Datasets:
Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +140 -0
- parse/dev/-P7G-8dmSh4/-P7G-8dmSh4_span.pdf +3 -0
- parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_layout.pdf +3 -0
- parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_origin.pdf +3 -0
- parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_span.pdf +3 -0
- parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt.md +432 -0
- parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt_content_list.json +0 -0
- parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt_middle.json +0 -0
- parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt_model.json +0 -0
- parse/dev/0xiJLKH-ufZ/0xiJLKH-ufZ_layout.pdf +3 -0
- parse/dev/0xiJLKH-ufZ/0xiJLKH-ufZ_span.pdf +3 -0
- parse/dev/1QQnYd02etI/1QQnYd02etI_layout.pdf +3 -0
- parse/dev/1QQnYd02etI/1QQnYd02etI_origin.pdf +3 -0
- parse/dev/1QQnYd02etI/1QQnYd02etI_span.pdf +3 -0
- parse/dev/1wVvweK3oIb/1wVvweK3oIb.md +474 -0
- parse/dev/1wVvweK3oIb/1wVvweK3oIb_content_list.json +0 -0
- parse/dev/1wVvweK3oIb/1wVvweK3oIb_middle.json +0 -0
- parse/dev/1wVvweK3oIb/1wVvweK3oIb_model.json +0 -0
- parse/dev/2EDqbSCnmF/2EDqbSCnmF.md +376 -0
- parse/dev/2EDqbSCnmF/2EDqbSCnmF_content_list.json +2070 -0
- parse/dev/2EDqbSCnmF/2EDqbSCnmF_middle.json +0 -0
- parse/dev/2EDqbSCnmF/2EDqbSCnmF_model.json +0 -0
- parse/dev/3RBY8fKjHeu/images/159d86a4fe017221206965fa98efc6ce35e16bebec7536f231b04a5fa470830b.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/17f2d11eee9937e70f62a1993623ebccd221887d067e71919c350fa57662f4d3.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/24fe88bf92baa43778d9defa3450750bc0d2c910fa9c12c5902630d7c2316e1e.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/29586778ea39548465206a088f601c3e740302c103692a574d7a70d8188b3983.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/372acb53fc3f9ab9f178baac319f8b0cb0c4ee1ca96f374bada7397c4ec23630.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/476b14497f73d983953dbe7a34cff454d210243e28b1aa97dafb54ae6a92e42f.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/69863294723843746383e47fe99dcd32744e499d1c526f51df4121c52ff99fe8.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/69e7a0dc11e3a7ecd812dec526777f2a39e7aed63605587ef789903e7f57fb8c.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/bbd9aa6b3f541685e1ecf9dd1c4451b92904b361a6547ee2e39414769cb64de4.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/bced559cf73caae8b60315e212417c82fda960dadfbf4c1dc28fc8ea7be19aee.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/bd74dd0968c35c4dc8ffeba6cc51d9f109521ca66008a561e30a7fcccfe148a6.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/d3f8da6b396f76d54da62528b9e91defa6543c22e48371ec0f3e6d668424058f.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/d59a106b3ad7e1f46762af4cb6c79589191a386db6dd18da2381b810f0e72627.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/db6cf431ae9355646aa06c810c30e311d8db38009707b4dea4bad788085ac2bb.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/dfce202941b6d7b7a3b4e91b152da625264b3b1c43837193ab53e137e11b01f3.jpg +3 -0
- parse/dev/3RBY8fKjHeu/images/e30f877426a1aa2686b70c08629a56889403c96105b017890f8ea57b7982c4a2.jpg +3 -0
- parse/dev/3jooF27-0Wy/3jooF27-0Wy_layout.pdf +3 -0
- parse/dev/3jooF27-0Wy/3jooF27-0Wy_origin.pdf +3 -0
- parse/dev/3jooF27-0Wy/3jooF27-0Wy_span.pdf +3 -0
- parse/dev/3mRwyG5one/3mRwyG5one_layout.pdf +3 -0
- parse/dev/3mRwyG5one/3mRwyG5one_origin.pdf +3 -0
- parse/dev/3mRwyG5one/3mRwyG5one_span.pdf +3 -0
- parse/dev/4NpoSrT8uU-/4NpoSrT8uU-_content_list.json +1264 -0
- parse/dev/4NpoSrT8uU-/4NpoSrT8uU-_middle.json +0 -0
- parse/dev/4NpoSrT8uU-/4NpoSrT8uU-_model.json +0 -0
- parse/dev/4WgqjmYacAf/4WgqjmYacAf_layout.pdf +3 -0
- parse/dev/4WgqjmYacAf/4WgqjmYacAf_origin.pdf +3 -0
- parse/dev/4WgqjmYacAf/4WgqjmYacAf_span.pdf +3 -0
.gitattributes
CHANGED
|
@@ -2119,3 +2119,143 @@ parse/dev/I1hQbx10Kxn/I1hQbx10Kxn_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
|
| 2119 |
parse/dev/I1hQbx10Kxn/I1hQbx10Kxn_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2120 |
parse/dev/puMfaHb1hY/puMfaHb1hY_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2121 |
parse/dev/yJE7iQSAep/yJE7iQSAep_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2119 |
parse/dev/I1hQbx10Kxn/I1hQbx10Kxn_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2120 |
parse/dev/puMfaHb1hY/puMfaHb1hY_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2121 |
parse/dev/yJE7iQSAep/yJE7iQSAep_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2122 |
+
parse/dev/Uczck6TlSZ/Uczck6TlSZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2123 |
+
parse/dev/ne6zeqLFCZ/ne6zeqLFCZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2124 |
+
parse/dev/gfwON7rAm4/gfwON7rAm4_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2125 |
+
parse/dev/luGXvawYWJ/luGXvawYWJ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2126 |
+
parse/dev/VD-AYtP0dve/VD-AYtP0dve_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2127 |
+
parse/dev/4WgqjmYacAf/4WgqjmYacAf_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2128 |
+
parse/dev/luGXvawYWJ/luGXvawYWJ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2129 |
+
parse/dev/VD-AYtP0dve/VD-AYtP0dve_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2130 |
+
parse/dev/luGXvawYWJ/luGXvawYWJ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2131 |
+
parse/dev/4WgqjmYacAf/4WgqjmYacAf_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2132 |
+
parse/dev/rBCvMG-JsPd/rBCvMG-JsPd_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2133 |
+
parse/dev/rBCvMG-JsPd/rBCvMG-JsPd_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2134 |
+
parse/dev/4WgqjmYacAf/4WgqjmYacAf_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2135 |
+
parse/dev/lXuByUeHhd/lXuByUeHhd_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2136 |
+
parse/dev/rBCvMG-JsPd/rBCvMG-JsPd_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2137 |
+
parse/dev/fAdMly4ki5/fAdMly4ki5_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2138 |
+
parse/dev/jtiQ26sCJi/jtiQ26sCJi_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2139 |
+
parse/dev/fn0FXlXkzL/fn0FXlXkzL_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2140 |
+
parse/dev/lXuByUeHhd/lXuByUeHhd_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2141 |
+
parse/dev/fAdMly4ki5/fAdMly4ki5_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2142 |
+
parse/dev/jtiQ26sCJi/jtiQ26sCJi_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2143 |
+
parse/dev/jtiQ26sCJi/jtiQ26sCJi_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2144 |
+
parse/dev/lXuByUeHhd/lXuByUeHhd_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2145 |
+
parse/dev/fn0FXlXkzL/fn0FXlXkzL_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2146 |
+
parse/dev/Mng8CQ9eBW/Mng8CQ9eBW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2147 |
+
parse/dev/jSorGn2Tjg/jSorGn2Tjg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2148 |
+
parse/dev/Mng8CQ9eBW/Mng8CQ9eBW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2149 |
+
parse/dev/fAdMly4ki5/fAdMly4ki5_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2150 |
+
parse/dev/jSorGn2Tjg/jSorGn2Tjg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2151 |
+
parse/dev/I1hQbx10Kxn/I1hQbx10Kxn_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2152 |
+
parse/dev/Mng8CQ9eBW/Mng8CQ9eBW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2153 |
+
parse/dev/fn0FXlXkzL/fn0FXlXkzL_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2154 |
+
parse/dev/puMfaHb1hY/puMfaHb1hY_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2155 |
+
parse/dev/0xiJLKH-ufZ/0xiJLKH-ufZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2156 |
+
parse/dev/Jbdc0vTOcol/Jbdc0vTOcol_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2157 |
+
parse/dev/ePgJfxYxl7m/ePgJfxYxl7m_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2158 |
+
parse/dev/ePgJfxYxl7m/ePgJfxYxl7m_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2159 |
+
parse/dev/Pkb5FA5AjqP/Pkb5FA5AjqP_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2160 |
+
parse/dev/0xiJLKH-ufZ/0xiJLKH-ufZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2161 |
+
parse/dev/Pkb5FA5AjqP/Pkb5FA5AjqP_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2162 |
+
parse/dev/HdYxZ_OVZG/HdYxZ_OVZG_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2163 |
+
parse/dev/g7U9jD_2CUr/g7U9jD_2CUr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2164 |
+
parse/dev/Jbdc0vTOcol/Jbdc0vTOcol_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2165 |
+
parse/dev/g7U9jD_2CUr/g7U9jD_2CUr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2166 |
+
parse/dev/ePgJfxYxl7m/ePgJfxYxl7m_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2167 |
+
parse/dev/HdYxZ_OVZG/HdYxZ_OVZG_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2168 |
+
parse/dev/g7U9jD_2CUr/g7U9jD_2CUr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2169 |
+
parse/dev/isPnnaTZaP5/isPnnaTZaP5_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2170 |
+
parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2171 |
+
parse/dev/HdYxZ_OVZG/HdYxZ_OVZG_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2172 |
+
parse/dev/isPnnaTZaP5/isPnnaTZaP5_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2173 |
+
parse/dev/isPnnaTZaP5/isPnnaTZaP5_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2174 |
+
parse/dev/qw674L9PfQE/qw674L9PfQE_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2175 |
+
parse/dev/Pkb5FA5AjqP/Pkb5FA5AjqP_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2176 |
+
parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2177 |
+
parse/dev/qw674L9PfQE/qw674L9PfQE_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2178 |
+
parse/dev/K0E_F0gFDgA/K0E_F0gFDgA_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2179 |
+
parse/dev/qw674L9PfQE/qw674L9PfQE_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2180 |
+
parse/dev/K0E_F0gFDgA/K0E_F0gFDgA_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2181 |
+
parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2182 |
+
parse/dev/K0E_F0gFDgA/K0E_F0gFDgA_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2183 |
+
parse/dev/VFBjuF8HEp/VFBjuF8HEp_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2184 |
+
parse/dev/De4FYqjFueZ/De4FYqjFueZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2185 |
+
parse/dev/IL5zJqfxAa/IL5zJqfxAa_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2186 |
+
parse/dev/IL5zJqfxAa/IL5zJqfxAa_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2187 |
+
parse/dev/De4FYqjFueZ/De4FYqjFueZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2188 |
+
parse/dev/-P7G-8dmSh4/-P7G-8dmSh4_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2189 |
+
parse/dev/De4FYqjFueZ/De4FYqjFueZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2190 |
+
parse/dev/VFBjuF8HEp/VFBjuF8HEp_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2191 |
+
parse/dev/IL5zJqfxAa/IL5zJqfxAa_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2192 |
+
parse/dev/ATiz_CDA66/ATiz_CDA66_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2193 |
+
parse/dev/tZmqS73_07/tZmqS73_07_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2194 |
+
parse/dev/tZmqS73_07/tZmqS73_07_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2195 |
+
parse/dev/URNZQmbxpwh/URNZQmbxpwh_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2196 |
+
parse/dev/tZmqS73_07/tZmqS73_07_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2197 |
+
parse/dev/ATiz_CDA66/ATiz_CDA66_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2198 |
+
parse/dev/ATiz_CDA66/ATiz_CDA66_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2199 |
+
parse/dev/VFBjuF8HEp/VFBjuF8HEp_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2200 |
+
parse/dev/oMI9PjOb9Jl/oMI9PjOb9Jl_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2201 |
+
parse/dev/URNZQmbxpwh/URNZQmbxpwh_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2202 |
+
parse/dev/URNZQmbxpwh/URNZQmbxpwh_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2203 |
+
parse/dev/B72HXs80q4/B72HXs80q4_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2204 |
+
parse/dev/JTmO2V9Xpz/JTmO2V9Xpz_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2205 |
+
parse/dev/JTmO2V9Xpz/JTmO2V9Xpz_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2206 |
+
parse/dev/oMI9PjOb9Jl/oMI9PjOb9Jl_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2207 |
+
parse/dev/wiBEFdAvl8L/wiBEFdAvl8L_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2208 |
+
parse/dev/wiBEFdAvl8L/wiBEFdAvl8L_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2209 |
+
parse/dev/oMI9PjOb9Jl/oMI9PjOb9Jl_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2210 |
+
parse/dev/B72HXs80q4/B72HXs80q4_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2211 |
+
parse/dev/PvOo1sHKzf/PvOo1sHKzf_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2212 |
+
parse/dev/B72HXs80q4/B72HXs80q4_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2213 |
+
parse/dev/PvOo1sHKzf/PvOo1sHKzf_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2214 |
+
parse/dev/akddwRG6EGi/akddwRG6EGi_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2215 |
+
parse/dev/PvOo1sHKzf/PvOo1sHKzf_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2216 |
+
parse/dev/akddwRG6EGi/akddwRG6EGi_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2217 |
+
parse/dev/JTmO2V9Xpz/JTmO2V9Xpz_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2218 |
+
parse/dev/3jooF27-0Wy/3jooF27-0Wy_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2219 |
+
parse/dev/akddwRG6EGi/akddwRG6EGi_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2220 |
+
parse/dev/wiBEFdAvl8L/wiBEFdAvl8L_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2221 |
+
parse/dev/3jooF27-0Wy/3jooF27-0Wy_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2222 |
+
parse/dev/VpYBxaPLaj-/VpYBxaPLaj-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2223 |
+
parse/dev/xp5VOBxTxZ/xp5VOBxTxZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2224 |
+
parse/dev/1QQnYd02etI/1QQnYd02etI_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2225 |
+
parse/dev/1QQnYd02etI/1QQnYd02etI_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2226 |
+
parse/dev/Qg2vi4ZbHM9/Qg2vi4ZbHM9_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2227 |
+
parse/dev/xp5VOBxTxZ/xp5VOBxTxZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2228 |
+
parse/dev/1QQnYd02etI/1QQnYd02etI_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2229 |
+
parse/dev/b9APFSTylGT/b9APFSTylGT_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2230 |
+
parse/dev/b9APFSTylGT/b9APFSTylGT_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2231 |
+
parse/dev/Qg2vi4ZbHM9/Qg2vi4ZbHM9_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2232 |
+
parse/dev/Qg2vi4ZbHM9/Qg2vi4ZbHM9_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2233 |
+
parse/dev/eLxADkHrBcR/eLxADkHrBcR_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2234 |
+
parse/dev/Bl8CQrx2Up4/Bl8CQrx2Up4_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2235 |
+
parse/dev/eLxADkHrBcR/eLxADkHrBcR_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2236 |
+
parse/dev/xp5VOBxTxZ/xp5VOBxTxZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2237 |
+
parse/dev/RRGVCN8kjim/RRGVCN8kjim_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2238 |
+
parse/dev/eLxADkHrBcR/eLxADkHrBcR_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2239 |
+
parse/dev/Bl8CQrx2Up4/Bl8CQrx2Up4_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2240 |
+
parse/dev/XVjTT1nw5z/XVjTT1nw5z_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2241 |
+
parse/dev/RRGVCN8kjim/RRGVCN8kjim_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2242 |
+
parse/dev/b9APFSTylGT/b9APFSTylGT_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2243 |
+
parse/dev/Bl8CQrx2Up4/Bl8CQrx2Up4_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2244 |
+
parse/dev/RRGVCN8kjim/RRGVCN8kjim_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2245 |
+
parse/dev/x6INXlnUGro/x6INXlnUGro_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2246 |
+
parse/dev/XVjTT1nw5z/XVjTT1nw5z_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2247 |
+
parse/dev/XVjTT1nw5z/XVjTT1nw5z_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2248 |
+
parse/dev/8qWazUd8Jm/8qWazUd8Jm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2249 |
+
parse/dev/NXHXoYMLIG/NXHXoYMLIG_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2250 |
+
parse/dev/NXHXoYMLIG/NXHXoYMLIG_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2251 |
+
parse/dev/8qWazUd8Jm/8qWazUd8Jm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2252 |
+
parse/dev/NXHXoYMLIG/NXHXoYMLIG_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2253 |
+
parse/dev/KWCZfuqshd/KWCZfuqshd_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2254 |
+
parse/dev/3mRwyG5one/3mRwyG5one_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2255 |
+
parse/dev/8qWazUd8Jm/8qWazUd8Jm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2256 |
+
parse/dev/3mRwyG5one/3mRwyG5one_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2257 |
+
parse/dev/3mRwyG5one/3mRwyG5one_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2258 |
+
parse/dev/KWCZfuqshd/KWCZfuqshd_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2259 |
+
parse/dev/KWCZfuqshd/KWCZfuqshd_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2260 |
+
parse/dev/VpYBxaPLaj-/VpYBxaPLaj-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 2261 |
+
parse/dev/3jooF27-0Wy/3jooF27-0Wy_span.pdf filter=lfs diff=lfs merge=lfs -text
|
parse/dev/-P7G-8dmSh4/-P7G-8dmSh4_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b4e5b19c049eef099dfffb145a018253ba2458f4d1d2c5889d0a519464161f59
|
| 3 |
+
size 1228476
|
parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6fad4b6901aaf3c7a784be562b7e0770f0ff9f73274505fbda51c593567f62f5
|
| 3 |
+
size 708117
|
parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1fe6bd21076d5008f8afff0f893c4ac36e45e12a807979931207a27f56bb6127
|
| 3 |
+
size 510872
|
parse/dev/-QHUWgkh1OY/-QHUWgkh1OY_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:051068aa56656a6fed5ef26d5473afc9b060a1519a6586400581094e19db2d02
|
| 3 |
+
size 732492
|
parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt.md
ADDED
|
@@ -0,0 +1,432 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# TEXTLESS PHRASE STRUCTURE INDUCTION FROM VISUALLY-GROUNDED SPEECH
|
| 2 |
+
|
| 3 |
+
Anonymous authors Paper under double-blind review
|
| 4 |
+
|
| 5 |
+
# ABSTRACT
|
| 6 |
+
|
| 7 |
+
We study phrase structure induction from visually-grounded speech without intermediate text or text pre-trained models. The core idea is to first segment the speech waveform into sequences of word segments, then induce phrase structure based on the inferred segment-level continuous representations. To this end, we present the Audio-Visual Neural Syntax Learner (AV-NSL) that learns non-trivial phrase structure by listening to audio and looking at images, without ever reading text. Experiments on SpokenCOCO, the spoken version of MSCOCO with paired images and spoken captions, show that AV-NSL infers meaningful phrase structures similar to those learned from naturally-supervised text parsing, quantitatively and qualitatively. The findings in this paper extend prior work in unsupervised language acquisition from speech and grounded grammar induction, and manifest one possibility of bridging the gap between the two fields.
|
| 8 |
+
|
| 9 |
+
# 1 INTRODUCTION
|
| 10 |
+
|
| 11 |
+
Toddlers learn their first language through listening, talking, and interacting with the world through multi-sensory inputs. Different levels of early language acquisition happen without supervisory feedback (Dupoux, 2018): phonetics, phonology, morphology, syntax, semantics, pragmatics. It is therefore crucial to think about learning language, from identifying lower-level phones or words to inducing high-level linguistic structure like grammar, in natural settings.1 To this end, there have been two ongoing efforts in parallel:
|
| 12 |
+
|
| 13 |
+
• Zero-resource speech processing, where speech models are constructed without any textual intermediates, with the goal of mimicking how children learn to speak before learning to read or write. The modeling tasks are constrained to unsupervised learning of subphones, phones, and words (Jansen et al., 2013). • Grammar induction, which aims to learn latent syntactic structures, including constituency trees and dependency trees, with no annotation of syntactic structures as supervision.
|
| 14 |
+
|
| 15 |
+
Notably in recent years, multi-modal induction has emerged as a promising and effective objective for both efforts. In speech, Harwath (2018) proposed to leverage parallel image-speech data to acquire associated words (Harwath & Glass, 2017) and phones (Harwath et al., 2020) from raw waveforms. In syntax induction, Shi et al. (2019) proposed to induce phrase-structure grammar from parallel image-text data. The above observations motivated us to build a computational model that leverages the visual modality to acquire low-level words up to high-level phrase-structure from raw speech waveforms, without any intermediate textual forms or any direct supervision.2
|
| 16 |
+
|
| 17 |
+
In this paper, we present the Audio-Visual Neural Syntax Learner (AV-NSL), an approach toward learning phrase structure from raw speech waveforms without relying on any kind of intermediate textual form or text pre-trained models (Figure 1). In a nutshell, AV-NSL trains a visually-grounded syntax learner directly on a sequence of continuous speech representations given by an audio-visual word segmentation model. We also introduce a self-training process and an unsupervised decoding method to improve the final output of in AV-NSL. To measure the effectiveness of AV-NSL, we compare it to text-based syntax learner VG-NSL (Shi et al., 2019) and further introduce a novel evaluation metric, SAIOU, that accounts for structure differences when the number of tree nodes are mismatched. To validate our design choice of AV-NSL, we construct several baselines and introduce alternative modeling choices, including acoustic compound-PCFG (Kim et al., 2019a). Qualitatively, we provide constituency recall analyses and the visualizations of the inferred word segmentation and tree structures.
