A question document is divided into predetermined areas, and it is judged whether each divided area is important, to thereby extract an important area. A reply example candidate likelihood value is calculated for each important area, the likelihood value indicating the degree representative of whether each reply example candidate corresponds to a question content. By using the reply example candidate likelihood value, important areas having similar meanings are combined to extract final important parts. A reply example candidate is selected for each important part from reply example candidates prepared beforehand. A reply example candidate reliability degree representative of certainty of each reply example candidate and a reply composition degree indicating whether it is necessary to compose a new reply are calculated, and by using these values, question documents are distributed to different operator terminals.