When using cosine similarity in information retrieval, document vector length and query vector length are used for normalization. So if TF-IDF is used as a weighting function, then using raw frequency in TF is equal to normalizing it by document's length (mathematically speaking ).
My question is:
I am using log(TF*SomeValue + 1) as a local weight component instead of TF, where "SomeValue" is a calculated statistical property of the term. Using normalized TF in the formula gave me worse results than using raw frequency! So it's not the same in this case. How can I normalize term local weight to avoid bias toward lengthy documents?