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
Figure 1: We study the process of inducing phrase structure, in the form of constituency parse tree, on unsupervised inferred word segments from raw speech waveform. No intermediate text tokens or ASR is needed. For illustration purpose, here we show the gold parse tree from the given text caption.
|
| 21 |
+
|
| 22 |
+
In summary, we present the first study on inducing phrase structure from visually-grounded speech without relying on text, introducing the AV-NSL model (§3) with comprehensive experiments (§4) and analysis (§5). As a by product, we improve over the previous state of the art in unsupervised word segmentation (§4.4).
|
| 23 |
+
|
| 24 |
+
# 2 RELATED WORK
|
| 25 |
+
|
| 26 |
+
# 2.1 UNSUPERVISED AND DISTANTLY SUPERVISED GRAMMAR INDUCTION
|
| 27 |
+
|
| 28 |
+
Much work has been proposed to induce grammar from different sources of distant supervision, including language modeling (Shen et al., 2018; 2019; Kim et al., 2019a;b), masked language modeling (Drozdov et al., 2019), natural language inference (Li et al., 2019), and, more recently, visual grounding via image-caption matching (Shi et al., 2019; Zhao & Titov, 2020; Hong et al., 2021; Wan et al., 2022, inter alia). There has also been extensive study directly targeting unsupervised constituency parsing (Klein & Manning, 2002; 2004; Bod, 2006; Spitkovsky et al., 2013, inter alia). To the best of our knowledge, existing work on grammar induction from distant supervision has been based almost exclusively on text input. The most relevant work to ours is MMC-PCFG (Zhang et al., 2021), where speech features are treated as an auxiliary input for video-text grammar induction. However, text data and an off-the-shelf automatic speech recognition (ASR) model are required. In contrast to them, AV-NSL induces constituency parse trees from raw speech bypassing text, with distant supervision from parallel audio-visual data.
|
| 29 |
+
|
| 30 |
+
# 2.2 UNSUPERVISED LANGUAGE ACQUISITION FROM SPEECH
|
| 31 |
+
|
| 32 |
+
The earliest work (de Sa, 1994; De Marcken, 1996; Roy & Pentland, 2002) on language acquisition from speech required phonetic lexicon/labels in the process. The idea of spoken term discovery, i.e., discovering repetitive patterns or keywords from unannotated speech, was first addressed by Park & Glass (2007). Thereafter, subsequent work improved upon the original (Zhang & Glass, 2009; Jansen & Van Durme, 2011; McInnes & Goldwater, 2011; Zhang, 2013, inter alia). Other related work has considered tasks like unsupervised word segmentation and unsupervised ASR, sometimes jointly with spoken term discovery (Lee & Glass, 2012; Lee et al., 2015; Kamper et al., 2015; 2017; Kamper & van Niekerk, 2021; Chorowski et al., 2021; Bhati et al., 2021; Kamper, 2022; Algayres et al., 2022) The discovery of lexical units was applied to text-free language modeling (Nguyen et al., 2020; Peng & Harwath, 2022a) and speech generation (Lakhotia et al., 2021; Polyak et al., 2021; Kharitonov et al., 2022). The ZeroSpeech challenges (Versteegh et al., 2015; Dunbar et al., 2017; 2019; 2020; Nguyen et al., 2020) have been a major driving force in the field.
|
| 33 |
+
|
| 34 |
+
Harwath (2018) opened up a new direction in visually grounded language acquisition, showing word-like (Harwath & Glass, 2017) and phone-like (Harwath et al., 2020) units are acquired from speech by analyzing audio-visual retrieval models. Numerous works have studied the characteristics of the linguistic information acquired in visually grounded speech models (Havard et al., 2019; Khorrami & Ras¨ anen, 2021; Olaleye & Kamper, 2021; Wang & Hasegawa-Johnson, 2021; ¨ Mitja Nikolaus, 2022). Peng & Harwath (2022b) shows that clear word segmentation and identification naturally emerge from a visually grounded, self-supervised speech model named VG-HuBERT, by analyzing the model’s self-attention heads. Unlike the above, AV-NSL acquires phrase structure, in the form of constituency parsing on top of unsupervised word segments.
|
| 35 |
+
|
| 36 |
+
# 2.3 SPEECH PARSING AND ITS APPLICATIONS
|
| 37 |
+
|
| 38 |
+
Early work on speech parsing can be traced back to the SParseval toolkit (Roark et al., 2006), for evaluating text parsers given (errorful) ASR output. Tran et al. (2018; 2019); Tran & Ostendorf (2021) explored the use of acoustic-prosodic features for text parsing with auxiliary speech input. Lou et al. (2019) trained a text parser (Kitaev & Klein, 2018) to detect speech disfluencies. In the past, syntax has also been studied in the context of speech prosody (Wagner & Watson, 2010; Kohn ¨ et al., 2018). The most relevant work to ours is Pupier et al. (2022), where a text dependency parser is trained from speech jointly with an ASR model. Moreover, text syntax parsing has been applied to prosody modeling in end-to-end text-to-speech (TTS; Guo et al., 2019; Tyagi et al., 2020; Kaiki et al., 2021). This work builds on top of pre-existing text parsing algorithms or pre-existing phrase structures from text, whereas we study phrase structure acquisition in the absence of text.
|
| 39 |
+
|
| 40 |
+
# 3 METHOD
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
Figure 2: Illustration of AV-NSL, which extends VG-NSL (Shi et al., 2019) to audio-visual inputs.
|
| 44 |
+
|
| 45 |
+
Given a set of paired spoken captions and images, the Audio-Visual Neural Syntax Learner (AVNSL) infers phrase structures from subsequences of raw speech segments without relying on text. The basis of AV-NSL is the Visually-Grounded Neural Syntax Learner (VG-NSL) (Shi et al., 2019). VG-NSL learns constituency parse trees by guiding a sequential tree sampling process with textimage matching. To extend VG-NSL to audio-visual inputs, the central challenge is extracting semantically-meaningful word segments from unannotated speech. We break down the problem into a two-step process: (1) obtaining sequences of word segments, and (2) extracting segment-level self-supervised representations. With these simple modifications, AV-NSL learns non-trivial phrase structure without ever reading text, instead by listening to speech and looking at images.
|
| 46 |
+
|
| 47 |
+
# 3.1 BACKGROUND: VISUALLY-GROUNDED NEURAL SYNTAX LEARNER
|
| 48 |
+
|
| 49 |
+
VG-NSL (Shi et al., 2019) is composed of a bottom-up text parser and a text-image embedding matching module. The parser consists of an embedding similarity scoring function score and an embedding cembeddings sively scorin ${ \cal { W } } = \{ w _ { i } ^ { 0 } \} _ { i = 1 } ^ { N }$ nction comof length g adjacent $N$ ne. Given a text caption, den, the parser synthesizes a consmbeddings at each step. At step ed by a sequence of wordtuency parse tree by recur-, VG-NSL (1) evaluates all $t$
|
| 50 |
+
|
| 51 |
+
consecutive pairs of embeddings $\langle w _ { i } ^ { t } , w _ { i + 1 } ^ { t } \rangle$ and assigns a scalar score to each with score, (2) selects a pair $\langle w _ { i ^ { \prime } } ^ { t } , w _ { i ^ { \prime } + 1 } ^ { t } \rangle$ based on the corresponding scores,3 and (3) combines the selected pair of embeddings via combine to form a new phrase embedding for the next step, copying the remaining ones to the next step. In VG-NSL, score is parameterized by a 2-layer ReLU-activated MLP, and combine is defined by the L2-normalized sum of the input embeddings. The resulting tree is inherently binary and there are $N - 1$ combining steps in total, as the tree parser must combine two nodes in each step.
|
| 52 |
+
|
| 53 |
+
The text-image embedding matching module of VG-NSL is based on the standard hinge-based triplet loss (Kiros et al., 2014), where the sentence-based loss is modified to a phrase-based one. Additionally, the loss function is adapted to estimate the visual concreteness of a text span: intuitively, the smaller the loss related to a candidate constituent $c$ , the larger the concreteness of $c$ , and vice versa. The concreteness of a constituent $c$ is defined as
|
| 54 |
+
|
| 55 |
+
$$
|
| 56 |
+
\mathbf { \nabla } \cdot e \left( \mathbf { c } ; \mathbf { i } \right) = \sum _ { \mathbf { c } ^ { \prime } } \left[ \cos \left( \mathbf { i } , \mathbf { c } \right) - \cos \left( \mathbf { i } , \mathbf { c } ^ { \prime } \right) - \delta \right] _ { + } + \sum _ { \mathbf { i } ^ { \prime } } \left[ \cos \left( \mathbf { i } ^ { \prime } , \mathbf { c } \right) - \cos \left( \mathbf { i } ^ { \prime } , \mathbf { c } \right) - \delta \right] _ { + } ,
|
| 57 |
+
$$
|
| 58 |
+
|
| 59 |
+
where c is the vector representation of $c$ ; i is the corresponding vector of the parallel image of $c ; \mathbf { c } ^ { \prime }$ is a candidate constituent from a sentence that is not in parallel with i; $\mathbf { i } ^ { \prime }$ is an image that is not in parallel with $c ; \delta$ is a constant margin. Here, $[ \cdot ] _ { + } : = \operatorname* { m a x } ( \cdot , 0 )$ . Finally, the estimated concreteness scores are passed back to the parser as rewards to the constituents. VG-NSL jointly optimizes the visual-semantic embedding loss, and trains the parser with REINFORCE (Williams, 1992).
|
| 60 |
+
|
| 61 |
+
# 3.2 AUDIO-VISUAL NEURAL SYNTAX LEARNER
|
| 62 |
+
|
| 63 |
+
AV-NSL extends VG-NSL by: (1) incorporating an audio-visual word segmentation model for obtaining sequences of word segments from unannotated speech, (2) jointly optimizing segment-level embeddings along with phrase structure induction, and (3) employing deeper score and combine function parameterization in the parsing module. We empirically found (3) necessary, mainly because speech embeddings are inherently richer, less clean, and semantically more ambiguous than word embeddings. In AV-NSL, score is parameterized by a 4-layer MLP with GELU nonlinearities (Hendrycks & Gimpel, 2016), and combine is a 5-layer MLP with GELUs. On the other hand, such parameterization may cause the text-based sampling procedure to favor sampling the visually-salient words (Shi et al., 2019; Kojima et al., 2020). We describe (1) and (2) in detail as follows.
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
Figure 3: Example of word segmentation from VG-HuBERT (top). We use the midpoints of adjacent attention boundaries (vertical blue dashed lines) as the word boundaries. We observe that function words are ignored by VG-HuBERT; to account for this, we introduce segment insertion (bottom): short segments are placed in long enough gaps between existing segments, such that function words are recovered. Inserted segments are marked with $\cdot _ { + } \cdot$ . Best viewed in color.
|
| 67 |
+
|
| 68 |
+
Audio-visual word segmentation: AV-NSL leverages VG-HuBERT Peng & Harwath (2022b) for word segmentation (Figure 2; bottom). VG-HuBERT is trained to associate spoken captions with natural images via retrieval training, without any textual supervision. After training, spoken word segmentation emerges via magnitude thresholding the self-attention heads of the model’s audio encoder: at layer $l$ , we threshold each CLS token attention weights over each temporal speech frame token to only show top $p \%$ of the magnitude. In Figure 3, we visualize the attention weights that each speech frame receives from the CLS token. Weights from different attention heads are plotted in different colors, and color transparency represents the magnitude of the attention weights.
|
| 69 |
+
|
| 70 |
+
However, an issue we observed with VG-HuBERT is that they tend to ignore function words such as $\mathbf { \ddot { a } } ^ { , , }$ , “the”, and “of”. While this is less of an issue for word segmentation and identification, it is problematic for our purpose, as the function words are critical for phrase induction. Therefore, we devise a simple heuristic to pick up function words’ segments – segment insertion. We insert a short word segment whenever there is a sufficiently long enough gap of $s$ seconds, and VGHuBERT fails to place an attention segment. See bottom of Figure 3. Since this could introduce false positives (inserting segments where there is no word spoken), we apply unsupervised voice activity detection (Tan et al., 2020) to further restrict segment insertion only in voiced regions. The length of the insertion gap $s$ , the VG-HuBERT segmentation layer $l$ , attention magnitude threshold $p \%$ , and model training snapshots over different random seeds and training steps, are all determined in an unsupervised fashion with minimal Bayes’ risk decoding, introduced in Section 3.4.
|
| 71 |
+
|
| 72 |
+
Speech segment representations: Given the word segments from the audio-visual segmentation model, segment representations are extracted as inputs for the tree sampling module. Ideally, these segments should be semantically-meaningful and mimic word embeddings method is speech discretization that converts the inputs into sequences of d $\mathbf { \bar { \mathit { W } } } = \{ w _ { i } ^ { 0 } \} _ { i = 1 } ^ { N }$ . A naive(Lakhotia et al., 2021). Yet, we are targeting word-level phrase structures, while speech discretization, namely acoustic unit discovery, are sub-phone level, which does not fit into our setup. Different from it, AVNSL is based on continuous segment-level self-supervised representations. Let’s denote the framelevel representation sequence as $R = \{ r _ { j } \} _ { j = 1 } ^ { T }$ , where $T$ is the speech sequence length. Audio-visual word segmentation returns an alignment $\bar { \boldsymbol { A } } ( i ) = \boldsymbol { r } _ { p : q }$ that maps the ith word segment to the $p$ th to $q$ th acoustic frames. The segment-level continuous representation for the ith word is simply,
|
| 73 |
+
|
| 74 |
+
$$
|
| 75 |
+
w _ { i } ^ { 0 } = \sum _ { t \in A ( i ) } \stackrel { } { a _ { i t } } r _ { i t }
|
| 76 |
+
$$
|
| 77 |
+
|
| 78 |
+
where $a _ { i t }$ is the attention weights over the segments specified by $A ( i )$ . By default in AV-NSL, $R$ is the layer representation from VG-HuBERT, and $a _ { i t }$ is the CLS token attention weights over frames within each segment. In some cases, visual grounding is not available in AV-NSL’s word segmentation, e.g. VG-HuBERT is not available. We instead take $R$ as the layer representation from a vanilla HuBERT (Hsu et al., 2021a), and $a _ { i t }$ is parameterized by a hidden layer that is jointly optimized with the tree sampling module. Despite its simplicity, AV-NSL learns meaningful phrase structures on these segment representation sequences.
|
| 79 |
+
|
| 80 |
+
# 3.3 SELF-TRAINING
|
| 81 |
+
|
| 82 |
+
A self-training procedure is introduced for AV-NSL to further improve its parsing capability. Previously, it has been shown that self-training consistently improves the performance of text-based unsupervised constituency parsing. In Shi et al. (2020), the self-training model was based on Benepar (Kitaev & Klein, 2018), a supervised neural constituency parser, which (1) takes a sentence as the input, (2) maps it to word representations, and (3) predicts a score for any constituency parse tree. In the inference stage, the model evaluates all possible tree structures and outputs the highest-scoring one using the CKY algorithm (Kasami, 1966; Younger, 1967; Cocke, 1969).
|
| 83 |
+
|
| 84 |
+
In this work, we introduce s-Benepar, which is based on the original Benepar, except the model input is the segment-level continuous HuBERT representations mean-pooled over unsupervised word segmentation from VG-HuBERT with segment insertion, and model output is AV-NSL’s inferred constituency parse from Section 3.2. We also removed part-of-speech tag prediction as in Benepar, as there is no textual supervision in our setting. To summarize, with paired speech $D _ { A }$ and image $D _ { V }$ data, the training scheme for AV-NSL with self-training is as follows:
|
| 85 |
+
|
| 86 |
+
1. Train an AV-NSL from audio-visual data $( D _ { A } , D _ { V } )$ and obtain the trained model $M _ { a v }$ .
|
| 87 |
+
2. Generate parse tree $T _ { 0 }$ with $M _ { a v }$ for $D _ { A }$ . Obtain audio-tree pairs $( D _ { A } , T _ { 0 } )$ . Set $T = T _ { 0 }$ .
|
| 88 |
+
3. Train an s-Benepar from $( D _ { A } , T )$ and obtain the trained model $M _ { s } ^ { i }$ .
|
| 89 |
+
4. Generate parse tree $T _ { i }$ with $M _ { s } ^ { i }$ for $D _ { A }$ . Obtain audio-tree pairs $( D _ { A } , T _ { i } )$ . Set $T = T _ { i }$ .
|
| 90 |
+
5. Go to Step 3 if we have not reached the desirable number of iterations; return $T$ otherwise.
|
| 91 |
+
|
| 92 |
+
We find it helpful to iterate s-Benepar training twice $( i = 2$ ), but the results plateau afterwards.
|
| 93 |
+
|
| 94 |
+
# 3.4 UNSUPERVISED DECODING
|
| 95 |
+
|
| 96 |
+
One key ingredient of AV-NSL is applying minimum Bayes risk (MBR) decoding (Bickel & Li, 1977) as the selection criterion for fully-unsupervised spoken word segmentation and phrasestructure induction.4 Specifically, this is in contrast to all prior unsupervised word segmentation work, in which ground truth word segments from a development set are required for decoding.
|
| 97 |
+
|
| 98 |
+
At a high level, given a loss function $\ell _ { M B R } ( O _ { 1 } , O _ { 2 } )$ between two outputs $O _ { 1 }$ and $O _ { 2 }$ , and a set of $k$ outputs $\mathcal { O } = \{ O _ { 1 } , \ldots , O _ { k } \}$ , we select the optimal output
|
| 99 |
+
|
| 100 |
+
$$
|
| 101 |
+
\hat { O } = \arg \operatorname* { m i n } _ { O ^ { \prime } \in { \mathcal O } } \sum _ { O ^ { \prime \prime } \in { \mathcal O } } \ell _ { M B R } ( O ^ { \prime } , O ^ { \prime \prime } ) .
|
| 102 |
+
$$
|
| 103 |
+
|
| 104 |
+
For word segmentation, we define the loss between two segmentation proposals $ { \boldsymbol { S } } _ { 1 }$ and $S _ { 2 }$ by $\ell _ { M B R } ( S _ { 1 } , S _ { 2 } ) ^ { - } = - \mathrm { M I O U } ( S _ { 1 } , S _ { 2 } )$ , where $\mathrm { { M I O U } } ( \cdot , \cdot )$ denotes the mean intersection over union ratio across all matched pairs of predicted word spans from $S _ { 1 }$ and $S _ { 2 }$ . We match the predicted word spans using the maximum weight matching algorithm (Galil, 1986), where word spans correspond to vertices, and we define edge weights by the temporal overlap between the corresponding spans.
|
| 105 |
+
|
| 106 |
+
For phrase structure induction, we define the loss function between two parse trees $\mathcal { T } _ { 1 }$ and $\mathcal { T } _ { 2 }$ by $\ell _ { M B R } ( \mathcal { T } _ { 1 } , \mathcal { T } _ { 2 } ) = 1 - F _ { 1 } ( \mathcal { T } _ { 1 } , \mathcal { T } _ { 2 } )$ , where $F _ { 1 } ( \cdot , \cdot )$ denotes the $F _ { 1 }$ score between two trees.
|
| 107 |
+
|
| 108 |
+
# 4 EXPERIMENTS
|
| 109 |
+
|
| 110 |
+
# 4.1 SETTING
|
| 111 |
+
|
| 112 |
+
Dataset: All models are evaluated on SpokenCOCO, the spoken version of MSCOCO (Lin et al., 2014) where the text captions are read out by MTurk users (Hsu et al., 2021b). It contains $8 3 \mathrm { k } / 5 \mathrm { k } / 5 \mathrm { k }$ images for training, validation, and test: each image has 5 corresponding spoken captions. SpokenCOCO totals 740h of read speech from $2 . 3 \mathrm { k }$ speakers, with an average utterance duration of about 4 seconds, covering 29K different word types.
|
| 113 |
+
|
| 114 |
+
Preprocessing: For oracle word segmentation, we ran an off-the-shelf English ASR from Montreal Force Aligner (McAuliffe et al., 2017) that was pre-trained on Librispeech and adapted to SpokenCOCO. We removed a few utterances that have mismatches in their ASR transcripts and their text captions. Following Shi et al. (2019), we included trivial spans in tree evaluation. Additionally, we ran an off-the-shelf English parser (Kitaev & Klein, 2018) on the ASR transcript (normalized text with punctuation removed) to generate the oracle trees for SpokenCOCO.
|
| 115 |
+
|
| 116 |
+
# 4.2 BASELINES AND TOPLINES
|
| 117 |
+
|
| 118 |
+
AV-NSL segments speech waveforms into word segments, then learns phrase structures on top of the learned segments. Both segmentation and structure induction are fully-unsupervised and visuallygrounded. To help us examine the role of each component in AV-NSL, we therefore further construct the following baselines and toplines. Their full descriptions are in Appendix A.1.
|
| 119 |
+
|
| 120 |
+
Trivial tree structures: Following (Shi et al., 2019), we include baselines without linguistic information: random binary trees, left-branching binary trees, and right-branching binary trees.
|
| 121 |
+
|
| 122 |
+
AV-cPCFG: We train compound probabilistic context free grammar (cPCFG) (Kim et al., 2019a) on word-level discrete speech tokens. Similar to AV-NSL, word segments and segment representations are based on VG-HuBERT. Different from AV-NSL, the segment representations are discretized via kmeans to obtain word-level discrete indices. In short, AV-cPCFG leverages visual cues only for segmentation and segment representations, but not for phrase structure induction.
|
| 123 |
+
|
| 124 |
+
DPDP-cPCFG: Instead of training cPCFG on audio-visual word segments and audio-visual segment representations, DPDP-cPCFG does not rely on any visual grounding throughout. Instead, DPDP (Kamper, 2022) and vanilla HuBERT representations are used. As in AV-cPCFG, kmeans is used for word-level discretization.
|
| 125 |
+
|
| 126 |
+
Oracle AV-NSL: To remove the uncertainty of unsupervised word segmentation, we directly train AV-NSL on top of oracle word segmentation via force alignment.
|
| 127 |
+
|
| 128 |
+
# 4.3 EVALUATION METRIC
|
| 129 |
+
|
| 130 |
+
Word segmentation. We use the standard word boundary prediction metrics (precision, recall and F1), which are calculated by comparing the temporal position between inferred word boundaries and force aligned word boundaries. In particular, following Peng & Harwath (2022b), when an inferred boundary is located within $\pm 2 0 m s$ of a force aligned boundary, we declare a successful prediction.
|
| 131 |
+
|
| 132 |
+
Parsing. For parsing with oracle word segmentation, we use EVALB to calculate the $F _ { 1 }$ score between the predicted and ground-truth parse trees.5 For parsing with inferred word segmentation, due to the mismatch in the number of nodes between the predicted and ground-truth parse trees, we introduce the structured average intersection-over-union ratio (SAIOU) as an additional metric.
|
| 133 |
+
|
| 134 |
+
SAIOU takes both word segmentation quality and temporal overlap between induced constituents into consutterance $\mathcal { T } _ { 1 } = \{ c _ { 1 , i } = ( \ell _ { 1 , i } , \dot { r } _ { 1 , i } ) \} _ { i = 1 } ^ { n _ { 1 } }$ pa t id $\mathcal { T } _ { 2 } = \{ c _ { 2 , j } = ( \ell _ { 2 , j } , r _ { 2 , j } ) \} _ { j = 1 } ^ { n _ { 2 } }$ s over the same speech, represented by a set of constituency tempthe constituents in l boand d, $\ell$ $r$ ignmen, where $\mathcal { T } _ { 1 }$ $\mathcal { T } _ { 2 }$ $\begin{array} { r } { \hat { \mathcal { A } } = \arg \operatorname* { m a x } _ { \nu a l i d \mathcal { A } } \sum _ { i = 1 } ^ { n _ { 1 } } \sum _ { j = 1 } ^ { n _ { 2 } } \mathcal { A } _ { i , j } \mathrm { I o U } ( c _ { 1 , i } , c _ { 2 , j } ) } \end{array}$ $A _ { i , j } = 1$ denotes $c _ { 1 , i }$ aligns with $c _ { 2 , j }$ , and $A _ { i , j } = 0$ otherwise; $\operatorname { I o U } ( \cdot , \cdot )$ denotes the intersection-over-union ratio between two spans. A valid alignment $\mathcal { A }$ is one that satisfies the following conditions:
|
| 135 |
+
|
| 136 |
+
1. Any constituent may be aligned with up to 1 constituent in the other tree;
|
| 137 |
+
|
| 138 |
+
2. For any pair of $i$ and $j$ where $A _ { i , j } = 1$ ,
|
| 139 |
+
|
| 140 |
+
• Any descendant of $c _ { 1 , i } , c _ { 1 , k }$ , may either align to a descendant of $c _ { 2 , j }$ or be left unaligned;
|
| 141 |
+
• Any ancestor of $c _ { 1 , i } , c _ { 1 , k ^ { \prime } }$ , may either align to a ancestor of $c _ { 2 , j }$ or be left unaligned;
|
| 142 |
+
• Any descendant of $c _ { 2 , j } , c _ { 2 , p }$ , may either align to a descendant of $c _ { 1 , i }$ or be left unaligned;
|
| 143 |
+
• Any ancestor of $c _ { 2 , j } , c _ { 2 , p ^ { \prime } }$ , may either align to a ancestor of $c _ { 1 , i }$ or be left unaligned.
|
| 144 |
+
|
| 145 |
+
Given the optimal alignment $\hat { A }$ , we calculate the structured average IOU between $\mathcal { T } _ { 1 }$ and $\mathcal { T } _ { 2 }$ by
|
| 146 |
+
|
| 147 |
+
$$
|
| 148 |
+
\operatorname { S A I o U } ( \mathcal { T } _ { 1 } , \mathcal { T } _ { 2 } ) = \frac { 2 } { n _ { 1 } + n _ { 2 } } \left( \sum _ { i = 1 } ^ { n _ { 1 } } \sum _ { j = 1 } ^ { n _ { 2 } } \hat { A } _ { i , j } \mathrm { I o U } ( c _ { 1 , i } , c _ { 2 , j } ) \right) .
|
| 149 |
+
$$
|
| 150 |
+
|
| 151 |
+
# 4.4 UNSUPERVISED WORD SEGMENTATION
|
| 152 |
+
|
| 153 |
+
We validate our decision of adopting VG-HuBERT to extract word-like units from raw speech waveforms for later phrase structure parsing. In particular, we investigate two questions: (1) How does segment insertion affect word segmentation performance? (2) how does MBR-based VG-HuBERT compare to supervised selected VG-HuBERT?
|
| 154 |
+
|
| 155 |
+
In Table 1, in addition to VG-HuBERT, we also list a speech-only word segmentation algorithm DPDP (Kamper, 2022). Note that audio-visual model VG-HuBERT significantly outperform DPDP. For question (1), by comparing the third row and the fourth row, as expected we see that performing segment insertion improves recall and hurts precision, and slightly improves F1. For question (2), by comparing the fourth row and the fifth row (second to last row), we see that MBR selection actually leads to better performance than supervised selection. The final MBR selection we adopted is based on the last row, where we first performed MBR selection on SpokenCOCO val set on all 405 candidates, and subsequently chose the 10 most selected combinations to perform another round of MBR decoding. Getting the top 10 most selected combinations does not require knowing the performance on segmentation, and therefore this process is still completely unsupervised. The reason for doing 2 iterations of MBR is because performing MBR on 405 candidates on SpokenCOCO training set is estimated to take 2 months, and MBR on 10 candidates can be done in 5 days. Comparing the last two rows, we observe that two iterations of MBR does not lead to worse results.
|
| 156 |
+
|
| 157 |
+
# 4.5 UNSUPERVISED PHRASE STRUCTURE INDUCTION
|
| 158 |
+
|
| 159 |
+
We quantitatively show that AV-NSL learns meaningful phrase structure given word segments. First, Table 2 is the main result of the fully-unsupervised AV-NSL on SpokenCOCO, evaluated with SAIOU. The best performing AV-NSL is based on our improved VG-HuBERT with MBR top 10 selection for word segmentation, attention-weighted mean-pool over VG-HuBERT layers as the segment representations, and another MBR decoding over all phrase structure induction hyperparameters. Comparing AV-NSL against AV-cPCFG and AV-cPCFG against DPDP-cPCFG, we empirically show the necessity of training AV-NSL on continuous segment representation instead of discretized speech tokens, and the effectiveness of visual-grounding in our overall model design.
|
| 160 |
+
|
| 161 |
+
Table 1: Word Segmentation Performance on SpokenCOCO validation set. Out. Sel. denotes output selection methods, and #Sel. Cand. denotes the number of candidate models to be selected. MBR (2iter) means we first run MBR on all 405 candidates, and then run MBR again on the $1 0 \ \mathrm { m o s t }$ selected candidates. Our improved VG-HuBERT with MBR achieves the best boundary $F _ { 1 }$ .
