In one embodiment, a method includes receiving an invoice. The invoice includes invoice positions and each of the invoice positions includes a position text. The method also includes converting each word and number of the position text of each invoice position to a word embedding vector, summing the word embedding vectors for each invoice position to generate a word vector for each invoice position, and concatenating the word vector and a number vector of each invoice position to generate a position vector for each invoice position. The method further includes generating a first combined position vector for a first invoice position, generating a second combined position vector for a second invoice position, generating an invoice vector by summing the first and second combined position vectors, comparing the invoice vector to a fraud detection parameter, and determining whether the invoice is indicative of fraud based on the comparison.