1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814
extern crate ibig;
extern crate num_traits;
extern crate rayon;
extern crate rustfst;
extern crate sesdiff;
extern crate simple_error;
use rayon::prelude::*;
use rustfst::prelude::*;
use sesdiff::shortest_edit_script;
use std::cmp::min;
use std::cmp::Ordering;
use std::collections::{BTreeMap, BTreeSet, HashMap};
use std::error::Error;
use std::fs::File;
use std::io::{BufRead, BufReader};
use std::str::FromStr;
use std::sync::Arc;
use std::time::SystemTime;
pub mod anahash;
pub mod cache;
pub mod confusables;
pub mod distance;
pub mod index;
pub mod iterators;
pub mod search;
pub mod test;
pub mod types;
pub mod vocab;
pub use crate::anahash::*;
pub use crate::cache::*;
pub use crate::confusables::*;
pub use crate::distance::*;
pub use crate::index::*;
pub use crate::iterators::*;
pub use crate::search::*;
pub use crate::types::*;
pub use crate::vocab::*;
/// An absolute maximum on the anagram distance, even for long inputs
const MAX_ANAGRAM_DISTANCE: u8 = 12;
/// An absolute maximum on the edit distance, even for long inputs
const MAX_EDIT_DISTANCE: u8 = 12;
/// The VariantModel is the most high-level model of analiticcl, it holds
/// all data required for variant matching.
pub struct VariantModel {
/// Maps Vocabulary IDs to their textual strings and other related properties
pub decoder: VocabDecoder,
/// Map strings to vocabulary IDs
pub encoder: VocabEncoder,
/// Defines the alphabet used for the variant model
pub alphabet: Alphabet,
///The main index, mapping anagrams to instances
pub index: AnaIndex,
///A secondary sorted index
///indices of the outer vector correspond to the length of an anagram (in chars) - 1
///Inner vector is always sorted
pub sortedindex: BTreeMap<u16, Vec<AnaValue>>,
/// Ngrams for simple context-sensitive language modelling
/// when finding the most probable sequence of variants
pub ngrams: HashMap<NGram, u32>,
///Total frequency, index corresponds to n-1 size, so this holds the total count for unigrams, bigrams, etc.
pub freq_sum: Vec<usize>,
/// Do we have frequency information for variant matching?
pub have_freq: bool,
/// Do we have an LM?
pub have_lm: bool,
/// Context rules
pub context_rules: Vec<ContextRule>,
/// Tags used by the context rules
pub tags: Vec<String>,
///Weights used in distance scoring
pub weights: Weights,
/// Stores the names of the loaded lexicons, they will be referenced by index from individual
/// items for provenance reasons
pub lexicons: Vec<String>,
/// Holds weighted confusable recipes that can be used in scoring and ranking
pub confusables: Vec<Confusable>,
///Process confusables before pruning by max_matches
pub confusables_before_pruning: bool,
pub debug: u8,
}
impl VariantModel {
/// Instantiate a new variant model
pub fn new(alphabet_file: &str, weights: Weights, debug: u8) -> VariantModel {
let mut model = VariantModel {
alphabet: Vec::new(),
encoder: HashMap::new(),
decoder: Vec::new(),
index: HashMap::new(),
sortedindex: BTreeMap::new(),
ngrams: HashMap::new(),
freq_sum: vec![0],
have_freq: false,
have_lm: false,
weights,
lexicons: Vec::new(),
confusables: Vec::new(),
confusables_before_pruning: false,
context_rules: Vec::new(),
tags: Vec::new(),
debug,
};
model
.read_alphabet(alphabet_file)
.expect("Error loading alphabet file");
init_vocab(&mut model.decoder, &mut model.encoder);
model
}
/// Instantiate a new variant model, explicitly passing an alphabet rather than loading one
/// from file.
pub fn new_with_alphabet(alphabet: Alphabet, weights: Weights, debug: u8) -> VariantModel {
let mut model = VariantModel {
alphabet: alphabet,
decoder: Vec::new(),
encoder: HashMap::new(),
index: HashMap::new(),
sortedindex: BTreeMap::new(),
ngrams: HashMap::new(),
freq_sum: vec![0],
have_freq: false,
have_lm: false,
weights,
lexicons: Vec::new(),
confusables: Vec::new(),
confusables_before_pruning: false,
context_rules: Vec::new(),
tags: Vec::new(),
debug,
};
init_vocab(&mut model.decoder, &mut model.encoder);
model
}
/// Configure the model to match against known confusables prior to pruning on maximum weight.
/// This may lead to better results but may have a significant performance impact.
pub fn set_confusables_before_pruning(&mut self) {
self.confusables_before_pruning = true;
}
/// Returns the size of the alphabet, this is typically +1 longer than the actual alphabet file
/// as it includes the UNKNOWN symbol.
pub fn alphabet_size(&self) -> CharIndexType {
self.alphabet.len() as CharIndexType + 1 //+1 for UNK
}
/// Get an item from the index or insert it if it doesn't exist yet
pub fn get_or_create_index<'a, 'b>(
&'a mut self,
anahash: &'b AnaValue,
) -> &'a mut AnaIndexNode {
if self.contains_key(anahash) {
self.index
.get_mut(anahash)
.expect("get_mut on node after check")
} else {
self.index.insert(
anahash.clone(),
AnaIndexNode {
instances: Vec::new(),
charcount: anahash.char_count(self.alphabet_size()),
},
);
self.index
.get_mut(&anahash)
.expect("get_mut on node after insert")
}
}
/// Build the anagram index (and secondary index) so the model
/// is ready for variant matching
pub fn build(&mut self) {
eprintln!("Computing anagram values for all items in the lexicon...");
// Hash all strings in the lexicon
// and add them to the index
let mut tmp_hashes: Vec<(AnaValue, VocabId)> = Vec::with_capacity(self.decoder.len());
for (id, value) in self.decoder.iter().enumerate() {
if value.vocabtype.check(VocabType::INDEXED) {
//get the anahash
let anahash = value.text.anahash(&self.alphabet);
if self.debug >= 2 {
eprintln!(
" -- Anavalue={} VocabId={} Text={}",
&anahash, id, value.text
);
}
tmp_hashes.push((anahash, id as VocabId));
}
}
eprintln!(" - Found {} instances", tmp_hashes.len());
eprintln!("Adding all instances to the index...");
self.index.clear();
for (anahash, id) in tmp_hashes {
//add it to the index
let node = self.get_or_create_index(&anahash);
node.instances.push(id);
}
eprintln!(" - Found {} anagrams", self.index.len());
eprintln!("Creating sorted secondary index...");
self.sortedindex.clear();
for (anahash, node) in self.index.iter() {
if !self.sortedindex.contains_key(&node.charcount) {
self.sortedindex.insert(node.charcount, Vec::new());
}
let keys = self
.sortedindex
.get_mut(&node.charcount)
.expect("getting sorted index (1)");
keys.push(anahash.clone()); //TODO: see if we can make this a reference later
}
eprintln!("Sorting secondary index...");
let mut sizes: Vec<u16> = self.sortedindex.keys().map(|x| *x).collect();
sizes.sort();
for size in sizes {
let keys = self
.sortedindex
.get_mut(&size)
.expect("getting sorted index (2)");
keys.sort();
eprintln!(" - Found {} anagrams of length {}", keys.len(), size);
}
eprintln!("Constructing Language Model...");
//extra unigrams extracted from n-grams that need to be added to the vocabulary decoder
let mut unseen_parts: Option<VocabEncoder> = Some(VocabEncoder::new());
for id in 0..self.decoder.len() {
if self
.decoder
.get(id)
.expect("item")
.vocabtype
.check(VocabType::LM)
{
//get the ngram and find any unseen parts
if let Ok(ngram) = self.into_ngram(id as VocabId, &mut unseen_parts) {
let freq = self.decoder.get(id).unwrap().frequency;
if ngram.len() > 1 {
//reserve the space for the total counts
for _ in self.freq_sum.len()..ngram.len() {
self.freq_sum.push(0);
}
//add to the totals for this order of ngrams
self.freq_sum[ngram.len() - 1] += freq as usize;
} else {
self.freq_sum[0] += freq as usize;
}
self.add_ngram(ngram, freq);
}
}
}
if let Some(unseen_parts) = unseen_parts {
//add collected unseen n-gram parts to the decoder
for (part, id) in unseen_parts {
self.add_ngram(NGram::UniGram(id), 1);
self.encoder.insert(part.clone(), id);
self.decoder.push(VocabValue::new(part, VocabType::LM));
}
}
if self.ngrams.is_empty() {
eprintln!(" - No language model provided");
self.have_lm = false;
} else {
eprintln!(
" - Found {} n-grams for language modelling",
self.ngrams.len()
);
self.have_lm = true;
}
}
/// Tests if the anagram value exists in the index
pub fn contains_key(&self, key: &AnaValue) -> bool {
self.index.contains_key(key)
}
///Get all anagram instances for a specific entry
pub fn get_anagram_instances(&self, text: &str) -> Vec<&VocabValue> {
let anavalue = text.anahash(&self.alphabet);
let mut instances: Vec<&VocabValue> = Vec::new();
if let Some(node) = self.index.get(&anavalue) {
for vocab_id in node.instances.iter() {
instances.push(
self.decoder
.get(*vocab_id as usize)
.expect("vocab from decoder"),
);
}
}
instances
}
///Get an exact item in the lexicon (if it exists)
pub fn get(&self, text: &str) -> Option<&VocabValue> {
for instance in self.get_anagram_instances(text) {
if instance.text == text {
return Some(instance);
}
}
None
}
///Tests if the lexicon has a specific entry, by text
pub fn has(&self, text: &str) -> bool {
for instance in self.get_anagram_instances(text) {
if instance.text == text {
return true;
}
}
false
}
///Resolves a vocabulary ID
pub fn get_vocab(&self, vocab_id: VocabId) -> Option<&VocabValue> {
self.decoder.get(vocab_id as usize)
}
/// Decomposes and decodes and anagram value into the characters that make it up.