|
| 162 |
+
|
| 163 |
+
<table><tr><td>Method</td><td>Insertion</td><td>Out. Sel.</td><td>#Sel. Cand.</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>DPDP (Kamper,2022)</td><td></td><td>supervised</td><td></td><td>17.37</td><td>9.00</td><td>11.85</td></tr><tr><td>VG-HuBERT (Peng & Harwath,2022b)</td><td></td><td>supervised</td><td></td><td>36.19</td><td>27.22</td><td>31.07</td></tr><tr><td rowspan="3">Improved VG-HuBERT (Ours)</td><td></td><td>supervised</td><td></td><td>34.34</td><td>29.85</td><td>31.94</td></tr><tr><td></td><td>MBR</td><td>405</td><td>33.83</td><td>34.37</td><td>34.10</td></tr><tr><td>√</td><td>MBR (2iter)</td><td>405→10</td><td>33.31</td><td>34.90</td><td>34.09</td></tr></table>
|
| 164 |
+
|
| 165 |
+
Table 2: Fully-unsupervised phrase structure induction results on SpokenCOCO. The best overall number and the best number produced by neural models are in boldface. Full table in Appendix 7.
|
| 166 |
+
|
| 167 |
+
<table><tr><td colspan="3">Model</td><td rowspan="2">Output Selection</td><td rowspan="2">SAIoU</td></tr><tr><td>Syntax Induction</td><td>Segmentation</td><td>Seg.Representation (continuous/discrete)</td></tr><tr><td>Right-Branching</td><td>VG-HuBERT+MBR10</td><td></td><td></td><td>0.546</td></tr><tr><td>Right-Branching</td><td>DPDP</td><td></td><td></td><td>0.478</td></tr><tr><td>AV-NSL</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o (continuous)</td><td>MBR</td><td>0.516</td></tr><tr><td>AV-NSL</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT10,11,12 (continuous)</td><td>MBR</td><td>0.521</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+4k km (discrete)</td><td>last ckpt.</td><td>0.499</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+8k km (discrete)</td><td>last ckpt.</td><td>0.481</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT2+2k km (discrete)</td><td>last ckpt.</td><td>0.465</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT1o+2k km (discrete)</td><td>last ckpt.</td><td>0.426</td></tr></table>
|
| 168 |
+
|
| 169 |
+
<table><tr><td rowspan="2">Model</td><td rowspan="2">Segmentation</td><td rowspan="2">Seg. Representation</td><td colspan="3">tree target</td><td rowspan="2">Output Selection</td><td rowspan="2">SAIoU</td></tr><tr><td>train</td><td>val</td><td>test</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT2</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.538</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT6</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.538</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT2,4,6.8,10,12</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>MBR</td><td>0.536</td></tr></table>
|
| 170 |
+
|
| 171 |
+
Table 3: Single round self-training in Section 3.3 improves the best AV-NSL from Table 2. We train s-Benepar on the trees from fully-unsupervised AV-NSL. Full table in Appendix 8.
|
| 172 |
+
|
| 173 |
+
Secondly, Table 3 shows that our proposed self-training with s-Benepar complements AV-NSL. Generally, a single round of self-training improves the SAIOU, and our best s-Benepar improves the best AV-NSL from 0.521 to 0.538. Thirdly, Table 4 isolates phrase structure induction from word segmentation quality with oracle AV-NSL. Different from Table 2, since there is no mismatch in the number of tree nodes, we can adopt $F _ { 1 }$ evaluation. With proper segment-level representations, unsupervised oracle AV-NSL matches or out-performs text-based VG-NSL. Similar to Tabel 3, selftraining with s-Benepar on oracle AV-NSL trees further improves the syntax induction results, almost matching that of right-branching tree. Last but not least, perhaps surprisingly, right-branching trees (RBT) on the given word segmentation reach the best SAIOU and $F _ { 1 }$ scores. We note that the rightbranching approach highly aligns with the head-initial property of English (Baker, 2001), especially in our setting where all punctuation marks were removed; thus, it is nontrivial for AV-NSL to reach the performance on par with RBT without inductive biases favoring any specific type of trees.
|
| 174 |
+
|
| 175 |
+
# 5 ANALYSES
|
| 176 |
+
|
| 177 |
+
Unsupervised Constituent Recall: Following Shi et al. (2019), we show the recall of specific types of constituents (Table 5). While VG-NSL benefits from the head-initial (HI) bias, where abstract words are encouraged to appear in the beginning of a constituent, it is worth noting that AV-NSL outperforms all variations of VG-NSL, without inductive biases favoring any specific types of trees.
|
| 178 |
+
|
| 179 |
+
Table 4: Phrase structure induction with oracle segmentation given. Full table in Appendix 9.
|
| 180 |
+
|
| 181 |
+
<table><tr><td colspan="2">Model</td><td rowspan="2">Output Selection</td><td rowspan="2">F1</td></tr><tr><td>Syntax Induction</td><td>Seg.Representation</td></tr><tr><td>Random</td><td></td><td></td><td>32.77</td></tr><tr><td>Left-Branching</td><td></td><td></td><td>24.56</td></tr><tr><td>Right-Branching VG-NSL</td><td></td><td>Supervised</td><td>57.39 53.11</td></tr><tr><td></td><td>word embeddings</td><td></td><td></td></tr><tr><td>oracle AV-NSL</td><td>log-Mel spectrogram</td><td>Supervised</td><td>42.01</td></tr><tr><td>oracle AV-NSL</td><td>HuBERT2</td><td>Supervised</td><td>55.51</td></tr><tr><td>oracle AV-NSL</td><td>HuBERT2</td><td>MBR</td><td>54.99</td></tr><tr><td>oracle AV-NSL</td><td>HuBERT2,4,6,8,10,12,24</td><td>MBR</td><td>55.96</td></tr><tr><td>oracle AV-NSL →s-Benepar</td><td>HuBERT2</td><td>MBR</td><td>57.24</td></tr><tr><td>oracle AV-NSL →s-Benepar</td><td>HuBERT12</td><td>MBR</td><td>57.33</td></tr></table>
|
| 182 |
+
|
| 183 |
+
Ablation Study: We present two ablations to examine the effectiveness of high-quality word segmentation and visual representation (Table 6). We train AV-NSL with the following modifications:
|
| 184 |
+
|
| 185 |
+
1. Fix the visual representations, but replace oracle segmentation with naive uniform word segmentation, where the number of words in each caption is given (uniform AV-NSL).
|
| 186 |
+
2. Fix the oracle word segmentation, but replace visual embeddings with random images, where each pixel is independently sampled from a uniform distribution.
|
| 187 |
+
|
| 188 |
+
We observe that there are significant performance drops in both settings, comparing to the AV-NSL trained with oracle segmentation and high-quality visual representation. This set of results complement Table 2, stressing that precise word segmentation and high-quality visual representations are both necessary for phrase structure induction from speech. Furthermore, we provide tree structure and word segmentation visualizations for qualitative analysis in the Appendix.
|
| 189 |
+
|
| 190 |
+
Table 6: Top rows: performance of AV-NSL with word segmentation in various quality and high-quality visual embeddings. Bottom rows: performance of AV-NSL with visual embeddings in various quality and highquality word segmentation. DINO: a selfsupervised model that produces high-quality visual representations (Caron et al., 2021).
|
| 191 |
+
|
| 192 |
+
Table 5: Recall of specific typed phrases, including noun phrases (NP), verb phrases (VP), prepositional phrases (PP) and adjective phrases (ADJP), and overall $F _ { 1 }$ score, evaluated on the SpokenCOCO test split. The VG-NSL numbers are taken from (Shi et al., 2019). AV-NSL here are trained on oracle segmentation with vanilla HuBERT as the layer representations.
|
| 193 |
+
|
| 194 |
+
<table><tr><td rowspan="2">Model</td><td rowspan="2">F1</td><td colspan="4">Constituent Recall</td></tr><tr><td>NP</td><td>VP</td><td>PP</td><td>ADJP</td></tr><tr><td>VG-NSL (Shi et al.,2019)</td><td>50.4</td><td>79.6</td><td>26.2</td><td>42.0</td><td>22.0</td></tr><tr><td>VG-NSL + HI</td><td>53.3</td><td>74.6</td><td>32.5</td><td>66.5</td><td>21.7</td></tr><tr><td>VG-NSL + HI+ FastText</td><td>54.4</td><td>78.8</td><td>24.4</td><td>65.6</td><td>22.0</td></tr><tr><td>oracle AV-NSL</td><td>55.6</td><td>55.5</td><td>68.1</td><td>66.6</td><td>22.1</td></tr></table>
|
| 195 |
+
|
| 196 |
+
<table><tr><td colspan="2">Model</td><td rowspan="2">Visual</td><td rowspan="2">F1</td></tr><tr><td>Syntax Induction</td><td>Seg.Repre.</td></tr><tr><td>oracle AV-NSL</td><td>HuBERT10</td><td>ResNet101</td><td>50.50</td></tr><tr><td>uniform AV-NSL</td><td>HuBERT10</td><td>ResNet101</td><td>36.62</td></tr><tr><td>oracle AV-NSL</td><td>HuBERT2</td><td></td><td>55.71</td></tr><tr><td></td><td></td><td>DINO</td><td></td></tr><tr><td>oracle AV-NSL</td><td>HuBERT2</td><td>random</td><td>31.23</td></tr></table>
|
| 197 |
+
|
| 198 |
+
# 6 CONCLUSION
|
| 199 |
+
|
| 200 |
+
In recent years, there have been fruitful progresses in multi-modal induction for zero-resource speech processing and grammar induction respectively. The idea of leveraging the visual modality to learn language competence, either lexicon units from speech or syntactic structure from text, is an attractive approach for modeling human language acquisition. Our study contributes to both lines of research, by presenting an unifying framework that learns phrase structure from visually-grounded speech, without any text. We show that our proposed model, AV-NSL, infers meaningful constituency parse trees on top of continuous word segment representations, both quantitatively and qualitatively. To justify our modeling design choices, we construct several baselines and introduce a novel evaluation metric. We envision our research as the first of many in textless structure learning.
|
| 201 |
+
|
| 202 |
+
# ETHICS STATEMENT
|
| 203 |
+
|
| 204 |
+
This work is scientific at its core, as the goal is to study the process of grammar induction from speech with visual grounding. The data used in this work is also publicly available. One potential concern is that the data and experiments are based on English, which does not represent the global human population. However, we believe that our proposed method is general enough to be applied to other spoken languages when the data is available, because we do not use any language specific speech processing techniques, and we do not have any built-in bias within the models.
|
| 205 |
+
|
| 206 |
+
# REPRODUCIBILITY STATEMENT
|
| 207 |
+
|
| 208 |
+
AV-NSL code, s-Benepar code, and SAIOU evaluation code will be made publicly available. AVNSL code is based on the VG-NSL codebase. s-Benepar code is based on the Benepar codebase. SpokenCOCO is publicly available to download. All models are trained on a single GPU. We also included as many experimental details as we can in the main content of the paper and in Appendix A.2.
|
| 209 |
+
|
| 210 |
+
# REFERENCES
|
| 211 |
+
|
| 212 |
+
Robin Algayres, Tristan Ricoul, Julien Karadayi, Hugo Laurenc¸on, Salah Zaiem, Abdelrahman Mohamed, Benoˆıt Sagot, and Emmanuel Dupoux. Dp-parse: Finding word boundaries from raw speech with an instance lexicon. arXiv preprint arXiv:2206.11332, 2022.
|
| 213 |
+
|
| 214 |
+
Mark C Baker. The atoms of language. Basic Books, 2001.
|
| 215 |
+
|
| 216 |
+
Saurabhchand Bhati, Jesus Villalba, Piotr ´ Zelasko, Laureano Moro-Velazquez, and Najim Dehak. ˙ Segmental contrastive predictive coding for unsupervised word segmentation. Interspeech, 2021.
|
| 217 |
+
|
| 218 |
+
Peter J Bickel and Bo Li. Mathematical statistics. In Test. Citeseer, 1977.
|
| 219 |
+
|
| 220 |
+
Rens Bod. An all-subtrees approach to unsupervised parsing. COLING-ACL, 2006.
|
| 221 |
+
|
| 222 |
+
Mathilde Caron, Hugo Touvron, Ishan Misra, Herve J ´ egou, Julien Mairal, Piotr Bojanowski, and ´ Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9650–9660, 2021.
|
| 223 |
+
|
| 224 |
+
Jan Chorowski, Grzegorz Ciesielski, Jarosław Dzikowski, Adrian Łancucki, Ricard Marxer, Ma- ´ teusz Opala, Piotr Pusz, Paweł Rychlikowski, and Michał Stypułkowski. Aligned contrastive predictive coding. Interspeech, 2021.
|
| 225 |
+
|
| 226 |
+
John Cocke. Programming languages and their compilers: Preliminary notes. New York University, 1969.
|
| 227 |
+
|
| 228 |
+
Carl De Marcken. Unsupervised language acquisition. PhD thesis, Massachusetts Institute of Technology, 1996.
|
| 229 |
+
|
| 230 |
+
Virginia R de Sa. Learning classification with unlabeled data. NeurIPS, 1994.
|
| 231 |
+
|
| 232 |
+
Andrew Drozdov, Pat Verga, Mohit Yadav, Mohit Iyyer, and Andrew McCallum. Unsupervised latent tree induction with deep inside-outside recursive autoencoders. NAACL-HLT, 2019.
|
| 233 |
+
|
| 234 |
+
Ewan Dunbar, Xuan Nga Cao, Juan Benjumea, Julien Karadayi, Mathieu Bernard, Laurent Besacier, Xavier Anguera, and Emmanuel Dupoux. The zero resource speech challenge 2017. ASRU, 2017.
|
| 235 |
+
|
| 236 |
+
Ewan Dunbar, Robin Algayres, Julien Karadayi, Mathieu Bernard, Juan Benjumea, Xuan-Nga Cao, Lucie Miskic, Charlotte Dugrain, Lucas Ondel, Alan W Black, et al. The zero resource speech challenge 2019: Tts without t. Interspeech, 2019.
|
| 237 |
+
|
| 238 |
+
Ewan Dunbar, Julien Karadayi, Mathieu Bernard, Xuan-Nga Cao, Robin Algayres, Lucas Ondel, Laurent Besacier, Sakriani Sakti, and Emmanuel Dupoux. The zero resource speech challenge 2020: Discovering discrete subword and word units. Interspeech, 2020.
|
| 239 |
+
|
| 240 |
+
Emmanuel Dupoux. Cognitive science in the era of artificial intelligence: A roadmap for reverseengineering the infant language-learner. Cognition, 173:43–59, 2018.
|
| 241 |
+
|
| 242 |
+
Zvi Galil. Efficient algorithms for finding maximum matching in graphs. ACM Comput. Surv., 18 (1):23–38, mar 1986. ISSN 0360-0300. doi: 10.1145/6462.6502. URL https://doi.org/ 10.1145/6462.6502.
|
| 243 |
+
|
| 244 |
+
Haohan Guo, Frank K Soong, Lei He, and Lei Xie. Exploiting syntactic features in a parsed tree to improve end-to-end tts. Interspeech, 2019.
|
| 245 |
+
|
| 246 |
+
David Harwath and James R Glass. Learning word-like units from joint audio-visual analysis. ACL, 2017.
|
| 247 |
+
|
| 248 |
+
David Harwath, Wei-Ning Hsu, and James Glass. Learning hierarchical discrete linguistic units from visually-grounded speech. ICLR, 2020.
|
| 249 |
+
|
| 250 |
+
David Frank Harwath. Learning spoken language through vision. PhD thesis, Massachusetts Institute of Technology, 2018.
|
| 251 |
+
|
| 252 |
+
William N. Havard, Jean-Pierre Chevrot, and Laurent Besacier. Word recognition, competition, and activation in a model of visually grounded speech. In CoNLL, 2019.
|
| 253 |
+
|
| 254 |
+
Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016.
|
| 255 |
+
|
| 256 |
+
Yining Hong, Qing Li, Song-Chun Zhu, and Siyuan Huang. Vlgrammar: Grounded grammar induction of vision and language. ICCV, 2021.
|
| 257 |
+
|
| 258 |
+
Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451–3460, 2021a.
|
| 259 |
+
|
| 260 |
+
Wei-Ning Hsu, David Harwath, Christopher Song, and James Glass. Text-free image-to-speech synthesis using learned segmental units. ACL, 2021b.
|
| 261 |
+
|
| 262 |
+
Aren Jansen and Benjamin Van Durme. Efficient spoken term discovery using randomized algorithms. ASRU, 2011.
|
| 263 |
+
|
| 264 |
+
Aren Jansen, Emmanuel Dupoux, Sharon Goldwater, Mark Johnson, Sanjeev Khudanpur, Kenneth Church, Naomi Feldman, Hynek Hermansky, Florian Metze, Richard Rose, et al. A summary of the 2012 jhu clsp workshop on zero resource speech technologies and models of early language acquisition. ICASSP, 2013.
|
| 265 |
+
|
| 266 |
+
Nobuyoshi Kaiki, Sakriani Sakti, and Satoshi Nakamura. Using local phrase dependency structure information in neural sequence-to-sequence speech synthesis. In 2021 24th Conference of the Oriental COCOSDA International Committee for the Co-ordination and Standardisation of Speech Databases and Assessment Techniques (O-COCOSDA), pp. 206–211. IEEE, 2021.
|
| 267 |
+
|
| 268 |
+
Herman Kamper. Word segmentation on discovered phone units with dynamic programming and self-supervised scoring. arXiv preprint arXiv:2202.11929, 2022.
|
| 269 |
+
|
| 270 |
+
Herman Kamper and Benjamin van Niekerk. Towards unsupervised phone and word segmentation using self-supervised vector-quantized neural networks. Interspeech, 2021.
|
| 271 |
+
|
| 272 |
+
Herman Kamper, Aren Jansen, and Sharon Goldwater. Fully unsupervised small-vocabulary speech recognition using a segmental bayesian model. In Sixteenth Annual Conference of the International Speech Communication Association, 2015.
|
| 273 |
+
|
| 274 |
+
Herman Kamper, Aren Jansen, and Sharon Goldwater. A segmental framework for fullyunsupervised large-vocabulary speech recognition. Computer Speech & Language, 46:154–174, 2017.
|
| 275 |
+
|
| 276 |
+
Tadao Kasami. An efficient recognition and syntax-analysis algorithm for context-free languages. Coordinated Science Laboratory Report no. R-257, 1966.
|
| 277 |
+
|
| 278 |
+
Eugene Kharitonov, Ann Lee, Adam Polyak, Yossi Adi, Jade Copet, Kushal Lakhotia, Tu-Anh Nguyen, Morgane Riviere, Abdelrahman Mohamed, Emmanuel Dupoux, et al. Text-free prosody- \` aware generative spoken language modeling. ACL, 2022.
|
| 279 |
+
|
| 280 |
+
Khazar Khorrami and Okko Johannes Ras¨ anen. Can phones, syllables, and words emerge as ¨ side-products of cross-situational audiovisual learning? - a computational investigation. ArXiv, abs/2109.14200, 2021.
|
| 281 |
+
|
| 282 |
+
Yoon Kim, Chris Dyer, and Alexander M Rush. Compound probabilistic context-free grammars for grammar induction. ACL, 2019a.
|
| 283 |
+
|
| 284 |
+
Yoon Kim, Alexander M Rush, Lei Yu, Adhiguna Kuncoro, Chris Dyer, and Gabor Melis. Unsuper- ´ vised recurrent neural network grammars. NAACL-HLT, 2019b.
|
| 285 |
+
|
| 286 |
+
Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014.
|
| 287 |
+
|
| 288 |
+
Nikita Kitaev and Dan Klein. Constituency parsing with a self-attentive encoder. In ACL, 2018.
|
| 289 |
+
|
| 290 |
+
Dan Klein and Christopher D Manning. A generative constituent-context model for improved grammar induction. ACL, 2002.
|
| 291 |
+
|
| 292 |
+
Dan Klein and Christopher D Manning. Corpus-based induction of syntactic structure: Models of dependency and constituency. ACL, 2004.
|
| 293 |
+
|
| 294 |
+
Arne Kohn, Timo Baumann, and Oskar D ¨ orfler. An empirical analysis of the correlation of syntax ¨ and prosody. Interspeech, 2018.
|
| 295 |
+
|
| 296 |
+
Noriyuki Kojima, Hadar Averbuch-Elor, Alexander M Rush, and Yoav Artzi. What is learned in visually grounded neural syntax acquisition. ACL, 2020.
|
| 297 |
+
|
| 298 |
+
Shankar Kumar and William Byrne. Minimum Bayes-risk decoding for statistical machine translation. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pp. 169–176, Boston, Massachusetts, USA, May 2 - May 7 2004. Association for Computational Linguistics. URL https://aclanthology.org/N04-1022.
|
| 299 |
+
|
| 300 |
+
Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, et al. On generative spoken language modeling from raw audio. Transactions of the Association for Computational Linguistics, 9:1336–1354, 2021.
|
| 301 |
+
|
| 302 |
+
Chia-ying Lee and James Glass. A nonparametric bayesian approach to acoustic model discovery. ACL, 2012.
|
| 303 |
+
|
| 304 |
+
Chia-ying Lee, Timothy J O’donnell, and James Glass. Unsupervised lexicon discovery from acoustic input. Transactions of the Association for Computational Linguistics, 3:389–403, 2015.
|
| 305 |
+
|
| 306 |
+
Bowen Li, Lili Mou, and Frank Keller. An imitation learning approach to unsupervised parsing. ACL, 2019.
|
| 307 |
+
|
| 308 |
+
Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ´ European conference on computer vision, pp. 740–755. Springer, 2014.
|
| 309 |
+
|
| 310 |
+
Paria Jamshid Lou, Yufei Wang, and Mark Johnson. Neural constituency parsing of speech transcripts. NAACL-HLT, 2019.
|
| 311 |
+
|
| 312 |
+
Jana M Mason. When do children begin to read: An exploration of four year old children’s letter and word reading competencies. Reading Research Quarterly, pp. 203–227, 1980.
|
| 313 |
+
|
| 314 |
+
Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. Montreal forced aligner: Trainable text-speech alignment using kaldi. Interspeech, 2017.
|
| 315 |
+
|
| 316 |
+
Fergus McInnes and Sharon Goldwater. Unsupervised extraction of recurring words from infantdirected speech. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 33, 2011.
|
| 317 |
+
|
| 318 |
+
Grzegorz Chrupała Mitja Nikolaus, Afra Alishahi. Learning english with peppa pig. TACL, 2022.
|
| 319 |
+
|
| 320 |
+
Letitia Naigles. Children use syntax to learn verb meanings. Journal of child language, 17(2): 357–374, 1990.
|
| 321 |
+
|
| 322 |
+
Tu Nguyen, Maureen de Seyssel, Patricia Roz’e, Morgane Riviere, Evgeny Kharitonov, Alexei \` Baevski, Ewan Dunbar, and Emmanuel Dupoux. The zero resource speech benchmark 2021: Metrics and baselines for unsupervised spoken language modeling. Self-Supervised Learning for Speech and Audio Processing NeurIPS Workshop, 2020.
|
| 323 |
+
|
| 324 |
+
Kayode Olaleye and Herman Kamper. Attention-based keyword localisation in speech using visual grounding. In Interspeech, 2021.
|
| 325 |
+
|
| 326 |
+
Alex S Park and James R Glass. Unsupervised pattern discovery in speech. IEEE Transactions on Audio, Speech, and Language Processing, 16(1):186–197, 2007.
|
| 327 |
+
|
| 328 |
+
Puyuan Peng and David Harwath. Self-supervised representation learning for speech using visual grounding and masked language modeling. Self-Supervised Learning for Speech and Audio Processing Workshop at AAAI, 2022a.
|
| 329 |
+
|
| 330 |
+
Puyuan Peng and David Harwath. Word discovery in visually grounded, self-supervised speech models. Interspeech, 2022b.
|
| 331 |
+
|
| 332 |
+
Adam Polyak, Yossi Adi, Jade Copet, Eugene Kharitonov, Kushal Lakhotia, Wei-Ning Hsu, Abdelrahman Mohamed, and Emmanuel Dupoux. Speech resynthesis from discrete disentangled self-supervised representations. Interspeech, 2021.
|
| 333 |
+
|
| 334 |
+
Adrien Pupier, Maximin Coavoux, Benjamin Lecouteux, and Jer´ ome Goulian. End-to-end depen- ˆ dency parsing of spoken french. In Interspeech, 2022.
|
| 335 |
+
|
| 336 |
+
Brian Roark, Mary Harper, Eugene Charniak, Bonnie Dorr, Mark Johnson, Jeremy G Kahn, Yang Liu, Mari Ostendorf, John Hale, Anna Krasnyanskaya, et al. Sparseval: Evaluation metrics for parsing speech. LREC, 2006.
|
| 337 |
+
|
| 338 |
+
Deb K Roy and Alex P Pentland. Learning words from sights and sounds: A computational model. Cognitive science, 26(1):113–146, 2002.
|
| 339 |
+
|
| 340 |
+
Yikang Shen, Zhouhan Lin, Chin-Wei Huang, and Aaron Courville. Neural language modeling by jointly learning syntax and lexicon. ICLR, 2018.
