With　the　explosion　of　information,　it　is　becoming　increasingly　difficult　to　get　what　is　really　wanted.　Dimensionality　reduction　is　the　first　step　in　efficient　processing　of　large　data.　Although　dimensionality　can　be　reduced　in　many　ways,　little　work　has　been　done　to　achieve　dimensionality　reduction　without　changing　the　inner　semantic　relationship　among　high　dimension　data.　To　remedy　this　problem,　we　introduced　a　manifold　learning　based　method,　named　Mutual　information　preserving　mapping　(MIPM),　to　explore　the　low-dimensional,　neighborhood　and　mutual　information　preserving　embeddings　of　high　dimensional　inputs.　Experimental　results　show　that　the　proposed　method　is　effective　for　the　text　dimensionality　reduction　task.　The　MIPM　was　used　to　develop　a　temporal　summarization　system　for　efficiently　monitoring　the　information　associated　with　an　event　over　time.　With　respect　to　the　established　baselines,　results　of　these　experiments　show　that　our　method　is　effective　in　the　temporal　summarization.