Putting Words in BERT’s Mouth: Navigating Contextualized Vector Spaces with Pseudowords

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10300–10313 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10300 Putting Words in BERT’s Mouth: Navigating Contextualized Vector Spaces with Pseudowords Taelin Karidi1 Yichu Zhou2 Nathan Schneider3 Omri Abend1 Vivek Srikumar2 1Hebrew University of Jerusalem, {taelin.karidi, omri.abend}@mail.huji.ac.il 2University of Utah, {flyaway, svivek}@cs.utah.edu 3Georgetown University, nathan.schneider@georgetown.edu Abstract We present a method for exploring regions around individual points in a contextualized vector space (particularly, BERT space), as a way to investigate how these regions corre- spond to word senses. By inducing a con- textualized “pseudoword” as a stand-in for a static embedding in the input layer, and then performing masked prediction of a word in the sentence, we are able to investigate the geom- etry of the BERT-space in a controlled man- ner around individual instances. Using our method on a set of carefully constructed sen- tences targeting ambiguous English words, we find substantial regularity in the contextualized space, with regions that correspond to distinct word senses; but between these regions there are occasionally “sense voids”—regions that do not correspond to any intelligible sense.1 1 Introduction Vector spaces defined over static word vectors are somewhat interpretable, as the points are limited to the vocabulary. Contextualized representations (CRs), by contrast, are mysterious because of the unbounded number of distinct contextualized em- beddings, and no obvious way to discover the word and context that would correspond to an arbitrary point in the space. Attempts have been made to characterize the information captured in contextu- alized representations (Rogers et al., 2020; Tenney et al., 2019; Liu et al., 2019), but some of the tech- niques used (e.g., probing classifiers) have been subject to criticism for their indirectness. We propose a new technique called Masked Pseudoword Probing (MaPP) that allows con- trolled exploration of the space of a contextualized masked LMs (specifically, English BERT; Devlin et al., 2019). MaPP takes advantage of the static embedding at the first layer of BERT and “hallu- 1Our code