|
| 341 |
+
|
| 342 |
+
Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. Ordered neurons: Integrating tree structures into recurrent neural networks. ICLR, 2019.
|
| 343 |
+
|
| 344 |
+
Freda Shi, Daniel Fried, Marjan Ghazvininejad, Luke Zettlemoyer, and Sida I Wang. Natural language to code translation with execution. arXiv preprint arXiv:2204.11454, 2022.
|
| 345 |
+
|
| 346 |
+
Haoyue Shi, Jiayuan Mao, Kevin Gimpel, and Karen Livescu. Visually grounded neural syntax acquisition. ACL, 2019.
|
| 347 |
+
|
| 348 |
+
Haoyue Shi, Karen Livescu, and Kevin Gimpel. On the role of supervision in unsupervised constituency parsing. In EMNLP. Association for Computational Linguistics, 2020.
|
| 349 |
+
|
| 350 |
+
Valentin I. Spitkovsky, Hiyan Alshawi, and Daniel Jurafsky. Breaking out of local optima with count transforms and model recombination: A study in grammar induction. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1983–1995, Seattle, Washington, USA, October 2013. Association for Computational Linguistics. URL https: //aclanthology.org/D13-1204.
|
| 351 |
+
|
| 352 |
+
Zheng-Hua Tan, Najim Dehak, et al. rvad: An unsupervised segment-based robust voice activity detection method. Computer speech & language, 59:1–21, 2020.
|
| 353 |
+
|
| 354 |
+
Trang Tran and Mari Ostendorf. Assessing the use of prosody in constituency parsing of imperfect transcripts. Interspeech, 2021.
|
| 355 |
+
|
| 356 |
+
Trang Tran, Shubham Toshniwal, Mohit Bansal, Kevin Gimpel, Karen Livescu, and Mari Ostendorf. Parsing speech: a neural approach to integrating lexical and acoustic-prosodic information. NAACL-HLT, 2018.
|
| 357 |
+
|
| 358 |
+
Trang Tran, Jiahong Yuan, Yang Liu, and Mari Ostendorf. On the role of style in parsing speech with neural models. Interspeech, 2019.
|
| 359 |
+
|
| 360 |
+
Shubhi Tyagi, Marco Nicolis, Jonas Rohnke, Thomas Drugman, and Jaime Lorenzo-Trueba. Dynamic prosody generation for speech synthesis using linguistics-driven acoustic embedding selection. Interspeech, 2020.
|
| 361 |
+
|
| 362 |
+
Maarten Versteegh, Roland Thiolliere, Thomas Schatz, Xuan Nga Cao, Xavier Anguera, Aren Jansen, and Emmanuel Dupoux. The zero resource speech challenge 2015. ISCA, 2015.
|
| 363 |
+
|
| 364 |
+
Michael Wagner and Duane G Watson. Experimental and theoretical advances in prosody: A review. Language and cognitive processes, 25(7-9):905–945, 2010.
|
| 365 |
+
|
| 366 |
+
Bo Wan, Wenjuan Han, Zilong Zheng, and Tinne Tuytelaars. Unsupervised vision-language grammar induction with shared structure modeling. ICLR, 2022.
|
| 367 |
+
|
| 368 |
+
Liming Wang and Mark Hasegawa-Johnson. A translation framework for visually grounded spoken unit discovery. In ACSSC, 2021.
|
| 369 |
+
|
| 370 |
+
Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229–256, 1992. URL https://link.springer.com/ content/pdf/10.1007/BF00992696.pdf.
|
| 371 |
+
|
| 372 |
+
Daniel H Younger. Recognition and parsing of context-free languages in time n3. Information and control, 10(2):189–208, 1967.
|
| 373 |
+
|
| 374 |
+
Hao Zhang and Daniel Gildea. Efficient multi-pass decoding for synchronous context free grammars. In Proceedings of ACL-08: HLT, pp. 209–217, Columbus, Ohio, June 2008. Association for Computational Linguistics. URL https://aclanthology.org/P08-1025.
|
| 375 |
+
|
| 376 |
+
Songyang Zhang, Linfeng Song, Lifeng Jin, Kun Xu, Dong Yu, and Jiebo Luo. Video-aided unsupervised grammar induction. NAACL-HLT, 2021.
|
| 377 |
+
|
| 378 |
+
Yaodong Zhang. Unsupervised speech processing with applications to query-by-example spoken term detection. PhD thesis, Massachusetts Institute of Technology, 2013.
|
| 379 |
+
|
| 380 |
+
Yaodong Zhang and James R Glass. Unsupervised spoken keyword spotting via segmental dtw on gaussian posteriorgrams. ASRU, 2009.
|
| 381 |
+
|
| 382 |
+
Yanpeng Zhao and Ivan Titov. Visually grounded compound pcfgs. EMNLP, 2020.
|
| 383 |
+
|
| 384 |
+
# A APPENDIX
|
| 385 |
+
|
| 386 |
+
A.1 BASELINES
|
| 387 |
+
|
| 388 |
+
AV-cPCFG: We train compound probabilistic context free grammar (cPCFG) (Kim et al., 2019a) on word-level discrete speech tokens. Similar to AV-NSL, word segments are obtained from VGHuBERT with segment insertion, and segment representations are extracted from VG-Hubert layer 10 with CLS attention weighted mean-pool. Different from AV-NSL, the segment representations are discretized via kmeans to obtain word-level discrete indices. Because the discretization is wordlevel instead of phone-level, we swept the number of kmeans cluster over $\left\{ 1 \mathrm { k } , 2 \mathrm { k } , 4 \mathrm { k } , 8 \mathrm { k } , 1 2 \mathrm { k } , 1 6 \mathrm { k } , \right.$ ， $2 0 \mathrm { k } \}$ , which corresponds to the dictionary size in cPCFG. In summary, AV-cPCFG leverages visual cues only for segmentation and segment representations, but not for phrase structure induction.
|
| 389 |
+
|
| 390 |
+
DPDP-cPCFG: Instead of training cPCFG on audio-visual word segments and audio-visual segment representations, DPDP-cPCFG does not rely on any visual grounding throughout. Instead, DPDP (Kamper, 2022), a recent speech-only word segmentation algorithm, and vanilla HuBERT representations mean-pooled over DPDP segments are used. We swept through HuBERT layer {2, 4, 6, 8, 10, 12}. As in AV-cPCFG, kmeans is used for word-level discretization.
|
| 391 |
+
|
| 392 |
+
oracle AV-NSL: To remove the uncertainty of unsupervised word segmentation, we directly train AV-NSL on top of oracle word segmentation via force alignment. The segment representations are based on learnable attention pooling over vanilla HuBERT layer $\{ 2 , 4 , 6 , 8 , 1 0 , 1 2 \}$ representations. We also tried log Mel spectrograms and HuBERT-L 300M to examine the effectiveness of different input representations. One note is that simpler score and combine parametrization suffices here6.
|
| 393 |
+
|
| 394 |
+
# A.2 HYPERPARAMETERS
|
| 395 |
+
|
| 396 |
+
For VG-HuBERT, we run MBR selection on the combination of insertion gap $\{ 0 . 1 , 0 . 2 , 0 . 3 \}$ seconds, segmentation layer $\{ 9 , 1 0 , 1 1 \}$ , attention magnitude threshold at top $\{ 3 0 \% , 2 0 \% , 1 0 \% \}$ , three training random seeds, and model snapshots at training step 20k, 30k, 40k, 50k, 60k. This gives 405 combinations in total.
|
| 397 |
+
|
| 398 |
+
# A.3 FULL RESULTS TABLE
|
| 399 |
+
|
| 400 |
+
# A.4 WORD SEGMENTATION VIZ
|
| 401 |
+
|
| 402 |
+
We show more examples of word segmentation generated by our improved VG-HuBERT in Figure 4. Segments marked with $" + "$ are inserted segments, and vertical blue dotted lines are inferred word boundaries.
|
| 403 |
+
|
| 404 |
+
# A.5 VISUALIZATION OF INDUCED TREES
|
| 405 |
+
|
| 406 |
+
We visualize the induced trees in Figure 5.
|
| 407 |
+
|
| 408 |
+
Table 7: Fully-unsupervised phrase structure induction results evaluated with SAIOU.
|
| 409 |
+
|
| 410 |
+
<table><tr><td colspan="3">Model</td><td rowspan="2">Output Selection</td><td rowspan="2">SAIoU</td></tr><tr><td>Syntax Induction</td><td>Segmentation</td><td>Seg.Representation (continuous/discrete)</td></tr><tr><td>Right-Branching</td><td>VG-HuBERT+MBR10</td><td></td><td></td><td>0.546</td></tr><tr><td>Right-Branching</td><td>DPDP</td><td></td><td></td><td>0.478</td></tr><tr><td>AV-NSL</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o (continuous)</td><td>MBR</td><td>0.516</td></tr><tr><td>AV-NSL</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT11 (continuous)</td><td>MBR</td><td>0.498</td></tr><tr><td>AV-NSL</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT12 (continuous)</td><td>MBR</td><td>0.492</td></tr><tr><td>AV-NSL</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT10,11,12 (continuous)</td><td>MBR</td><td>0.521</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+1k km (discrete)</td><td>last ckpt.</td><td>0.454</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+2k km (discrete)</td><td>last ckpt.</td><td>0.444</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+4k km (discrete)</td><td>last ckpt.</td><td>0.499</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+8k km (discrete)</td><td>last ckpt.</td><td>0.481</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+12k km (discrete)</td><td>last ckpt.</td><td>0.473</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+16k km (discrete)</td><td>last ckpt.</td><td>0.471</td></tr><tr><td>AV-cPCFG</td><td>VG-HuBERT+MBR10</td><td>VG-HuBERT1o+20k km (discrete)</td><td>last ckpt.</td><td>0.454</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT2+1k km (discrete)</td><td>last ckpt.</td><td>0.434</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT2+2k km (discrete)</td><td>last ckpt.</td><td>0.465</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT2+4k km (discrete)</td><td>last ckpt.</td><td>0.444</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT2+8k km (discrete)</td><td>last ckpt.</td><td>0.387</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT2+12k km (discrete)</td><td>last ckpt.</td><td>0.447</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT2+16k km (discrete)</td><td>last ckpt.</td><td>0.360</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT1o+1k km (discrete)</td><td>last ckpt.</td><td>0.403</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT1o+2k km (discrete)</td><td>last ckpt.</td><td>0.426</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT1o+4k km (discrete)</td><td>last ckpt.</td><td>0.415</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT1o+8k km (discrete)</td><td>last ckpt.</td><td>0.367</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT1o+12k km (discrete)</td><td>last ckpt.</td><td>0.415</td></tr><tr><td>DPDP-cPCFG</td><td>DPDP</td><td>HuBERT1o+16k km (discrete)</td><td>last ckpt.</td><td>0.414</td></tr></table>
|
| 411 |
+
|
| 412 |
+
Table 8: Self-training results evaluated with SAIOU.
|
| 413 |
+
|
| 414 |
+
<table><tr><td rowspan="2">Model</td><td rowspan="2">Segmentation</td><td rowspan="2">Seg.Representation</td><td colspan="3">tree target</td><td rowspan="2">Output Selection</td><td rowspan="2">SAIoU</td></tr><tr><td>train</td><td>val</td><td>test</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT2</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.538</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT4</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.536</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT6</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.538</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT8</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.532</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT10</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.537</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT12</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>last ckpt.</td><td>0.536</td></tr><tr><td>s-Benepar</td><td>VG-HuBERT+MBR10</td><td>HuBERT2,4,6,8,10,12</td><td>AV-NSL</td><td>AV-NSL</td><td>oracle</td><td>MBR</td><td>0.536</td></tr></table>
|
| 415 |
+
|
| 416 |
+
Table 9: Phrase structure induction with oracle segmentation given results evaluated with $F _ { 1 }$ .
|
| 417 |
+
|
| 418 |
+
<table><tr><td colspan="3">Model</td><td rowspan="2">Output Selection</td><td rowspan="2">F1</td></tr><tr><td>Syntax Induction</td><td>Segmentation</td><td>Seg.Representation</td></tr><tr><td>Random</td><td>oracle</td><td></td><td></td><td>32.77</td></tr><tr><td>Left-Branching</td><td>oracle</td><td></td><td></td><td>24.56</td></tr><tr><td>Right-Branching</td><td>oracle</td><td></td><td></td><td>57.39</td></tr><tr><td>VG-NSL</td><td></td><td>word embeddings</td><td>Supervised</td><td>53.11</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>log-Mel spectrogram</td><td>Supervised</td><td>42.01</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT2</td><td>Supervised</td><td>55.51</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT-L24</td><td></td><td>54.63</td></tr><tr><td></td><td></td><td></td><td>Supervised</td><td></td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT2</td><td>MBR</td><td>54.99</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT4</td><td>MBR</td><td>53.25</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT6</td><td>MBR</td><td>53.46</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT8</td><td>MBR</td><td>53.14</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT10</td><td>MBR</td><td>36.67</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT12</td><td>MBR</td><td>48.51</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT-L24</td><td>MBR</td><td>54.39</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT2,4,6,8,10,12</td><td>MBR</td><td>55.56</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT2,4,6,8,10,12,24</td><td>MBR</td><td>55.96</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td>AV-NSL →s-Benepar</td><td>oracle</td><td>HuBERT2</td><td>MBR</td><td>57.24</td></tr><tr><td>AV-NSL→s-Benepar</td><td>oracle</td><td>HuBERT4</td><td>MBR</td><td>57.08</td></tr><tr><td>AV-NSL→s-Benepar</td><td>oracle</td><td>HuBERT6</td><td>MBR</td><td>56.81</td></tr><tr><td>AV-NSL →s-Benepar</td><td>oracle</td><td>HuBERT8</td><td>MBR</td><td>56.94</td></tr><tr><td>AV-NSL→s-Benepar</td><td>oracle</td><td>HuBERT10</td><td>MBR</td><td>57.16</td></tr><tr><td>AV-NSL →s-Benepar</td><td>oracle</td><td>HuBERT12</td><td>MBR</td><td>57.33</td></tr></table>
|
| 419 |
+
|
| 420 |
+
Table 10: Recall of specific typed phrases, and overall $F _ { 1 }$ score, evaluated on the SpokenCOCO test split. VG-NSL numbers are taken directly from (Shi et al., 2019). AV-NSL here are trained on oracle segmentation with vanilla HuBERT as the layer representations.
|
| 421 |
+
|
| 422 |
+
<table><tr><td rowspan="2">Model</td><td rowspan="2">F1</td><td colspan="4">Constituent Recall</td></tr><tr><td>NP</td><td>VP</td><td>PP</td><td>ADJP</td></tr><tr><td>VG-NSL (Shi et al.,2019)</td><td>50.4</td><td>79.6</td><td>26.2</td><td>42.0</td><td>22.0</td></tr><tr><td>VG-NSL + HI</td><td>53.3</td><td>74.6</td><td>32.5</td><td>66.5</td><td>21.7</td></tr><tr><td>VG-NSL +HI+FastText</td><td>54.4</td><td>78.8</td><td>24.4</td><td>65.6</td><td>22.0</td></tr><tr><td>AV-NSL (oracle seg.+ HuBERT2)</td><td>55.6</td><td>55.5</td><td>68.1</td><td>66.6</td><td>22.1</td></tr><tr><td>AV-NSL (oracle seg.+HuBERT4)</td><td>53.7</td><td>57.4</td><td>56.8</td><td>61.3</td><td>21.3</td></tr><tr><td>AV-NSL (oracle seg.+HuBERT6)</td><td>53.9</td><td>59.4</td><td>55.4</td><td>59.3</td><td>21.2</td></tr><tr><td>AV-NSL (oracle seg.+HuBERT8)</td><td>53.9</td><td>56.0</td><td>58.0</td><td>64.9</td><td>22.5</td></tr><tr><td>AV-NSL (oracle seg.+HuBERT10)</td><td>50.6</td><td>55.8</td><td>48.1</td><td>57.0</td><td>20.5</td></tr><tr><td>AV-NSL (oracle seg. + HuBERT12)</td><td>49.0</td><td>62.5</td><td>34.4</td><td>45.0</td><td>17.4</td></tr></table>
|
| 423 |
+
|
| 424 |
+
Table 11: Top rows: Impact of segmentation quality for AV-NSL with number of words segments known in advance. Bottom rows: Impact of visual embedding for AV-NSL
|
| 425 |
+
|
| 426 |
+
<table><tr><td colspan="3">Model</td><td rowspan="2">Visual Embedding</td><td rowspan="2">F1</td></tr><tr><td>Syntax Induction</td><td>Segmentation</td><td>Seg.Representation</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT2</td><td>ResNet101</td><td>55.51</td></tr><tr><td>AV-NSL</td><td>uniform</td><td>HuBERT2</td><td>ResNet101</td><td>48.97</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT10</td><td>ResNet101</td><td>50.50</td></tr><tr><td>AV-NSL</td><td>uniform</td><td>HuBERT10</td><td>ResNet101</td><td>36.62</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT2</td><td>DINO</td><td>55.71</td></tr><tr><td>AV-NSL</td><td>oracle</td><td>HuBERT2</td><td>random</td><td>31.23</td></tr></table>
|
| 427 |
+
|
| 428 |
+
|
| 429 |
+
Figure 4: Examples of attention segments generated by VG-HuBERT. Inserted segments are marked with $" + "$ . Vertical blue dotted lines are inferred word boundaries.
|
| 430 |
+
|
| 431 |
+
|
| 432 |
+
Figure 5: Visualization of an example produced by AV-NSL (best viewed in color). Top (red and green): the ground-truth parse tree; bottom (blue and yellow): the generated parse tree. In each tree, a parent segment adjacently covers its two children segments.
|
parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/0c2SbGJ3Lt/0c2SbGJ3Lt_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/0xiJLKH-ufZ/0xiJLKH-ufZ_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5673a7e11ea1cd8762ad8fe7ace57afe12d1065f3d3580a0b647cbbe7552f847
|
| 3 |
+
size 8596058
|
parse/dev/0xiJLKH-ufZ/0xiJLKH-ufZ_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e7f01defb7587f3aec33bb53c0c31c4276867276f278dcd356c5f2098ae09d11
|
| 3 |
+
size 8600909
|
parse/dev/1QQnYd02etI/1QQnYd02etI_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6bb4ddf28f6680e6ad1da88595b2fd2860d2c0461727a44920b2c10c6b301e12
|
| 3 |
+
size 1954944
|
parse/dev/1QQnYd02etI/1QQnYd02etI_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fce373ad721daa577fc2e5d38f13cbfcd03240e1ec7f2b4db1a3d3d059aab7e2
|
| 3 |
+
size 1785998
|
parse/dev/1QQnYd02etI/1QQnYd02etI_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cf30b4c76e72c63dcca718de02a1eda3c0e59dbcec15c2880410e2bd96959574
|
| 3 |
+
size 1960116
|
parse/dev/1wVvweK3oIb/1wVvweK3oIb.md
ADDED
|
@@ -0,0 +1,474 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SIMPLE GNN REGULARISATION FOR 3D MOLECULARPROPERTY PREDICTION & BEYOND
|
| 2 |
+
|
| 3 |
+
Jonathan Godwin, Michael Schaarschmidt, Alexander Gaunt, Alvaro Sanchez-Gonzales, Yulia Rubanova, Petar Velickovi ˇ c,´ James Kirkpatrick & Peter Battaglia
|
| 4 |
+
|
| 5 |
+
DeepMind, London {jonathangodwin}@deepmind.com
|
| 6 |
+
|
| 7 |
+
# ABSTRACT
|
| 8 |
+
|
| 9 |
+
In this paper we show that simple noisy regularisation can be an effective way to address oversmoothing. We argue that regularisers addressing oversmoothing should both penalise node latent similarity and encourage meaningful node representations. From this observation we derive “Noisy Nodes”, a simple technique in which we corrupt the input graph with noise, and add a noise correcting node-level loss. The diverse node level loss encourages latent node diversity, and the denoising objective encourages graph manifold learning. Our regulariser applies well-studied methods in simple, straightforward ways which allow even generic architectures to overcome oversmoothing and achieve state of the art results on quantum chemistry tasks, and improve results significantly on Open Graph Benchmark (OGB) datasets. Our results suggest Noisy Nodes can serve as a complementary building block in the GNN toolkit.
|
| 10 |
+
|
| 11 |
+
# 1 INTRODUCTION
|
| 12 |
+
|
| 13 |
+
Graph Neural Networks (GNNs) are a family of neural networks that operate on graph structured data by iteratively passing learned messages over the graph’s structure (Scarselli et al., 2009; Bronstein et al., 2017; Gilmer et al., 2017; Battaglia et al., 2018; Shlomi et al., 2021). While Graph Neural Networks have demonstrated success in a wide variety of tasks (Zhou et al., 2020a; Wu et al., 2020; Bapst et al., 2020; Schütt et al., 2017; Klicpera et al., 2020a), it has been proposed that in practice “oversmoothing” limits their ability to benefit from overparametrization.
|
| 14 |
+
|
| 15 |
+
Oversmoothing is a phenomenon where a GNN’s latent node representations become increasing indistinguishable over successive steps of message passing (Chen et al., 2019). Once these representations are oversmoothed, the relational structure of the representation is lost, and further message-passing cannot improve expressive capacity. We argue that the challenges of overcoming oversmoothing are two fold. First, finding a way to encourage node latent diversity; second, to encourage the diverse node latents to encode meaningful graph representations. Here we propose a simple noise regulariser, Noisy Nodes, and demonstrate how it overcomes these challenges across a range of datasets and architectures, achieving top results on OC20 IS2RS & IS2RE direct, QM9 and OGBG-PCQM4Mv1.
|
| 16 |
+
|
| 17 |
+
Our “Noisy Nodes” method is a simple technique for regularising GNNs and associated training procedures. During training, our noise regularisation approach corrupts the input graph’s attributes with noise, and adds a per-node noise correction term. We posit that our Noisy Nodes approach is effective because the model is rewarded for maintaining and refining distinct node representations through message passing to the final output, which causes it to resist oversmoothing. Like denoising autoencoders, it encourages the model to explicitly learn the manifold on which the uncorrupted input graph’s features lie, analogous to a form of representation learning. When applied to 3D molecular prediction tasks, it encourages the model to distinguish between low and high energy states. We find that applying Noisy Nodes reduces oversmoothing for shallower networks, and allows us to see improvements with added depth, even on tasks for which depth was assumed to be unhelpful.
|
| 18 |
+
|
| 19 |
+
This study’s approach is to investigate the combination of Noisy Nodes with generic, popular baseline GNN architectures. For 3D Molecular prediction we use a standard architecture working on 3D point clouds developed for particle fluid simulations, the Graph Net Simulator (GNS) (Sanchez-Gonzalez\* et al., 2020), which has also been used for molecular property prediction (Hu et al., 2021b). Without using Noisy Nodes the GNS is not a competitive model, but using Noisy Nodes allows the GNS to achieve top performance on three 3D molecular property prediction tasks: the OC20 IS2RE direct task by $43 \%$ over previous work, $12 \%$ on OC20 IS2RS direct, and top results on 3 out of 12 of the QM9 tasks. For non-spatial GNN benchmarks we test a MPNN (Gilmer et al., 2017) on OGBG-MOLPCBA and OGBG-PCQM4M (Hu et al., 2021a) and again see significant improvements. Finally, we applied Noisy Nodes to a GCN (Kipf & Welling, 2016), arguably the most popular and simple GNN, trained on OGBN-Arxiv and see similar results. These results suggest Noisy Nodes can serve as a complementary GNN building block.
|
| 20 |
+
|
| 21 |
+
# 2 PRELIMINARIES: GRAPH PREDICTION PROBLEM
|
| 22 |
+
|
| 23 |
+
Let $G = ( V , E , g )$ be an input graph. The nodes are $V = \{ v _ { 1 } , \ldots , v _ { | V | } \}$ , where $v _ { i } \in \mathbb { R } ^ { d _ { v } }$ . The directed, attributed edges are $E = \left\{ e _ { 1 } , \dots , e _ { | E | } \right\}$ : each edge includes a sender node index, receiver node index, and edge attribute, $\boldsymbol { e } _ { k } = \left( \boldsymbol { s } _ { k } , r _ { k } , \boldsymbol { e } _ { k } \right)$ , respectively, where $s _ { k } , r _ { k } \in \{ 1 , \dots , | V | \}$ and $e _ { k } \in \mathbb { R } ^ { d _ { e } }$ . The graph-level property is $g \in \mathbb { R } ^ { d _ { g } }$ .
|
| 24 |
+
|
| 25 |
+
The goal is to predict a target graph, $G ^ { \prime }$ , with the same structure as $G$ , but different node, edge, and/or graph-level attributes. We denote $\hat { G } ^ { \prime }$ as a model’s prediction of $G ^ { \prime }$ . Some error metric defines quality of $\hat { G } ^ { \prime }$ with respect to the target $G ^ { \prime }$ , $\mathrm { E r r o r } ( \hat { G } ^ { \prime } , G ^ { \prime } )$ , which the training loss terms are defined to optimize. In this paper the phrase “message passing steps” is synonymous with “GNN layers”.