/// Mostly intended for debugging purposes.
pub fn decompose_anavalue(&self, av: &AnaValue) -> Vec<&str> {
let mut result = Vec::new();
for c in av.iter(self.alphabet_size()) {
result.push(
self.alphabet
.get(c.0.charindex as usize)
.expect("alphabet item must exist")
.get(0)
.unwrap()
.as_str(),
);
}
result
}
///Read the alphabet from a TSV file
///The file contains one alphabet entry per line, but may
///consist of multiple tab-separated alphabet entries on that line, which
///will be treated as the identical.
///The alphabet is not limited to single characters but may consist
///of longer string, a greedy matching approach will be used so order
///matters (but only for this)
pub fn read_alphabet(&mut self, filename: &str) -> Result<(), std::io::Error> {
if self.debug >= 1 {
eprintln!("Reading alphabet from {}...", filename);
}
let f = File::open(filename)?;
let f_buffer = BufReader::new(f);
for line in f_buffer.lines() {
if let Ok(line) = line {
if !line.is_empty() {
let fields = line
.split("\t")
.filter_map(|x| match x {
"\\s" => Some(" ".to_owned()),
"\\t" => Some("\t".to_owned()),
"\\n" => Some("\n".to_owned()),
_ => {
if x.trim().is_empty() {
None
} else {
Some(x.trim().to_owned())
}
}
})
.collect();
self.alphabet.push(fields);
}
}
}
if self.debug >= 2 {
eprintln!(" -- Read alphabet of size {}", self.alphabet.len());
for (i, items) in self.alphabet.iter().enumerate() {
let av = AnaValue::character(i as CharIndexType);
eprintln!(" -- #{} -> {} - {:?}", i, av, items);
}
} else if self.debug >= 1 {
eprintln!(" -- Read alphabet of size {}", self.alphabet.len());
}
Ok(())
}
///Read a confusiblelist from a TSV file
///Contains edit scripts in the first columned (formatted in sesdiff style)
///and optionally a weight in the second column.
///favourable confusables have a weight > 1.0, unfavourable ones are < 1.0 (penalties)
///Weight values should be relatively close to 1.0 as they are applied to the entire score
pub fn read_confusablelist(&mut self, filename: &str) -> Result<(), std::io::Error> {
if self.debug >= 1 {
eprintln!("Reading confusables from {}...", filename);
}
let f = File::open(filename)?;
let f_buffer = BufReader::new(f);
for line in f_buffer.lines() {
if let Ok(line) = line {
if !line.is_empty() {
let fields: Vec<&str> = line.split("\t").collect();
let weight = if fields.len() >= 2 {
fields
.get(1)
.unwrap()
.parse::<f64>()
.expect("score should be a float")
} else {
1.0
};
self.add_to_confusables(fields.get(0).unwrap(), weight)?;
}
}
}
if self.debug >= 1 {
eprintln!(" -- Read {} confusables", self.confusables.len());
}
Ok(())
}
/// Add a confusable
pub fn add_to_confusables(
&mut self,
editscript: &str,
weight: f64,
) -> Result<(), std::io::Error> {
let confusable = Confusable::new(editscript, weight)?;
self.confusables.push(confusable);
Ok(())
}
/// Add a (weighted) variant to the model, referring to a reference that already exists in
/// the model.
/// Variants will be added
/// to the lexicon automatically when necessary. Set VocabType::TRANSPARENT
/// if you want variants to only be used as an intermediate towards items that
/// have already been added previously through a more authoritative lexicon.
pub fn add_variant(
&mut self,
ref_id: VocabId,
variant: &str,
score: f64,
freq: Option<u32>,
params: &VocabParams,
) -> bool {
let variantid = self.add_to_vocabulary(variant, freq, ¶ms);
self.add_variant_by_id(ref_id, variantid, score)
}
/// Add a (weighted) variant to the model, referring to a reference that already exists in
/// the model.
/// Variants will be added
/// to the lexicon automatically when necessary. Set VocabType::TRANSPARENT
/// if you want variants to only be used as an intermediate towards items that
/// have already been added previously through a more authoritative lexicon.
pub fn add_variant_by_id(&mut self, ref_id: VocabId, variantid: VocabId, score: f64) -> bool {
if variantid != ref_id {
//link reference to variant
if let Some(vocabvalue) = self.decoder.get_mut(ref_id as usize) {
let variantref = VariantReference::ReferenceFor((variantid, score));
if vocabvalue.variants.is_none() {
vocabvalue.variants = Some(vec![variantref]);
} else if let Some(variantrefs) = vocabvalue.variants.as_mut() {
//only add if it doesn't already exists (only first mention counts, regardless of score)
if !variantrefs.iter().any(|x| match x {
VariantReference::ReferenceFor((y, _)) => variantid == *y,
_ => false,
}) {
variantrefs.push(variantref);
}
}
}
//link variant to reference
if let Some(vocabvalue) = self.decoder.get_mut(variantid as usize) {
let variantref = VariantReference::VariantOf((ref_id, score));
if vocabvalue.variants.is_none() {
vocabvalue.variants = Some(vec![variantref]);
} else if let Some(variantrefs) = vocabvalue.variants.as_mut() {
//only add if it doesn't already exists (only first mention counts, regardless of score)
if !variantrefs.iter().any(|x| match x {
VariantReference::VariantOf((y, _)) => variantid == *y,
_ => false,
}) {
variantrefs.push(variantref);
}
}
}
true
} else {
false
}
}
///Read vocabulary (a lexicon or corpus-derived lexicon) from a TSV file
///May contain frequency information
///The parameters define what value can be read from what column
pub fn read_vocabulary(
&mut self,
filename: &str,
params: &VocabParams,
) -> Result<(), std::io::Error> {
if self.debug >= 1 {
eprintln!(
"Reading vocabulary #{} from {} ({:?})...",
self.lexicons.len() + 1,
filename,
params.vocab_type
);
}
let beginlen = self.decoder.len();
let f = File::open(filename)?;
let f_buffer = BufReader::new(f);
let mut params = params.clone();
params.index = self.lexicons.len() as u8;
for line in f_buffer.lines() {
if let Ok(line) = line {
if !line.is_empty() {
let fields: Vec<&str> = line.split("\t").collect();
let text = fields
.get(params.text_column as usize)
.expect("Expected text column not found");
let frequency = if let Some(freq_column) = params.freq_column {
if params.vocab_type.check(VocabType::INDEXED) {
self.have_freq = true;
}
fields
.get(freq_column as usize)
.unwrap_or(&"1")
.parse::<u32>()
.expect("frequency should be a valid integer")
} else {
1
};
self.add_to_vocabulary(text, Some(frequency), ¶ms);
}
}
}
if self.debug >= 1 {
eprintln!(
" - Read vocabulary of size {}",
self.decoder.len() - beginlen
);
}
self.lexicons.push(filename.to_string());
Ok(())
}
pub fn read_contextrules(&mut self, filename: &str) -> Result<(), std::io::Error> {
if self.debug >= 1 {
eprintln!("Reading context rules {}...", filename);
}
let f = File::open(filename)?;
let f_buffer = BufReader::new(f);
let mut linenr = 0;
for line in f_buffer.lines() {
if let Ok(line) = line {
linenr += 1;
if !line.is_empty() && !line.starts_with('#') {
let fields: Vec<&str> = line.split("\t").collect();
if fields.len() < 2 {
return Err(std::io::Error::new(
std::io::ErrorKind::Other,
format!(
"Expected at least two columns in context rules file {}, line {}",
filename, linenr
),
));
}
let pattern: &str = fields.get(0).unwrap();
if pattern.is_empty() {
continue;
}
let score = fields.get(1).unwrap().parse::<f32>();
if let Err(_) = score {
return Err(std::io::Error::new(std::io::ErrorKind::Other, format!("context rule score should be a floating point value above or below 1.0, got {} ({}, line {})", fields.get(1).unwrap(), filename,linenr)));
}
let score = score.unwrap();
let tag: Vec<&str> = match fields.get(2) {
Some(s) => s
.split(";")
.filter_map(|w| {
let w = w.trim();
if w.is_empty() {
None
} else {
Some(w)
}
})
.collect(),
None => Vec::new(),
};
let mut tagoffset: Vec<&str> = match fields.get(3) {
Some(s) => s
.split(";")
.filter_map(|w| {
let w = w.trim();
if w.is_empty() {
None
} else {
Some(w)
}
})
.collect(),
None => Vec::new(),
};
if tag.len() == 1 && tagoffset.len() == 0 {
tagoffset.push("0:");
} else if tag.len() != tagoffset.len() {
return Err(std::io::Error::new(std::io::ErrorKind::Other, format!("Multiple tags are specified for a context rule, expected the same number of tag offsets! (semicolon separated) ({}, line {})", filename, linenr)));
}
if let Err(error) = self.add_contextrule(pattern, score, tag, tagoffset) {
return Err(std::io::Error::new(
std::io::ErrorKind::Other,
format!(
"Error adding context rule: {} ({}, line {})",
error, filename, linenr
),
));
}
}
}
}
if self.debug >= 1 {
eprintln!(" -- Read {} context rules", self.context_rules.len());
}
Ok(())
}
pub fn add_contextrule(
&mut self,
pattern: &str,
score: f32,
tag: Vec<&str>,
tagoffset: Vec<&str>,
) -> Result<(), std::io::Error> {
let expressions: Vec<&str> = pattern.split(";").map(|s| s.trim()).collect();
let mut pattern: Vec<PatternMatch> = Vec::new();
for expr in expressions {
match PatternMatch::parse(expr, &self.lexicons, &self.encoder) {
Ok(pm) => pattern.push(pm),
Err(err) => {
return Err(std::io::Error::new(
std::io::ErrorKind::Other,
format!("Error parsing context rule: {}", err),
))
}
}
}
let mut errmsg: Option<&str> = None;
let tag: Vec<u16> = tag
.iter()
.map(|tag| {
if tag.is_empty() {
errmsg = Some("tag is empty");
}
let mut pos = None;
for (i, t) in self.tags.iter().enumerate() {
if t == tag {
pos = Some(i as u16);
break;
}
}
if pos.is_none() {
self.tags.push(tag.to_string());
(self.tags.len() - 1) as u16
} else {
pos.unwrap()
}
})
.collect();
if let Some(errmsg) = errmsg {
return Err(std::io::Error::new(std::io::ErrorKind::Other, errmsg));
}
let mut error: Option<&str> = None;
let mut tagoffset: Vec<(u8, u8)> = tagoffset
.iter()
.map(|s| {
let fields: Vec<&str> = s.split(":").collect();
let tagbegin: u8 = if let Some(tagbegin) = fields.get(0) {
if tagbegin.is_empty() {
0
} else {
match tagbegin.parse::<u8>() {
Ok(x) => x,
Err(_) => {
error = Some("tag offset should be an integer");
0
}
}
}
} else {
0
};
let taglength: u8 = if let Some(taglength) = fields.get(1) {
if taglength.is_empty() {
pattern.len() as u8 - tagbegin
} else {
match taglength.parse::<u8>() {
Ok(x) => x,
Err(_) => {
error = Some("tag length should be an integer");
0
}
}
}
} else {
pattern.len() as u8 - tagbegin
};
(tagbegin, taglength)
})
.collect();
if let Some(error) = error {
return Err(std::io::Error::new(std::io::ErrorKind::Other, error));
}
while tagoffset.len() < tag.len() {
tagoffset.push((0, pattern.len() as u8));
}
if !pattern.is_empty() {
self.context_rules.push(ContextRule {
pattern,
score,
tag,
tagoffset,
});
}
Ok(())
}
///Read a weighted variant list from a TSV file. Contains a canonical/reference form in the
///first column, and variants with score (two columns) in the following columns. May also
///contain frequency information (auto detected), in which case the first column has the
///canonical/reference form, the second column the frequency, and all further columns hold
///variants, their score and their frequency (three columns).
