├── LICENSE ├── README.md ├── data ├── CADEC │ └── __init__.py ├── ShARe13 │ └── __init__.py └── examples │ ├── dev.txt │ ├── test.txt │ └── train.txt ├── emb └── __init__.py ├── framework ├── core │ ├── module.py │ └── seq2seq.py ├── main.py └── utils │ ├── common.py │ ├── data_util.py │ ├── eval.py │ └── util.py └── output └── exp1 ├── checkpoint └── __init__.py └── log └── __init__.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | {one line to give the program's name and a brief idea of what it does.} 635 | Copyright (C) {year} {name of author} 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | {project} Copyright (C) {year} {fullname} 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Discontinuous NER with Pointer Network 2 | 3 | Codes for the AAAI 2021 paper: [Rethinking Boundaries: End-To-End Recognition of Discontinuous Mentions with Pointer Networks](https://ojs.aaai.org/index.php/AAAI/article/view/17513). 4 | 5 | ------------------- 6 | 7 | 8 | # Requirement 9 | 10 | ```bash 11 | python>=1.6 12 | numpy>=1.13.3 13 | torch>=0.4.0 14 | ``` 15 | 16 | # Datasets 17 | 18 | Two benchmark datasets for discontinuous NER. 19 | Download them and put at `./data` folds. 20 | 21 | - [CADEC](https://data.csiro.au/collection/csiro:10948) 22 | - ShARe13 23 | 24 | 25 | Data format preprocessing. 26 | Please process the annotation as following format: 27 | 28 | ```bash 29 | Upset stomach and the feeling that I may need to throw up . 30 | 0,1 ADR|10,11 ADR 31 | ``` 32 | 33 | See the example data in [./data/examples](data%2Fexamples). 34 | 35 | 36 | 37 | # Experiments 38 | 39 | To train the parser, run the following script: 40 | 41 | ```bash 42 | python ./framework/main.py 43 | ``` 44 | 45 | Change the parameters for training, testing. 46 | 47 | 48 | # Citation 49 | 50 | ``` 51 | @inproceedings{FeiDisNERAAAI21, 52 | author = {Hao Fei and Donghong Ji and Bobo Li and 53 | Yijiang Liu and Yafeng Ren and Fei Li}, 54 | title = {Rethinking Boundaries: End-To-End Recognition of Discontinuous Mentions with Pointer Networks}, 55 | booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence}, 56 | pages = {12785--12793}, 57 | year = {2021}, 58 | } 59 | ``` 60 | 61 | 62 | # License 63 | 64 | The code is released under Apache License 2.0 for Non-commercial use only. -------------------------------------------------------------------------------- /data/CADEC/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scofield7419/DisNER-PtrNet/5c748d09c36c42b4828338fba6c5b7ba3f326c04/data/CADEC/__init__.py -------------------------------------------------------------------------------- /data/ShARe13/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scofield7419/DisNER-PtrNet/5c748d09c36c42b4828338fba6c5b7ba3f326c04/data/ShARe13/__init__.py -------------------------------------------------------------------------------- /data/examples/dev.txt: -------------------------------------------------------------------------------- 1 | Upset stomach and the feeling that I may need to throw up . 2 | 0,1 ADR|10,11 ADR 3 | 4 | The pain in my stomach is in the middle of my stomach above my belly button and its a deep constant pain . 5 | 19,21 ADR|1,4 ADR 6 | 7 | I also feel like I may need to go to the restroom , but when I do I can ' t . 8 | 4,20 ADR 9 | 10 | Swelling & weight increase . 11 | 0,0 ADR|2,3 ADR 12 | 13 | uncontrollable Diarrhea , kidney function . 14 | 1,1 ADR 15 | 16 | This product is the best medicine for arthritis pain I have ever taken , however after taking it for approx . 10 years it started affecting my kidneys , my stomach ( acid reflux ) and the diarrhea got worse , short term , I had no problems , this would be a great product if the manufacturer could get rid of the side effects . 17 | 25,25,29,34 ADR|25,27 ADR|37,37 ADR 18 | 19 | dizzines dioreah stomach cramps disorientation . 20 | 0,0 ADR|1,1 ADR|2,3 ADR|4,4 ADR 21 | 22 | followed by ulceration of the osophragus , and stomach , and possibly the gut . 23 | 2,3,8,8 ADR|2,5 ADR 24 | 25 | Now I dont know how I am going to get back to work , it has left me feeling exausted , and depressed . 26 | 18,18,22,22 ADR|18,19 ADR 27 | 28 | NSAID gastritis , severe stomach upset after prolonged use . 29 | 1,1 ADR|3,5 ADR 30 | 31 | false full feeling , very poor appetite . 32 | 0,2 ADR|4,6 ADR 33 | 34 | increased menstruation , 2 - 3 periods a month instead of once a month , menstrual cramps present with or without vaginal bleeding . 35 | 0,1 ADR|15,16,18,18,21,22 ADR|3,13 ADR|15,16,20,22 ADR 36 | 37 | -------------------------------------------------------------------------------- /data/examples/test.txt: -------------------------------------------------------------------------------- 1 | Upset stomach and the feeling that I may need to throw up . 2 | 0,1 ADR|10,11 ADR 3 | 4 | The pain in my stomach is in the middle of my stomach above my belly button and its a deep constant pain . 5 | 19,21 ADR|1,4 ADR 6 | 7 | I also feel like I may need to go to the restroom , but when I do I can ' t . 8 | 4,20 ADR 9 | 10 | Swelling & weight increase . 11 | 0,0 ADR|2,3 ADR 12 | 13 | uncontrollable Diarrhea , kidney function . 14 | 1,1 ADR 15 | 16 | This product is the best medicine for arthritis pain I have ever taken , however after taking it for approx . 10 years it started affecting my kidneys , my stomach ( acid reflux ) and the diarrhea got worse , short term , I had no problems , this would be a great product if the manufacturer could get rid of the side effects . 17 | 25,25,29,34 ADR|25,27 ADR|37,37 ADR 18 | 19 | dizzines dioreah stomach cramps disorientation . 