|
| 26 |
+
|
| 27 |
+
# 3 OVERSMOOTHING
|
| 28 |
+
|
| 29 |
+
“Oversmoothing” is when the node latent vectors of a GNN become very similar after successive layers of message passing. Once nodes are identical there is no relational information contained in the nodes, and no higher-order latent graph representations can be learned. It is easiest to see this effect with the update function of a Graph Convolutional Network with no adjacency normalization $\begin{array} { r } { v _ { i } ^ { k } = \sum _ { j } W v _ { j } ^ { k - 1 } } \end{array}$ with $j \in N e i g h b o r h o o d _ { v _ { i } }$ , $W \in \mathbb { R } ^ { d _ { g } \times d _ { g } }$ and $k$ the layer index. As the number of applications increases, the averaging effect of the summation forces the nodes to become almost identical. However, as soon as residual connections are added we can construct a network that need not suffer from oversmoothing by setting the residual updates to zero at a similarity threshold. Similarly, multi-head attention Vaswani et al. (2017); Velickovi ˇ c et al. (2018) and GNNs with edge ´ updates (Battaglia et al., 2018; Gilmer et al., 2017) can modulate node updates. As such for modern GNNs oversmoothing is primarily a “training” problem - i.e. how to choose model architectures and regularisers to encourage and preserve meaningful latent relational representations.
|
| 30 |
+
|
| 31 |
+
We can discern two desiderata for a regulariser or loss that addresses oversmoothing. First, it should penalise identical node latents. Second, it should encourage meaningful latent representations of the data. One such example may be the auto-regressive loss of transformer based language models (Brown et al. (2020)). In this case, each word (equivalent to node) prediction must be distinct, and the auto-regressive loss encourages relational dependence upon prior words. We can take inspiration from this observation to derive auxiliary losses that both have diverse node targets and encourage relational representation learning. In the following section we derive one such regulariser, Noisy Nodes.
|
| 32 |
+
|
| 33 |
+
# 4 NOISY NODES
|
| 34 |
+
|
| 35 |
+
Noisy Nodes tackles the oversmoothing problem by adding a diverse noise correction target, modifying the original graph prediction problem definition in several ways. It introduces a graph corrupted by noise, $\bar { \tilde { G } } = \tilde { ( V , E , g ) }$ , where $\tilde { v } _ { i } \in \tilde { V }$ is constructed by adding noise, $\sigma _ { i }$ , to the input nodes, $\tilde { v } _ { i } = v _ { i } + \sigma _ { i }$ . The edges, $\tilde { E }$ , and graph-level attribute, $\tilde { g }$ , can either be uncorrupted by noise (i.e., $\tilde { E } = E , \tilde { g } = g )$ , calculated from the noisy nodes (for example in a nearest neighbors graph), or corrupted independent of the nodes—these are minor choices that can be informed by the specific problem setting.
|
| 36 |
+
|
| 37 |
+
$$
|
| 38 |
+
\begin{array} { c } { { \displaystyle \binom { \zeta _ { i } } { \upsilon _ { i } } { \cdots } _ { { \bf \bar { \Phi } } _ { i } } ; } } \\ { { + \Delta _ { i } \left[ \begin{array} { c } { { { \bf \bar { \Phi } } _ { \bar { i } } } } \\ { { { \bf \bar { \Phi } } _ { \bar { i } } } } \end{array} \right] ^ { \prime } { \bf \bar { \Phi } } _ { \bar { { { + } } } \Delta _ { i } } - \sigma _ { i } } } \\ { { \displaystyle \binom { \bf \bar { \bf \Phi } _ { \bar { i } } } { \bf \bar { \Phi } } ^ { \prime } { \bf \bar { \Phi } } ^ { \prime } { \bf \bar { \Phi } } _ { \bar { { { + } } } \Delta _ { i } } } } \end{array}
|
| 39 |
+
$$
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
Figure 2: Per layer node latent diversity, measured by MAD on a 16 layer MPNN trained on OGBGMOLPCBA. Noisy Nodes maintains a higher level of diversity throughout the network than competing methods.
|
| 43 |
+
|
| 44 |
+
Figure 1: Noisy Node mechanics during training. Input positions are corrupted with noise $\sigma$ , and the training objective is the node-level difference between target positions and the noisy inputs.
|
| 45 |
+
|
| 46 |
+
Our method requires a noise correction target to prevent oversmoothing by enforcing diversity in the last layers of the GNN, which can be achieved with an auxiliary denoising autoencoder loss. For example, where the Error is defined with respect to graph-level predictions (e.g., predict the minimum energy value of some molecular system), a second output head can be added to the GNN architecture which requires denoising the inputs as targets. Alternatively, if the inputs and targets are in the same real domain as is the case for physical simulations we can adjust the target for the noise. Figure 1 demonstrates this Noisy Nodes set up. The auxiliary loss is weighted by a constant coefficient $\lambda \in \mathbb { R }$
|
| 47 |
+
|
| 48 |
+
In Figure 2 we illustrate the impact of Noisy Nodes on oversmoothing by plotting the Mean Absolute Distance (MAD) (Chen et al., 2020) of the residual updates of each layer of an MPNN trained on the QM9 (Ramakrishnan et al., 2014) dataset, and compare it to alternative methods DropEdge (Rong et al., 2019) and DropNode (Do et al., 2021). MAD is a measure of the diversity of graph node features, often used to quantify oversmoothing, the higher the number the more diverse the node features, the lower the number the less diverse. In this plot we can see that for Noisy Nodes the node updates remain diverse for all of the layers, whereas without Noisy Nodes diversity is lost. Further analysis of MAD across seeds and with sorted layers can be seen in Appendix Figures 7 and 6 for models applied to 3D point clouds.
|
| 49 |
+
|
| 50 |
+
The Graph Manifold Learning Perspective. By using an implicit mapping from corrupted data to clean data, the Noisy Nodes objective encourages the model to learn the manifold on which the clean data lies— we speculate that the GNN learns to go from low probability graphs to high probability graphs. In the autoencoder case the GNN learns the manifold of the input data. When node targets are provided, the GNN learns the manifold of the target data (e.g. the manifold of atoms at equilibrium). We speculate that such a manifold may include commonly repeated substructures that are useful for downstream prediction tasks. A similar motivation can be found for denoising in (Vincent et al., 2010; Song & Ermon, 2019).
|
| 51 |
+
|
| 52 |
+
The Energy Perspective for Molecular Property Prediction. Local, random distortions of the geometry of a molecule at a local energy minimum are almost certainly higher energy configurations. As such, a task that maps from a noised molecule to a local energy minimum is learning a mapping from high energy to low energy. Data such as QM9 contains molecules at local minima.
|
| 53 |
+
|
| 54 |
+
Some problems have input data that is already high energy, and targets that are at equilibrium. For these datasets we can generate new high energy states by adding noise to the inputs but keeping the equilibrium target the same, Figure 1 demonstrates this approach. To preserve translation invariance we use displacements between input and target $\Delta$ , the corrected target after noise is $\Delta - \sigma$ .
|
| 55 |
+
|
| 56 |
+
# 5 RELATED WORK
|
| 57 |
+
|
| 58 |
+
Oversmoothing. Recent work has aimed to understand why it is challenging to realise the benefits of training deeper GNNs (Wu et al., 2020). Since first being noted in ((Li et al., 2018)) oversmoothing has been studied extensively and regularisation techniques have been suggested to overcome it (Chen et al., 2019; Cai & Wang, 2020; Rong et al., 2019; Zhou et al., 2020b; Yang et al., 2020; Do et al., 2021; Zhao & Akoglu, 2020). A recent paper, (Li et al., 2021), finds, as in previous work, (Li et al., 2019; 2020), the optimal depth for some datasets they evaluate on to be far lower (5 for OGBN-Arxiv from the Open Graph Benchmark (Hu et al., 2020a), for example) than the 1000 layers possible.
|
| 59 |
+
|
| 60 |
+
Denoising & Noise Models. Training neural networks with noise has a long history (Sietsma & Dow, 1991; Bishop, 1995). Of particular relevance are Denoising Autoencoders (Vincent et al., 2008) in which an autoencoder is trained to map corrupted inputs $\tilde { \mathbf { x } }$ to uncorrupted inputs $\mathbf { X }$ . Denoising Autoencoders have found particular success as a form of pre-training for representation learning (Vincent et al., 2010). More recently, in research applying GNNs to simulation (Sanchez-Gonzalez et al., 2018; Sanchez-Gonzalez\* et al., 2020; Pfaff et al., 2020) Gaussian noise is added during training to input positions of a ground truth simulator to mimic the distribution of errors of the learned simulator. Pre-training methods (Devlin et al., 2019; You et al., 2020; Thakoor et al., 2021) are another similar approach; most similarly to our method Hu et al. (2020b) apply a reconstruction loss to graphs with masked nodes to generate graph embeddings for use in downstream tasks. FLAG (Kong et al., 2020) adds adversarial noise during training to input node features as a form of data augmentation for GNNs that demonstrates improved performance for many tasks. It does not add an additional auxiliary loss, which we find is essential for addressing oversmoothing. In other related GNN work, (Sato et al., 2021) use random input features to improve generalisation of graph neaural networks. Adding noise to help input node disambiguation has also been covered in (Dasoulas et al., 2019; Loukas, 2020; Vignac et al., 2020; Murphy et al., 2019), but there is no auxiliary loss.
|
| 61 |
+
|
| 62 |
+
Finally, we take inspiration from (Vincent et al., 2008; 2010; Vincent, 2011; Song & Ermon, 2019) which use the observation that noised data lies off the data manifold for representation learning and generative modelling.
|
| 63 |
+
|
| 64 |
+
Machine Learning for 3D Molecular Property Prediction. One application of GNNs is to speed up quantum chemistry calculations which operate on 3D positions of a molecule (Duvenaud et al., 2015; Gilmer et al., 2017; Schütt et al., 2017; Hu et al., 2021b). Common goals are the prediction of molecular properties (Ramakrishnan et al., 2014), forces (Chmiela et al., 2017), energies (Chanussot\* et al., 2020) and charges (Unke & Meuwly, 2019).
|
| 65 |
+
|
| 66 |
+
A common approach to embed physical symmetries is to design a network that predicts a rotation and translation invariant energy (Schütt et al., 2017; Klicpera et al., 2020a; Liu et al., 2021). The input features of such models include distances (Schütt et al., 2017), angles (Klicpera et al., 2020b;a) or torsions and higher order terms (Liu et al., 2021). An alternative approach to embedding symmetries is to design a rotation equivariant neural network that use equivariant representations (Thomas et al., 2018; Köhler et al., 2019; Kondor et al., 2018; Fuchs et al., 2020; Batzner et al., 2021; Anderson et al., 2019; Satorras et al., 2021).
|
| 67 |
+
|
| 68 |
+
Machine Learning for Bond and Atom Molecular Graphs. Predicting properties from molecular graphs without 3D points, such as graphs of bonds and atoms, is studied separately and often used to benchmark generic graph property prediction models such as GCNs (Hu et al., 2020a) or GATs (Velickovi ˇ c et al., 2018). Models developed for 3D molecular property prediction cannot be applied ´ to bond and atom graphs. Common datasets that contain such data are OGBG-MOLPCBA and OGBG-MOLHIV.
|
| 69 |
+
|
| 70 |
+
# 6 3D MOLECULAR PROPERTY PREDICTION EXPERIMENTS AND RESULTS
|
| 71 |
+
|
| 72 |
+
In this section we evaluate how a popular, simple model, the GNS (Sanchez-Gonzalez\* et al., 2020) performs on 3D molecular prediction tasks when combined with Noisy Nodes. The GNS was originally developed for particle fluid simulations, but has recently been adapted for molecular property prediction (Hu et al., 2021b). We find that Without Noisy Nodes the GNS architecture is not competitive, but by using Noisy Nodes we see improved performance comparable to the use of specialised architectures.
|
| 73 |
+
|
| 74 |
+
We made minor changes to the GNS architecture. We featurise the distance input features using radial basis functions. We group layer weights, similar to grouped layers used in Jumper et al. (2021) for reduced parameter counts; for a group size of $n$ the first $n$ layer weights are repeated, i.e. the first layer with a group size of 10 has the same weights as the $1 1 ^ { t h }$ , $2 1 ^ { s t }$ , $3 1 ^ { s t }$ layers and so on. $n$ contiguous blocks of layers are considered a single group. Finally we find that decoding the intermediate latents and adding a loss after each group aids training stability. The decoder is shared across groups.
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
Figure 3: Validation curves, OC20 IS2RE ID. A) Without any node targets our model has poor performance and realises no benefit from depth. B) After adding a position node loss, performance improves as depth increases. C) As we add Noisy Nodes and parameters the model achieves SOTA, even with 3 layers, and stops overfitting. D) Adding Noisy Nodes allows a model with even fully shared weights to achieve SOTA.
|
| 78 |
+
|
| 79 |
+
We tested this architecture on three challenging molecular property prediction benchmarks: OC20 (Chanussot\* et al., 2020) IS2RS & IS2RE, and QM9 (Ramakrishnan et al., 2014). These benchmarks are detailed below, but as general distinctions, OC20 tasks use graphs $2 \mathrm { - } 2 0 \mathrm { x }$ larger than QM9. While QM9 always requires graph-level prediction, one of OC20’s two tasks (IS2RS) requires node-level predictions while the other (IS2RE) requires graph-level predictions. All training details may be found in the Appendix.
|
| 80 |
+
|
| 81 |
+
# 6.1 OPEN CATALYST 2020
|
| 82 |
+
|
| 83 |
+
Dataset. The OC20 dataset (Chanussot\* et al., 2020) (CC Attribution 4.0) describes the interaction of a small molecule (the adsorbate) and a large slab (the catalyst), with total systems consisting of 20-200 atoms simulated until equilibrium is reached.
|
| 84 |
+
|
| 85 |
+
We focus on two tasks; the Initial Structure to Resulting Energy (IS2RE) task which takes the initial structure of the simulation and predicts the final energy, and the Initial Structure to Resulting Structure (IS2RS) which takes the initial structure and predicts the relaxed structure. Note that we train the more common “direct” prediction task that map directly from initial positions to target in a single forward pass, and compare against other models trained for direct prediction.
|
| 86 |
+
|
| 87 |
+
Models are evaluated on 4 held out test sets. Four canonical validation datasets are also provided. Test sets are evaluated on a remote server hosted by the dataset authors with a very limited number of submissions per team.
|
| 88 |
+
|
| 89 |
+
Noisy Nodes in this case consists of a random jump between the initial position and relaxed position. During training we first sample uniformly from a point in the relaxation trajectory or interpolate uniformly between the initial and final positions $( v _ { i } - \tilde { v } _ { i } ) \gamma , \gamma \sim \mathrm { U } ( 0 , 1 )$ , and then add I.I.D Gaussian noise with mean zero and $\sigma = 0 . 3$ . The Noisy Node target is the relaxed structure.
|
| 90 |
+
|
| 91 |
+
Table 1: OC20 ISRE Validation, eV MAE, ↓. “GNS-Shared” indicates shared weights. “GNS-10” indicates a group size of 10.
|
| 92 |
+
|
| 93 |
+
<table><tr><td>Model</td><td>Layers</td><td>OOD Both</td><td>OOD Adsorbate</td><td>OOD Catalyst</td><td>ID</td></tr><tr><td>GNS</td><td>50</td><td>0.59 ±0.01</td><td>0.65 ±0.01</td><td>0.55 ±0.00</td><td>0.54 ±0.00</td></tr><tr><td>GNS-Shared + Noisy Nodes</td><td>50</td><td>0.49 ±0.00</td><td>0.54 ±0.00</td><td>0.51 ±0.01</td><td>0.51 ±0.01</td></tr><tr><td>GNS + Noisy Nodes</td><td>50</td><td>0.48 ±0.00</td><td>0.53 ±0.00</td><td>0.49 ±0.01</td><td>0.48 ±0.00</td></tr><tr><td>GNS-10 + Noisy Nodes</td><td>100</td><td>0.46±0.00</td><td>0.51 ±0.00</td><td>0.48 ±0.00</td><td>0.47 ±0.00</td></tr></table>
|
| 94 |
+
|
| 95 |
+
Table 2: Results OC20 IS2RE Test
|
| 96 |
+
|
| 97 |
+
<table><tr><td colspan="6">eV MAE↓</td></tr><tr><td></td><td>SchNet</td><td>DimeNet++</td><td>SpinConv</td><td>SphereNet</td><td>GNS + Noisy Nodes</td></tr><tr><td>OOD Both</td><td>0.704</td><td>0.661</td><td>0.674</td><td>0.638</td><td>0.465 (-24.0%)</td></tr><tr><td>OOD Adsorbate</td><td>0.734</td><td>0.725</td><td>0.723</td><td>0.703</td><td>0.565 (-22.8%)</td></tr><tr><td>OOD Catalyst</td><td>0.662</td><td>0.576</td><td>0.569</td><td>0.571</td><td>0.437 (-17.2%)</td></tr><tr><td>ID</td><td>0.639</td><td>0.562</td><td>0.558</td><td>0.563</td><td>0.422 (-18.8%)</td></tr><tr><td colspan="6">Average Energy within Threshold (AEwT) ↑</td></tr><tr><td></td><td>SchNet</td><td>DimeNet++</td><td>SpinConv</td><td>SphereNet</td><td>GNS + Noisy Nodes</td></tr><tr><td>OOD Both</td><td>0.0221</td><td>0.0241</td><td>0.0233</td><td>0.0241</td><td>0.047 (+95.8%)</td></tr><tr><td>OOD Adsorbate</td><td>0.0233</td><td>0.0207</td><td>0.026</td><td>0.0229</td><td>0.035 (+89.5%)</td></tr><tr><td>OOD Catalyst</td><td>0.0294</td><td>0.0410</td><td>0.0382</td><td>0.0409</td><td>0.080 (+95.1%)</td></tr><tr><td>ID</td><td>0.0296</td><td>0.0425</td><td>0.0408</td><td>0.0447</td><td>0.091 (+102.0%)</td></tr></table>
|
| 98 |
+
|
| 99 |
+
We first convert to fractional coordinates (i.e. use the periodic unit cell as the basis) which render the predictions of our model invariant to rotations, and append the following rotation and translation invariant vector $( \alpha \beta ^ { T } , \beta \gamma ^ { T } , \alpha \gamma ^ { T } , | \alpha | , | \beta | , | \gamma | ) \in \mathbb { R } ^ { 6 }$ to the edge features where $\alpha , \beta , \gamma$ are vectors of the unit cell. This additional vector provides rotation invariant angular and extent information to the GNN.
|
| 100 |
+
|
| 101 |
+
IS2RE Results. In Figure 3 we show how using Noisy Nodes allows the GNS to achieve state of the art performance. Figure $_ { 3 \mathrm { ~ A ~ } }$ shows that without any auxiliary node target, an IS2RE GNS achieves poor performance even with increased depth. The fact that increased depth does not result in improvement supports the hypothesis that GNS suffers from oversmoothing. As we add a node level position target in B) we see better performance, and improvement as depth increases, validating our hypothesis that node level targets are key to addressing oversmoothing. In C) we add noisy nodes and parameters, and see that the increased diversity of the node level predictions leads to very significant improvements and SOTA, even for a shallow 3 layer network. D) demonstrates this effect is not just due to increased parameters - SOTA can still be achieve with shared layer weights .
|
| 102 |
+
|
| 103 |
+
In Table 1 we conduct an ablation on our hyperparameters, and again demonstrate the improved performance of using Noisy Nodes. Results were averaged over 3 seeds and standard errors on the best obtained checkpoint show little sensitivity to initialisation. All results in the table are reported using sampling states from trajectories. We conducted an ablation on ID comparing sampling from a relaxation trajectory and interpolating between initial & final positions which found that interpolation improved our score from 0.47 to 0.45.
|
| 104 |
+
|
| 105 |
+
Our best hyperparameter setting was 100 layers which achieved a $9 5 . 6 \%$ relative performance improvement against SOTA results (Table 2) on the AEwT benchmark. Due to limited permitted test submissions, results presented here were from one test upload of our best performing validation seed.
|
| 106 |
+
|
| 107 |
+
IS2RS Results. In Table 4 we see that GNS $^ +$ Noisy Nodes is significantly better than the only other reported IS2RS direct result, ForceNet, itself a GNS variant.
|
| 108 |
+
|
| 109 |
+
Table 3: OC20 IS2RS Validation, ADwT, ↑
|
| 110 |
+
|
| 111 |
+
<table><tr><td>Model</td><td>Layers</td><td>OOD Both</td><td>OOD Adsorbate</td><td>OOD Catalyst</td><td>ID</td></tr><tr><td>GNS</td><td>50</td><td>43.0%±0.0</td><td>38.0%±0.0</td><td>37.5% 0.0</td><td>40.0%±0.0</td></tr><tr><td>GNS + Noisy Nodes</td><td>50</td><td>50.1%±0.0</td><td>44.3%±0.0</td><td>44.1%±0.0</td><td>46.1% ±0.0</td></tr><tr><td>GNS-10 + Noisy Nodes</td><td>50</td><td>52.0%±0.0</td><td>46.2%±0.0</td><td>46.1% ±0.0</td><td>48.3% ±0.0</td></tr><tr><td>GNS-10 + Noisy Nodes + Pos only</td><td>100</td><td>54.3%±0.0</td><td>48.3%±0.0</td><td>48.2% ±0.0</td><td>50.0% ±0.0</td></tr></table>
|
| 112 |
+
|
| 113 |
+
Table 4: OC20 IS2RS Test, ADwT, ↑
|
| 114 |
+
|
| 115 |
+
<table><tr><td>Model</td><td>OOD Both</td><td>OOD Adsorbate</td><td>OOD Catalyst</td><td>ID</td></tr><tr><td>ForceNet</td><td>46.9%</td><td>37.7%</td><td>43.7%</td><td>44.9%</td></tr><tr><td>GNS + Noisy Nodes</td><td> 52.7%</td><td>43.9%</td><td>48.4%</td><td> 50.9%</td></tr><tr><td>Relative Improvement</td><td>+12.4%</td><td>+16.4%</td><td>+10.7%</td><td>+13.3%</td></tr></table>
|
| 116 |
+
|
| 117 |
+
# 6.2 QM9
|
| 118 |
+
|
| 119 |
+
Dataset. The QM9 benchmark (Ramakrishnan et al., 2014) contains $1 3 4 \mathrm { k }$ molecules in equilibrium with up to 9 heavy C, O, N and F atoms, targeting 12 associated chemical properties (License: CCBY 4.0). We use 114k molecules for training, 10k for validation and 10k for test. All results are on the test set. We subtract a fixed per atom energy from the target values computed from linear regression to reduce variance. We perform training in $\mathbf { e V }$ units for energetic targets, and evaluate using MAE. We summarise the results across the targets using mean standardised MAE (std. MAE) in which MAEs are normalised by their standard deviation, and mean standardised logMAE. Std. MAE is dominated by targets with high relative error such as $\Delta \epsilon$ , whereas logMAE is sensitive to outliers such as $\left. R ^ { 2 } \right.$ . As is standard for this dataset, a model is trained separately for each target.
|
| 120 |
+
|
| 121 |
+
For this dataset we add I.I.D Gaussian noise with mean zero and $\sigma = 0 . 0 2$ to the input atom positions.
|
| 122 |
+
A denoising autoencoder loss is used.
|
| 123 |
+
|
| 124 |
+
Results In Table 6 we can see that adding Noisy Nodes significantly improves results by $2 3 . 1 \%$ relative for GNS, making it competitive with specialised architectures. To understand the effect of adding a denoising loss, we tried just adding noise and found no where near the same improvement (Table 6).
|
| 125 |
+
|
| 126 |
+
A GNS- $1 0 +$ Noisy Nodes with 30 layers achieves top results on 3 of the 12 targets and comparable performance on the remainder (Table 6). On the std. MAE aggregate metric $\mathrm { G N S + }$ Noisy Nodes performs better than all other reported results, showing that Noisy Nodes can make even a generic model competitive with models hand-crafted for molecular property prediction. The same trend is repeated for an rotation invariant version of this network that uses the principle axes of inertia ordered by eigenvalue as the co-ordinate frame (Table 5).
|
| 127 |
+
|
| 128 |
+
$\left. R ^ { 2 } \right.$ , the electronic spatial extent, is an outlier for GNS + Noisy Nodes. Interestingly, we found that without noise GNS- $^ { 1 0 + }$ Noisy Nodes achieves 0.33 for this target. We speculate that this target is particularly sensitive to noise, and the best noise value for this target would be significantly lower than for the dataset as a whole.
|
| 129 |
+
|
| 130 |
+
Table 5: QM9, Impact of Noisy Nodes on GNS architecture.
|
| 131 |
+
|
| 132 |
+
<table><tr><td></td><td>Layers</td><td>std. MAE</td><td>% Change</td><td>logMAE</td></tr><tr><td>GNS</td><td>10</td><td>1.17</td><td>=</td><td>-5.39</td></tr><tr><td>GNS + Noise But No Node Target</td><td>10</td><td>1.16</td><td>-0.9%</td><td>-5.32</td></tr><tr><td>GNS + Noisy Nodes</td><td>10</td><td>0.90</td><td>-23.1%</td><td>-5.58</td></tr><tr><td>GNS-10 + Noisy Nodes</td><td>20</td><td>0.89</td><td>-23.9%</td><td>-5.59</td></tr><tr><td>GNS-1O + Noisy Nodes + Invariance</td><td>30</td><td>0.92</td><td>-21.4%</td><td>-5.57</td></tr><tr><td>GNS-10 + Noisy Nodes</td><td>30</td><td>0.88</td><td>-24.8%</td><td>-5.60</td></tr></table>
|
| 133 |
+
|
| 134 |
+
Table 6: QM9, Test MAE, Mean & Standard Deviation of 3 Seeds Reported.
