### Table 2. F1 measures for all models (Naive Bayes, EM, and Hierarchical Dirichlet) on the eight benchmark tax- onomies. The Hierarchical Dirichlet results are presented both for the best smoothing parameter , and for a xed for all datasets.

2005

"... In PAGE 7: ... tors were also used to initialize both the standard EM and our regularized EM algorithms. In Table2 we present the classi cation results of the di erent approaches on the eight datasets. We observe that the standard EM algorithm performs poorly on several datasets.... In PAGE 7: ... Our smoothed variant of the EM algorithm per- forms signi cantly better than both the standard EM and the Naive Bayes classi er on almost all datasets. In Table2 we present the classi cation results for the best value of on each dataset as well as the results with a xed = 2. We observe that even when the chosen is not the best for the dataset, smoothed es- timates of the classi cation parameters result in con- siderably better accuracy than without smoothing.... ..."

Cited by 7

### TABLE I HIERARCHICAL TOPIC MODEL VS. FLAT MODEL

### Table 2. Process summary of hierarchical model

1999

Cited by 11

### Table 1. Terms in different topics. (The order of the topics is not relevant).

2000

"... In PAGE 50: ... In this way we can estimate the topic structure of the data, and also the topic probabilities si and topic-attribute probabili- ties q(i; A) of A in topic i. Comparing to the true probabil- ities, the mean squared errors (MSE) of topic probabilities and the MSEs of topic-attribute probabilities are listed in Table1 for quot; = 0, 0:01 and 0:1. These gures are averages of 10 experiments.... In PAGE 50: ...01 1:04 10 4 1:02 10 3 0.1 1:01 10 4 1:03 10 3 Table1 : Mean squared errors of estimated topic and topic-attribute probabilities in the ratio algorithm. In our varying-probability topic model, the topic proba- bilities si are randomly drawn for each document, and the ratio algorithm is not applicable.... In PAGE 62: ... We computed the lift statistics between all term pairs and used hierarchical average linkage clustering based on the inverses of lifts. Table1 shows how the terms are clustered into topics. The number of clusters (21) was chosen based on the distance between clusters being merged in the process of hierarchical clustering: until these 21 clusters, the intercluster distances were quite small but distances between the final 21 clusters were large.... ..."

Cited by 26

### Table 1. Estimated Dirichlet parameters.

"... In PAGE 6: ...33, corresponding to the weight of the region B and region A respectively. Table1 presents the Dirichlet parameters and their estimates of the two modes. Although the estimated values are near from the true parameter values, we note that this does not have to happen necessarily, since the same distribution can be modelled with difierent MODs.... ..."

### Table 1. Terms in different topics. (The order of the topics is not relevant).

2003

"... In PAGE 9: ... We computed the lift statistics between all term pairs and used hierarchical average linkage clustering based on the inverses of lifts. Table1 shows how the terms are clustered into topics. The number of clusters (21) was chosen based on the distance between clusters being merged in the process of hierarchical clustering: until these 21 clusters, the intercluster distances were quite small but distances between the final 21 clusters were large.... ..."

Cited by 1

### Table 1. Terms in different topics. (The order of the topics is not relevant).

2003

"... In PAGE 9: ... We computed the lift statistics between all term pairs and used hierarchical average linkage clustering based on the inverses of lifts. Table1 shows how the terms are clustered into topics. The number of clusters (21) was chosen based on the distance between clusters being merged in the process of hierarchical clustering: until these 21 clusters, the intercluster distances were quite small but distances between the final 21 clusters were large.... ..."

Cited by 1

### Table 2. Cluster-based retrieval using the CQL model and five query-specific hierarchical agglomerative clustering algorithms. The cluster language model parameter is set to Dirichlet smoothing at 1000 for both collections. The results at best threshold (t) are shown. Performance is measured in average precision.

2004

Cited by 35