Generally, the sentence cluster number and the dimensionality in LSA are determined by the prediction risk.
Through the clustering process, each state can be modeled by a set of sentence clusters ( D{S}_n=left{{S}_{n,1},{S}_{n,2},dots, {S}_{n,{k}_n}right} ), where ( {S}_{n,{k}_n} ) represents the k n -th sentence cluster in DS n.
The methods for sentence clustering and error-tolerant sentence generation could be applied to improve the detection performance.
For each DS, first, we manually select k unique sentences as the centroids of the sentence clusters.
After sentence clustering, a total of 494 clusters were obtained and used as an alternative approach to statistically characterize the sentence patterns.
Clearly, joint estimation of LSA-based mapping between linguistic features and DSs as well as sentence clustering outperformed the individual approach and the traditional keyword-spotting method.
Figure 11 illustrates the detection performance for each individual DS (DS 1 ~ 37) using (1) keyword spotting, (2) the proposed approach using the verification data, and the proposed approach (3) without/(4) with sentence clustering.
Fig. 11 Detection performance for each individual DS. a DS 1~19 and b DS 20~37 using (1) keyword spotting, (2) the proposed approach using the verification data, and the proposed approach (3) without/(4) with sentence clustering.
During the phase of theme analysis, the entire text was reread and phrases and sentence clusters that seemed to be thematic were marked as meaning units.
In the selective reading, the entire text was reread and phrases and sentence clusters that seemed to be thematic were marked as meaning units.
Then, too, his non sequiturs lead, sooner or later, to a few consecutive sentences that cluster like a slippery ball of meaning.