|
| 135 |
+
|
| 136 |
+
<table><tr><td>Target</td><td>Unit</td><td>SchNet</td><td>E(n)GNN</td><td>DimeNet++</td><td>SphereNet</td><td>PaiNN</td><td>GNS + Noisy Nodes</td></tr><tr><td>μ</td><td>D</td><td>0.033</td><td>0.029</td><td>0.030</td><td>0.027</td><td>0.012</td><td>0.025 ±0.01</td></tr><tr><td>α</td><td>a03</td><td>0.235</td><td>0.071</td><td>0.043</td><td>0.047</td><td>0.045</td><td>0.052 ±0.00</td></tr><tr><td>EHOMO</td><td>meV</td><td>41</td><td>29.0</td><td>24.6</td><td>23.6</td><td>27.6</td><td>20.4 ±0.2</td></tr><tr><td>ELUMO</td><td>meV</td><td>34</td><td>25.0</td><td>19.5</td><td>18.9</td><td>20.4</td><td>18.6 ±0.4</td></tr><tr><td>△</td><td>meV</td><td>63</td><td>48.0</td><td>32.6</td><td>32.3</td><td>45.7</td><td>28.6 ±0.1</td></tr><tr><td>(R²></td><td>a02</td><td>0.07</td><td>0.11</td><td>0.33</td><td>0.29</td><td>0.07</td><td>0.70 ±0.01</td></tr><tr><td>ZPVE</td><td>meV</td><td>1.7</td><td>1.55</td><td>1.21</td><td>1.12</td><td>1.28</td><td>1.16 ±0.01</td></tr><tr><td>Uo</td><td>meV</td><td>14.00</td><td>11.00</td><td>6.32</td><td>6.26</td><td>5.85</td><td>7.30 ±0.12</td></tr><tr><td>U</td><td>meV</td><td>19.00</td><td>12.00</td><td>6.28</td><td>7.33</td><td>5.83</td><td>7.57 ±0.03</td></tr><tr><td>H</td><td>meV</td><td>14.00</td><td>12.00</td><td>6.53</td><td>6.40</td><td>5.98</td><td>7.43±0.06</td></tr><tr><td>G</td><td>meV cal</td><td>14.00</td><td>12.00</td><td>7.56</td><td>8.0</td><td>7.35</td><td>8.30 ±0.14</td></tr><tr><td>Cv</td><td>molK</td><td>0.033</td><td>0.031</td><td>0.023</td><td>0.022</td><td>0.024</td><td>0.025 ±0.00</td></tr><tr><td>std. MAE</td><td>%</td><td>1.76</td><td>1.22</td><td>0.98</td><td>0.94</td><td>1.00</td><td>0.88</td></tr><tr><td>logMAE</td><td></td><td>-5.17</td><td>-5.43</td><td>-5.67</td><td>-5.68</td><td>-5.85</td><td>-5.60</td></tr></table>
|
| 137 |
+
|
| 138 |
+
Table 7: OGBG-PCQM4M Results
|
| 139 |
+
|
| 140 |
+
<table><tr><td>Model</td><td>Number of Layers</td><td>Using I Noisy Nodes</td><td>MAE</td></tr><tr><td>MPNN + Virtual Node</td><td>16</td><td>Yes</td><td>0.1249 ± 0.0003</td></tr><tr><td>MPNN+Virtual Node</td><td>50</td><td>No</td><td>0.1236 ± 0.0001</td></tr><tr><td>Graphormer (Ying et al., 2021)</td><td>1</td><td>1</td><td>0.1234</td></tr><tr><td>MPNN + Virtual Node</td><td>50</td><td>Yes</td><td>0.1218 ± 0.0001</td></tr></table>
|
| 141 |
+
|
| 142 |
+
# 7 NON-SPATIAL TASKS
|
| 143 |
+
|
| 144 |
+
The previous experiments use the 3D geometries of atoms, and models that operate on 3D points. However, the recipe of adding a denoising auxiliary loss can be applied to other graphs with different types of features. In this section we apply Noisy Nodes to additional datasets with no 3D points, using different GNNs, and show analagous effects to the 3D case. Details of the hyperparameters, models and training details can be found in the appendix.
|
| 145 |
+
|
| 146 |
+
# 7.1 OGBG-PCQM4M
|
| 147 |
+
|
| 148 |
+
This dataset from the OGB benchmarks consists of molecular graphs which consist of bonds and atom types, and no 3D or 2D coordinates. To adapt Noisy Nodes to this setting, we randomly flip node and edge features at a rate of $5 \%$ and add a reconstruction loss. We evaluate Noisy Nodes using an MPNN $^ +$ Virtual Node (Gilmer et al., 2017). The test set is not currently available for this dataset.
|
| 149 |
+
|
| 150 |
+
In Table 7 we see that for this task Noisy Nodes enables a 50 layer MPNN to reach state of the art results. Before adding Noisy Nodes, adding capacity beyond 16 layers did not improve results.
|
| 151 |
+
|
| 152 |
+
# 7.2 OGBG-MOLPCBA
|
| 153 |
+
|
| 154 |
+
The OGBG-MOLPCBA dataset contains molecular graphs with no 3D points, with the goal of classifying 128 biological activities. On the OGBG-MOLPCBA dataset we again use an $\mathbf { M P N N + }$ Virtual Node and random flipping noise. In Figure 4 we see that adding Noisy Nodes improves the performance of the base model, accentuated for deeper networks. Our 16 layer MPNN improved from $2 7 . 6 \% \pm 0 . 0 0 4$ to $2 8 . 1 \% \pm 0 . 0 0 2$ Mean Average Precision (“Mean AP”). Figure 5 demonstrates how Noisy Nodes improves performance during training. Of the reported results, our MPNN is most similar to $\mathrm { G C N ^ { 1 } \Sigma + }$ Virtual Node and $\mathrm { G I N } +$ Virtual Node (Xu et al., 2018) which report results of $2 4 . 2 \% \pm 0 . 0 0 3$ and $2 7 . 0 3 \% \pm 0 . 0 0 3$ respectively. We evaluate alternative methods for oversmoothing, DropNode and DropEdge in Figure 2 and find that Noisy Nodes is more effective at address oversmoothing, although all 3 methods can be combined favourably (results in appendix).
|
| 155 |
+
|
| 156 |
+
|
| 157 |
+
Figure 4: Adding Noisy Nodes with random flipping of input categories improves the performance of MPNNs, and the effect is accentuated with depth.
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
Figure 5: Validation curve comparing with and without noisy nodes. Using Noisy Nodes leads to a consistent improvement.
|
| 161 |
+
|
| 162 |
+
# 7.3 OGBN-ARXIV
|
| 163 |
+
|
| 164 |
+
The above results use models with explicit edge updates, and are reported for graph prediction. To test the effectiveness with Noisy Nodes with GCNs, arguably the simplest and most popular GNN, we use OGBN-ARXIV, a citation network with the goal of predicting the arxiv category of each paper. Adding Noisy Nodes, with noise as input dropout of 0.1, to 4 layer GCN with residual connections improves from $7 2 . 3 9 \% \pm 0 . 0 0 2$ accuracy to $7 2 . 5 2 \% \pm 0 . 0 0 3$ accuracy. A baseline 4 layer GCN on this dataset reports $7 1 . 7 1 \% \pm 0 . 0 0 2$ . The SOTA for this dataset is $7 4 . 3 1 \%$ (Sun & Wu, 2020).
|
| 165 |
+
|
| 166 |
+
# 7.4 LIMITATIONS
|
| 167 |
+
|
| 168 |
+
We have not demonstrated the effectiveness of Noisy Nodes in small data regimes, which may be important for learning from experimental data. The representation learning perspective requires access to a local minimum configuration, which is not the case for all quantum modeling datasets. We have also not demonstrated the combination of Noisy Nodes with more sophisticated 3D molecular property prediction models such as DimeNet++(Klicpera et al., 2020a), such models may require an alternative reconstruction loss to position change, such as pairwise interatomic distances. We leave this to future work.
|
| 169 |
+
|
| 170 |
+
Noisy Nodes requires careful selection of the form of noise, and a balance between the auxiliary and primary losses. This can require hyper parameter tuning, and models can be sensitive to the choice of these parameters. Noisy Nodes has a particular effect for deep GNNs, but depth is not always an advantage. There are situations, for example molecular dynamics, which place a premium on very fast inference time. However even at 3 layers (a comparable depth to alternative architectures) the GNS architecture achieves state of the art validation OC20 IS2RE predictions (Figure 3). Finally, returns diminish as depth increases indicating depth is not the only answer (Table 1).
|
| 171 |
+
|
| 172 |
+
# 8 CONCLUSIONS
|
| 173 |
+
|
| 174 |
+
In this work we present Noisy Nodes, a novel regularisation technique for GNNs with particular focus on 3D molecular property prediction. Noisy nodes helps address common challenges around oversmoothed node representations, shows benefits for GNNs of all depths, but in particular improves performance for deeper GNNs. We demonstrate results on challenging 3D molecular property prediction tasks, and some generic GNN benchmark datasets. We believe these results demonstrate Noisy Nodes could be a useful building block for GNNs for molecular property prediction and beyond.
|
| 175 |
+
|
| 176 |
+
# 9 REPRODUCIBILITY STATEMENT
|
| 177 |
+
|
| 178 |
+
Code for reproducing OGB-PCQM4M results using Noisy Nodes is available on github, and was prepared as part of a leaderboard submission. https://github.com/deepmind/ deepmind-research/tree/master/ogb_lsc/pcq.
|
| 179 |
+
|
| 180 |
+
We provide detailed hyper parameter settings for all our experiments in the appendix, in addition to formulae for computing the encoder and decoder stages of the GNS.
|
| 181 |
+
|
| 182 |
+
# 10 ETHICS STATEMENT
|
| 183 |
+
|
| 184 |
+
Who may benefit from this work? Molecular property prediction with GNNs is a fast-growing area with applications across domains such as drug design, catalyst discovery, synthetic biology, and chemical engineering. Noisy Nodes could aid models applied to these domains. We also demonstrate on OC20 that our direct state prediction approach is nearly as accurate as learned relaxed approaches at a small fraction of the computational cost, which may support material design which requires many predictions.
|
| 185 |
+
|
| 186 |
+
Finally, Noisy Nodes could be adapted and applied to many areas in which GNNs are used—for example, knowledge base completion, physical simulation or traffic prediction.
|
| 187 |
+
|
| 188 |
+
Potential negative impact and reflection. Noisy Nodes sees improved performance from depth, but the training of very deep GNNs could contribute to global warming. Care should be taken when utilising depth, and we note that Noisy Nodes settings can be calibrated at shallow depth.
|
| 189 |
+
|
| 190 |
+
# REFERENCES
|
| 191 |
+
|
| 192 |
+
Brandon M. Anderson, T. Hy, and R. Kondor. Cormorant: Covariant molecular neural networks. In NeurIPS, 2019.
|
| 193 |
+
|
| 194 |
+
Igor Babuschkin, Kate Baumli, Alison Bell, Surya Bhupatiraju, Jake Bruce, Peter Buchlovsky, David Budden, Trevor Cai, Aidan Clark, Ivo Danihelka, Claudio Fantacci, Jonathan Godwin, Chris Jones, Tom Hennigan, Matteo Hessel, Steven Kapturowski, Thomas Keck, Iurii Kemaev, Michael King, Lena Martens, Vladimir Mikulik, Tamara Norman, John Quan, George Papamakarios, Roman Ring, Francisco Ruiz, Alvaro Sanchez, Rosalia Schneider, Eren Sezener, Stephen Spencer, Srivatsan Srinivasan, Wojciech Stokowiec, and Fabio Viola. The DeepMind JAX Ecosystem, 2020. URL http://github.com/deepmind.
|
| 195 |
+
|
| 196 |
+
V. Bapst, T. Keck, Agnieszka Grabska-Barwinska, C. Donner, E. D. Cubuk, S. Schoenholz, A. Obika, Alexander W. R. Nelson, T. Back, D. Hassabis, and P. Kohli. Unveiling the predictive power of static structure in glassy systems. Nature Physics, 16:448–454, 2020.
|
| 197 |
+
|
| 198 |
+
P. Battaglia, Jessica B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, A. Santoro, R. Faulkner, Çaglar Gülçehre, H. Song, A. J. Ballard, J. Gilmer, George E. Dahl, Ashish Vaswani, Kelsey R. Allen, Charlie Nash, Victoria Langston, Chris Dyer, N. Heess, Daan Wierstra, P. Kohli, M. Botvinick, Oriol Vinyals, Y. Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. ArXiv, abs/1806.01261, 2018.
|
| 199 |
+
|
| 200 |
+
Simon Batzner, T. Smidt, L. Sun, J. Mailoa, M. Kornbluth, N. Molinari, and B. Kozinsky. Se(3)- equivariant graph neural networks for data-efficient and accurate interatomic potentials. ArXiv, abs/2101.03164, 2021.
|
| 201 |
+
|
| 202 |
+
Charles M. Bishop. Training with noise is equivalent to tikhonov regularization. Neural Computation, 7:108–116, 1995.
|
| 203 |
+
|
| 204 |
+
James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URL http://github.com/google/jax.
|
| 205 |
+
|
| 206 |
+
Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.
|
| 207 |
+
|
| 208 |
+
T. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, J. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. Henighan, R. Child, A. Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, J. Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020.
|
| 209 |
+
|
| 210 |
+
Chen Cai and Yusu Wang. A note on over-smoothing for graph neural networks. CoRR, abs/2006.13318, 2020. URL https://arxiv.org/abs/2006.13318.
|
| 211 |
+
|
| 212 |
+
Lowik Chanussot\*, Abhishek Das\*, Siddharth Goyal\*, Thibaut Lavril\*, Muhammed Shuaibi\*, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, Aini Palizhati, Anuroop Sriram, Brandon Wood, Junwoong Yoon, Devi Parikh, C. Lawrence Zitnick, and Zachary Ulissi. Open catalyst 2020 (oc20) dataset and community challenges. ACS Catalysis, 0(0): 6059–6072, 2020. doi: 10.1021/acscatal.0c04525. URL https://doi.org/10.1021/ acscatal.0c04525.
|
| 213 |
+
|
| 214 |
+
Deli Chen, Yankai Lin, Wei Li, Peng Li, Jie Zhou, and Xu Sun. Measuring and relieving the oversmoothing problem for graph neural networks from the topological view. CoRR, abs/1909.03211, 2019. URL http://arxiv.org/abs/1909.03211.
|
| 215 |
+
|
| 216 |
+
Deli Chen, Yankai Lin, W. Li, Peng Li, J. Zhou, and Xu Sun. Measuring and relieving the oversmoothing problem for graph neural networks from the topological view. In AAAI, 2020.
|
| 217 |
+
|
| 218 |
+
Stefan Chmiela, A. Tkatchenko, H. E. Sauceda, I. Poltavsky, Kristof T. Schütt, and K. Müller. Machine learning of accurate energy-conserving molecular force fields. Science Advances, 3, 2017.
|
| 219 |
+
|
| 220 |
+
George Dasoulas, Ludovic Dos Santos, Kevin Scaman, and Aladin Virmaux. Coloring graph neural networks for node disambiguation. ArXiv, abs/1912.06058, 2019.
|
| 221 |
+
|
| 222 |
+
J. Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019.
|
| 223 |
+
|
| 224 |
+
Tien Huu Do, Duc Minh Nguyen, Giannis Bekoulis, Adrian Munteanu, and N. Deligiannis. Graph convolutional neural networks with node transition probability-based message passing and dropnode regularization. Expert Syst. Appl., 174:114711, 2021.
|
| 225 |
+
|
| 226 |
+
David Duvenaud, Dougal Maclaurin, Jorge Aguilera-Iparraguirre, Rafael Gómez-Bombarelli, Timothy Hirzel, Alán Aspuru-Guzik, and Ryan P. Adams. Convolutional networks on graphs for learning molecular fingerprints. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, NIPS’15, pp. 2224–2232, Cambridge, MA, USA, 2015. MIT Press.
|
| 227 |
+
|
| 228 |
+
F. Fuchs, Daniel E. Worrall, Volker Fischer, and M. Welling. Se(3)-transformers: 3d roto-translation equivariant attention networks. ArXiv, abs/2006.10503, 2020.
|
| 229 |
+
|
| 230 |
+
J. Gilmer, S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. ArXiv, abs/1704.01212, 2017.
|
| 231 |
+
|
| 232 |
+
Jonathan Godwin\*, Thomas Keck\*, Peter Battaglia, Victor Bapst, Thomas Kipf, Yujia Li, Kimberly Stachenfeld, Petar Velickovi ˇ c, and Alvaro Sanchez-Gonzalez. Jraph: A library for graph neural ´ networks in jax., 2020. URL http://github.com/deepmind/jraph.
|
| 233 |
+
|
| 234 |
+
Tom Hennigan, Trevor Cai, Tamara Norman, and Igor Babuschkin. Haiku: Sonnet for JAX, 2020. URL http://github.com/deepmind/dm-haiku.
|
| 235 |
+
|
| 236 |
+
Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. ArXiv, abs/2005.00687, 2020a.
|
| 237 |
+
|
| 238 |
+
Weihua Hu, Bowen Liu, Joseph Gomes, M. Zitnik, Percy Liang, V. Pande, and J. Leskovec. Strategies for pre-training graph neural networks. arXiv: Learning, 2020b.
|
| 239 |
+
|
| 240 |
+
Weihua Hu, Matthias Fey, Hongyu Ren, Maho Nakata, Yuxiao Dong, and Jure Leskovec. Ogb-lsc: A large-scale challenge for machine learning on graphs. arXiv preprint arXiv:2103.09430, 2021a.
|
| 241 |
+
|
| 242 |
+
Weihua Hu, Muhammed Shuaibi, Abhishek Das, Siddharth Goyal, Anuroop Sriram, J. Leskovec, Devi Parikh, and C. L. Zitnick. Forcenet: A graph neural network for large-scale quantum calculations. ArXiv, abs/2103.01436, 2021b.
|
| 243 |
+
|
| 244 |
+
John M. Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Zídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A A Kohl, Andy Ballard, Andrew Cowie, Bernardino RomeraParedes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David A. Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis. Highly accurate protein structure prediction with alphafold. Nature, 596:583 – 589, 2021.
|
| 245 |
+
|
| 246 |
+
Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015.
|
| 247 |
+
|
| 248 |
+
Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. CoRR, abs/1609.02907, 2016. URL http://arxiv.org/abs/1609.02907.
|
| 249 |
+
|
| 250 |
+
Johannes Klicpera, Shankari Giri, Johannes T. Margraf, and Stephan Günnemann. Fast and uncertainty-aware directional message passing for non-equilibrium molecules. CoRR, abs/2011.14115, 2020a. URL https://arxiv.org/abs/2011.14115.
|
| 251 |
+
|
| 252 |
+
Johannes Klicpera, Janek Groß, and Stephan Günnemann. Directional message passing for molecular graphs. ArXiv, abs/2003.03123, 2020b.
|
| 253 |
+
|
| 254 |
+
Risi Kondor, Hy Truong Son, Horace Pan, Brandon M. Anderson, and Shubhendu Trivedi. Covariant compositional networks for learning graphs. CoRR, abs/1801.02144, 2018. URL http:// arxiv.org/abs/1801.02144.
|
| 255 |
+
|
| 256 |
+
Kezhi Kong, Guohao Li, Mucong Ding, Zuxuan Wu, Chen Zhu, Bernard Ghanem, G. Taylor, and T. Goldstein. Flag: Adversarial data augmentation for graph neural networks. ArXiv, abs/2010.09891, 2020.
|
| 257 |
+
|
| 258 |
+
Jonas Köhler, Leon Klein, and Frank Noé. Equivariant flows: sampling configurations for multi-body systems with symmetric energies, 2019.
|
| 259 |
+
|
| 260 |
+
G. Li, M. Müller, Ali K. Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep as cnns? 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9266–9275, 2019.
|
| 261 |
+
|
| 262 |
+
Guohao Li, C. Xiong, Ali K. Thabet, and Bernard Ghanem. Deepergcn: All you need to train deeper gcns. ArXiv, abs/2006.07739, 2020.
|
| 263 |
+
|
| 264 |
+
Guohao Li, Matthias Müller, Bernard Ghanem, and Vladlen Koltun. Training graph neural networks with 1000 layers. CoRR, abs/2106.07476, 2021. URL https://arxiv.org/abs/2106. 07476.
|
| 265 |
+
|
| 266 |
+
Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.
|
| 267 |
+
|
| 268 |
+
Yi Liu, Limei Wang, Meng Liu, Xuan Zhang, Bora Oztekin, and Shuiwang Ji. Spherical message passing for 3d graph networks. arXiv preprint arXiv:2102.05013, 2021.
|
| 269 |
+
|
| 270 |
+
Andreas Loukas. How hard is to distinguish graphs with graph neural networks? arXiv: Learning, 2020.
|
| 271 |
+
|
| 272 |
+
Ryan L. Murphy, Balasubramaniam Srinivasan, Vinayak A. Rao, and Bruno Ribeiro. Relational pooling for graph representations. In ICML, 2019.
|
| 273 |
+
T. Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and P. Battaglia. Learning mesh-based simulation with graph networks. ArXiv, abs/2010.03409, 2020.
|
| 274 |
+
R. Ramakrishnan, Pavlo O. Dral, M. Rupp, and O. A. von Lilienfeld. Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data, 1, 2014.
|
| 275 |
+
Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. The truly deep graph convolutional networks for node classification. CoRR, abs/1907.10903, 2019. URL http://arxiv.org/ abs/1907.10903.
|
| 276 |
+
Alvaro Sanchez-Gonzalez, N. Heess, Jost Tobias Springenberg, J. Merel, Martin A. Riedmiller, R. Hadsell, and P. Battaglia. Graph networks as learnable physics engines for inference and control. ArXiv, abs/1806.01242, 2018.
|
| 277 |
+
Alvaro Sanchez-Gonzalez\*, Jonathan Godwin\*, Tobias Pfaff\*, Rex Ying\*, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In Hal Daumé III and Aarti Singh (eds.), Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pp. 8459–8468. PMLR, 13–18 Jul 2020. URL http://proceedings.mlr.press/v119/sanchez-gonzalez20a.html.
|
| 278 |
+
R. Sato, Makoto Yamada, and Hisashi Kashima. Random features strengthen graph neural networks. In SDM, 2021.
|
| 279 |
+
Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks, 2021.
|
| 280 |
+
Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2009. doi: 10.1109/TNN.2008.2005605.
|
| 281 |
+
Kristof Schütt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, A. Tkatchenko, and K. Müller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. In NIPS, 2017.
|
| 282 |
+
Jonathan Shlomi, Peter Battaglia, and Jean-Roch Vlimant. Graph neural networks in particle physics. Machine Learning: Science and Technology, 2(2):021001, Jan 2021. ISSN 2632-2153. doi: 10.1088/2632-2153/abbf9a. URL http://dx.doi.org/10.1088/2632-2153/abbf9a.
|
| 283 |
+
J. Sietsma and Robert J. F. Dow. Creating artificial neural networks that generalize. Neural Networks, 4:67–79, 1991.
|
| 284 |
+
Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. ArXiv, abs/1907.05600, 2019.
|
| 285 |
+
Chuxiong Sun and Guoshi Wu. Adaptive graph diffusion networks with hop-wise attention. ArXiv, abs/2012.15024, 2020.
|
| 286 |
+
Shantanu Thakoor, C. Tallec, M. G. Azar, R. Munos, Petar Velivckovi’c, and Michal Valko. Bootstrapped representation learning on graphs. ArXiv, abs/2102.06514, 2021.
|
| 287 |
+
Nathaniel Thomas, Tess Smidt, Steven M. Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds. CoRR, abs/1802.08219, 2018. URL http://arxiv.org/abs/1802.08219.
|
| 288 |
+
Oliver T. Unke and Markus Meuwly. Physnet: A neural network for predicting energies, forces, dipole moments, and partial charges. Journal of Chemical Theory and Computation, 15(6):3678–3693, May 2019. ISSN 1549-9626. doi: 10.1021/acs.jctc.9b00181. URL http://dx.doi.org/10. 1021/acs.jctc.9b00181.
|
| 289 |
+
Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. ArXiv, abs/1706.03762, 2017.
|
| 290 |
+
Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua ´ Bengio. Graph attention networks, 2018.
|
| 291 |
+
Cl’ement Vignac, Andreas Loukas, and Pascal Frossard. Building powerful and equivariant graph neural networks with structural message-passing. arXiv: Learning, 2020.
|
| 292 |
+
Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23:1661–1674, 2011.
|
| 293 |
+
Pascal Vincent, H. Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML ’08, 2008.
|
| 294 |
+
Pascal Vincent, H. Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. J. Mach. Learn. Res., 11:3371–3408, 2010.
|
| 295 |
+
Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 2020.
|
| 296 |
+
Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? CoRR, abs/1810.00826, 2018. URL http://arxiv.org/abs/1810.00826.