///Consumes much more memory than equally weighted variants.
pub fn read_variants(
&mut self,
filename: &str,
params: Option<&VocabParams>,
transparent: bool,
) -> Result<(), std::io::Error> {
let params = if let Some(params) = params {
let mut p = params.clone();
p.index = self.lexicons.len() as u8;
p
} else {
VocabParams {
index: self.lexicons.len() as u8,
..Default::default()
}
};
let transparent_params = if transparent {
let mut p = params.clone();
p.vocab_type |= VocabType::TRANSPARENT;
p
} else {
params.clone()
};
if self.debug >= 1 {
eprintln!("Reading variants from {}...", filename);
}
let mut count = 0;
let mut has_freq = None;
let f = File::open(filename)?;
let f_buffer = BufReader::new(f);
for (linenr, line) in f_buffer.lines().enumerate() {
let linenr = linenr + 1;
if let Ok(line) = line {
if !line.is_empty() {
let fields: Vec<&str> = line.split("\t").collect();
let reference = fields.get(0).expect(
format!(
"reference item (line {}, column 1, of {})",
linenr, filename
)
.as_str(),
);
let freq = if has_freq.is_none() {
//autodetect whether we have frequency information or not
if (fields.len() - 2) % 3 == 0 {
let freq = fields.get(1).expect("second field");
match freq.parse::<u32>() {
Ok(freq) => {
has_freq = Some(true);
Some(freq)
}
_ => None,
}
} else {
//number of columns not consistent with holding frequency information
has_freq = Some(false);
None
}
} else if has_freq == Some(true) {
let freq = fields.get(1).expect("score of reference item");
Some(
freq.parse::<u32>().expect(
format!(
"Frequency must be an integer (line {}, column 2, of {})",
linenr, filename
)
.as_str(),
),
)
} else {
None
};
let ref_id = self.add_to_vocabulary(reference, freq, ¶ms);
let mut iter = fields.iter();
if has_freq == Some(true) {
iter.next();
iter.next();
while let (Some(variant), Some(score), Some(freq)) =
(iter.next(), iter.next(), iter.next())
{
let score = score.parse::<f64>().expect(format!("Variant scores must be a floating point value (line {} of {}, got {} instead), also parsing frequency", linenr, filename, score).as_str());
let freq = freq.parse::<u32>().expect(format!("Variant frequency must be an integer (line {} of {}), got {} instead", linenr, filename, freq).as_str());
if self.add_variant(
ref_id,
variant,
score,
Some(freq),
if transparent {
&transparent_params
} else {
¶ms
},
) {
count += 1;
}
}
} else {
iter.next();
while let (Some(variant), Some(score)) = (iter.next(), iter.next()) {
let score = score.parse::<f64>().expect(format!("Variant scores must be a floating point value (line {} of {}, got {}), no frequency information", linenr, filename, score).as_str());
if self.add_variant(
ref_id,
variant,
score,
None,
if transparent {
&transparent_params
} else {
¶ms
},
) {
count += 1;
}
}
}
}
}
}
if self.debug >= 1 {
eprintln!(" - Read weighted variants list, added {} references", count);
}
self.lexicons.push(filename.to_string());
Ok(())
}
/// Adds an entry in the vocabulary
pub fn add_to_vocabulary(
&mut self,
text: &str,
frequency: Option<u32>,
params: &VocabParams,
) -> VocabId {
let frequency = frequency.unwrap_or(1);
if self.debug >= 2 {
eprintln!(" -- Adding to vocabulary: {} ({})", text, frequency);
}
if let Some(vocab_id) = self.encoder.get(text) {
let item = self.decoder.get_mut(*vocab_id as usize).expect(&format!(
"Retrieving existing vocabulary entry {}",
vocab_id
));
match params.freq_handling {
FrequencyHandling::Sum => {
item.frequency += frequency;
}
FrequencyHandling::Max => {
if frequency > item.frequency {
item.frequency = frequency;
};
}
FrequencyHandling::Min => {
if frequency < item.frequency {
item.frequency = frequency;
};
}
FrequencyHandling::Replace => {
item.frequency = frequency;
}
}
if vocab_id == &BOS || vocab_id == &EOS || vocab_id == &UNK {
item.vocabtype = VocabType::LM; //by definition
} else if item.vocabtype.check(VocabType::TRANSPARENT)
&& !params.vocab_type.check(VocabType::TRANSPARENT)
{
//we can lose the transparency flag if a later lexicon doesn't provide it
item.vocabtype ^= VocabType::TRANSPARENT;
}
item.lexindex |= 1 << params.index;
if self.debug >= 3 {
eprintln!(
" (updated) freq={}, lexindex+={}",
item.frequency, params.index
);
}
*vocab_id
} else {
//item is new
self.encoder
.insert(text.to_string(), self.decoder.len() as u64);
self.decoder.push(VocabValue {
text: text.to_string(),
norm: text.normalize_to_alphabet(&self.alphabet),
frequency: frequency,
tokencount: text.chars().filter(|c| *c == ' ').count() as u8 + 1,
lexindex: 1 << params.index,
variants: None,
vocabtype: params.vocab_type,
});
if self.debug >= 3 {
eprintln!(" (new) lexindex={}", params.index);
}
self.decoder.len() as VocabId - 1
}
}
/// Find variants in the vocabulary for a given string (in its totality), returns a vector of vocabulary ID and score pairs
/// Returns a vector of three-tuples (VocabId, distance_score, freq_score)
/// The resulting vocabulary Ids can be resolved through `get_vocab()`
pub fn find_variants(&self, input: &str, params: &SearchParameters) -> Vec<VariantResult> {
if self.index.is_empty() {
eprintln!("ERROR: Model has not been built yet! Call build() before find_variants()");
return vec![];
}
//Compute the anahash
let normstring = input.normalize_to_alphabet(&self.alphabet);
let anahash = input.anahash(&self.alphabet);
let max_anagram_distance: u8 = match params.max_anagram_distance {
DistanceThreshold::Ratio(x) => min(
(normstring.len() as f32 * x).floor() as u8,
MAX_ANAGRAM_DISTANCE, //absolute maximum as a safeguard
),
DistanceThreshold::RatioWithLimit(x, limit) => {
min((normstring.len() as f32 * x).floor() as u8, limit)
}
DistanceThreshold::Absolute(x) => min(
x,
(normstring.len() as f64 / 2.0).floor() as u8, //we still override the absolute threshold when dealing with very small inputs
),
};
//Compute neighbouring anahashes and find the nearest anahashes in the model
let anahashes =
self.find_nearest_anahashes(&anahash, max_anagram_distance, params.stop_criterion);
let max_edit_distance: u8 = match params.max_edit_distance {
DistanceThreshold::Ratio(x) => min(
(normstring.len() as f32 * x).floor() as u8,
MAX_EDIT_DISTANCE, //absolute maximum as a safeguard
),
DistanceThreshold::RatioWithLimit(x, limit) => {
min((normstring.len() as f32 * x).floor() as u8, limit)
}
DistanceThreshold::Absolute(x) => min(
x,
(normstring.len() as f64 / 2.0).floor() as u8, //we still override the absolute threshold when dealing with very small inputs
),
};
//Get the instances pertaining to the collected hashes, within a certain maximum distance
//and compute distances
let variants = self.gather_instances(&anahashes, &normstring, input, max_edit_distance);
self.score_and_rank(
variants,
input,
normstring.len(),
params.max_matches,
params.score_threshold,
params.cutoff_threshold,
params.freq_weight,
)
}
///Auxiliary function used by [`learn_variants()`], abstracts over strict mode
fn find_variants_for_learning<'a>(
&self,
inputstr: &'a str,
params: &SearchParameters,
strict: bool,
) -> Vec<(&'a str, VariantResult)> {
if strict {
self.find_variants(inputstr, params)
.into_iter()
.map(|result| (inputstr, result))
.collect()
} else {
self.find_all_matches(inputstr, params)
.iter()
.filter_map(|result_match| {
if let Some(variants) = &result_match.variants {
if let Some(selected) = result_match.selected {
if let Some(result) = variants.get(selected) {
return Some((result_match.text, result.clone()));
}
}
}
None
})
.collect()
}
}
/// Processes input and finds variants (like [`find_variants()`]), but all variants that are found (which meet
/// the set thresholds) will be stored in the model rather than returned. Unlike `find_variants()`, this is
/// invoked with an iterator over multiple inputs and returns no output by itself. It
/// will automatically apply parallellisation.
