. Any advantages gained by more complex retrieval strate...

Citation Context

...en the proportionality and the degree of relevance and found that the number of relevant passages in a document corresponded directly with the degree of relevance awarded to the document by the user (=-=Maglaughlin, Meho, Yang, & Tang, 1998-=-).sIf the proportionality of relevance is an important factor in determining the relevance of a document, then the relevance levels used by the system should be finely graded to better reflect the use...