|
| 297 |
+
Chaoqi Yang, Ruijie Wang, Shuochao Yao, Shengzhong Liu, and Tarek Abdelzaher. Revisiting" over-smoothing" in deep gcns. arXiv preprint arXiv:2003.13663, 2020.
|
| 298 |
+
Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform bad for graph representation? ArXiv, abs/2106.05234, 2021.
|
| 299 |
+
Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. ArXiv, abs/2010.13902, 2020.
|
| 300 |
+
L. Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. ArXiv, abs/1909.12223, 2020.
|
| 301 |
+
Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. Graph neural networks: A review of methods and applications. AI Open, 1:57–81, 2020a.
|
| 302 |
+
Kuangqi Zhou, Yanfei Dong, Wee Sun Lee, Bryan Hooi, Huan Xu, and Jiashi Feng. Effective training strategies for deep graph neural networks. CoRR, abs/2006.07107, 2020b. URL https: //arxiv.org/abs/2006.07107.
|
| 303 |
+
|
| 304 |
+
# A APPENDIX
|
| 305 |
+
|
| 306 |
+
The following sections include details on training setup, hyper-parameters, input processing, as well as additional experimental results.
|
| 307 |
+
|
| 308 |
+
# A.1 ADDITIONAL METRICS FOR OPEN CATALYST IS2RS TEST SET
|
| 309 |
+
|
| 310 |
+
Relaxation approaches to IS2RS minimise forces with respect to positions, with the expectation that forces at the minimum are close to zero. One metric of such a model’s success is to evaluate the forces at the converged structure using ground truth Density Functional Theory calculations and see how close they are to zero. Two metrics are provided by OC20 (Chanussot\* et al., 2020) on the IS2RS test set: Force below Threshold (FbT), which is the percentage of structures that have forces below 0.05 eV/Angstrom, and Average Force below Threshold (AFbT) which is FbT calculated at multiple thresholds.
|
| 311 |
+
|
| 312 |
+
The OC20 project computes test DFT calculations on the evaluation server and presents a summary result for all IS2RS position predictions. Such calculations take 10-12 hours and they are not available for the validation set. Thus, we are not able to analyse the results in Tables 8 and 9 in any further detail. Before application to catalyst screening further work may be needed for direct approaches to ensure forces do not explode from atoms being too close together.
|
| 313 |
+
|
| 314 |
+
Table 8: OC20 IS2RS Test, Average Force below Threshold $\%$ , ↑
|
| 315 |
+
|
| 316 |
+
<table><tr><td>Model</td><td>Method</td><td>OOD Both</td><td>OOD Adsorbate</td><td>OOD Catalyst</td><td>ID</td></tr><tr><td>Noisy Nodes</td><td>Direct</td><td>0.09%</td><td>0.00%</td><td>0.29%</td><td>0.54%</td></tr></table>
|
| 317 |
+
|
| 318 |
+
Table 9: OC20 IS2RS Test, Force below Threshold %, ↑
|
| 319 |
+
|
| 320 |
+
<table><tr><td>Model</td><td>Method</td><td>OOD Both</td><td>OOD Adsorbate</td><td>OOD Catalyst</td><td>ID</td></tr><tr><td>Noisy Nodes</td><td>Direct</td><td>0.0%</td><td>0.0%</td><td>0.0%</td><td>0.0%</td></tr></table>
|
| 321 |
+
|
| 322 |
+
A.2 MORE DETAILS ON GNS ADAPTATIONS FOR MOLECULAR PROPERTY PREDICTION.
|
| 323 |
+
|
| 324 |
+
# Encoder.
|
| 325 |
+
|
| 326 |
+
The node features are a learned embedding lookup of the atom type, and in the case of OC20 two additional binary features representing whether the atom is part of the adsorbate or catalyst and whether the atom remains fixed during the quantum chemistry simulation.
|
| 327 |
+
|
| 328 |
+
The edge features, $e _ { k }$ are the distances $| d |$ featurised using $c$ Radial Bessel basis functions, $\tilde { e } _ { R B F , c } =$ ${ \sqrt { \frac { 2 } { R } } } { \frac { \sin ( { \frac { c \pi } { R } } d ) } { d } }$ , and the edge vector displacements, $d$ , normalised by the edge distance:
|
| 329 |
+
|
| 330 |
+
$$
|
| 331 |
+
e _ { k } = { \mathrm { C o n c a t } } ( { \tilde { e } } _ { R B F , 1 } ( | d | ) , . . . , { \tilde { e } } _ { R B F , c } ( | d | ) , \frac { d } { | d | } )
|
| 332 |
+
$$
|
| 333 |
+
|
| 334 |
+
Our conversion to fractional coordinates only applied to the vector quantities, i.e. $\frac { d } { | d | }$
|
| 335 |
+
|
| 336 |
+
# Decoder
|
| 337 |
+
|
| 338 |
+
The decoder consists of two parts, a graph-level decoder which predicts a single output for the input graph, and a node-level decoder which predicts individual outputs for each node. The graph-level decoder implements the following equation:
|
| 339 |
+
|
| 340 |
+
$$
|
| 341 |
+
y = W ^ { \mathrm { P r o c } } \sum _ { i = 1 } ^ { | V | } \mathrm { M L P } _ { \mathrm { P r o c } } ( a _ { i } ^ { \mathrm { P r o c } } ) + b ^ { \mathrm { P r o c } } + W ^ { \mathrm { E n c } } \sum _ { i = 1 } ^ { | V | } \mathrm { M L P } _ { \mathrm { E n c } } ( a _ { i } ^ { \mathrm { E n c } } ) + b ^ { \mathrm { E n c } }
|
| 342 |
+
$$
|
| 343 |
+
|
| 344 |
+
Where $a _ { i } ^ { \mathrm { P r o c } }$ are node latents from the Processor, $a _ { i } ^ { \mathrm { E n c } }$ are node latents from the Encoder, $W ^ { \mathrm { E n c } }$ and $W ^ { \mathrm { P r o c } }$ are linear layers, $b ^ { \mathrm { E n c } }$ and $b ^ { \mathrm { P r o c } }$ are biases, and $| V |$ is the number of nodes. The node-level decoder is simply an MLP applied to each $a _ { i } ^ { \mathrm { P r o c } }$ which predicts $a _ { i } ^ { \Delta }$ .
|
| 345 |
+
|
| 346 |
+
# A.3 MORE DETAILS ON MPNN FOR OGBG-PCQM4M AND OGBG-MOLPCBA
|
| 347 |
+
|
| 348 |
+
Our MPNN follows the blueprint of Gilmer et al. (2017). We use $\vec { h } _ { v } ^ { ( t ) }$ to denote the latent vector of node $v$ at message passing step $t$ , and $\vec { m } _ { u v } ^ { ( t ) }$ to be the computed message vector for the edge between nodes $u$ and $v$ at message passing step $t$ . We define the update functions as:
|
| 349 |
+
|
| 350 |
+
$$
|
| 351 |
+
\begin{array} { l } { { \displaystyle { \vec { m } } _ { u v } ^ { ( t + 1 ) } = \psi _ { t + 1 } \left( { \vec { h } } _ { u } ^ { ( t ) } , { \vec { h } } _ { v } ^ { ( t ) } , { \vec { m } } _ { u v } ^ { ( t ) } + { \vec { m } } _ { u v } ^ { ( t - 1 ) } \right) } } \\ { { \displaystyle { \vec { h } } _ { u } ^ { ( t + 1 ) } = \phi _ { t + 1 } \left( \vec { h } _ { u } ^ { ( t ) } , \sum _ { u \in \mathcal { N } _ { v } } { \vec { m } } _ { v u } ^ { ( t + 1 ) } , \sum _ { v \in \mathcal { N } _ { u } } { \vec { m } } _ { u v } ^ { ( t + 1 ) } \right) + { \vec { h } } _ { u } ^ { t } } } \end{array}
|
| 352 |
+
$$
|
| 353 |
+
|
| 354 |
+
Where the message function $\psi _ { t + 1 }$ and the update function $\phi _ { t + 1 }$ are MLPs. We use a “Virtual Node” which is connected to all other nodes to enable long range communication. Out readout function is an MLP. No spatial features are used.
|
| 355 |
+
|
| 356 |
+
|
| 357 |
+
Figure 6: GNS Unsorted MAD per Layer Averaged Over 3 Random Seeds. Evidence of oversmoothing is clear. Model trained on QM9.
|
| 358 |
+
|
| 359 |
+
|
| 360 |
+
Figure 7: GNS Sorted MAD per Layer Averaged Over 3 Random Seeds. The trend is clearer when the MAD values have been sorted. Model trained on QM9.
|
| 361 |
+
|
| 362 |
+
# A.4 EXPERIMENT SETUP FOR 3D MOLECULAR MODELING
|
| 363 |
+
|
| 364 |
+
Open Catalyst. All training experiments were ran on a cluster of TPU devices. For the Open Catalyst experiments, each individual run (i.e. a single random seed) utilised 8 TPU devices on 2 hosts (4 per host) for training, and 4 V100 GPU devices for evaluation (1 per dataset).
|
| 365 |
+
|
| 366 |
+
Each Open Catalyst experiment was ran until convergence for up to 200 hours. Our best result, the large 100 layer model requires 7 days of training using the above setting. Each configuration was run at least 3 times in this hardware configuration, including all ablation settings.
|
| 367 |
+
|
| 368 |
+
We further note that making effective use of our regulariser requires sweeping noise values. These sweeps are dataset dependent and can be carried out using few message passing steps.
|
| 369 |
+
|
| 370 |
+
QM9. Experiments were also run on TPU devices. Each seed was run using 8 TPU devices on a single host for training, and 2 V100 GPU devices for evaluation. QM9 targets were trained between 12-24 hours per experiment.
|
| 371 |
+
|
| 372 |
+
Following Klicpera et al. (2020b) we define std. MAE as :
|
| 373 |
+
|
| 374 |
+
$$
|
| 375 |
+
\mathrm { s t d . ~ } \mathrm { M A E } = \frac { 1 } { M } \sum _ { m = 1 } ^ { M } \left( \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \frac { | f _ { \theta } ^ { ( m ) } ( X _ { i } , z _ { i } ) - \hat { t } _ { i } ^ { ( m ) } | } { \sigma _ { m } } \right)
|
| 376 |
+
$$
|
| 377 |
+
|
| 378 |
+
and logMAE as:
|
| 379 |
+
|
| 380 |
+
$$
|
| 381 |
+
\log \mathrm { M A E } = \frac { 1 } { M } \sum _ { m = 1 } ^ { M } \log \left( \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \frac { | f _ { \theta } ^ { ( m ) } ( X _ { i } , z _ { i } ) - \hat { t } _ { i } ^ { ( m ) } | } { \sigma _ { m } } \right)
|
| 382 |
+
$$
|
| 383 |
+
|
| 384 |
+
with target index $m$ , number of targets $M = 1 2$ , dataset size $N$ , ground truth values $\hat { t } ^ { ( m ) }$ , model $f _ { \theta } ^ { ( m ) }$ , inputs $X _ { i }$ and $z _ { i }$ , and standard deviation $\sigma _ { m }$ of $\hat { t } ^ { ( m ) }$ .
|
| 385 |
+
|
| 386 |
+
# A.5 OVER SMOOTHING ANALYSIS FOR GNS
|
| 387 |
+
|
| 388 |
+
In addition to Figure 2, we repeat the analysis with a mean MAD over 3 seeds 7. Furthermore we remove the sorting layer by MAD value and find the trend holds.
|
| 389 |
+
|
| 390 |
+
# A.6 NOISE ABLATIONS FOR OGBG-MOLPCBA
|
| 391 |
+
|
| 392 |
+
We conduct a noise ablation on the random flipping noise for OGBG-MOLPCBA with an 8 layer MPNN $^ +$ Virtual Node, and find that our model is not very sensitive to the noise value (Table 10), but degrades from 0.1.
|
| 393 |
+
|
| 394 |
+
Table 10: OGBG-MOLPCBA Noise Ablation
|
| 395 |
+
|
| 396 |
+
<table><tr><td>Flip Probability</td><td>Mean AP</td></tr><tr><td>0.01</td><td>27.8% +- 0.002</td></tr><tr><td>0.03</td><td>27.9% +- 0.003</td></tr><tr><td>0.05</td><td>28.1% +- 0.001</td></tr><tr><td>0.1</td><td>28.0% +- 0.003</td></tr><tr><td>0.2</td><td>27.7% +- 0.002</td></tr></table>
|
| 397 |
+
|
| 398 |
+
Table 11: OGBG-MOLPCBA DropEdge Ablation
|
| 399 |
+
|
| 400 |
+
<table><tr><td></td><td>Mean AP</td></tr><tr><td>MPNN Without DropEdge</td><td>27.4% ± 0.002</td></tr><tr><td>MPNN With DropEdge</td><td>27.5% ± 0.001</td></tr><tr><td>MPNN + DropEdge + Noisy Nodes</td><td>27.8% ± 0.002</td></tr></table>
|
| 401 |
+
|
| 402 |
+
A.7 DROPEDGE & DROPNODE ABLATIONS FOR OGBG-MOLPCBA
|
| 403 |
+
|
| 404 |
+
We conduct an ablation with our 16 layer MPNN using DropEdge at a rate of 0.1 as an alternative approach to improving oversmoothing and find it does not improve performance for ogbg-molpcba (Table 11), similarly we find DropNode (Table 12) does not improve performance. In addition, we find that these two methods can’t be combined well together, reaching a performance of $2 7 . 0 \% \pm$ 0.003. However, both methods can be combined advantageously with Noisy Nodes.
|
| 405 |
+
|
| 406 |
+
We also measure the MAD of the node latents for each layer and find the indeed Noisy Nodes is more effective at addressing oversmoothing in Figure 8.
|
| 407 |
+
|
| 408 |
+
A.8 TRAINING CURVES FOR OC20 NOISY NODES ABLATIONS DEMONSTRATING OVERFITTING
|
| 409 |
+
|
| 410 |
+
Figure 9
|
| 411 |
+
|
| 412 |
+
Table 12: OGBG-MOLPCBA DropNode Ablation
|
| 413 |
+
|
| 414 |
+
<table><tr><td>一</td><td>Mean AP</td></tr><tr><td>MPNN With DropNode</td><td>27.5% ± 0.001</td></tr><tr><td>MPNN Without DropNode</td><td>27.5% ± 0.004</td></tr><tr><td>MPNN + DropNode + Noisy Nodes</td><td>28.2% ±0.005</td></tr></table>
|
| 415 |
+
|
| 416 |
+
|
| 417 |
+
Figure 8: Comparison of the effect of techniques to address oversmoothing on MPNNs. Whilst Some effect can be seen from DropEdge and DropNode, Noisy Nodes is significantly better at preserving per node diversity.
|
| 418 |
+
|
| 419 |
+
A.9 PSEUDOCODE FOR 3D MOLECULAR PREDICTION TRAINING STEP
|
| 420 |
+
|
| 421 |
+
<table><tr><td>Algorithm 1: Noisy Nodes Training Step</td></tr><tr><td>G=(V,E,g) // Input graph</td></tr><tr><td>G=G// Initialize noisy graph λ// Noisy Nodes Weight</td></tr><tr><td>if not_provided(V') then</td></tr><tr><td>-V←V end</td></tr><tr><td>if predict_differences then</td></tr><tr><td> △={u'- vili ∈1,...,|Vl}</td></tr><tr><td>end for each i∈1,...,|V| do</td></tr><tr><td></td></tr><tr><td>Oi = sample_node_noise(shape_of(ui));</td></tr><tr><td>Vi=Ui+Oi；</td></tr><tr><td>if predict_differences then</td></tr><tr><td>△i=△i-Oi;</td></tr><tr><td>end</td></tr><tr><td>endfor</td></tr><tr><td></td></tr><tr><td>E = recompute_edges(V);</td></tr><tr><td>G' = GNN(G);</td></tr><tr><td></td></tr><tr><td>if predict_differences then</td></tr><tr><td>V'=△i;</td></tr><tr><td>end</td></tr><tr><td></td></tr><tr><td></td></tr><tr><td>Loss = λ NoisyNodesLoss(G', V') + PrimaryLos(G', V/'); Loss.minimise()</td></tr></table>
|
| 422 |
+
|
| 423 |
+
|
| 424 |
+
Figure 9: Training curves to accompany Figure 3. This demonstrates that even as the validation performance is getting worse, training loss is going down, indicating overfitting.
|
| 425 |
+
|
| 426 |
+
Table 13: Open Catalyst training parameters.
|
| 427 |
+
|
| 428 |
+
<table><tr><td>Parameter</td><td>Value or description</td></tr><tr><td>Optimiser</td><td>Adam with warm up and cosine cycling</td></tr><tr><td>β1</td><td>0.9</td></tr><tr><td>β</td><td>0.95</td></tr><tr><td>Warm up steps</td><td>5e5</td></tr><tr><td>Warm up start learning rate</td><td>le-5</td></tr><tr><td>Warm up/cosine max learning rate</td><td>le-4</td></tr><tr><td>Cosine cycle length</td><td>5e6</td></tr><tr><td>Loss type</td><td>Mean squared error</td></tr><tr><td>Batch size</td><td>Dynamic to max edge/node/graph count</td></tr><tr><td>Max nodes in batch</td><td>1024</td></tr><tr><td>Max edges in batch</td><td>12800</td></tr><tr><td>Max graphs in batch</td><td>10</td></tr><tr><td>MLP number of layers</td><td>3</td></tr><tr><td>MLP hidden sizes</td><td>512</td></tr><tr><td>Number Bessel Functions</td><td>512</td></tr><tr><td>Activation</td><td>shifted softplus</td></tr><tr><td>message passing layers</td><td>50</td></tr><tr><td>Group size</td><td>10</td></tr><tr><td>Node/Edge latent vector sizes</td><td>512</td></tr><tr><td>Position noise</td><td>Gaussian (μ = O,σ = 0.3)</td></tr><tr><td>Parameter update</td><td>Exponentially moving average (EMA) smoothing</td></tr><tr><td>EMA decay</td><td>0.9999</td></tr><tr><td>Position Loss Co-efficient</td><td>1.0</td></tr></table>
|
| 429 |
+
|
| 430 |
+
# A.10 TRAINING DETAILS
|
| 431 |
+
|
| 432 |
+
Our code base is implemented in JAX using Haiku and Jraph for GNNs, and Optax for training (Bradbury et al., 2018; Babuschkin et al., 2020; Godwin\* et al., 2020; Hennigan et al., 2020). Model selection used early stopping.
|
| 433 |
+
|
| 434 |
+
All results reported as an average of 10 random seeds. OGBG-PCQM4M & OGBG-MOLPCBA were trained with 16 TPUs and evaluated with a single V100 GPU. OGBN-Arxiv was trained and evalated with a single TPU
|
| 435 |
+
|
| 436 |
+
# 3D Molecular Prediction
|
| 437 |
+
|
| 438 |
+
We minimise the mean squared error loss on mean and standard deviation normalised targets and use the Adam (Kingma & Ba, 2015) optimiser with warmup and cosine decay. For OC20 IS2RE energy prediction we subtract a learned reference energy, computed using an MLP with atom types as input.
|
| 439 |
+
|
| 440 |
+
For the GNS model the node and edge latents as well as MLP hidden layers were sized 512, with 3 layers per MLP and using shifted softplus activations throughout. OC20 & QM9 Models were trained on 8 TPU devices and evaluated on a single V100 GPUs. We provide the full set of hyper-parameters and computational resources used separately for each dataset in the Appendix. All noise levels were determined by sweeping a small range of values $( \approx 1 0 )$ ) informed by the noised feature covariance.
|
| 441 |
+
|
| 442 |
+
# Non Spatial Tasks
|
| 443 |
+
|
| 444 |
+
# A.11 HYPER-PARAMETERS
|
| 445 |
+
|
| 446 |
+
Open Catalyst. We list the hyper-parameters used to train the default Open Catalyst experiment. If not specified otherwise (e.g. in ablations of these parameters), experiments were ran with this configuration.
|
| 447 |
+
|
| 448 |
+
Table 14: QM9 training parameters.
|
| 449 |
+
|
| 450 |
+
<table><tr><td>Parameter</td><td>Value or description</td></tr><tr><td>Optimiser</td><td>Adam with warm up and cosine cycling</td></tr><tr><td>β1</td><td>0.9</td></tr><tr><td>β</td><td>0.95</td></tr><tr><td>Warm up steps</td><td>1e4</td></tr><tr><td>Warm up start learning rate</td><td>3e-7</td></tr><tr><td>Warm up/cosine max learning rate</td><td>le-4</td></tr><tr><td>Cosine cycle length</td><td>2e6</td></tr><tr><td>Loss type</td><td>Mean squared error</td></tr><tr><td>Batch size</td><td>Dynamic to max edge/node/graph count</td></tr><tr><td>Max nodes in batch</td><td>256</td></tr><tr><td>Max edges in batch</td><td>4096</td></tr><tr><td>Max graphs in batch</td><td>8</td></tr><tr><td>MLP number of layers</td><td>3</td></tr><tr><td>MLPhidden sizes</td><td>1024</td></tr><tr><td>Number Bessel Funtions</td><td>512</td></tr><tr><td>Activation</td><td>shifted softplus</td></tr><tr><td>message passing layers</td><td>10</td></tr><tr><td>Group Size</td><td>10</td></tr><tr><td>Node/Edge latent vector sizes</td><td>512</td></tr><tr><td>Position noise</td><td>Gaussian (μ= O,σ = 0.02)</td></tr><tr><td>Parameter update</td><td>Exponentially moving average (EMA) smoothing</td></tr><tr><td>EMA decay</td><td>0.9999</td></tr><tr><td>Position Loss Coefficient</td><td>0.1</td></tr></table>
|
| 451 |
+
|
| 452 |
+
Dynamic batch sizes refers to constructing batches by specifying maximum node, edge and graph counts (as opposed to only graph counts) to better balance computational load. Batches are constructed until one of the limits is reached.
|
| 453 |
+
|
| 454 |
+
Parameter updates were smoothed using an EMA for the current training step with the current decay value computed through $d e c a y = m \bar { i } n ( d e c a y , ( 1 . 0 + s t e p ) / ( 1 0 . 0 + \bar { s t e p } )$ . As discussed in the evaluation, best results on Open Catalyst were obtained by utilising a 100 layer network with group size 10.
|
| 455 |
+
|
| 456 |
+
QM9 Table 14 lists QM9 hyper-parameters which primarily reflect the smaller dataset and geometries with fewer long range interactions. For $U _ { 0 }$ , $U$ , $H$ and $G$ we use a slightly larger number of graphs per batch - 16 - and a smaller position loss co-efficient of 0.01.
|
| 457 |
+
|
| 458 |
+
OGBG-PCQM4M Table 15 provides the hyper parameters for OGBG-PCQM4M.
|
| 459 |
+
|
| 460 |
+
OGBG-MOLPCBA Table 16 provides the hyper parameters for the OGBG-MOLPCBA experiments
|
| 461 |
+
|
| 462 |
+
OGBN-ARXIV Table 17 provides the hyper parameters for the OGBN-Arxiv experiments.
|
| 463 |
+
|
| 464 |
+
Table 15: OGBG-PCQM4M Training Parameters.
|
| 465 |
+
|
| 466 |
+
<table><tr><td>Parameter</td><td>Value or description</td></tr><tr><td>Optimiser</td><td>Adam with warm up and cosine cycling</td></tr><tr><td>β1</td><td>0.9</td></tr><tr><td>β</td><td>0.95</td></tr><tr><td>Warm up steps</td><td>5e4</td></tr><tr><td>Warm up start learning rate</td><td>le-5</td></tr><tr><td>Warm up/cosine max learning rate</td><td>le-4</td></tr><tr><td>Cosine cycle length</td><td>5e5</td></tr><tr><td>Loss type</td><td>Mean absolute error</td></tr><tr><td>Reconstruction type</td><td>Softmax Cross Entropy</td></tr><tr><td>Batch size</td><td>Dynamic to max edge/node/graph count</td></tr><tr><td>Max nodes in batch</td><td>20,480</td></tr><tr><td>Max edges in batch</td><td>8,192</td></tr><tr><td>Max graphs in batch</td><td>512</td></tr><tr><td>MLP number of layers</td><td>2</td></tr><tr><td>MLP hidden sizes</td><td>512</td></tr><tr><td>Activation</td><td>relu</td></tr><tr><td>Node/Edge latent vector sizes</td><td>512</td></tr><tr><td>Noisy Nodes Category Flip Fate</td><td>0.05</td></tr><tr><td>Parameter update</td><td>Exponentially moving average (EMA) smoothing</td></tr><tr><td>EMA decay</td><td>0.999</td></tr><tr><td>Reconstruction Loss Coefficient</td><td>0.1</td></tr></table>
|
| 467 |
+
|
| 468 |
+
Table 16: OGBG-MOLPCBA Training Parameters.