20 | 0,0 ADR|1,1 ADR|2,3 ADR|4,4 ADR 21 | 22 | followed by ulceration of the osophragus , and stomach , and possibly the gut . 23 | 2,3,8,8 ADR|2,5 ADR 24 | 25 | Now I dont know how I am going to get back to work , it has left me feeling exausted , and depressed . 26 | 18,18,22,22 ADR|18,19 ADR 27 | 28 | NSAID gastritis , severe stomach upset after prolonged use . 29 | 1,1 ADR|3,5 ADR 30 | 31 | false full feeling , very poor appetite . 32 | 0,2 ADR|4,6 ADR 33 | 34 | increased menstruation , 2 - 3 periods a month instead of once a month , menstrual cramps present with or without vaginal bleeding . 35 | 0,1 ADR|15,16,18,18,21,22 ADR|3,13 ADR|15,16,20,22 ADR 36 | 37 | -------------------------------------------------------------------------------- /data/examples/train.txt: -------------------------------------------------------------------------------- 1 | Upset stomach and the feeling that I may need to throw up . 2 | 0,1 ADR|10,11 ADR 3 | 4 | The pain in my stomach is in the middle of my stomach above my belly button and its a deep constant pain . 5 | 19,21 ADR|1,4 ADR 6 | 7 | I also feel like I may need to go to the restroom , but when I do I can ' t . 8 | 4,20 ADR 9 | 10 | Swelling & weight increase . 11 | 0,0 ADR|2,3 ADR 12 | 13 | uncontrollable Diarrhea , kidney function . 14 | 1,1 ADR 15 | 16 | This product is the best medicine for arthritis pain I have ever taken , however after taking it for approx . 10 years it started affecting my kidneys , my stomach ( acid reflux ) and the diarrhea got worse , short term , I had no problems , this would be a great product if the manufacturer could get rid of the side effects . 17 | 25,25,29,34 ADR|25,27 ADR|37,37 ADR 18 | 19 | dizzines dioreah stomach cramps disorientation . 20 | 0,0 ADR|1,1 ADR|2,3 ADR|4,4 ADR 21 | 22 | followed by ulceration of the osophragus , and stomach , and possibly the gut . 23 | 2,3,8,8 ADR|2,5 ADR 24 | 25 | Now I dont know how I am going to get back to work , it has left me feeling exausted , and depressed . 26 | 18,18,22,22 ADR|18,19 ADR 27 | 28 | NSAID gastritis , severe stomach upset after prolonged use . 29 | 1,1 ADR|3,5 ADR 30 | 31 | false full feeling , very poor appetite . 32 | 0,2 ADR|4,6 ADR 33 | 34 | increased menstruation , 2 - 3 periods a month instead of once a month , menstrual cramps present with or without vaginal bleeding . 35 | 0,1 ADR|15,16,18,18,21,22 ADR|3,13 ADR|15,16,20,22 ADR 36 | 37 | -------------------------------------------------------------------------------- /emb/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scofield7419/DisNER-PtrNet/5c748d09c36c42b4828338fba6c5b7ba3f326c04/emb/__init__.py -------------------------------------------------------------------------------- /framework/core/module.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.autograd as autograd 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | import torch.optim as optim 7 | from framework.utils.common import * 8 | 9 | 10 | class WordEmbeddings(nn.Module): 11 | def __init__(self, vocab_size, embed_dim, pre_trained_embed_matrix, drop_out_rate): 12 | super(WordEmbeddings, self).__init__() 13 | self.embeddings = nn.Embedding(vocab_size, embed_dim, padding_idx=0) 14 | self.embeddings.weight.data.copy_(torch.from_numpy(pre_trained_embed_matrix)) 15 | self.dropout = nn.Dropout(drop_out_rate) 16 | 17 | def forward(self, words_seq): 18 | word_embeds = self.embeddings(words_seq) 19 | word_embeds = self.dropout(word_embeds) 20 | return word_embeds 21 | 22 | def weight(self): 23 | return self.embeddings.weight 24 | 25 | 26 | class CharEmbeddings(nn.Module): 27 | def __init__(self, vocab_size, embed_dim, drop_out_rate): 28 | super(CharEmbeddings, self).__init__() 29 | self.embeddings = nn.Embedding(vocab_size, embed_dim, padding_idx=0) 30 | self.dropout = nn.Dropout(drop_out_rate) 31 | 32 | def forward(self, words_seq): 33 | char_embeds = self.embeddings(words_seq) 34 | char_embeds = self.dropout(char_embeds) 35 | return char_embeds 36 | 37 | 38 | class Encoder(nn.Module): 39 | 40 | def __init__(self, input_dim, hidden_dim, 41 | word_vocab_len, char_vocab_len, word_embed_matrix, 42 | layers, is_bidirectional, drop_out_rate): 43 | super(Encoder, self).__init__() 44 | self.input_dim = input_dim 45 | self.hidden_dim = hidden_dim 46 | self.layers = layers 47 | self.is_bidirectional = is_bidirectional 48 | self.drop_rate = drop_out_rate 49 | self.word_embeddings = WordEmbeddings(word_vocab_len, word_embed_dim, word_embed_matrix, drop_rate) 50 | self.char_embeddings = CharEmbeddings(char_vocab_len, char_embed_dim, drop_rate) 51 | self.pos_embeddings = nn.Embedding(max_positional_idx, positional_embed_dim, padding_idx=0) 52 | if enc_type == 'LSTM': 53 | self.lstm = nn.LSTM(self.input_dim, self.hidden_dim, self.layers, batch_first=True, 54 | bidirectional=self.is_bidirectional) 55 | self.dropout = nn.Dropout(self.drop_rate) 56 | self.conv1d = nn.Conv1d(char_embed_dim, char_feature_size, conv_filter_size) 57 | self.max_pool = nn.MaxPool1d(max_word_len + conv_filter_size - 1, max_word_len + conv_filter_size - 1) 58 | 59 | 60 | def forward(self, words, chars, pos_seq, is_training=False): 61 | src_word_embeds = self.word_embeddings(words) 62 | pos_embeds = self.dropout(self.pos_embeddings(pos_seq)) 63 | char_embeds = self.char_embeddings(chars) 64 | char_embeds = char_embeds.permute(0, 2, 1) 65 | 66 | char_feature = torch.tanh(self.max_pool(self.conv1d(char_embeds))) 67 | char_feature = char_feature.permute(0, 2, 1) 68 | 69 | words_input = torch.cat((src_word_embeds, char_feature, pos_embeds), -1) 70 | if enc_type == 'LSTM': 71 | outputs, hc = self.lstm(words_input) 72 | 73 | outputs = self.dropout(outputs) 74 | return outputs, words_input 75 | 76 | 77 | class Pointer(nn.Module): 78 | def __init__(self, input_dim): 79 | super(Pointer, self).