Overlapping rollapply on any matrix in R

I am working on creating feature vectors out of a matrix of words. The features I am looking at are the n words before and after the current word. I have a matrix where each row has the original word, the lemma and part-of-speech. This needs to be converted into a new matrix that contains the previous n lemmas and their part-of-speech together with the current lemma and part-of-speech. I tried to use rollapply from the zoo package, but either I didn't uderstand it properly or it isn't ment to be doing this sort of stuff.

Introduction

So what I want to do is to take a matrix and roll over it creating a new feature matrix that for each row uses the row at the original matrix and the surrouding rows to create each feature row.

Example

The type of matrix that I am interested in is one like the following one:

a	1	y
b	2	z
c	1	x

I want to take this matrix and convert it into a matrix that looks like this:

-	b	-	2	-	z	a	1	y
a	c	1	1	y	x	b	2	z
b	-	2	-	z	-	c	1	x

Here I use - as a placeholder when there is a value missing. What I have done is that for each row in the feature matrix the two first columns are the first column of the previous and the next row, the 3rd and 4th column is the second column of the previous and next row, the 4th and 5th column is the third column of the previous and next row and, the last 3 columns in a row in the feature matrix is the data from that row.

The function

To do all this we need a higher-order function. A function that takes a function and the data that we are interested in and applys a transformation on that data.

In this case we will take a matrix and a function that operates on the vector of the current row and the matrix of the rows around it.

rollapply <- function(m, FUN, n=2, padding=c("-", "-", "-")) {
    ## Create and add padding that goes around the input matrix.
    padding <- matrix(rep(padding, each=n), ncol=ncol(m))
    dat <- rbind(padding, m, padding)

    ## Remove column names from the matrix as those will just confuse.
    colnames(dat) <- NULL

    ## Create a vector we can use as a template to get the surrounding
    ## rows from the data matrix.
    surr <- c(-n:n)[-(n+1)]

    ## Start with the first row, and since we added some padding to
    ## the `dat` matrix the first row is the row at the size of the
    ## padding + 1.
    i <- n+1

    ## Since we do not know how long the resulting vector from the
    ## function is just yet we need to run the function on the first
    ## row before we create the resulting matrix.
    head <- FUN(dat[i,], dat[surr+i,])

    ## We create the resulting matrix. It has the same amount of rows
    ## that the input matrix had, but it might have more or less
    ## columns. We can get the number of columns from the length of
    ## the result from applying the function to the first row and its
    ## surrounding rows.
    ret <- matrix(nrow=nrow(m), ncol=length(head))
    ret[1,] <- head

    ## For the rest of the rows we do the same thing as we did for the
    ## first row and add it to the correct place in the return matrix,
    ## except for the last n rows as those are padding and shouldn't
    ## be a part of the result matrix.
    for (i in (i+1):(nrow(dat)-n)) {
        ## Since the return matrix doesn't have the padding we also
        ## need to remove n from from the index.
        ret[i-n,] <- FUN(dat[i,], dat[surr+i,])
    }
    ret
}

There probably are faster and easier ways of doing this, but this was the way that I discovered when I was working on this problem.