pub fn learn_variants<'a, I>(
&mut self,
input: I,
params: &SearchParameters,
strict: bool,
auto_build: bool,
) -> usize
where
I: IntoParallelIterator<Item = &'a String> + IntoIterator<Item = &'a String>,
{
if self.debug >= 1 {
eprintln!("(Learning variants)");
}
let vocabparams = VocabParams::default()
.with_vocab_type(VocabType::TRANSPARENT)
.with_freq_handling(FrequencyHandling::Max);
let mut all_variants: Vec<Vec<(&'a str, VariantResult)>> = Vec::new();
if params.single_thread {
all_variants.extend(input.into_iter().map(|inputstr| {
self.find_variants_for_learning(inputstr.as_str(), params, strict)
}));
} else {
all_variants.par_extend(input.into_par_iter().map(|inputstr| {
self.find_variants_for_learning(inputstr.as_str(), params, strict)
}));
}
if self.debug >= 1 {
eprintln!(
"(adding variants over {} input items to the model)",
all_variants.len()
);
}
let mut count = 0;
let mut prev = None;
for (inputstr, result) in all_variants.into_iter().flatten() {
//get a vocabulary id for the input string;
//adding it to the vocabulary if it does not exist yet
let vocab_id = if let Some(vocab_id) = self.encoder.get(inputstr) {
//item exists
let vocabitem = self
.decoder
.get_mut(*vocab_id as usize)
.expect("item must exist");
//is this the first occurrence in a consecutive sequence?
if prev != Some(inputstr) {
//then increment the frequency
vocabitem.frequency += 1;
}
*vocab_id
} else {
//item is new
self.add_to_vocabulary(inputstr, Some(1), &vocabparams)
};
if result.vocab_id != vocab_id {
//ensure we don't add exact matches
if self.add_variant_by_id(result.vocab_id, vocab_id, result.dist_score) {
count += 1;
}
}
prev = Some(inputstr);
}
if self.debug >= 1 {
eprintln!("(added {} variants)", count);
}
if auto_build {
if self.debug >= 1 {
eprintln!("((re)building the model)");
}
self.build();
}
count
}
/// Find the nearest anahashes that exists in the model (computing anahashes in the
/// neigbhourhood if needed).
pub(crate) fn find_nearest_anahashes<'a>(
&'a self,
focus: &AnaValue,
max_distance: u8,
stop_criterion: StopCriterion,
) -> BTreeSet<&'a AnaValue> {
let mut nearest: BTreeSet<&AnaValue> = BTreeSet::new();
let begintime = if self.debug >= 2 {
eprintln!("(finding nearest anagram matches for focus anavalue {}, max_distance={}, stop_criterion={:?})", focus, max_distance, stop_criterion);
Some(SystemTime::now())
} else {
None
};
if let Some((matched_anahash, node)) = self.index.get_key_value(focus) {
//the easiest case, this anahash exists in the model!
if self.debug >= 2 {
eprintln!(" (found exact match)");
}
nearest.insert(matched_anahash);
if StopCriterion::StopAtExactMatch == stop_criterion {
for vocab_id in node.instances.iter() {
if let Some(_) = self.decoder.get(*vocab_id as usize) {
if self.debug >= 2 {
eprintln!(" (stopping early)");
}
return nearest;
}
}
}
}
let (focus_upper_bound, focus_charcount) = focus.alphabet_upper_bound(self.alphabet_size());
let focus_alphabet_size = focus_upper_bound + 1;
// Gather lookups to match against the secondary index
// keys correspond to the number of characters
// We first gather all lookups rather than doing them immediately,
// so we need to iterate over the secondary index only once, which
// has a slight performance benefit
let mut lookups: HashMap<u8, Vec<AnaValue>> = HashMap::new();
//Find anagrams reachable through insertions within the the maximum distance
for distance in 1..=max_distance {
let search_charcount = focus_charcount + distance as u16;
if let Some(lookups) = lookups.get_mut(&(search_charcount as u8)) {
lookups.push(focus.clone());
} else {
lookups.insert(search_charcount as u8, vec![focus.clone()]);
}
if self.debug >= 3 {
eprintln!(
" (scheduling finding insertion at distance {}, charcount {})",
distance, search_charcount
);
}
}
let searchparams = SearchParams {
max_distance: Some(max_distance as u32),
breadthfirst: true,
allow_empty_leaves: false,
allow_duplicates: false,
..Default::default()
};
/*let iterator = if let Some(cache) = cache {
focus.iter_recursive_external_cache(focus_alphabet_size+1, &searchparams, cache)
} else {*/
let iterator = focus.iter_recursive(focus_alphabet_size + 1, &searchparams);
/*};*/
// Do a breadth first search for deletions
for (deletion, distance) in iterator {
if self.debug >= 3 {
eprintln!(
" (testing deletion at distance {}, charcount {}: anavalue {})",
distance,
focus_charcount as u32 - distance,
deletion.value
);
if self.debug >= 4 {
let decomposed: String = self.decompose_anavalue(&deletion.value).join("");
eprintln!(" (anavalue decomposition: {})", decomposed);
}
}
if let Some((matched_anahash, _node)) = self.index.get_key_value(&deletion) {
if self.debug >= 3 {
eprintln!(" (deletion matches; anagram exists in index)");
}
//This deletion exists in the model
nearest.insert(matched_anahash);
}
let deletion_charcount = focus_charcount - distance as u16;
if self.debug >= 3 {
eprintln!(
" (scheduling search for insertions from deletion result anavalue {})",
deletion.value
);
}
//Find possible insertions starting from this deletion
for search_distance in 1..=(max_distance as u16 - distance as u16) {
let search_charcount = deletion_charcount + search_distance;
if self.debug >= 3 {
eprintln!(
" (search_distance={}, search_charcount={})",
search_distance, search_charcount
);
}
if let Some(lookups) = lookups.get_mut(&(search_charcount as u8)) {
lookups.push(deletion.value.clone());
} else {
lookups.insert(search_charcount as u8, vec![deletion.value.clone()]);
}
}
}
if self.debug >= 2 {
eprintln!("(finding all insertions)");
}
let mut count = 0;
let beginlength = nearest.len();
for (search_charcount, anavalues) in lookups.iter() {
if let Some(sortedindex) = self.sortedindex.get(&(*search_charcount as u16)) {
for candidate in sortedindex.iter() {
for av in anavalues {
if candidate.contains(&av) {
//this is where the magic happens
count += 1;
nearest.insert(candidate);
break;
}
}
}
}
}
if self.debug >= 2 {
eprintln!(
" (added {} out of {} candidates, preventing duplicates)",
nearest.len() - beginlength,
count
);
}
if self.debug >= 2 {
let endtime = SystemTime::now();
let duration = endtime
.duration_since(begintime.expect("begintime"))
.expect("clock can't go backwards")
.as_micros();
eprint!(
"(found {} anagram matches in total (in {} μs) for focus anavalue {}: ",
nearest.len(),
duration,
focus
);
for av in nearest.iter() {
eprint!(" {}", av);
}
eprintln!(")");
}
nearest
}
/// Gather instances with their edit distances and frequency, given a search string (normalised to the alphabet) and anagram hashes
pub(crate) fn gather_instances(
&self,
nearest_anagrams: &BTreeSet<&AnaValue>,
querystring: &[u8],
query: &str,
max_edit_distance: u8,
) -> Vec<(VocabId, Distance)> {
let mut found_instances = Vec::new();
let mut pruned_instances = 0;
let begintime = if self.debug >= 2 {
Some(SystemTime::now())
} else {
None
};
for anahash in nearest_anagrams {
let node = self
.index
.get(anahash)
.expect("all anahashes from nearest_anagrams must occur in the index");
for vocab_id in node.instances.iter() {
let vocabitem = self
.decoder
.get(*vocab_id as usize)
.expect("vocabulary id must exist in the decoder");
if self.debug >= 4 {
eprintln!(
" (comparing query {} with instance {})",
query, vocabitem.text
)
}
if let Some(ld) =
damerau_levenshtein(querystring, &vocabitem.norm, max_edit_distance)
{
if self.debug >= 4 {
eprintln!(" (ld={})", ld);
}
//we only get here if we make the max_edit_distance cut-off
let distance = Distance {
ld: ld,
lcs: if self.weights.lcs > 0.0 {
longest_common_substring_length(querystring, &vocabitem.norm)
} else {
0
},
prefixlen: if self.weights.prefix > 0.0 {
common_prefix_length(querystring, &vocabitem.norm)
} else {
0
},
suffixlen: if self.weights.suffix > 0.0 {
common_suffix_length(querystring, &vocabitem.norm)
} else {
0
},
samecase: if self.weights.case > 0.0 {
vocabitem
.text
.chars()
.next()
.expect("first char")
.is_lowercase()
== query.chars().next().expect("first char").is_lowercase()
} else {
true
},
};
//match will be added to found_instances at the end of the block (we
//need to borrow the distance for a bit still)
//add the original match
found_instances.push((*vocab_id, distance));
} else {
if self.debug >= 4 {
eprintln!(" (exceeds max_edit_distance {})", max_edit_distance);
}
pruned_instances += 1;
}
}
}
//found_instances.sort_unstable_by_key(|k| k.1 ); //sort by distance, ascending order
if self.debug >= 2 {
let endtime = SystemTime::now();
let duration = endtime
.duration_since(begintime.expect("begintime"))
.expect("clock can't go backwards")
.as_micros();
eprintln!("(found {} instances (pruned {} above max_edit_distance {}) over {} anagrams in {} μs)", found_instances.len(), pruned_instances, max_edit_distance, nearest_anagrams.len(), duration);
}
found_instances
}
/// Rank and score all variants, returns a vector of three-tuples: (VocabId, distance score, frequency score)
pub(crate) fn score_and_rank(
&self,
instances: Vec<(VocabId, Distance)>,
input: &str,
input_length: usize,
max_matches: usize,
score_threshold: f64,
cutoff_threshold: f64,
freq_weight: f32,
) -> Vec<VariantResult> {
let mut results: Vec<VariantResult> = Vec::new();
let mut max_freq = 0.0;
let mut has_expandable_variants = false;
let weights_sum = self.weights.sum();
assert!(input_length > 0);
let begintime = if self.debug >= 2 {
eprintln!("(scoring and ranking {} instances)", instances.len());
Some(SystemTime::now())
} else {
None
};
//Compute scores
for (vocab_id, distance) in instances.iter() {
if let Some(vocabitem) = self.decoder.get(*vocab_id as usize) {
//all scores are expressed in relation to the input length
let distance_score: f64 = if distance.ld as usize > input_length {
0.0
} else {
1.0 - (distance.ld as f64 / input_length as f64)
};
let lcs_score: f64 = distance.lcs as f64 / input_length as f64;
let prefix_score: f64 = distance.prefixlen as f64 / input_length as f64;