and dataset are available at https://github. com/tai314159/PWIBM-Putting-Words-in-Bert-s-Mouth 1 BERT(The event is in October.) Learn pseudoword in place of that is customized to reconstruct . static inputs: contextualized outputs: Run masked prediction with the modified input vector. BERT(The event is in [MASK].) October ✓ July ✓ winter ✓ London ✗ ##ic ✗ Examine top predictions for sense match. Run BERT for a sentence. ➋ ➊ ➌ ➍ <latexit sha1_base64="rtRl6Yut6LufEGEe8AOUfx1c0dU=" >ADkHicfZJbaxQxFMfTHa1vW3roy/BUhCRZXba6u6TFX0QWyh2xasmSyZy40mQxJZi8O+RC+BX8Lj74Ln4NXxUzuy3Ldlo PBA7/3z+Xc3LCnKfa+P7PlYZ36/bqnbW7zXv3Hzx81FrfONKyUAz6THKpTkKqgacZ9E1qOJzkCqgIORyH528rfjwCpVOZHZpDmeC xlkapYwaJw1aCRHUJGFUfraDkhiYmPIwAWsJrgMYQWauR6m+Qc+sxWQojR60Nv12r+fv7Oxiv73rB0HQdYm/HXR7Hdxp+7PY3Fs9+ PX965dv+4P1RstZIVwlzJOtT7t+Lk5K6kyKeNgm6TQkFN2TmMoZ42weMtJQxJ5VZm8Exd8lUPzI2YOCs2qQCNK2XJQoXWUxE6R 4X0VaJ17HTwkTdM1dwXhjI2PwtUcGxkbhqPB6mCpjhU5dQplJXBGYJVZQZ9z1N8g5ckQo+unM/5aCokep5SaiKBZ1YV3RMXlTZ/4 yu2RdGlzW3SAZjJoWg2bAkY1te/s7YXoWj0YKOajRawKgG4wWMazBZwKQGXVUklHxYNVvyclIzTJcNU2vdDF0OCr45OQranZft4MA NU4DmsYaeoKfoGeqgV2gPvUf7qI8Y+oF+oz/or7fhdb3X3pu5tbFysecxWgrvwz/q7zpu</latexit>zThe zevent zis zin ... <latexit sha1_base64="yEw52jcYOQbo5QxAJyMNMUtfnw=" >ADkHicfZJPaxQxGMbTHa1/dNtPXoJloKILDN7aLcnK3oQWyh2xaZclk35kJTSZDktk/DPkQXvwKfhcP3sWv4VUx021Zdrc 1MPDw/J5k8r540JwY8Pw51ojuHd/cHGw+ajx0+ebra2tk+NKjWDHlNC6fOYGhA8h57lVsB5oYHKWMBZfPmu5mcj0Iar/MROC+hL muY84Yxabw1aGZHUZnFSTdygIhYmtjrJwDmCVwGMILe3I27u8HPnMBkqawatnbAdhmEURbgW0f5e6MXBQbcTdXFUI792DtePf3/+ uXb0WCr0fIbWSn9T5mgxlxEYWH7FdWMwGuSUoDBWXNIXqhEO73priBOl/ZdbfOUu5OoLFlZOfBRbLsHg2lmIUGnMVMY+USOz GrzNnZR2qTb9wUXpYWcze6SlAJbhevG4yHXwKyYekGZ5r4IzDKqKbP+eZrkPfgiNXzy534uQFOr9KuKUJ1KOnG+6JS8rtX/gr7Z10 GvmrskhzFTUtJ8WJGxq25eZ+yW4Wg0p6MVmsxhsgLTOUxXYDaH2Qr0VZFYiWHdbCX8zCwHpouBqXN+hm4GBd8tTjvtaK/dOfbD1EG ztYGeoxfoJYrQPjpEH9AR6iGfqDf6A/6G2wH3eBN8HYWbaxd73mGFlbw8R+edTpT</latexit>xThe xevent xis xin ... <latexit sha1_base64="fX1JdZ4z0q0CfXLwZ4IE4RzvY=" >ADQXicfZLfatRAFMan8V9drd3WK/FmsBRKkWUSdN29K+iFN2ILblto1mUymWSHzmTizGR3Yxh8AJ/DO9GH8CnEJ/CueOuNk27 Lso16IHD4fl/C+U5OlHOmDULfV7xr12/cvLV6u3Xn7tq9fbG5qGWhSJ0QCSX6jCmnKW0YFhtPjXFEsIk6PotPnNT+aUKWZzN6Y MqdDgdOMJYxg46R+0EosBlHSfXevt0dVaGhM1OxzFo4am+hTr/XRU8CiDoI9QLUdc1T5Pf9PvSdUtfW3ubOj4+f363tjza8dhLU giaGcKx1ic+ys2wsowqlthYWmOSanOKXV+eQWbjspholU7skMPFeXfPV0uREzZ4WGCaphrSxZsNC6FJFz1EhfZbX4N3ZSmKQ3d