__init__() 80 | self.input_dim = input_dim 81 | self.linear_info = nn.Linear(dec_hidden_size, self.input_dim, bias=False) 82 | self.linear_ctx = nn.Linear(self.input_dim, self.input_dim, bias=False) 83 | self.linear_query = nn.Linear(self.input_dim, self.input_dim, bias=True) 84 | self.projection = nn.Sequential( 85 | nn.Linear(3 * self.input_dim, 250), 86 | nn.Dropout(drop_rate), 87 | nn.ReLU(True), 88 | nn.Linear(250, 1) 89 | ) 90 | 91 | def forward(self, s_prev, enc_hs, cur_men_rep, src_mask): 92 | src_time_steps = enc_hs.size()[1] 93 | 94 | dh = self.linear_info(cur_men_rep) 95 | uh = self.linear_ctx(enc_hs) 96 | wq = self.linear_query(s_prev) 97 | pointer_input = torch.cat((uh, wq.repeat(1, src_time_steps, 1), dh.repeat(1, src_time_steps, 1)), 2) 98 | wquh = torch.tanh(pointer_input) 99 | attn_weights = self.projection(wquh).squeeze() 100 | attn_weights.data.masked_fill_(src_mask.squeeze().byte().data, -float('inf')) 101 | attn_weights = F.softmax(attn_weights, dim=-1) 102 | attn_weights_ = attn_weights.unsqueeze(0).unsqueeze(-1) 103 | ctx = (enc_hs * attn_weights_).sum(dim=1) 104 | return attn_weights, ctx 105 | 106 | 107 | class SelfAttention1(nn.Module): 108 | def __init__(self, input_dim): 109 | super(SelfAttention1, self).__init__() 110 | self.input_dim = input_dim 111 | self.projection = nn.Sequential( 112 | nn.Linear(self.input_dim, 50), 113 | nn.Dropout(drop_rate), 114 | nn.ReLU(True), 115 | nn.Linear(50, 1) 116 | ) 117 | 118 | def forward(self, hid_rep): 119 | hid_rep = torch.stack(hid_rep) 120 | hid_rep = hid_rep.unsqueeze(0) 121 | energy = self.projection(hid_rep) 122 | weights = F.softmax(energy.squeeze(-1), dim=1) 123 | att_rep = (hid_rep * weights.unsqueeze(-1)).sum(dim=1) 124 | 125 | return att_rep 126 | 127 | 128 | 129 | class SelfAttention2(nn.Module): 130 | def __init__(self, input_dim, drop_out_rate): 131 | super(SelfAttention2, self).__init__() 132 | self.input_dim = input_dim 133 | self.drop_rate = drop_out_rate 134 | self.pointer_lstm = nn.LSTM(2 * self.input_dim, self.input_dim, 1, batch_first=True, 135 | bidirectional=True) 136 | 137 | self.pointer_lin = nn.Linear(2 * self.input_dim, 1) 138 | self.dropout = nn.Dropout(self.drop_rate) 139 | 140 | 141 | def forward(self, hid_rep): 142 | src_time_steps = hid_rep.size()[1] 143 | 144 | pointer_lstm_out, phc = self.pointer_lstm(hid_rep) 145 | pointer_lstm_out = self.dropout(pointer_lstm_out) 146 | ptr_ = self.pointer_lin(pointer_lstm_out).squeeze() 147 | ptr_weights = F.softmax(ptr_, dim=-1) 148 | att_rep = torch.bmm(ptr_weights, hid_rep ) # .squeeze() 149 | 150 | return att_rep 151 | 152 | 153 | 154 | class Decoder(nn.Module): 155 | def __init__(self, input_dim, hidden_dim, layers, drop_out_rate, max_length): 156 | super(Decoder, self).__init__() 157 | self.input_dim = input_dim 158 | self.hidden_dim = hidden_dim 159 | self.layers = layers 160 | self.drop_rate = drop_out_rate 161 | self.max_length = max_length 162 | 163 | self.mPointer = Pointer(input_dim) 164 | self.lstm = nn.LSTMCell(self.input_dim + enc_hidden_size + enc_inp_size, self.hidden_dim) 165 | 166 | self.pointer_lstm = nn.LSTM(2 * self.input_dim, self.input_dim, 1, batch_first=True, 167 | bidirectional=True) 168 | 169 | self.pointer_lin = nn.Linear(2 * self.input_dim, 1) 170 | self.dropout = nn.Dropout(self.drop_rate) 171 | 172 | self.out_lin = nn.Linear(3 * self.input_dim, max_src_len) 173 | 174 | 175 | def forward(self, dec_inp, h_prev, enc_hs, cur_men_rep, src_mask, is_training=False): 176 | src_time_steps = enc_hs.size()[1] 177 | hidden, cell_state = self.lstm(dec_inp, h_prev) 178 | hidden = self.dropout(hidden) 179 | ptr_weights, ctx = self.mPointer(hidden, enc_hs, cur_men_rep, src_mask) 180 | out_input = torch.cat((ctx, hidden, cur_men_rep), -1) 181 | att_out_lin = self.out_lin(out_input) 182 | att_out_lin.data.masked_fill_(src_mask.squeeze().byte().data, -float('inf')) 183 | att_out_ = F.softmax(att_out_lin.squeeze(), dim=-1) 184 | 185 | return (hidden, cell_state), att_out_ 186 | 187 | 188 | -------------------------------------------------------------------------------- /framework/core/seq2seq.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.autograd as autograd 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | import torch.optim as optim 6 | from framework.core.module import * 7 | from framework.utils.common import * 8 | 9 | 10 | class Seq2SeqModel(nn.Module): 11 | def __init__(self, word_vocab_len, char_vocab_len, word_embed_matrix): 12 | super(Seq2SeqModel, self).__init__() 13 | self.encoder = Encoder(enc_inp_size, int(enc_hidden_size / 2), word_vocab_len, char_vocab_len, 14 | word_embed_matrix, 1, True, drop_rate) 15 | self.decoder = Decoder(dec_inp_size, dec_hidden_size, 1, drop_rate, max_trg_len) 16 | self.dropout = nn.Dropout(drop_rate) 17 | 18 | init_wt = np.ones_like(np.arange(max_src_len)) 19 | init_wt[0] = 50 20 | weit = torch.FloatTensor(init_wt) 21 | if torch.cuda.is_available(): 22 | weit = weit.cuda() 23 | self.myLoss = nn.NLLLoss(weight=weit, ignore_index=-1) 24 | self.att_a = SelfAttention1(dec_hidden_size) 25 | self.att_b = SelfAttention1(dec_hidden_size) 26 | self.lin_projection = nn.Sequential( 27 | nn.Linear(6 * dec_hidden_size, dec_hidden_size), 28 | nn.Dropout(drop_rate), 29 | nn.ReLU(True), 30 | ) 31 | 32 | def generate_rep_via_att(self, cur_pred_mention_reps_, cur_pred_mention_embs_): 33 | 34 | men_rep_ = self.att_a(cur_pred_mention_reps_) 35 | men_reps = torch.cat((cur_pred_mention_reps_[0], men_rep_.squeeze(1), cur_pred_mention_reps_[-1]), -1) 36 | men_emb_ = self.att_a(cur_pred_mention_embs_) 37 | men_embs = torch.cat((cur_pred_mention_embs_[0], men_emb_.squeeze(1), cur_pred_mention_embs_[-1]), -1) 38 | 39 | men_repre = torch.cat((men_reps, men_embs), -1) 40 | men_repre = self.lin_projection(men_repre) 41 | 42 | return men_repre 43 | 44 | def forward(self, src_words_seq, src_words_seq_len, src_mask, src_char_seq, pos_seq, trg_seq, gd_y, 45 | is_training=False): 46 | 47 | batch_len = src_words_seq.size()[0] # == 1 48 | src_time_steps = src_words_seq_len 49 | 50 | EOM_FLAG = 0 51 | NEXT_FLAG = src_time_steps - 1 52 | 53 | total_dec_time_steps = 0 54 | total_loss = 0.0 55 | 56 | ptr_trace = [] 57 | ptr_input_trace = [] 58 | 59 | pred_mentions = [] 60 | cur_pred_mention_ = [] 61 | 62 | cur_pred_mention_reps_ = [] 63 | cur_pred_mention_embs_ = [] 64 | 65 | enc_hs, src_word_embeds = self.encoder(src_words_seq, src_char_seq, pos_seq, is_training) 66 | 67 | h0 = autograd.Variable(torch.FloatTensor(torch.zeros(batch_len, dec_hidden_size))) 68 | c0 = autograd.Variable(torch.FloatTensor(torch.zeros(batch_len, dec_hidden_size))) 69 | if torch.cuda.is_available(): 70 | h0 = h0.cuda() 71 | c0 = c0.cuda() 72 | 73 | dec_hid = (h0, c0) 74 | cur_zero_men_rep = autograd.Variable(torch.FloatTensor(torch.zeros(batch_len, dec_hidden_size))) 75 | cur_men_rep = autograd.Variable(torch.FloatTensor(torch.zeros(batch_len, dec_hidden_size))) 76 | psedu_emb = autograd.Variable(torch.FloatTensor(torch.zeros(batch_len, enc_inp_size))) 77 | psedu_prev_enc_hid = autograd.Variable(torch.FloatTensor(torch.zeros(batch_len, enc_hidden_size))) 78 | if torch.cuda.is_available(): 79 | cur_zero_men_rep = cur_zero_men_rep.cuda() 80 | cur_men_rep = cur_men_rep.cuda() 81 | psedu_emb = psedu_emb.cuda() 82 | psedu_prev_enc_hid = psedu_prev_enc_hid.cuda() 83 | 84 | prev_dec_hid = torch.cat((cur_men_rep, psedu_emb, psedu_prev_enc_hid), -1) 85 | 86 | dec_outs = self.decoder(prev_dec_hid, dec_hid, enc_hs, cur_men_rep, src_mask, is_training) 87 | 88 | ptr_input_trace.append(-1) 89 | 90 | dec_hid = dec_outs[0] 91 | 92 | ptr_weights = dec_outs[1] 93 | ptr_index = ptr_weights.argmax(0).squeeze().data.cpu().numpy() 94 | last_ptr_index = ptr_index 95 | ptr_trace.append(ptr_index) 96 | 97 | if is_training: 98 | loss_ = self.myLoss(torch.log(ptr_weights.unsqueeze(0) + 1e-10), gd_y[:, 0]) 99 | total_loss += loss_ 100 | 101 | total_dec_time_steps += 1 102 | 103 | last_head_time_step = 0 104 | last_is_NEXT = True 105 | last_is_EOM = False 106 | 107 | cur_enc_time_step = trg_seq[1] 108 | first_NEXT = 1 109 | 110 | while cur_enc_time_step != src_time_steps - 1: 111 | ptr_input_trace.append(cur_enc_time_step) 112 | 113 | if is_training: 114 | gold_cur_enc_time_step = trg_seq[total_dec_time_steps] 115 | prev_dec_hid = torch.cat((cur_men_rep, src_word_embeds[:, gold_cur_enc_time_step, :], 116 | enc_hs[:, gold_cur_enc_time_step, :]), -1) + prev_dec_hid 117 | else: 118 | prev_dec_hid = torch.cat((cur_men_rep, src_word_embeds[:, cur_enc_time_step, :], 119 | enc_hs[:, cur_enc_time_step, :]), -1) + prev_dec_hid 120 | 121 | dec_outs = self.decoder(prev_dec_hid, dec_hid, enc_hs, cur_men_rep, src_mask, is_training) 122 | 123 | dec_hid = dec_outs[0] 124 | ptr_weights = dec_outs[1] 125 | 126 | if is_training: 127 | loss_ = self.myLoss(torch.log(ptr_weights.unsqueeze(0) + 1e-10), gd_y[:, total_dec_time_steps]) 128 | total_loss += loss_ 129 | 130 | ptr_index = ptr_weights.argmax(0).squeeze().data.cpu().numpy() 131 | 132 | if is_training: 133 | # teacher forcing on the route trace path 134 | ptr_index = gd_y[:, total_dec_time_steps].data.cpu().numpy().squeeze() 135 | 136 | ptr_trace.append(ptr_index) 137 | 138 | # path analyzing 139 | if ptr_index != NEXT_FLAG: 140 | 141 | if last_is_NEXT: 142 | last_head_time_step = cur_enc_time_step 143 | 144 | if (NEXT_FLAG not in ptr_trace): 145 | if first_NEXT == 2: 146 | last_head_time_step = last_ptr_non_NEXT_index 147 | 148 | if last_is_NEXT or last_is_EOM: 149 | cur_pred_mention_.append(cur_enc_time_step) 150 | else: 151 | cur_pred_mention_.append(last_ptr_index) 152 | 153 | if last_is_EOM: 154 | cur_pred_mention_.pop(len(cur_pred_mention_) - 1) 155 | 156 | if ptr_index != EOM_FLAG: 157 | cur_pred_mention_reps_.append(dec_hid[0]) 158 | cur_pred_mention_embs_.append(enc_hs[:, cur_enc_time_step, :]) 159 | 160 | last_is_EOM = False 161 | 162 | elif ptr_index == EOM_FLAG: 163 | last_is_EOM = True 164 | pred_mentions.append(','.join([str(item) for item in cur_pred_mention_])) 165 | cur_pred_mention_ = [] 166 | cur_pred_mention_reps_ = [] 167 | cur_pred_mention_embs_ = [] 168 | 169 | cur_enc_time_step = ptr_index 170 | 171 | if len(cur_pred_mention_reps_) > 0: 172 | men_repre = self.generate_rep_via_att(cur_pred_mention_reps_, cur_pred_mention_embs_) 173 | cur_men_rep = men_repre 174 | else: 175 | cur_men_rep = cur_zero_men_rep 176 | 177 | if first_NEXT == 1: 178 | last_ptr_non_NEXT_index = ptr_index 179 | last_is_NEXT = False 180 | first_NEXT += 1 181 | 182 | elif ptr_index == NEXT_FLAG: 183 | if last_is_EOM: 184 | cur_enc_time_step = (last_head_time_step + 1) 185 | else: 186 | cur_enc_time_step += 1 187 | 188 | last_is_EOM = False 189 | last_is_NEXT = True 190 | 191 | cur_pred_mention_reps_ = [] 192 | cur_pred_mention_embs_ = [] 193 | cur_men_rep = cur_zero_men_rep 194 | 195 | total_dec_time_steps += 1 196 | last_ptr_index = ptr_index 197 | 198 | if is_training or True: 199 | cur_enc_time_step = trg_seq[total_dec_time_steps] 200 | 201 | if is_training: 202 | return total_loss 203 | else: 204 | return pred_mentions, ptr_input_trace, ptr_trace 