On p.25 of book 1 (in the latest available online version dated back June 2023), in Section 1.5.4.2, it is stated that we often normalize each row of the TF-IDF matrix. According to the definition of TF-IDF in the book, i.e., $(TF-IDF)_{ij}$ refers to the frequency of the $i$-th term in the $j$-th document, normalizing each row corresponds to comparing (the occurrences of) all the words on the same scale.
Just wonder whether we actually want to normalize each column, instead of each row, of TF-IDF? This corresponds to comparing all the documents on the same scale, regardless of their lengths.
Also, there is some minute notation inconsistency in the following Sec. 1.5.4.3. Previous, the size of the vocabulary was denoted by $D$ (as what we do in most of the book), while here we switch to the undefined $V$.
On p.25 of book 1 (in the latest available online version dated back June 2023), in Section 1.5.4.2, it is stated that we often normalize each row of the TF-IDF matrix. According to the definition of TF-IDF in the book, i.e., $(TF-IDF)_{ij}$ refers to the frequency of the $i$-th term in the $j$-th document, normalizing each row corresponds to comparing (the occurrences of) all the words on the same scale.
Just wonder whether we actually want to normalize each column, instead of each row, of TF-IDF? This corresponds to comparing all the documents on the same scale, regardless of their lengths.
Also, there is some minute notation inconsistency in the following Sec. 1.5.4.3. Previous, the size of the vocabulary was denoted by $D$ (as what we do in most of the book), while here we switch to the undefined $V$.