let suffix_score: f64 = distance.suffixlen as f64 / input_length as f64;
//simple weighted linear combination (arithmetic mean to normalize it again) over all normalized distance factors
//expresses a similarity score, sensitive to the length of the input string, and where an exact match by default is 1.0
let score = (self.weights.ld * distance_score
+ self.weights.lcs * lcs_score
+ self.weights.prefix * prefix_score
+ self.weights.suffix * suffix_score
+ if distance.samecase {
self.weights.case
} else {
0.0
})
/ weights_sum;
let freq_score: f64 = if self.have_freq {
//absolute frequency, normalisation in later pass
vocabitem.frequency as f64
} else {
1.0
};
if freq_score > max_freq {
max_freq = freq_score;
}
if !has_expandable_variants && vocabitem.variants.is_some() {
has_expandable_variants = true;
}
if score.is_nan() {
//should never happen
panic!(
"Invalid score (NaN) computed for variant={}, distance={:?}, score={}",
vocabitem.text, distance, score
);
}
if score >= score_threshold {
results.push(VariantResult {
vocab_id: *vocab_id,
dist_score: score,
freq_score,
via: None,
});
if self.debug >= 3 {
eprintln!(
" (variant={}, distance={:?}, score={}, transparent={})",
vocabitem.text,
distance,
score,
vocabitem.vocabtype.check(VocabType::TRANSPARENT)
);
}
} else {
if self.debug >= 3 {
eprintln!(
" (PRUNED variant={}, distance={:?}, score={}, transparent={})",
vocabitem.text,
distance,
score,
vocabitem.vocabtype.check(VocabType::TRANSPARENT)
);
}
}
}
}
//rescore with confusable weights (EARLY)
if !self.confusables.is_empty() && self.confusables_before_pruning {
self.rescore_confusables(&mut results, input);
}
if has_expandable_variants {
results = self.expand_variants(results);
//Collect maximum frequency after expansion
for result in results.iter() {
if result.freq_score > max_freq {
max_freq = result.freq_score;
}
}
}
//normalize frequency score
if max_freq > 0.0 {
for result in results.iter_mut() {
result.freq_score = result.freq_score / max_freq;
}
}
//Sort the results by distance score, descending order
self.rank_results(&mut results, freq_weight);
if has_expandable_variants {
//remove duplicates (can only occur when variant expansion was performed)
results.dedup_by_key(|x| x.vocab_id);
}
//Crop the results at max_matches or cut off at the cutoff threshold
if max_matches > 0 && results.len() > max_matches {
let last_score = results
.get(max_matches - 1)
.expect("get last score")
.score(freq_weight);
let cropped_score = results
.get(max_matches)
.expect("get cropped score")
.score(freq_weight);
if cropped_score < last_score {
if self.debug >= 2 {
eprintln!(
" (truncating {} matches to {})",
results.len(),
max_matches
);
}
//simplest case, crop at the max_matches
results.truncate(max_matches);
} else {
//cropping at max_matches comes at arbitrary point of equal scoring items,
//we crop earlier instead:
let mut early_cutoff = 0;
let mut late_cutoff = 0;
for (i, result) in results.iter().enumerate() {
if result.dist_score == cropped_score && early_cutoff == 0 {
early_cutoff = i;
}
if result.dist_score < cropped_score {
late_cutoff = i;
break;
}
}
if early_cutoff > 0 {
if self.debug >= 2 {
eprintln!(
" (truncating {} matches (early) to {})",
results.len(),
early_cutoff + 1
);
}
results.truncate(early_cutoff + 1);
} else if late_cutoff > 0 {
if self.debug >= 2 {
eprintln!(
" (truncating {} matches (late) to {})",
results.len(),
late_cutoff + 1
);
}
results.truncate(late_cutoff + 1);
}
}
}
//rescore with confusable weights (LATE, default)
if !self.confusables.is_empty() && !self.confusables_before_pruning {
self.rescore_confusables(&mut results, input);
self.rank_results(&mut results, freq_weight);
}
// apply the cutoff threshold
let mut cutoff = 0;
let mut bestscore = None;
if cutoff_threshold >= 1.0 {
for (i, result) in results.iter().enumerate() {
if let Some(bestscore) = bestscore {
if result.score(freq_weight) <= bestscore / cutoff_threshold {
cutoff = i;
break;
}
} else {
bestscore = Some(result.score(freq_weight));
}
}
}
if cutoff > 0 {
let l = results.len();
results.truncate(cutoff);
if self.debug >= 2 {
eprintln!(
" (truncating {} matches to {} due to cutoff value)",
l,
results.len()
);
}
}
if self.debug >= 2 {
for (i, result) in results.iter().enumerate() {
if let Some(vocabitem) = self.decoder.get(result.vocab_id as usize) {
eprintln!(
" (ranked #{}, variant={}, score={}, distance_score={}, freq_score={})",
i + 1,
vocabitem.text,
result.score(freq_weight),
result.dist_score,
result.freq_score
);
}
}
}
if self.debug >= 2 {
let endtime = SystemTime::now();
let duration = endtime
.duration_since(begintime.expect("begintime"))
.expect("clock can't go backwards")
.as_micros();
eprintln!(
" (scored and ranked {} results in {} μs)",
results.len(),
duration
);
}
results
}
/// Rescore results according to confusables
pub fn rescore_confusables(&self, results: &mut Vec<VariantResult>, input: &str) {
if self.debug >= 2 {
eprintln!(" (rescoring with confusable weights)");
}
for result in results.iter_mut() {
result.dist_score *= self.compute_confusable_weight(input, result.vocab_id);
}
}
/// Sorts a result vector of (VocabId, distance_score, freq_score)
/// in decreasing order (best result first)
pub fn rank_results(&self, results: &mut Vec<VariantResult>, freq_weight: f32) {
results.sort_by(|a, b| a.rank_cmp(&b, freq_weight).expect("ordering"));
}
/// Expand variants, adding all references for variants
/// In case variants are 'transparent', only the references will be retained
/// as results.
/// The results list does not need to be sorted yet. This function may yield
/// duplicates. For performance, call this only when you know there are variants that
/// may be expanded.
pub fn expand_variants(&self, mut results: Vec<VariantResult>) -> Vec<VariantResult> {
if self.debug >= 3 {
eprintln!(" (expanding variants, resolving transparency)");
}
let mut new_results = Vec::with_capacity(results.len());
let mut count = 0;
for result in results.drain(..) {
count += 1;
let vocabitem = self
.decoder
.get(result.vocab_id as usize)
.expect("vocabitem must exist");
if let Some(variantrefs) = &vocabitem.variants {
for variantref in variantrefs.iter() {
if let VariantReference::VariantOf((target_id, variant_dist_score)) = variantref
{
new_results.push(VariantResult {
vocab_id: *target_id,
dist_score: result.dist_score * variant_dist_score,
freq_score: {
//take the minimum frequency of the item we refer to and the one of this variant
//note: frequency score is still absolute (not-normalised) at this point
let targetitem = self
.decoder
.get(*target_id as usize)
.expect("vocabitem must exist");
if (targetitem.frequency as f64) < result.freq_score {
targetitem.frequency as f64
} else {
result.freq_score
}
},
via: Some(result.vocab_id),
});
}
}
}
if !vocabitem.vocabtype.check(VocabType::TRANSPARENT) {
//add the original item
new_results.push(result);
}
}
if self.debug >= 3 {
eprintln!(
" (expanded {} instances to {})",
count,
new_results.len()
);
}
new_results
}
/// compute weight over known confusables
/// Should return 1.0 when there are no known confusables
/// < 1.0 when there are unfavourable confusables
/// > 1.0 when there are favourable confusables
pub fn compute_confusable_weight(&self, input: &str, candidate: VocabId) -> f64 {
let mut weight = 1.0;
if let Some(candidate) = self.decoder.get(candidate as usize) {
let editscript = shortest_edit_script(input, &candidate.text, false, false, false);
if self.debug >= 3 {
eprintln!(
" (editscript {} -> {}: {:?})",
input, candidate.text, editscript
);
}
for confusable in self.confusables.iter() {
if confusable.found_in(&editscript) {
if self.debug >= 3 {
eprintln!(
" (input {} with candidate {} instantiates {:?})",
input, &candidate.text, confusable
);
}
weight *= confusable.weight;
}
}
}
weight
}
///Adds the input item to the reverse index, as instantiation of the given vocabulary id
pub fn add_to_reverse_index(
&self,
reverseindex: &mut ReverseIndex,
input: &str,
matched_vocab_id: VocabId,
score: f64,
) {
let variant = match self.encoder.get(input) {
Some(known_vocab_id) => {
if *known_vocab_id == matched_vocab_id {
//item is an exact match, add all
return;
}
Variant::Known(*known_vocab_id)
}
_ => Variant::Unknown(input.to_string()),
};
if self.debug >= 2 {
eprintln!(
" (adding variant {:?} to reverse index for match {})",
variant, matched_vocab_id
);
}
if let Some(existing_variants) = reverseindex.get_mut(&matched_vocab_id) {
existing_variants.push((variant, score));
} else {
reverseindex.insert(matched_vocab_id, vec![(variant, score)]);
}
}
///Searches a text and returns all highest-ranking variants found in the text
pub fn find_all_matches<'a>(&self, text: &'a str, params: &SearchParameters) -> Vec<Match<'a>> {
let mut matches = Vec::new();
if text.is_empty() {
return matches;
}
if self.debug >= 1 {
eprintln!("(finding all matches in text: {})", text);
}
if self.index.is_empty() {
eprintln!(
"ERROR: Model has not been built yet! Call build() before find_all_matches()"
);
return matches;
}
//Find the boundaries and classify their strength
let boundaries = find_boundaries(text);
let strengths = classify_boundaries(&boundaries);
if self.debug >= 2 {
eprintln!(" (boundaries: {:?})", boundaries);
eprintln!(" ( strenghts: {:?})", strengths);
}
let mut begin: usize = 0;
let mut begin_index: usize = 0;
//Compose the text into batches, each batch ends where a hard boundary is found
for (i, (strength, boundary)) in strengths.iter().zip(boundaries.iter()).enumerate() {
if *strength == BoundaryStrength::Hard && boundary.offset.begin != begin {
let text_current = &text[begin..boundary.offset.begin];
let boundaries = &boundaries[begin_index..i + 1];
if self.debug >= 2 {
eprintln!(
" (found hard boundary at {}:{}: {})",
boundary.offset.begin, boundary.offset.end, text_current
);
for boundary in boundaries.iter() {
eprintln!(
" (inner boundary {}:{})",