FnzwtCMzGdJCg6NhPWmYMwUJYaXrsFEMRcCkjFWmBi3z1b4grqQir5y32dU4WNVLtViFUq8My60Gn4uO7+Z3R7vjC6rUdZnRKpB A4i6twaqvLXzO1V+FksqCTBk0WMGnAdAHTBhwv4LgBXaowkjyuly15NWsYymVDa27octDgf9uDoO3+0EB+6YAjCvVfAQPAI7wAf PwB54CfbBABDwAXwCX8BX75v30zvzfs2t3srFO/fBUnm/wD5ARof</latexit>z⇤ in <latexit sha1_base64="PFrDajDHYwuGYWrQW/ks/ZLSBpw=" >ADP3icfZLNbtNAFIWnLpQSfpqWBQs2I6pKCKHINi0ku0qwYINoJdJW6kTReDx2Rp3xWDPjJMah2DC7BgAw/BljdAPA7xJY d4QoSg1cydLR+Y6te69vlHOmje9/XfPWr1zduLZ5vXj5q3bW+3tnRMtC0Von0gu1VmENeUso3DKdnuaJYRJyeRhfPan46pkoz mb02ZU4HAqcZSxjBxlnD9l0ksBlFSfXGDitk6NRUTFsLh+1dv9Pr+fv7B9DvHPhGHad8B+H3V4Ag4/q93DjeNvX969/XA03PbaK JakEDQzhGOtzwM/N4MK8MIp7aFCk1zTC5wSqtZ3xbuOSuGiVTuyQycuSu5urfciKmLQsME1bB2ViJYaF2KyCVqpC+z2vwbOy9M0 h1ULMsLQzMy7yUpODQS1nuCMVOUGF46gYlibghIRlhYtw2W+g5dUMq+tJ91VOFTZSPawQVqnAU+uGTtGjWv0vyLJF0KnWHsrohE ghcBZXaGKrxY+Z2MtwPF7ScYMmS5g0YLqEaQOlnDUgG4qFEke18uWvJo2AuVqoLTW3dDiUOC/xUnYCZ50wmN3TCGY1ya4B+6DByA AT8EheAGOQB8QYMF78BF8j57370f3s951Fv784dsFLer9/QGBqL</latexit>zis <latexit sha1_base64="2dLuqRUZLU0BWNzWO4niWO+3Cc=" >ADQnicfZLPbtNAEMa3BkoJ/9Jyg8uKqhJCyHJMS5NbJThwQbQSaSvVUbRej51Vd73W7tqJsSzxClx4BDjCQ3DiDRBvwA1x5cA 6IYrSACOtNPp+365mZifMONPG876uOZcuX1m/unGtdf3GzVu325tbx1rmikKfSi7VaUg0cJZC3zD4TRTQETI4SQ8f9rwkwKUZjJ9 ZcoMBoIkKYsZJcZKw/bdQBAzCuPqdT2sAgMTU0EBqalrPGxve26v5+3u7mHP3fN83+/axHvsd3sd3HG9aWwfrB9+/Lu7YfD4abTD iJc2HvU060Put4mRlURBlGOdStINeQEXpOEqimpd4x0oRjqWyJzV4qi75mvIyIybWig0ToHGjLFmI0LoUoXU0SF9kjfg3dpabu DuoWJrlBlI6qyXOTYSN6PCEVNADS9tQqhitglMR0QRauxAW8EzsE0qeGHfZmBIkaqh1VAVCLIpLZNJ8GjJvufkaVzo81aO0EKYy qFIGlUBeO6mv/NuL4Ii2JBixUaL2C8ApMFTFbgaAFHK9B2FYSR82wJa8mK4Zy2VDWtd2h+aLgfyfHvt54vpHdpl8NIsNdA/dRw9 QB+2jA/QcHaI+ougNeo8+ok/OZ+e78P5ObM6a3/u3EFL4fz6DeavG+8=</latexit>zevent <latexit sha1_base64="8jtPwyDCxZJMBcUgZX3ahbSMONA=" >ADQHicfZLPa9RAFMenUWtdf3SrF8HLYCmIyJLEVndvBT14EVvotoVmWSazL8nQmUyYmexuDPF/8OJfIHjSP8Kj/4H4D3gTr56 c7Los26gPAl+n2/Cey8vzDjTxnW/rjmXLl9Zv7pxrX9xs1bm+2t28da5opCn0ou1WlINHCWQt8w+E0U0BEyOEkPH9W85MxKM1k emSKDAaCxCmLGCXGWsP23UAQk4R+boaloGBqSmPEqgqPGxvu51ez93d3cNuZ8/1fb9rhfvY7/Y87HXcW3vrx9+/Lu7YeD4ZbTD kaS5gJSQznR+sxzMzMoiTKMcqhaQa4hI/ScxFDOGq/wjrVGOJLKPqnBM3clVzeXGTG1UWyYAI1rZyVChNaFCG2iRvoiq82/sbPcR N1BydIsN5DSeS9RzrGRuF4UHjEF1PDCkIVs0NgmhBFqLHrbAXPwQ6p4KX97qsMFDFSPSwDomJBpUdOg4e1ep/QZYugla1doIUJl