205 | -------------------------------------------------------------------------------- /framework/main.py: -------------------------------------------------------------------------------- 1 | import sys 2 | import os 3 | import numpy as np 4 | import warnings 5 | from collections import OrderedDict 6 | import datetime 7 | import json 8 | from tqdm import tqdm 9 | import math 10 | import torch 11 | import torch.nn as nn 12 | import torch.nn.functional as F 13 | import torch.optim as optim 14 | 15 | from framework.utils.data_util import * 16 | from framework.utils.util import * 17 | from framework.utils.eval import * 18 | from framework.utils.common import * 19 | from framework.core.seq2seq import Seq2SeqModel 20 | 21 | torch.backends.cudnn.deterministic = True 22 | 23 | warnings.filterwarnings('ignore') 24 | 25 | 26 | def get_model(model_id): 27 | if model_id == 1: 28 | return Seq2SeqModel(word_vocab_len, char_vocab_len, word_embed_matrix) 29 | 30 | 31 | def train_model(model_id, train_samples, dev_samples, best_model_file): 32 | train_size = len(train_samples) 33 | train_samples = train_samples[:int(train_size / 10)] 34 | train_size = len(train_samples) 35 | batch_count = int(math.ceil(train_size / batch_size)) 36 | custom_print(batch_count) 37 | model = get_model(model_id) 38 | pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad) 39 | custom_print('Parameters size:', pytorch_total_params) 40 | 41 | custom_print(model) 42 | if torch.cuda.is_available(): 43 | model.cuda() 44 | 45 | custom_print('weight factor:', wf) 46 | optimizer = optim.Adam(model.parameters(), lr=learning_rate, weight_decay=lr_rate_decay) 47 | custom_print(optimizer) 48 | 49 | best_dev_f = -1.0 50 | best_epoch_idx = -1 51 | for epoch_idx in range(0, num_epoch): 52 | model.train() 53 | model.zero_grad() 54 | custom_print('Epoch:', epoch_idx + 1) 55 | cur_seed = random_seed + epoch_idx + 1 56 | 57 | set_random_seeds(cur_seed) 58 | cur_shuffled_train_data = shuffle_data(train_samples) 59 | start_time = datetime.datetime.now() 60 | train_loss_val = 0.0 61 | 62 | optimizer = decay_learning_rate(optimizer, epoch_idx, learning_rate) 63 | 64 | for batch_idx in tqdm(range(0, batch_count)): 65 | batch_start = batch_idx * batch_size 66 | batch_end = min(len(cur_shuffled_train_data), batch_start + batch_size) 67 | cur_batch = cur_shuffled_train_data[batch_start:batch_end] 68 | 69 | outputs_loss = 0.0 70 | 71 | for ele in range(len(cur_batch)): 72 | src_words_seq_, src_words_mask_, src_chars_seq_, positional_seq_, \ 73 | trg_seq_, gd_y_, gold_mention_, src_words_seq_len_ = get_variablized_data(cur_batch[ele], word_vocab, 74 | char_vocab) 75 | 76 | outputs_loss += model(src_words_seq_, src_words_seq_len_, src_words_mask_, src_chars_seq_, 77 | positional_seq_, 78 | trg_seq_, gd_y_, True) 79 | 80 | outputs_loss.backward() 81 | torch.nn.utils.clip_grad_norm_(model.parameters(), 10.0) 82 | if (batch_idx + 1) % update_freq == 0: 83 | optimizer.step() 84 | model.zero_grad() 85 | train_loss_val += outputs_loss.item() 86 | 87 | train_loss_val /= batch_count 88 | end_time = datetime.datetime.now() 89 | custom_print('\nTraining loss:', train_loss_val) 90 | custom_print('Training time:', end_time - start_time) 91 | 92 | custom_print('\nDev Results\n') 93 | set_random_seeds(random_seed) 94 | dev_metrics = predict(train_samples, model) 95 | 96 | dev_p, dev_r, dev_f = dev_metrics 97 | 98 | if dev_f >= best_dev_f: 99 | best_epoch_idx = epoch_idx + 1 100 | best_epoch_seed = cur_seed 101 | custom_print('\nmodel saved......') 102 | best_dev_f = dev_f 103 | torch.save(model.state_dict(), best_model_file) 104 | 105 | custom_print('\n\n') 106 | if epoch_idx + 1 - best_epoch_idx >= early_stop_cnt: 107 | break 108 | 109 | custom_print('*******') 110 | custom_print('Best Epoch:', best_epoch_idx) 111 | custom_print('Best Dev F1:', best_dev_f) 112 | 113 | 114 | def predict(samples, model): 115 | model.eval() 116 | set_random_seeds(random_seed) 117 | start_time = datetime.datetime.now() 118 | 119 | gold_mentions = [] 120 | pred_mentions = [] 121 | 122 | gd_ptr_input_trace = [] 123 | prd_ptr_input_trace = [] 124 | gd_ptr_trace = [] 125 | prd_ptr_trace = [] 126 | for instance_idx in tqdm(range(0, len(samples))): 127 | src_words_seq_, src_words_mask_, src_chars_seq_, positional_seq_, \ 128 | trg_seq_, gd_y_, gold_mention_, src_words_seq_len_ = get_variablized_data(samples[instance_idx], word_vocab, 129 | char_vocab) 130 | gold_mentions.append(gold_mention_) 131 | 132 | with torch.no_grad(): 133 | pred_mentions_, ptr_input_trace, ptr_trace = model(src_words_seq_, src_words_seq_len_, src_words_mask_, 134 | src_chars_seq_, positional_seq_, 135 | trg_seq_, gd_y_, False) 136 | 137 | pred_mentions.append(pred_mentions_) 138 | prd_ptr_input_trace.append(ptr_input_trace) 139 | prd_ptr_trace.append(ptr_trace) 140 | gd_ptr_input_trace.append(trg_seq_) 141 | gd_ptr_trace.append(gd_y_.squeeze().data.cpu().numpy()) 142 | model.zero_grad() 143 | 144 | end_time = datetime.datetime.now() 145 | custom_print('\nPrediction time:', end_time - start_time) 146 | metrics = get_metrics(gold_mentions, pred_mentions) 147 | return metrics 148 | 149 | 150 | def custom_print(*msg): 151 | for i in range(0, len(msg)): 152 | if i == len(msg) - 1: 153 | print(msg[i]) 154 | logger.write(str(msg[i]) + '\n') 155 | else: 156 | print(msg[i], ' ', end='') 157 | logger.write(str(msg[i])) 158 | 159 | 160 | if __name__ == "__main__": 161 | 162 | n_gpu = torch.cuda.device_count() 163 | set_random_seeds(random_seed) 164 | os.environ['CUDA_VISIBLE_DEVICES'] = '-1' 165 | 166 | if job_mode == 'train': 167 | logger = open(os.path.join(trg_data_folder, 'training.log'), 'w') 168 | custom_print('loading data......') 169 | model_file_name = os.path.join(trg_data_folder, 'model.h5py') 170 | 171 | train_file = os.path.join(src_data_folder, 'train.txt') 172 | train_data = read_data(train_file, 1) 173 | 174 | dev_file = os.path.join(src_data_folder, 'dev.txt') 175 | dev_data = read_data(dev_file, 2) 176 | 177 | custom_print('Training data size:', len(train_data)) 178 | custom_print('Development data size:', len(dev_data)) 179 | custom_print("preparing vocabulary......") 