boundary.offset.begin, boundary.offset.end
);
}
}
//Gather all segments for this batch
let mut batch_matches: Vec<Match<'a>> = Vec::new();
for order in 1..=params.max_ngram {
//Find all n-grams of this order
let mut currentorder_matches: Vec<Match<'a>> = find_match_ngrams(
text,
boundaries,
order,
begin,
Some(boundary.offset.begin),
);
if self.debug >= 2 {
eprintln!(
" (processing {} {}-grams: {:?})",
currentorder_matches.len(),
order,
currentorder_matches
);
}
//find variants for all segments of the current order in this batch
//for higher order matches, we first check if the match is not redundant
//(if the score of the unigrams isn't perfect already)
//so we don't needlessly look up variants we won't use anyway
if params.single_thread {
currentorder_matches.iter_mut().for_each(|segment| {
if order == 1 || !redundant_match(segment, &batch_matches) {
if self.debug >= 1 {
eprintln!(
" (----------- finding variants for: {} -----------)",
segment.text
);
}
let variants = self.find_variants(&segment.text, params);
if self.debug >= 1 {
eprintln!(" (found {} variants)", variants.len());
}
segment.variants = Some(variants);
} else if self.debug >= 2 {
eprintln!(" (skipping redundant match: {})", segment.text);
}
});
} else {
//(in parallel)
currentorder_matches.par_iter_mut().for_each(|segment| {
if order == 1 || !redundant_match(segment, &batch_matches) {
if self.debug >= 1 {
eprintln!(
" (----------- finding variants for: {} -----------)",
segment.text
);
}
let variants = self.find_variants(&segment.text, params);
if self.debug >= 1 {
eprintln!(" (found {} variants)", variants.len());
}
segment.variants = Some(variants);
} else if self.debug >= 2 {
eprintln!(" (skipping redundant match: {})", segment.text);
}
});
}
batch_matches.extend(currentorder_matches.into_iter());
}
/*if params.context_weight > 0.0 {
self.rescore_input_context(&mut batch_matches, &boundaries, params);
}*/
let l = matches.len();
//consolidate the matches, finding a single segmentation that has the best (highest
//scoring) solution
if params.max_ngram > 1 || self.have_lm || !self.context_rules.is_empty() {
//(debug will be handled in the called method)
matches.extend(
self.most_likely_sequence(
batch_matches,
boundaries,
begin,
boundary.offset.begin,
params,
text_current,
)
.into_iter(),
);
} else {
if self.debug >= 1 {
eprintln!(" (returning matches directly, no need to find most likely sequence for unigrams)");
}
matches.extend(batch_matches.into_iter().map(|mut m| {
m.selected = Some(0); //select the first (highest ranking) option
m
}));
}
if self.debug >= 1 {
eprintln!(" (added sequence of {} matches)", matches.len() - l);
}
begin = boundary.offset.end; //(the hard boundary itself is not included in any variant/sequence matching)
begin_index = i + 1
}
}
if self.debug >= 1 {
eprintln!("(returning {} matches)", matches.len());
if self.debug >= 2 {
eprintln!(" (MATCHES={:?})", matches);
}
}
if params.unicodeoffsets {
if self.debug >= 1 {
eprintln!("(remapping UTF-8 offsets to unicodepoints)");
}
remap_offsets_to_unicodepoints(text, matches)
} else {
matches
}
}
/*
fn set_match_boundaries<'a>(&self, matches: &mut Vec<Match<'a>>, boundaries: &[Match<'a>]) {
for m in matches.iter_mut() {
for (i, boundary) in boundaries.iter().enumerate() {
if m.offset.begin == boundary.offset.end {
m.prevboundary = Some(i)
} else if m.offset.end == boundary.offset.begin {
m.nextboundary = Some(i)
}
}
m.n = if let Some(prevboundary) = m.prevboundary {
m.nextboundary.expect("next boundary must exist") - prevboundary
} else {
m.nextboundary.expect("next boundary must exist") + 1
};
}
}
/// Find the unigram context from the input for all matches
fn find_input_context<'a>(&self, matches: &Vec<Match<'a>>) -> Vec<(usize,Context<'a>)> {
let mut results = Vec::with_capacity(matches.len());
for (i, m) in matches.iter().enumerate() {
let mut left = None;
let mut right = None;
for mcontext in matches.iter() {
if let Some(prevboundary) = m.prevboundary {
if mcontext.nextboundary == Some(prevboundary) && mcontext.n == 1{
left = Some(mcontext.text);
}
}
if let Some(nextboundary) = m.nextboundary {
if mcontext.prevboundary == Some(nextboundary) && mcontext.n == 1{
right = Some(mcontext.text);
}
}
}
results.push(
(i, Context {
left,
right,
})
);
}
results
}
*/
/*
/// Rescores variants by incorporating a language model component in the variant score.
/// For simplicity, however, this component is based on the original
/// input text rather than corrected output from other parts.
fn rescore_input_context<'a>(&self, matches: &mut Vec<Match<'a>>, boundaries: &[Match<'a>], params: &SearchParameters) {
if self.debug >= 2 {
eprintln!(" (rescoring variants according to input context)");
}
self.set_match_boundaries(matches, boundaries);
let matches_with_context = self.find_input_context(matches);
assert_eq!(matches_with_context.len(), matches.len());
let mut tokens: Vec<Option<VocabId>> = Vec::new();
let mut perplexities: Vec<f64> = Vec::new();
for (i, context) in matches_with_context.iter() {
let m = matches.get(*i).expect("match must exist");
let left = match context.left {
Some(text) => self.encoder.get(text).map(|x| *x),
None => Some(BOS)
};
let right = match context.right {
Some(text) => self.encoder.get(text).map(|x| *x),
None => Some(BOS)
};
perplexities.clear();
let mut best_perplexity = 99999.0; //to be minimised
if let Some(variants) = &m.variants {
for (variant, _dist_score, _freq_score) in variants.iter() {
if let Ok(mut ngram) = self.into_ngram(*variant, &mut None) {
tokens.clear();
tokens.push(left);
loop {
match ngram.pop_first() {
NGram::Empty => break,
unigram => tokens.push(unigram.first())
}
}
tokens.push(right);
let (_lm_logprob, perplexity) = self.lm_score_tokens(&tokens);
if perplexity < best_perplexity {
best_perplexity = perplexity;
}
perplexities.push(perplexity);
}
}
}
if self.debug >= 2 {
eprintln!(" (processing {} variants for match {}, best_perplexity={})", perplexities.len(), i+1, best_perplexity);
}
let m = matches.get_mut(*i).expect("match must exist");
for (j, perplexity) in perplexities.iter().enumerate() {
let variants = &mut m.variants.as_mut().expect("variants must exist");
let (vocab_id, score, freq_score) = variants.get_mut(j).expect("variant must exist");
//compute a weighted geometric mean between language model score
//and variant model score
//first normalize the perplexity where the best one corresponds to 1.0, and values decrease towards 0 as perplexity increases, the normalisation is technically not needed for geometric mean but we do need to invert the scale (minimisation of perplexity -> maximisation of score)
let lmscore = best_perplexity / perplexity;
//then the actual computation is done in log-space for more numerical stability,
//and cast back afterwards
let oldscore = *score;
*score = ((score.ln() + params.context_weight as f64 * lmscore.ln()) / (1.0 + params.context_weight) as f64).exp();
// fixed weight for variant model ------------------^
if self.debug >= 3 {
if let Some(vocabitem) = self.decoder.get(*vocab_id as usize) {
eprintln!(" (leftcontext={:?}, variant={}, rightcontext={:?}, oldscore={}, score={}, norm_lm_score={}, perplexity={})", context.left, vocabitem.text, context.right, oldscore, score, lmscore, perplexity);
}
}
}
}
}
*/
/// Find the solution that maximizes the variant scores, decodes using a Weighted Finite State Transducer
fn most_likely_sequence<'a>(
&self,
matches: Vec<Match<'a>>,
boundaries: &[Match<'a>],
begin_offset: usize,
end_offset: usize,
params: &SearchParameters,
input_text: &str,
) -> Vec<Match<'a>> {
if self.debug >= 2 {
eprintln!(
"(building FST for finding most likely sequence in range {}:{})",
begin_offset, end_offset
);
}
//Build a finite state transducer
let mut fst = VectorFst::<TropicalWeight>::new();
let mut symtab_in = SymbolTable::new(); //only used for drawing the FST in debug mode
let mut symtab_out = SymbolTable::new(); //only used for drawing the FST in debug mode
//add initial state
let start = fst.add_state();
fst.set_start(start).expect("set start state");
//adds states for all boundaries
let mut final_found = false;
let states: Vec<u32> = boundaries
.iter()
.map(|boundary| {
let state = fst.add_state();
if boundary.offset.begin == end_offset || boundary.offset.end == end_offset {
final_found = true;
fst.set_final(state, 0.0).expect("set end state");
}
state
})
.collect();
if !final_found {
//sanity check
panic!("no final state found");
}
if self.debug >= 2 {
eprintln!(
" (added {} states ({} boundaries), not including the start state)",
states.len(),
boundaries.len()
);
}
let mut output_symbols: Vec<OutputSymbol> = vec![
OutputSymbol {
vocab_id: 0,
symbol: 0,
match_index: 0,
variant_index: None,
boundary_index: 0,
}, //first entry is a dummy entry because the 0 symbol is reserved for epsilon
];
//add transitions between the boundary states
for (match_index, m) in matches.iter().enumerate() {
if self.debug >= 2 {
symtab_in.add_symbol(m.text); //symbol_index = match_index + 1
}
let mut prevboundary: Option<usize> = None;
let mut nextboundary: Option<usize> = None;
let input_symbol = (match_index + 1) as u32;
for (i, boundary) in boundaries.iter().enumerate() {
if m.offset.begin == boundary.offset.end {
prevboundary = Some(i)
} else if m.offset.end == boundary.offset.begin {
nextboundary = Some(i)
}
}
let n;
let prevstate = if let Some(prevboundary) = prevboundary {
n = nextboundary.expect("next boundary must exist") - prevboundary;
*states.get(prevboundary).expect("prev state must exist")
} else {
n = nextboundary.expect("next boundary must exist") + 1;
start
};
let nextstate = *states
.get(nextboundary.expect("next boundary must exist"))
.expect("next state must exist");
if m.variants.is_some() && !m.variants.as_ref().unwrap().is_empty() {
for (variant_index, variantresult) in
m.variants.as_ref().unwrap().iter().enumerate()
{
let output_symbol = output_symbols.len() as u32;
