A large number of real-world observations by social sensors all over the world can be obtained from various social networking services. Especially, observations covering miscellaneous areas of interest are posted to Twitter as short text messages. Our goal is to extract a wide range of observations related to the target of interest specified by the user from Twitter regardless of their popularity. Assuming that the related observations are likely to contain words people often associate with each other, the associative relations among words are learned from the past messages. When a user gives a keyword representing his/her current interest, recent related observations are extracted based on their word frequency distributions in terms of the associative relations to the keyword.