This paper provides an overview of some techniques in our implementing a Web image indexing system. Our methodologies are based on the associated text of web images. The potential semantics of the image are derived from terms or concepts appearing in the associated text with respect to their frequencies. However, the relevance of a term to an image is often sensitive to its relative distance to the image. To solve this problem, in our approach, the whole associated text is partitioned into a sequence of semantic blocks (SB). The semantic fragmentation is based on the nested structures of the text with respect to the image. And the most internal HTML element which includes the image links or command is assigned with a highest relevant factor, and the most external HTML element which contains the image is assigned the lowest factor in deriving semantics of the embedded images. The total semantic relevance of a term is then calculated as the sum of its local semantic contributions through all of the semantic blocks. Then, an inverted index is constructed based on the semantic relevances of terms. Experiment shows that the semantic extraction methods based on our fragmentations can effectively improve recalls and precisions of Web image searches.