output_symbols.push(OutputSymbol {
vocab_id: variantresult.vocab_id,
symbol: output_symbol,
match_index,
variant_index: Some(variant_index),
boundary_index: nextboundary.expect("next boundary must exist"),
});
if self.debug >= 3 {
let mut variant_text = String::new();
variant_text += self
.decoder
.get(variantresult.vocab_id as usize)
.expect("variant_text")
.text
.as_str();
variant_text += format!(" ({})", output_symbol).as_str(); //we encode the output symbol in the text otherwise the symbol table returns the old match
eprintln!(
" (transition state {}->{}: {} ({}) -> {} and variant score {})",
prevstate,
nextstate,
m.text,
input_symbol,
variant_text,
-1.0 * variantresult.score(params.freq_weight).ln() as f32
);
let osym = symtab_out.add_symbol(variant_text);
assert!(osym == output_symbol);
}
//each transition gets a base cost of n (the number of input tokens it covers)
//on top of that cost in the range 0.0 (best) - 1.0 (worst) expresses the
//distance score (inversely)
let cost: f32 =
n as f32 + (1.0 - variantresult.score(params.freq_weight) as f32);
fst.add_tr(
prevstate,
Tr::new(input_symbol, output_symbol, cost, nextstate),
)
.expect("adding transition");
}
} else if n == 1 {
//only for unigrams
let output_symbol = output_symbols.len() as u32;
output_symbols.push(OutputSymbol {
vocab_id: 0, //0 vocab_id means we have an Out-of-Vocabulary word to copy from input
symbol: output_symbol,
match_index,
variant_index: None,
boundary_index: nextboundary.expect("next boundary must exist"),
});
//OOV emission cost
let cost: f32 = n as f32 + 1.0;
if self.debug >= 3 {
eprintln!(
" (transition state {}->{}: {} ({}) -> OOV ({}) and score {})",
prevstate, nextstate, m.text, input_symbol, output_symbol, cost
);
let mut variant_text = String::from_str(m.text).expect("from str");
variant_text += format!(" ({})", output_symbol).as_str(); //we encode the output symbol in the text otherwise the symbol table returns the old match
let osym = symtab_out.add_symbol(&variant_text);
if osym != output_symbol {
panic!(
"Output symbol out of sync: {} vs {}, variant_text={}",
osym, output_symbol, variant_text
);
}
}
fst.add_tr(
prevstate,
Tr::new(input_symbol, output_symbol, cost, nextstate),
)
.expect("adding transition");
}
}
//failsafe: add high-cost epsilon transitions between boundaries to ensure the graph always has a complete path
for i in 0..boundaries.len() {
let nextboundary = i;
let prevstate = if i == 0 {
start
} else {
*states.get(i - 1).expect("prev state must exist")
};
let nextstate = *states.get(nextboundary).expect("next state must exist");
fst.add_tr(prevstate, Tr::new(0, 0, 100.0, nextstate))
.expect("adding transition");
}
if output_symbols.len() == 1 {
if self.debug >= 2 {
eprintln!(" (no output symbols found, FST not needed, aborting)");
}
//we have no output symbols, building an FST is not needed, just return the input
return matches;
}
//find the n most likely sequences, note that we only consider the distance scores here,
//language modelling (considering context) is applied in a separate step later
if self.debug >= 3 {
eprintln!(" (computed FST: {:?})", fst);
eprintln!(" (symtab_in={:?})", symtab_in);
eprintln!(" (symtab_out={:?})", symtab_out);
eprintln!(" (finding shortest path)");
fst.set_input_symbols(Arc::new(symtab_in));
fst.set_output_symbols(Arc::new(symtab_out));
let input_text_filename = input_text
.replace(" ", "_")
.replace("\"", "")
.replace("'", "")
.replace(".", "")
.replace("/", "")
.replace("?", ""); //strip filename unfriendly chars
let mut config = DrawingConfig::default();
config.portrait = true;
config.title = input_text.to_owned();
if let Err(e) = fst.draw(
format!("/tmp/analiticcl.{}.fst.dot", input_text_filename.as_str()),
&config,
) {
panic!("FST draw error: {}", e);
}
}
let fst: VectorFst<TropicalWeight> = shortest_path_with_config(
&fst,
ShortestPathConfig::default().with_nshortest(params.max_seq),
)
.expect("computing shortest path fst");
let mut sequences: Vec<Sequence> = Vec::new();
let mut best_lm_perplexity: f64 = 999999.0; //to be minimised
let mut best_variant_cost: f32 = (boundaries.len() - 1) as f32 * 2.0; //worst score, to be improved (to be minimised)
let mut best_context_score: f64 = 0.0; //to be maximised
for (i, path) in fst.paths_iter().enumerate() {
//iterates over the n shortest path hypotheses (does not return them in weighted order)
let variant_cost: f32 = *path.weight.value();
let mut sequence = Sequence::new(variant_cost);
if self.debug >= 3 {
eprintln!(" (#{}, path: {:?})", i + 1, path);
}
for output_symbol in path.olabels.iter() {
let output_symbol = output_symbols
.get(*output_symbol as usize)
.expect("expected valid output symbol");
sequence.output_symbols.push(output_symbol.clone());
}
if self.have_lm && params.lm_weight > 0.0 {
//Apply the language model, considers context
let (lm_logprob, perplexity) = self.lm_score(&sequence, &boundaries);
sequence.lm_logprob = lm_logprob;
sequence.perplexity = perplexity;
if sequence.perplexity < best_lm_perplexity {
best_lm_perplexity = sequence.perplexity;
}
}
if !self.context_rules.is_empty() {
//Apply context rules and apply tags (if any), considers context
let (context_score, sequence_results) = self.test_context_rules(&sequence);
sequence.context_score = context_score;
sequence.tags = sequence_results
.into_iter()
.map(|vecpm| {
vecpm
.into_iter()
.filter_map(|pm| {
if pm.tag.is_some() {
Some((pm.tag.unwrap(), pm.seqnr))
} else {
None
}
})
.collect()
})
.collect();
if self.debug >= 3 && sequence.context_score != 1.0 {
eprintln!(" (context_score: {})", sequence.context_score);
}
}
if variant_cost < best_variant_cost {
best_variant_cost = variant_cost;
}
if sequence.context_score > best_context_score {
best_context_score = sequence.context_score;
}
sequences.push(sequence);
}
let mut debug_ranked: Option<Vec<(Sequence, f64, f64, f64, f64)>> = if self.debug >= 1 {
Some(Vec::new())
} else {
None
};
//Compute the normalized scores
let mut best_score: f64 = -99999999.0; //to be maximised
let mut best_sequence: Option<Sequence> = None;
for sequence in sequences.into_iter() {
//we normalize both LM and variant model scores so the best score corresponds with 1.0 (in non-logarithmic terms, 0.0 in logarithmic space). We take the natural logarithm for more numerical stability and easier computation.
let norm_lm_score: f64 = if self.have_lm && params.lm_weight > 0.0 {
(best_lm_perplexity / sequence.perplexity).ln()
} else {
0.0
};
let norm_variant_score: f64 =
(best_variant_cost as f64 / sequence.variant_cost as f64).ln();
let norm_context_score: f64 = (sequence.context_score / best_context_score).ln();
//then we interpret the score as a kind of pseudo-probability and minimize the joint
//probability (the product; addition in log-space)
let score = if (!self.have_lm || params.lm_weight == 0.0)
&& (self.context_rules.is_empty() || params.contextrules_weight == 0.0)
{
//no need for full computation, take a shortcut:
norm_variant_score
} else {
(params.lm_weight as f64 * norm_lm_score
+ params.variantmodel_weight as f64 * norm_variant_score
+ params.contextrules_weight as f64 * norm_context_score)
/ (params.lm_weight as f64
+ params.variantmodel_weight as f64
+ params.contextrules_weight as f64) //note: the denominator isn't really relevant for finding the best score but normalizes the output for easier interpretability (=geometric mean)
};
if self.debug >= 1 {
debug_ranked.as_mut().unwrap().push((
sequence.clone(),
norm_lm_score,
norm_variant_score,
norm_context_score,
score,
));
}
if score > best_score || best_sequence.is_none() {
best_score = score;
best_sequence = Some(sequence);
}
}
if self.debug >= 1 {
//debug mode: output all candidate sequences and their scores in order
debug_ranked
.as_mut()
.unwrap()
.sort_by(|a, b| b.4.partial_cmp(&a.4).unwrap_or(Ordering::Equal)); //sort by score
for (i, (sequence, norm_lm_score, norm_variant_score, norm_context_score, score)) in
debug_ranked.unwrap().into_iter().enumerate()
{
eprintln!(" (#{}, final_score={}, norm_lm_score={} (perplexity={}, logprob={}, weight={}), norm_variant_score={} (variant_cost={}, weight={}), norm_context_score={} (context_score={}, weight={})", i+1, score.exp(), norm_lm_score.exp(), sequence.perplexity, sequence.lm_logprob, params.lm_weight, norm_variant_score.exp(), sequence.variant_cost, params.variantmodel_weight, norm_context_score.exp(), sequence.context_score, params.contextrules_weight);
let mut inputtext: String = String::new();
let mut text: String = String::new();
for (j, output_symbol) in sequence.output_symbols.iter().enumerate() {
let m = matches
.get(output_symbol.match_index)
.expect("match index must exist");
inputtext += m.text;
if output_symbol.vocab_id > 0 {
text += self
.decoder
.get(output_symbol.vocab_id as usize)
.expect("vocab")
.text
.as_str();
} else {
text += m.text;
}
if !sequence.tags.is_empty() {
if let Some(tags) = sequence.tags.get(j) {
for (tagindex, seqnr) in tags.iter() {
text += format!(
"[#{}:{}]",
self.tags.get(*tagindex as usize).expect("Tag must exist"),
seqnr
)
.as_str();
}
}
}
inputtext += " | ";
text += " | ";
}
eprintln!(" (text_out={})", text);
eprintln!(" (text_in={})", inputtext);
}
}
//return matches corresponding to best sequence
let best_sequence = best_sequence.expect("there must be a best sequence");
best_sequence
.output_symbols
.iter()
.enumerate()
.map(|(i, osym)| {
let m = matches
.get(osym.match_index)
.expect("match should be in bounds");
let mut m = m.clone();
m.selected = osym.variant_index;
if !best_sequence.tags.is_empty() {
if let Some(tags) = best_sequence.tags.get(i) {
m.tag = tags.iter().map(|x| x.0).collect();
m.seqnr = tags.iter().map(|x| x.1).collect();
}
}
m
})
.collect()
}
/// Favours or penalizes certain combinations of lexicon matches. matching words X and Y
/// respectively with lexicons A and B might be favoured over other combinations.