180 | save_vocab = os.path.join(trg_data_folder, 'vocab.pkl') 181 | 182 | word_vocab, char_vocab, word_embed_matrix = build_vocab(train_data, save_vocab, embedding_file) 183 | word_vocab_len = len(word_vocab) 184 | char_vocab_len = len(char_vocab) 185 | # NEXT_FLAG = word_vocab[''] 186 | # EOM_FLAG = word_vocab[''] 187 | 188 | custom_print("Training started......") 189 | train_model(model_name, train_data, dev_data, model_file_name) 190 | logger.close() 191 | 192 | if job_mode == 'test': 193 | logger = open(os.path.join(trg_data_folder, 'test.log'), 'w') 194 | custom_print(sys.argv) 195 | custom_print("loading word vectors......") 196 | vocab_file_name = os.path.join(trg_data_folder, 'vocab.pkl') 197 | word_vocab, char_vocab = load_vocab(vocab_file_name, char_vocab) 198 | # NEXT_FLAG = word_vocab[''] 199 | # EOM_FLAG = word_vocab[''] 200 | 201 | word_embed_matrix = np.zeros((len(word_vocab), word_embed_dim), dtype=np.float32) 202 | custom_print('vocab size:', len(word_vocab)) 203 | 204 | model_file = os.path.join(trg_data_folder, 'model.h5py') 205 | 206 | best_model = get_model(model_name) 207 | custom_print(best_model) 208 | if torch.cuda.is_available(): 209 | best_model.cuda() 210 | if n_gpu > 1: 211 | best_model = torch.nn.DataParallel(best_model) 212 | best_model.load_state_dict(torch.load(model_file)) 213 | 214 | custom_print('\nTest Results\n') 215 | test_file = os.path.join(src_data_folder, 'test.txt') 216 | test_data = read_data(test_file, 3) 217 | 218 | print('Test size:', len(test_data)) 219 | set_random_seeds(random_seed) 220 | test_metrics = predict(test_data, best_model) 221 | -------------------------------------------------------------------------------- /framework/utils/common.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | S = '' 4 | S_pointer = -1 5 | EOM = '' 6 | EOM_pointer = 0 7 | NEXT = '' 8 | 9 | random_seed = 42 10 | 11 | src_data_folder = r'./data/examples' 12 | trg_data_folder = r'../output/exp1' 13 | if not os.path.exists(trg_data_folder): 14 | os.mkdir(trg_data_folder) 15 | 16 | model_name = 1 17 | job_mode = 'train' 18 | batch_size = 1 19 | num_epoch = 30 20 | lr_rate_decay = 0.05 21 | learning_rate = 0.00025 22 | 23 | max_src_len = 100 24 | max_trg_len = 300 25 | w2v_file_path = r'../emb' 26 | embedding_file = os.path.join(w2v_file_path, 'w2v.txt') 27 | update_freq = 1 28 | wf = 1.0 29 | att_type = 0 30 | 31 | use_hadamard = False 32 | gcn_num_layers = 3 33 | enc_type = 'LSTM' 34 | 35 | word_embed_dim = 300 36 | word_min_freq = 1 37 | 38 | char_embed_dim = 50 39 | char_feature_size = 50 40 | conv_filter_size = 3 41 | max_word_len = 10 42 | positional_embed_dim = 20 43 | max_positional_idx = 100 44 | 45 | enc_inp_size = word_embed_dim + char_feature_size + positional_embed_dim 46 | enc_hidden_size = word_embed_dim 47 | dec_inp_size = enc_hidden_size 48 | dec_hidden_size = dec_inp_size 49 | l1_type_embed_dim = 50 50 | 51 | drop_rate = 0.3 52 | layers = 2 53 | early_stop_cnt = 5 54 | sample_cnt = 0 55 | -------------------------------------------------------------------------------- /framework/utils/data_util.py: -------------------------------------------------------------------------------- 1 | from framework.utils.common import * 2 | from framework.utils.util import * 3 | import torch 4 | import torch.autograd as autograd 5 | from recordclass import recordclass 6 | from collections import OrderedDict 7 | import pickle 8 | 9 | import os, sys 10 | import numpy as np 11 | 12 | Sample = recordclass("Sample", "Id SrcSeqs TrgSeqs TrgOutput Mention") 13 | 14 | 15 | def read_data(data_file, datatype): 16 | ''' 17 | :param data_file: 18 | :param datatype: 1: training data 19 | :return: 20 | ''' 21 | # mode = {1:'train.txt', 2:'dev.txt', 3:'test.txt'} 22 | 23 | # with open(data_file, 'r') as reader: 24 | # data_lines = reader.readlines() 25 | data_lines = make_data(data_file) 26 | 27 | samples = [] 28 | uid = 1 29 | for i in range(0, len(data_lines)): 30 | src_line = data_lines[i].strip().split('|||') 31 | src_words = src_line[0].strip().split() 32 | 33 | tgt_pointers = src_line[1].strip().split() # 第一个: -1 34 | tgt_pointers = [int(item) for item in tgt_pointers] 35 | 36 | tgt_output_pointer = src_line[2].strip().split() # 37 | tgt_output_pointer = [int(item) for item in tgt_output_pointer] 38 | 39 | mentions = src_line[3].strip().split('>') # 40 | mentions = [item for item in mentions if item != ''] 41 | # mentions = '|'.join(mentions) 42 | 43 | if datatype == 1 and (len(src_words) > max_src_len or len(tgt_pointers) > max_trg_len): 44 | continue 45 | 46 | sample = Sample(Id=uid, SrcSeqs=src_words, TrgSeqs=tgt_pointers, 47 | TrgOutput=tgt_output_pointer, Mention=mentions) 48 | samples.append(sample) 49 | uid += 1 50 | 51 | return samples 52 | 53 | 54 | def make_data(data_folder): 55 | max_len = 0 56 | max_tgt_len = [] 57 | lens = [] 58 | 59 | id_ = 0 60 | instances = [] 61 | with open(data_folder, "r") as f: 62 | for sentence in f: 63 | # print(sentence) 64 | tokens = [t for t in sentence.strip().split()] 65 | tokens.insert(0, EOM) 66 | tokens.append(NEXT) 67 | 68 | lens.append(len(tokens)) 69 | if len(tokens) > max_len: 70 | max_len = len(tokens) 71 | 72 | annotations = next(f).strip() 73 | # actions = self.parse.mention2actions(annotations, len(tokens)) 74 | # oracle_mentions = [str(s) for s in self.parse.parse(actions, len(tokens))] 75 | gold_mentions = annotations.split("|") if len(annotations) > 0 else [] 76 | gold_mentions_ = [] 77 | # flag = False 78 | 79 | head_index = [] 80 | for men in gold_mentions: 81 | indexe_ = men.strip().split()[0].strip().split(',') 82 | indexes = [int(item) + 1 for item in indexe_] 83 | assert len(indexes) % 2 == 0 84 | indexes = list(set(indexes)) 85 | indexes.sort() 86 | head_index.append(indexes[0]) 87 | # if len(indexes) > 4: flag = True 88 | gold_mentions_.append(indexes) 89 | 90 | gold_mentions_.sort() 91 | head_index = list(set(head_index)) 92 | head_index.sort() 93 | # if flag: 94 | # print(file_n, gold_mentions_) 95 | 96 | NEXT_pointer = len(tokens) - 1 97 | tgt_seq_pointers = [] 98 | tgt_output_pointers = [] 99 | tgt_seq_pointers.append(S_pointer) 100 | tgt_output_pointers.append(NEXT_pointer) 101 | for idx, item in enumerate(tokens): 102 | if idx == 0: continue # skip the EOM token (as S) 103 | if idx in head_index: 104 | for cur_men in gold_mentions_: 105 | if cur_men[0] == idx: 106 | for posi_ in range(0, len(cur_men) - 1): 107 | tgt_seq_pointers.append(cur_men[posi_]) 108 | tgt_output_pointers.append(cur_men[posi_ + 1]) 109 | tgt_seq_pointers.append(cur_men[-1]) 110 | tgt_output_pointers.append(EOM_pointer) 111 | 112 | tgt_seq_pointers.append(EOM_pointer) 113 | tgt_output_pointers.append(idx) 114 | 115 | tgt_output_pointers.pop(len(tgt_output_pointers) - 1) # replace as NEXT 116 | tgt_output_pointers.append(NEXT_pointer) 117 | 118 | else: 119 | tgt_seq_pointers.append(idx) 120 | tgt_output_pointers.append(NEXT_pointer) 121 | 122 | tokens_str = ' '.join(tokens) 123 | tgt_seq_pointers = ' '.join([str(item) for item in tgt_seq_pointers]) 124 | tgt_output_pointers = ' '.join([str(item) for item in tgt_output_pointers]) 125 | gold_mentions_str = '>'.join([','.join([str(item) for item in men]) for men in gold_mentions_]) 126 | # assemble: 127 | instances.append( 128 | tokens_str + '|||' + tgt_seq_pointers + '|||' + tgt_output_pointers + '|||' + gold_mentions_str) 129 | 130 | assert len(next(f).strip()) == 0, f.readline() 131 | 132 | max_tgt_len.append(len(tgt_seq_pointers)) 133 | 134 | id_ += 1 135 | 136 | return instances 137 | 138 | 139 | def get_variablized_data(sample, word_vocab, char_vocab): 140 | src_words_seq_ = torch.from_numpy(np.array([get_words_index_seq(sample.SrcSeqs, word_vocab)])) 141 | src_words_mask_ = torch.from_numpy(np.array([get_padded_mask(len(sample.SrcSeqs))])) 142 | src_chars_seq_ = torch.from_numpy(np.array([get_char_seq(sample.SrcSeqs, char_vocab)])) 143 | positional_seq_ = torch.from_numpy(np.array([get_positional_index(len(sample.SrcSeqs))])) 144 | gd_y_ = torch.from_numpy(np.array([get_pointers(sample.TrgOutput)])) 145 | trg_seq_ = get_pointers(sample.TrgSeqs) 146 | 147 | if torch.cuda.is_available(): 148 | src_words_seq_ = src_words_seq_.cuda() 149 | src_words_mask_ = src_words_mask_.cuda() 150 | src_chars_seq_ = src_chars_seq_.cuda() 151 | positional_seq_ = positional_seq_.cuda() 152 | gd_y_ = gd_y_.cuda() 153 | 154 | src_words_seq_ = autograd.Variable(src_words_seq_) 155 | src_words_mask_ = autograd.Variable(src_words_mask_) 156 | src_chars_seq_ = autograd.Variable(src_chars_seq_) 157 | positional_seq_ = autograd.Variable(positional_seq_) 158 | gd_y_ = autograd.Variable(gd_y_) 159 | 160 | gold_mentions = sample.Mention 161 | 162 | src_words_seq_len_ = len(sample.SrcSeqs) 163 | 164 | return src_words_seq_, src_words_mask_, src_chars_seq_, positional_seq_, \ 165 | trg_seq_, gd_y_, gold_mentions, src_words_seq_len_ 166 | 167 | 168 | def shuffle_data(data): 169 | # custom_print(len(data)) 170 | data.sort(key=lambda x: x.Id) 171 | # num_batch = int(len(data) / batch_size) 172 | # rand_idx = random.sample(len(data))#range(num_batch), num_batch) 173 | new_data = random.sample(data, len(data)) 174 | return new_data 175 | 176 | 177 | def get_words_index_seq(words, word_vocab): 178 | seq = list() 179 | for word in words: 180 | if word in word_vocab: 181 | seq.append(word_vocab[word]) 182 | else: 183 | seq.append(word_vocab['']) 184 | pad_len = max_src_len - len(words) 185 | for i in range(0, pad_len): 186 | seq.append(word_vocab['']) 187 | return seq 188 | 189 | 190 | def get_char_seq(words, char_vocab): 191 | char_seq = list() 192 | for i in range(0, conv_filter_size - 1): 193 | char_seq.append(char_vocab['']) 194 | for word in words: 195 | for c in word[0:min(len(word), max_word_len)]: 196 | if c in char_vocab: 197 | char_seq.append(char_vocab[c]) 198 | else: 199 | char_seq.append(char_vocab['']) 200 | pad_len = max_word_len - len(word) 201 | for i in range(0, pad_len): 202 | char_seq.append(char_vocab['']) 203 | for i in range(0, conv_filter_size - 1): 204 | char_seq.append(char_vocab['']) 205 | 206 | pad_len = max_src_len - len(words) 207 | for i in range(0, pad_len): 208 | for i in range(0, max_word_len + conv_filter_size - 1): 209 | char_seq.append(char_vocab['']) 210 | 211 | return char_seq 212 | 213 | 214 | def get_padded_mask(cur_len): 215 | mask_seq = list() 216 | for i in range(0, cur_len): 217 | mask_seq.append(0) 218 | pad_len = max_src_len - cur_len 219 | for i in range(0, pad_len): 220 | mask_seq.append(1) 221 | return mask_seq 222 | 223 | 224 | def get_positional_index(sent_len): 225 | index_seq = [min(i + 1, max_positional_idx - 1) for i in range(sent_len)] 226 | index_seq += [0 for _ in range(max_src_len - sent_len)] 227 | return index_seq 228 | 229 | 230 | def get_pointers(seqs): 231 | pointer_list = [] 232 | for item in seqs: 233 | pointer_list.append(item) 234 | return pointer_list 235 | 236 | 237 | def load_word_embedding(embed_file, vocab): 238 | # custom_print('vocab length:', len(vocab)) 239 | embed_vocab = OrderedDict() 240 | embed_matrix = list() 241 | 242 | embed_vocab[''] = 0 243 | embed_matrix.append(np.zeros(word_embed_dim, dtype=np.float32)) 244 | 245 | embed_vocab[''] = 0 246 | embed_matrix.append(np.random.uniform(-0.25, 0.25, word_embed_dim)) 247 | 248 | word_idx = 2 249 | with open(embed_file, "r") as f: 250 | for line in f: 251 | parts = line.split() 252 | if len(parts) < word_embed_dim + 1: 253 | continue 254 | word = parts[0] 255 | if word in vocab and vocab[word] >= word_min_freq: 256 | vec = [np.float32(val) for val in parts[1:]] 257 | embed_matrix.append(vec) 258 | embed_vocab[word] = word_idx 259 | word_idx += 1 260 | 261 | for word in vocab: 262 | if word not in embed_vocab and vocab[word] >= word_min_freq: 263 | embed_matrix.append(np.random.uniform(-0.25, 0.25, word_embed_dim)) 264 | embed_vocab[word] = word_idx 265 | word_idx += 1 266 | 267 | # custom_print('embed dictionary length:', len(embed_vocab)) 268 | return embed_vocab, np.array(embed_matrix, dtype=np.float32) 269 | 270 | 271 | def build_vocab(data, save_vocab, embedding_file): 272 | vocab = OrderedDict() 273 | char_v = OrderedDict() 274 | char_v[''] = 0 275 | char_v[''] = 1 276 | char_idx = 2 277 | for d in data: 278 | for word in d.SrcSeqs: 279 | if word not in vocab: 280 | vocab[word] = 1 281 | else: 282 | vocab[word] += 1 283 | 284 | for c in word: 285 | if c not in char_v: 286 | char_v[c] = char_idx 287 | char_idx += 1 288 | 289 | word_v, embed_matrix = load_word_embedding(embedding_file, vocab) 290 | output = open(save_vocab, 'wb') 291 | pickle.dump([word_v, char_v], output) 292 | output.close() 293 | return word_v, char_v, embed_matrix 294 | 295 | 296 | def load_vocab(vocab_file, char_vocab): 297 | with open(vocab_file, 'rb') as f: 298 | embed_vocab, char_vocab = pickle.load(f) 299 | return embed_vocab, char_vocab 300 | 301 | # if __name__ == '__main__': 302 | # main() 303 | -------------------------------------------------------------------------------- /framework/utils/eval.py: -------------------------------------------------------------------------------- 1 | from framework.utils.common import * 2 | from framework.utils.util import * 3 | 4 | 5 | def get_metrics(gold_mentions, pred_mentions): 6 | gt_ = 0 7 | pred_ = 0 8 | correct_ = 0 9 | 10 | for gts, pds in zip(gold_mentions, pred_mentions): 11 | gt_ += len(gts) 12 | pred_ += len(pds) 13 | for pd in pds: 14 | if pd in gts: 15 | correct_ += 1 16 | 17 | print(pred_, '\t|\t', gt_, '\t|\t', correct_) 18 | p = float(correct_) / (pred_ + 1e-18) 19 | r = float(correct_) / (gt_ + 1e-18) 20 | F1 = (2 * p * r) / (p + r + 1e-18) 21 | print('P:', p, 'R:', r, 'F1:', F1) 22 | 23 | return p, r, F1 24 | -------------------------------------------------------------------------------- /framework/utils/util.py: -------------------------------------------------------------------------------- 1 | import framework.main as main 2 | import random 3 | import numpy as np 4 | import torch 5 | from framework.utils.common import * 6 | 7 | 8 | def set_random_seeds(seed): 9 | random.seed(seed) 10 | np.random.seed(seed) 11 | torch.manual_seed(seed) 12 | # if main.n_gpu > 1: 13 | # torch.cuda.manual_seed_all(seed) 14 | 15 | 16 | def save_output_dev(gold_mentions, pred_mentions, outfile): 17 | writer = open(outfile, 'w') 18 | for i in range(0, len(gold_mentions)): 19 | gold_mention_ = gold_mentions[i] 20 | pred_mention_ = pred_mentions[i] 21 | 22 | cur_str = 'golds: ' + '\t'.join('[' + item + ']' for item in gold_mention_) 23 | cur_str += "\n" 24 | cur_str += 'preds: ' + '\t'.join('[' + item + ']' for item in pred_mention_) 25 | cur_str += "\n\n" 26 | 27 | writer.write(cur_str) 28 | writer.close() 29 | 30 | 31 | def save_trace_dev(trg_seq, prd_ptr_input_trace, gd_ptr_trace, prd_ptr_trace, outfile): 32 | writer = open(outfile, 'w') 33 | for i in range(0, len(trg_seq)): 34 | trg_seq_ = trg_seq[i] 35 | prd_trg_seq_ = prd_ptr_input_trace[i] 36 | 37 | gd_output_ = gd_ptr_trace[i] 38 | prd_output_ = prd_ptr_trace[i] 39 | 40 | cur_str = 'input golds: ' + '\t'.join('[' + str(item) + ']' for item in trg_seq_) 41 | cur_str += "\n" 42 | cur_str += 'input preds: ' + '\t'.join('[' + str(item) + ']' for item in prd_trg_seq_) 43 | cur_str += "\n------\n" 44 | cur_str += 'output golds: ' + '\t'.join('[' + str(item) + ']' for item in gd_output_) 45 | cur_str += "\n" 46 | cur_str += 'output preds: ' + '\t'.join('[' + str(item) + ']' for item in prd_output_) 47 | cur_str += "\n\n" 48 | 49 | writer.write(cur_str) 50 | writer.close() 51 | 52 | 53 | def decay_learning_rate(optimizer, iter, init_lr): 54 | lr = init_lr / (1 + lr_rate_decay * iter) 55 | for param_group in optimizer.param_groups: 56 | param_group['lr'] = lr 57 | return optimizer 58 | -------------------------------------------------------------------------------- /output/exp1/checkpoint/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scofield7419/DisNER-PtrNet/5c748d09c36c42b4828338fba6c5b7ba3f326c04/output/exp1/checkpoint/__init__.py -------------------------------------------------------------------------------- /output/exp1/log/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scofield7419/DisNER-PtrNet/5c748d09c36c42b4828338fba6c5b7ba3f326c04/output/exp1/log/__init__.py --------------------------------------------------------------------------------