QKQdJRGUyqcvFnJtVFOB4v6bhBoyWMGjBewrgBkyVMGtBOFYSj+plS15OG4FiNVBUlb2hxaHgf4tjv+M96fiH9ph8NK8NdA/dRw+ Qh56ifQCHaA+ougNeo8+ok/OZ+e78P5OY86a3/euYNWyvn1G75QGto=</latexit>zThe <latexit sha1_base64="TKSOYPqgwH8PsP6BDeOeCBsisK0=" >ADP3icfZLNbtQwFIXdAKUMf1NYsGBjUVCEVJaGFmVwkWbBCtxLSV6tHIcZyMVTuObGdmQuSHYMLsGAD8GWN0A8ADvElh3 OTKvRNMCVIl2d7yS65+bGBWfaBMG3Ne/S5SvrVzeuda7fuHnrdnfzqGWpSJ0QCSX6jGmnKW04FhtPjQlEsYk6P4tPnDT+aUKWZ zN+YqBDgbOcpYxg46R9x4S2IzjtH5rRzUydGZqlsLR92twO/3g52dXRj4u0EURT3XBE+iXj+EoR/Ma2tv/eD71/fvPu6PNr0uS iQpBc0N4VjrkzAozLDGyjDCqe2gUtMCk1Oc0Xo+t4XbTkpgKpV7cgPn6oqvma0wYuas0DBNWyUFQsWlcido4G6YusEf/GTkqT9 oYua1EampPFLGnJoZGw2RNMmKLE8Mo1mCjmQkAyxgoT47bZQS+oC6noK/fd1wV2Ej1qEZYZQLPrAudocdN9z+j2/OZ0XWdbZTKZ FC4Dyp0dTW5z9mai/CyWRJy2aLmHagtkSZi04XsJxC7pUKJY8aZYteT1rGapVQ2Wtu6HzQ4H/bg4jP3zqRwfumCKwqA1wHzwAD0E InoE98BLsgwEgwIP4BP47H3xfng/vV8Lq7d29s5dsFLe7z/COhqG</latexit>zin <latexit sha1_base64="A3RE+X5OivYqN72L3DbznsihaMQ=" >ADNHicfZLNbtNAFIWnboESfpqWDRKbESkSqlA0tiAku0qwYIMoEmkr1SEaT6dUWc81sw4ibHM07CFJe+CxA6x5RkYt6mimJ8 rWTo637F17/WNMsGNJeTbhre5de36je2brVu379zdae/uHRuVawZDpoTSpxE1IHgKQ8utgNMA5WRgJPo/EXNT2agDVfpO1tkMJI0 SXnMGbXOGrfv74eS2mkUlx+q9wfj0MLCljyt9sftDukO+j3yNMCkS0g/ID0nhF/4A+w75y6OmhZR+NdbyucKJZLSC0T1Jgzn2R2V FJtORNQtcLcQEbZOU2gvGi8wo+cNcGx0u5JLb5w13J1b5mVCxfFlkswa5RKYwoZOVjn/mC1+Td2ltu4P3JTZrmFlF2EecCW4XrH eEJ18CsKJygTHPXP2ZTqimzbpOt8CW4+TS8dt9k4GmVumDMqQ6kXRuXmT8Emt/hd0G14GnWqFKcyZkpKmkzKcV+XVL5lXDTabre CsCeMVi5sWbGkyaYrNm0yN08YKTGp16xEuWjyYp0XVeXu5uo48L/FcdD1e93gbdA5DJYXtI0eoIfoMfLRc3SIXqEjNEQMfUSf0Gf 0xfvqfd+eD8vo97G8p17aK28X78BO2gTKA=</latexit>z⇤ in <latexit sha1_base64="6X6bdxFGrSfzht/pBWXyr6qhurQ=" >ADMnicfZJNb9NAEIa3boESPpqCxIXLihQJIRTZPrTprRIcuCKRNpKdRStN2Nn1V2vtTtOYrn+M1zhyJ+BG+LKj2DdpopiPkZ a6dU8765mZifOpbDo+982vM2tW7fvbN/t3Lv/4OFOd/fRidWF4TDkWmpzFjMLUmQwRIESznIDTMUSTuOL1w0/nYGxQmcfscxhpFia iURwhi417j7ZixTDaZxUi3ocISywElm9N+72/L7v+0EQ0EYEB/u+E4eHgzAY0KBLnpkGcfjXW8rmheKMiQS2bteDnOKqYQcEl1 J2osJAzfsFSqK7Krulzl5rQRBt3MqRX2TVfU1mOauGsFIUCu0aZsrZUsYON7w/WJP/GzgtMBiPXZV4gZPy6jKSQFDVtJkQnwgBHW TrBuBGufsqnzDCObo6d6A24/gy8c+z8Ew1OZlFTGTKraoXb9p9KpR/zO6CS+NTnWiDOZcK8WySRXN6+rmQ+Z1i81mKzhrw2TFkj ZLVyxts+mKTdvM9RPFWk6aMWvpNqTFy3Ve1rXbm5vloP8WJ2E/2O+H8LeUbjcoG3ylDwjL0hADsgReUuOyZBwck+kc/ki/fV+7 98H5eW72N5Z3HZC28X78BPSASeg=</latexit>xin Figure 1: Illustration of the MaPP method as used in the …

Content truncated to 500 words