/// This returns either a bonus or penalty (number slightly above/below 1.0) score/
/// for the sequence as a whole.
pub fn test_context_rules<'a>(
&self,
sequence: &Sequence,
) -> (f64, Vec<Vec<PatternMatchResult>>) {
let sequence: Vec<(VocabId, u32)> = sequence
.output_symbols
.iter()
.map(|output_symbol| {
if output_symbol.vocab_id == 0 {
(output_symbol.vocab_id, 0)
} else {
if let Some(vocabvalue) = self.decoder.get(output_symbol.vocab_id as usize) {
(output_symbol.vocab_id, vocabvalue.lexindex)
} else {
(output_symbol.vocab_id, 0)
}
}
})
.collect();
//The sequence will flag which items in the sequence have been covered by matches (non-empty vec)
//and if so, what context rule scores and tags (inner vec) apply to that match. It's later used to compute
//the final score
let mut sequence_results: Vec<Vec<PatternMatchResult>> = vec![vec!(); sequence.len()];
let mut found = false;
for begin in 0..sequence.len() {
for context_rule in self.context_rules.iter() {
if context_rule.matches(&sequence, begin, &mut sequence_results) {
found = true;
if self.debug >= 2 {
let text: Vec<&str> = sequence
.iter()
.map(|(vocab_id, _)| {
if *vocab_id == 0 {
"<UNK>"
} else {
if let Some(vocabvalue) = self.decoder.get(*vocab_id as usize) {
vocabvalue.text.as_ref()
} else {
"<UNK>"
}
}
})
.collect();
eprintln!(
" Context rule matches: {:?} <-- \"{}\" --> {:?}",
context_rule,
text.join(" | "),
sequence_results
);
}
}
}
}
if !found {
(1.0, sequence_results) //just a shortcut to prevent unnecessary computation
} else {
(
//compute sum score
sequence_results
.iter()
.map(|x| {
if !x.is_empty() {
x[0].score //score is equal for all subelements (only tags differ), just grab first one
} else {
1.0
}
})
.sum::<f32>() as f64
/ sequence.len() as f64,
sequence_results,
)
}
}
/// Computes the logprob and perplexity for a given sequence as produced in
/// most_likely_sequence()
pub fn lm_score<'a>(&self, sequence: &Sequence, boundaries: &[Match<'a>]) -> (f32, f64) {
//step 1: collect all tokens in the sequence
let mut tokens: Vec<Option<VocabId>> =
Vec::with_capacity(sequence.output_symbols.len() + 5); //little bit of extra space to prevent needing to reallocate too quickly and to hold the BOS/EOS markers
tokens.push(Some(BOS));
for output_symbol in sequence.output_symbols.iter() {
let next_boundary = boundaries
.get(output_symbol.boundary_index)
.expect("boundary should be in bounds");
if output_symbol.vocab_id == 0 {
//out of vocabulary (copied from input)
tokens.push(None);
} else {
if let Ok(mut ngram) = self.into_ngram(output_symbol.vocab_id, &mut None) {
loop {
match ngram.pop_first() {
NGram::Empty => break,
unigram => tokens.push(unigram.first()),
}
}
}
}
//add boundary as a token too
if !next_boundary.text.trim().is_empty() {
if let Some(vocab_id) = self.encoder.get(next_boundary.text.trim()) {
if let Ok(mut ngram) = self.into_ngram(*vocab_id, &mut None) {
loop {
match ngram.pop_first() {
NGram::Empty => break,
unigram => tokens.push(unigram.first()),
}
}
}
} else {
//out of vocabulary boundary tokens (copied from input)
tokens.push(None);
}
}
}
tokens.push(Some(EOS));
//Compute the score over the tokens
self.lm_score_tokens(&tokens)
}
/// Computes the logprob and perplexity for a given sequence of tokens.
/// The tokens are either in the vocabulary or are None if out-of-vocabulary.
pub fn lm_score_tokens<'a>(&self, tokens: &Vec<Option<VocabId>>) -> (f32, f64) {
//move a sliding window over the tokens
let mut logprob = 0.0;
let mut n = 0;
for i in 1..=tokens.len() - 1 {
if let Ok(bigram) = NGram::from_option_list(&tokens[i - 1..i + 1]) {
let prior = NGram::from_option_list(&tokens[i - 1..i]).expect("extracting prior");
let priorcount = if let Some(priorcount) = self.ngrams.get(&prior) {
*priorcount
} else {
1
};
//Do we have a joint probability for the bigram that forms the transition?
if let Some(jointcount) = self.ngrams.get(&bigram) {
if priorcount < *jointcount {
//sanity check, shouldn't be the case, correct:
logprob += (*jointcount as f32).ln()
} else {
logprob += (*jointcount as f32 / priorcount as f32).ln()
}
} else {
logprob += TRANSITION_SMOOTHING_LOGPROB
}
n += 1;
} else {
//if we have an out of vocabulary bigram or prior we fall back to add-one smoothing
//simply setting the count of that ngram/prior to 1
//for the perplexity computation this means the score doesn't change, but n does
//increase, so we end up with a lower perplexity
n += 1;
logprob += TRANSITION_SMOOTHING_LOGPROB
}
}
//PP(W) = (1/P(w1...wN))^(1/N)
// in logspace: PP(W) = -1.0/N * Log(P(w1...Wn))
let perplexity = -1.0 / (n as f64) * logprob as f64;
(logprob, perplexity)
}
/// Add an ngram for language modelling
pub fn add_ngram(&mut self, ngram: NGram, frequency: u32) {
if let Some(ngram) = self.ngrams.get_mut(&ngram) {
//update the count for this ngram
*ngram += frequency;
} else {
//add the new ngram
self.ngrams.insert(ngram, frequency);
}
}
/// Decompose a known vocabulary Id into an Ngram
fn into_ngram(
&self,
word: VocabId,
unseen_parts: &mut Option<VocabEncoder>,
) -> Result<NGram, Box<dyn Error>> {
let word_dec = self
.decoder
.get(word as usize)
.expect("word does not exist in decoder");
let mut iter = word_dec.text.split(" ");
match word_dec.tokencount {
0 => Ok(NGram::Empty),
1 => Ok(NGram::UniGram(self.encode_token(
iter.next().expect("ngram part"),
true,
unseen_parts,
))),
2 => Ok(NGram::BiGram(
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
)),
3 => Ok(NGram::TriGram(
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
)),
4 => Ok(NGram::QuadGram(
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
)),
5 => Ok(NGram::QuintGram(
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
self.encode_token(iter.next().expect("ngram part"), true, unseen_parts),
)),
_ => simple_error::bail!("Can only deal with n-grams up to order 5"),
}
}
/// Encode one token, optionally returning either UNK or putting it in ``unseen`` if it is new.
/// Use in ngram construction
fn encode_token(
&self,
token: &str,
use_unk: bool,
unseen: &mut Option<VocabEncoder>,
) -> VocabId {
if let Some(vocab_id) = self.encoder.get(token) {
*vocab_id
} else if use_unk {
UNK
} else if let Some(unseen) = unseen.as_mut() {
if let Some(vocab_id) = unseen.get(token) {
*vocab_id
} else {
let vocab_id: VocabId = self.decoder.len() as VocabId + unseen.len() as VocabId;
unseen.insert(token.to_string(), vocab_id);
vocab_id
}
} else {
panic!("Token does not exist in vocabulary (and returning unknown tokens or adding new ones was not set)");
}
}
/// Gives the text representation for this match, always uses the solution (if any) and falls
/// back to the input text only when no solution was found.
pub fn match_to_str<'a>(&'a self, m: &Match<'a>) -> &'a str {
if let Some(vocabvalue) = self.match_to_vocabvalue(m) {
vocabvalue.text.as_str()
} else {
m.text
}
}
/// Gives the vocabitem for this match, always uses the solution (if any) and falls
/// back to the input text only when no solution was found.
pub fn match_to_vocabvalue<'a>(&'a self, m: &Match<'a>) -> Option<&'a VocabValue> {
if let Some(result) = m.solution() {
self.decoder.get(result.vocab_id as usize)
} else {
None
}
}
/// Turns the ngram into a tokenised string; the tokens in the ngram will be separated by a space.
pub fn ngram_to_str(&self, ngram: &NGram) -> String {
let v: Vec<&str> = ngram
.to_vec()
.into_iter()
.map(|v| {
self.decoder
.get(v as usize)
.expect("ngram must contain valid vocab ids")
.text
.as_str()
})
.collect();
v.join(" ")
}
/// Converts a match to an NGram representation, this only works if all tokens in the ngram are
/// in the vocabulary.
pub fn match_to_ngram<'a>(
&'a self,
m: &Match<'a>,
boundaries: &[Match<'a>],
) -> Result<NGram, String> {
let internal = m.internal_boundaries(boundaries);
let parts = find_match_ngrams(m.text, internal, 1, 0, None);
let mut ngram = NGram::Empty;
for part in parts {
if let Some(vocabid) = self.encoder.get(part.text) {
ngram.push(*vocabid);
} else {
return Err(format!(
"unable to convert match to ngram, contains out-of-vocabulary token: {}",
part.text
));
}
}
Ok(ngram)
}
}