Learning word representations by jointly modeling syntagmatic and paradigmatic relations