|
| 469 |
+
|
| 470 |
+
<table><tr><td>Parameter</td><td>Value or description</td></tr><tr><td>Optimiser β</td><td>Adam with warm up and cosine cycling 0.9</td></tr><tr><td>β2</td><td>0.95</td></tr><tr><td>Warm up steps</td><td>1e4</td></tr><tr><td></td><td></td></tr><tr><td>Warm up start learning rate</td><td>1e-5</td></tr><tr><td>Warm up/cosine max learning rate</td><td>le-4</td></tr><tr><td>Cosine cycle length</td><td>1e5</td></tr><tr><td>Loss type</td><td>Softmax Cross Entropy</td></tr><tr><td>Reconstruction loss type</td><td>Softmax Cross Entropy</td></tr><tr><td>Batch size</td><td>Dynamic to max edge/node/graph count</td></tr><tr><td>Max nodes in batch</td><td>20,480</td></tr><tr><td>Max edges in batch</td><td>8,192</td></tr><tr><td>Max graphs in batch</td><td>512</td></tr><tr><td></td><td>2</td></tr><tr><td>MLP number of layers MLP hidden sizes</td><td>512</td></tr><tr><td>Activation</td><td>relu</td></tr><tr><td>BatchNormalization</td><td>Yes,after every hidden layer</td></tr><tr><td>Node/Edge latent vector sizes</td><td>512</td></tr><tr><td>Dropnode Rate</td><td></td></tr><tr><td></td><td>0.1 0.1</td></tr><tr><td>Dropout Rate</td><td></td></tr><tr><td>Noisy Nodes Category Flip Fate Parameter update</td><td>0.05 Exponentially moving average (EMA) smoothing</td></tr><tr><td>EMA decay</td><td>0.999</td></tr><tr><td>Reconstruction Loss Coefficient</td><td>0.1</td></tr></table>
|
| 471 |
+
|
| 472 |
+
Table 17: OGBG-ARXIV Training Parameters.
|
| 473 |
+
|
| 474 |
+
<table><tr><td>Parameter</td><td>Value or description</td></tr><tr><td>Optimiser</td><td>Adam with warm up and cosine cycling</td></tr><tr><td>β</td><td>0.9</td></tr><tr><td>β</td><td>0.95</td></tr><tr><td>Warm up steps</td><td>50</td></tr><tr><td>Warm up start learning rate</td><td>le-5</td></tr><tr><td>Warm up/cosine max learning rate</td><td>1e-3</td></tr><tr><td>Cosine cycle length</td><td>12,000</td></tr><tr><td>Loss type</td><td>Softmax Cross Entropy</td></tr><tr><td>Reconstruction loss type</td><td>Mean Squared Error</td></tr><tr><td>Batch size</td><td>Full graph</td></tr><tr><td>MLP number of layers</td><td>1</td></tr><tr><td>Activation</td><td>relu</td></tr><tr><td>Batch Normalization</td><td>Yes,after every hidden layer</td></tr><tr><td>Node/Edge latent vector sizes</td><td>256</td></tr><tr><td>Dropout Rate</td><td>0.5</td></tr><tr><td>Noisy Nodes Input Dropout</td><td>0.05</td></tr><tr><td>Reconstruction Loss Coefficient</td><td>0.1</td></tr></table>
|
parse/dev/1wVvweK3oIb/1wVvweK3oIb_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/1wVvweK3oIb/1wVvweK3oIb_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/1wVvweK3oIb/1wVvweK3oIb_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/2EDqbSCnmF/2EDqbSCnmF.md
ADDED
|
@@ -0,0 +1,376 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Any-to-Any Generation via Composable Diffusion
|
| 2 |
+
|
| 3 |
+
# Zineng Tang1∗
|
| 4 |
+
|
| 5 |
+
# Mohit Bansal1†
|
| 6 |
+
|
| 7 |
+
Ziyi Yang2† Chenguang ${ \bf Z } { \bf h } { \bf u } ^ { 2 \ddagger }$ Michael Zeng2 1University of North Carolina at Chapel Hill 2Microsoft Azure Cognitive Services Research https://codi-gen.github.io
|
| 8 |
+
|
| 9 |
+
# Abstract
|
| 10 |
+
|
| 11 |
+
We present Composable Diffusion (CoDi), a novel generative model capable of generating any combination of output modalities, such as language, image, video, or audio, from any combination of input modalities. Unlike existing generative AI systems, CoDi can generate multiple modalities in parallel and its input is not limited to a subset of modalities like text or image. Despite the absence of training datasets for many combinations of modalities, we propose to align modalities in both the input and output space. This allows CoDi to freely condition on any input combination and generate any group of modalities, even if they are not present in the training data. CoDi employs a novel composable generation strategy which involves building a shared multimodal space by bridging alignment in the diffusion process, enabling the synchronized generation of intertwined modalities, such as temporally aligned video and audio. Highly customizable and flexible, CoDi achieves strong joint-modality generation quality, and outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis. The project page with demonstrations and code is at https://codi-gen.github.io/
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
Figure 1: CoDi can generate various (joint) combinations of output modalities from diverse (joint) sets of inputs: video, image, audio, and text (example combinations depicted by the colored arrows).
|
| 15 |
+
|
| 16 |
+
# 1 Introduction
|
| 17 |
+
|
| 18 |
+
Recent years have seen the rise of powerful cross-modal models that can generate one modality from another, e.g. text-to-text [6, 37], text-to-image [13, 19, 22, 41, 44], or text-to-audio [23, 33]. However, these models are restricted in their real-world applicability where multiple modalities coexist and interact. While one can chain together modality-specific generative models in a multi-step generation setting, the generation power of each step remains inherently limited, and a serial, multistep process can be cumbersome and slow. Moreover, independently generated unimodal streams will not be consistent and aligned when stitched together in a post-processing way (e.g., synchronized video and audio). The development of a comprehensive and versatile model that can generate any combination of modalities from any set of input conditions has been eagerly anticipated, as it would more accurately capture the multimodal nature of the world and human comprehension, seamlessly consolidate information from a wide range of sources, and enable strong immersion in human-AI interactions (for example, by generating coherent video, audio, and text description at the same time).
|
| 19 |
+
|
| 20 |
+
In pursuit of this goal, we propose Composable Diffusion, or CoDi, the first model capable of simultaneously processing and generating arbitrary combinations of modalities as shown in Fig. 1. Training a model to take any mixture of input modalities and flexibly generate any mixture of outputs presents significant computational and data requirements, as the number of combinations for the input and output modalities scales exponentially. Also aligned training data for many groups of modalities is scarce or even non-existent, making it infeasible to train with all possible input-output combinations. To address this challenge, we propose to align multiple modalities in both the input conditioning (Section 3.2) and generation diffusion step (Section 3.4). Furthermore, a proposed “Bridging Alignment” strategy for contrastive learning (Section 3.2) allows us to efficiently model the exponential number of input-output combinations with a linear number of training objectives.
|
| 21 |
+
|
| 22 |
+
Building a model with any-to-any generation capacity with exceptional generation quality requires comprehensive model design and training on diverse data resources. Therefore, we build CoDi in an integrative way. First, we train a latent diffusion model (LDM) for each modality, e.g., text, image, video, and audio. These models can be trained in parallel independently, ensuring exceptional singlemodality generation quality using widely available modality-specific training data (i.e., data with one or more modalities as input and one modality as output). For conditional cross-modality generation, such as generating images using audio+language prompts, the input modalities are projected into a shared feature space (Section 3.2), and the output LDM attends to the combination of input features. This multimodal conditioning mechanism prepares the diffusion model to condition on any modality or combination of modalities without directly training for such settings.
|
| 23 |
+
|
| 24 |
+
The second stage of training enables the model to handle many-to-many generation strategies that involve simultaneously generating arbitrary combinations of output modalities. To the best of our knowledge, CoDi is the first AI model with this capability. This is achieved by adding a crossattention module to each diffuser, and an environment encoder $V$ to project the latent variable of different LDMs into a shared latent space (Section 3.4). Next, we freeze the parameters of the LDM, training only the cross-attention parameters and $V$ . Since the environment encoder of different modalities are aligned, an LDM can cross-attend with any group of co-generated modalities by interpolating the representation’s output by $V$ . This enables CoDi to seamlessly generate any group of modalities, without training on all possible generation combinations. This reduces the number of training objectives from exponential to linear.
|
| 25 |
+
|
| 26 |
+
We demonstrate the any-to-any generation capability of CoDi, including single-to-single modality generation, multi-condition generation, and the novel capacity of joint generation of multiple modalities. For example, generating synchronized video and audio given the text input prompt; or generating video given a prompt image and audio. We also provide a quantitative evaluation of CoDi using eight multimodal datasets. As the latest work from Project i-Code [55] towards Composable AI, CoDi exhibits exceptional generation quality across assorted scenarios, with synthesis quality on par or even better than single to single modality SOTA, e.g., audio generation and audio captioning.
|
| 27 |
+
|
| 28 |
+
# 2 Related Works
|
| 29 |
+
|
| 30 |
+
Diffusion models (DMs) learn the data distribution by denoising and recovering the original data. Deep Diffusion Process (DDP) [45] adopts a sequence of reversible diffusion steps to model image probability distribution. It uses a reversible encoder to map the input image to a latent space and a decoder to map the latent variables to an output image. Denoising diffusion probabilistic model (DDPM) [20] uses a cascade of diffusion processes to gradually increase the complexity of the probability density function model. At each step, the model adds noise to the input image and estimates the corresponding noise level using an autoregressive model. This allows the model to capture the dependencies between adjacent pixels and generate high-quality images. Score-based generative models (SOG) [46] use the score function to model the diffusion process. [40] generates high-fidelity images conditioned on CLIP representations of text prompts. Latent diffusion model (LDM) [41] uses a VAE to encode inputs into latent space to reduce modeling dimension and improves efficiency. The motivation is that image compression can be separated into semantic space by a diffusion model and perceptual space by an autoencoder. By incorporating temporal modeling modules and cascading model architectures, video diffusion models have been built upon image diffusers to generate temporally consistent and inherent frames[14, 19, 21, 44]. Diffusion models have also been applied to other domains, such as generating audio from text and vision prompts[23, 33].
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
Figure 2: CoDi model architecture: (a) We first train individual diffusion model with aligned prompt encoder by “Bridging Alignment”; (b) Diffusion models learn to attend with each other via “Latent Alignment”; (c) CoDi achieves any-to-any generation with a linear number of training objectives.
|
| 34 |
+
|
| 35 |
+
Multimodal modeling has experienced rapid advancement recently, with researchers striving to build uniform representations of multiple modalities using a single model to achieve more comprehensive cross-modal understanding. Vision transformers [11], featuring diverse model architectures and training techniques, have been applied to various downstream tasks such as vision Q&A and image captioning. Multimodal encoders have also proven successful in vision-language [1, 8, 57], videoaudio [47] and video-speech-language [55, 56] domains. Aligning data from different modalities is an active research area [12, 38], with promising applications in cross-modality retrieval and building uniform multimodal representations [33, 35, 41].
|
| 36 |
+
|
| 37 |
+
# 3 Methodology
|
| 38 |
+
|
| 39 |
+
# 3.1 Preliminary: Latent Diffusion Model
|
| 40 |
+
|
| 41 |
+
Diffusion models (DM) represent a class of generative models that learn data distributions $p ( { \pmb x } )$ by simulating the diffusion of information over time. During training, random noise is iteratively added to $_ { \textbf { \em x } }$ , while the model learns to denoise the examples. For inference, the model denoises data points sampled from simple distributions such as Gaussian. Latent diffusion models (LDM) [41] learn the distribution of the latent variable $_ { z }$ corresponding to $_ { \textbf { \em x } }$ , significantly reducing computational cost by decreasing the data dimension.
|
| 42 |
+
|
| 43 |
+
In LDM, an autoencoder is first trained to reconstruct $_ { \textbf { \em x } }$ , i.e., $\hat { \pmb { x } } = D ( E ( \pmb { x } ) )$ , where $E$ and $D$ denote the encoder and decoder, respectively. The latent variable $z = E ( { \pmb x } )$ is iteratively diffused over time steps $t$ based on a variance schedule $\beta _ { 1 } , \ldots , \beta _ { T }$ , i.e., $q ( z _ { t } | z _ { t - 1 } ) = \mathcal { N } ( z _ { t } ; \sqrt { 1 - \beta _ { t } } z _ { t - 1 } , \beta _ { t } I )$ [20, 45].
|
| 44 |
+
|
| 45 |
+
The forward process allows the random sampling of ${ \boldsymbol { z } } _ { t }$ at any timestep in a closed form [20, 45]: $\boldsymbol { z } _ { t } = \alpha _ { t } \boldsymbol { z } + \sigma _ { t } \boldsymbol { \epsilon }$ , where $\epsilon \sim \mathcal { N } ( 0 , I )$ , $\alpha _ { t } : = 1 - \beta _ { t }$ and $\begin{array} { r } { \sigma _ { t } : = \dot { 1 } - \prod _ { s = 1 } ^ { t } \dot { \alpha } _ { s } } \end{array}$ . The diffuser learns how to denoise from $\left\{ { z } _ { t } \right\}$ to recover $_ z$ . Following the reparameterization method proposed in [20], the denoising training objective can be expressed as [41]:
|
| 46 |
+
|
| 47 |
+
$$
|
| 48 |
+
\begin{array} { r } { \mathcal { L } _ { D } = \mathbb { E } _ { z , \epsilon , t } \Vert \epsilon - \epsilon _ { \theta } ( z _ { t } , t , C ( \pmb { y } ) ) \Vert _ { 2 } ^ { 2 } . } \end{array}
|
| 49 |
+
$$
|
| 50 |
+
|
| 51 |
+
In data generation, the denoising process can be realized through reparameterized Gaussian sampling:
|
| 52 |
+
|
| 53 |
+
$$
|
| 54 |
+
p ( z _ { t - 1 } | z _ { t } ) = \mathcal { N } \left( z _ { t - 1 } ; \frac { 1 } { \sqrt { \alpha _ { t } } } \left( z _ { t } - \frac { \beta _ { t } } { \sqrt { \sigma _ { t } } } \epsilon _ { \theta } \right) , \beta _ { t } I \right) .
|
| 55 |
+
$$
|
| 56 |
+
|
| 57 |
+
In $\mathcal { L } _ { D }$ , the diffusion time step $t \sim \mathcal { U } [ 1 , T ]$ ; $\epsilon _ { \theta }$ is a denoising model with UNet backbone parameterized by $\theta ; { \boldsymbol { y } }$ represents the conditional variable that can be used to control generation; $C$ is the prompt encoder. The conditioning mechanism is implemented by first featurizing $\textbf { { y } }$ into $C ( \boldsymbol { y } )$ , then the UNet $\epsilon _ { \theta }$ conditions on $C ( \boldsymbol { y } )$ via cross-attention, as described in [41]. Distinct from previous works, our model can condition on any combinations of modalities of text, image, video and audio. Details are presented in the following section.
|
| 58 |
+
|
| 59 |
+
# 3.2 Composable Multimodal Conditioning
|
| 60 |
+
|
| 61 |
+
To enable our model to condition on any combination of input/prompt modalities, we align the prompt encoder of text, image, video and audio (denoted by $C _ { t }$ , $C _ { i }$ , $C _ { v }$ , and $C _ { a }$ , respectively) to project the input from any modality into the same space. Multimodal conditioning can then be conveniently achieved by interpolating the representations of each modality $m$ : $\begin{array} { r } { C ( x _ { t } , \bar { x _ { i } } , x _ { v } , x _ { a } ) = \sum _ { m } \alpha _ { m } C ( \bar { m } ) } \end{array}$ for $m \in \ b { x } _ { t } , \ b { x } _ { i } , \ b { x } _ { v } , \ b { x } _ { a }$ , with $\textstyle \sum _ { m } \alpha _ { m } = 1$ . Through simple weighted interpolation of aligned embeddings, we enable models trained with single-conditioning (i.e., with only one input) to perform zero-shot multi-conditioning (i.e., with multiple inputs). This process is illustrated in Fig. 2 (a)(2).
|
| 62 |
+
|
| 63 |
+
Optimizing all four prompt encoders simultaneously in a combinatorial manner is computationally heavy, with $\mathcal { O } ( n ^ { 2 } )$ pairs. Additionally, for certain dual modalities, well-aligned paired datasets are limited or unavailable e.g., image-audio pairs. To address this challenge, we propose a simple and effective technique called "Bridging Alignment" to efficiently align conditional encoders. As shown in Fig. 2 (a)(1), we choose the text modality as the "bridging" modality due to its ubiquitous presence in paired data, such as text-image, text-video, and text-audio pairs. We begin with a pretrained text-image paired encoder, i.e., CLIP [38]. We then train audio and video prompt encoders on audio-text and video-text paired datasets using contrastive learning, with text and image encoder weights frozen.
|
| 64 |
+
|
| 65 |
+
In this way, all four modalities are aligned in the feature space. As shown in Section 5.2, CoDi can effectively leverage and combine the complementary information present in any combination of modalities to generate more accurate and comprehensive outputs. The high generation quality remains unaffected with respect to the number of prompt modalities. As we will discuss in subsequent sections, we continue to apply Bridging Alignment to align the latent space of LDMs with different modalities to achieve joint multimodal generation.
|
| 66 |
+
|
| 67 |
+
# 3.3 Composable Diffusion
|
| 68 |
+
|
| 69 |
+
Training an end-to-end anything-to-anything model requires extensive learning on various data resources. The model also needs to maintain generation quality for all synthesis flows. To address these challenges, CoDi is designed to be composable and integrative, allowing individual modalityspecific models to be built independently and then smoothly integrated later. Specifically, we start by independently training image, video, audio, and text LDMs. These diffusion models then efficiently learn to attend across modalities for joint multimodal generation (Section 3.4) by a novel mechanism named “latent alignment”.
|
| 70 |
+
|
| 71 |
+
Image Diffusion Model. The image LDM follows the same structure as Stable Diffusion 1.5 [41] and is initialized with the same weights. Reusing the weights transfers the knowledge and exceptional generation fidelity of Stable Diffusion trained on large-scale high-quality image datasets to CoDi.
|
| 72 |
+
|
| 73 |
+
Video Diffusion Model. To model the temporal properties of videos and simultaneously maintain vision generation quality, we construct the video diffuser by extending the image diffuser with temporal modules. Specifically, we insert pseudo-temporal attention before the residual block [13]. However, we argue that pseudo-temporal attention only enables video frames to globally attend to each other by flattening the pixels (height, width dimension) to batch dimension, resulting in a lack of cross-frame interaction between local pixels. We argue that this results in the common temporal-inconsistency issue in video generation that locations, shapes, colors, etc. of objects can be inconsistent across generated frames. To address this problem, we propose adapting the latent shift method [2] that performs temporal-spatial shifts on latent features in accordance with temporal attention. We divide the video by the hidden dimension into $k = 8$ chunks, and for each chunk $i = 0$ to 7, we shift the temporal dimension forward by $i$ positions. Further details will be provided in the appendix.
|
| 74 |
+
|
| 75 |
+
Audio Diffusion Model. To enable flexible cross-modality attention in joint generation, the audio diffuser is designed to have a similar architecture to vision diffusers, where the mel-spectrogram can be naturally viewed as an image with 1 channel. We use a VAE encoder to encode the melspectrogram of audio to a compressed latent space. In audio synthesis, a VAE decoder maps the latent variable to the mel-spectrogram, and a vocoder generates the audio sample from the mel-spectrogram. We employ the audio VAE from [33] and the vocoder from [27].
|
| 76 |
+
|
| 77 |
+
Text Diffusion Model. The VAE of the text LDM is OPTIMUS [29], and its encoder and decoder are [9] and GPT-2 [39], respectively. For the denoising UNet, unlike the one in image diffusion, the 2D convolution in residual blocks is replaced with 1D convolution [53].
|
| 78 |
+
|
| 79 |
+
# 3.4 Joint Multimodal Generation by Latent Alignment
|
| 80 |
+
|
| 81 |
+
The final step is to enable cross-attention between diffusion flows in joint generation, i.e., generating two or more modalities simultaneously. This is achieved by adding cross-modal attention sublayers to the UNet $\epsilon _ { \theta }$ (Fig. 2 (b)(2)). Specifically, consider a diffusion model of modality $A$ that cross-attends with another modality $B$ . Let the latent variables of modalities $m _ { A }$ and $m _ { B }$ at diffusion step $t$ be denoted as $ { \boldsymbol { z } } _ { t } ^ { A }$ and $\hat { z _ { t } ^ { B } }$ , respectively. The proposed “Latent Alignment” technique is such that a modality-specific environment encoder $V _ { B }$ first projects $ { \boldsymbol { z } } _ { t } ^ { B }$ into a shared latent space for different modalities. Then, in each layer of the UNet for modality $A$ , a cross-attention sublayer attends to $V _ { B } \big ( z _ { t } ^ { B } \big )$ . For the diffusion model of modality $A$ , the training objective in Eq. (1) now becomes:
|
| 82 |
+
|
| 83 |
+
$$
|
| 84 |
+
\mathcal { L } _ { C r o s s } ^ { A } = \mathbb { E } _ { z , \epsilon , t } \Vert \epsilon - \epsilon _ { \theta _ { c } } ( z _ { t } ^ { A } , V _ { B } ( z _ { t } ^ { B } ) , t , C ( \pmb { y } ) ) \Vert _ { 2 } ^ { 2 } ,
|
| 85 |
+
$$
|
| 86 |
+
|
| 87 |
+
where $\theta _ { c }$ denotes the weights of cross-attention modules in the UNet.
|
| 88 |
+
|
| 89 |
+
The training objective of $A + B$ joint generation is $\mathcal { L } _ { C r o s s } ^ { A } + \mathcal { L } _ { C r o s s } ^ { B }$ . $V ( \cdot )$ of different modalities are trained to be aligned with contrastive learning. Since $z _ { t } ^ { A }$ and $z _ { t } ^ { B }$ at any time step can be sampled with closed form in the diffusion process Section 3.1, one can conveniently train the contrastive learning together with $\mathcal { L } _ { C r o s s }$ . The purpose of $V$ is to achieve the generation of any combination of modalities (in polynomial) by training on a linear number of joint-generation tasks. For example, if we have trained the joint generation of modalities $A , B$ , and $B$ , $C$ independently, then we have $V _ { A } ( z _ { t } ^ { A } )$ , $V _ { B } \big ( z _ { t } ^ { B } \big )$ , and $V _ { C } ( z _ { t } ^ { C } )$ aligned. Therefore, CoDi can seamlessly achieve joint generation of modalities $A$ and $C$ without any additional training. Moreover, such design automatically effortlessly enables joint generation of modalities $A$ , $B$ , and $C$ concurrently. Specifically, UNet of $A$ can cross-attend with the interpolation of $V _ { B } \big ( z _ { t } ^ { B } \big )$ , and $V _ { C } ( z _ { t } ^ { C } )$ , although CoDi has not been trained with such task.
|
| 90 |
+
|
| 91 |
+
As shown in Fig. 2(b)(3), we follow similar designs to the "Bridging Alignment" in training joint generation: (1) We first train the cross-attention weights in the image and text diffusers, as well as their environment encoders $V$ , on text-image paired data. (2) We freeze the weights of the text diffuser and train the environment encoder and cross-attention weights of the audio diffuser on text-audio paired data. (3) Finally we freeze the audio diffuser and its environment encoder, and train the joint generation of the video modality on audio-video paired data. As demonstrated in Section 5.3, although only trained on three paired joint generation tasks (i.e, Text $^ +$ Audio, Text+Image, and Video+Audio), CoDi is capable of generating assorted combinations of modalities simultaneously that are unseen in training, e.g., joint image-text-audio generation in Fig. 5.
|
| 92 |
+
|
| 93 |
+
Table 1: Training tasks (CT stands for “contrastive learning” to align prompt encoders) and datasets with corresponding statistics. \* denotes the number of accessible examples in the original datasets.
|
| 94 |
+
|
| 95 |
+
<table><tr><td>Categories</td><td>Tasks</td><td>Datasets</td><td># of samples</td><td>Domain</td></tr><tr><td>Image + Text</td><td>Image-→Text,Text-→Image Text-→Image+Text</td><td>Laion400M [42]</td><td>400M</td><td>Open</td></tr><tr><td>Audio + Text</td><td>Text→Audio,Audio-→Text, Text-→Audio+Text,Audio-Text CT</td><td>AudioSet [16] AudioCaps [24] Freesound 500K BBC Sound Effect</td><td>900K* 46K 2.5M 30K</td><td>YouTube YouTube Public audio samples Authentic natural sound</td></tr><tr><td>Audiovisual</td><td>Image→Audio,Image→Video+Audio</td><td>AudioSet SoundNet [3]</td><td>900K* 1.0M*</td><td>YouTube Flickr, natural sound</td></tr><tr><td>Video</td><td>Text-→Video,Image→Video, Video-Text CT</td><td>Webvid10M[4] HD-Villa-100M [54]</td><td>10.7M 100M</td><td>Short videos YouTube</td></tr></table>
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
Figure 3: Single-to-single modality generation. Clockwise from top left: text image, image text, image video, audio image.
|
| 99 |
+
|
| 100 |
+
# 4 Experiments
|
| 101 |
+
|
| 102 |
+
# 4.1 Training Objectives and Datasets
|
| 103 |
+
|
| 104 |
+
We list training tasks of CoDi in Table 1, including single modality synthesis, joint multimodal generation, and contrastive learning to align prompt encoders. Table 1 provides an overview of the datasets, tasks, number of samples, and domain. Datasets are from the following domains: image $^ +$ text (e.g. image with caption), audio $^ +$ text (e.g. audio with description), audio $^ +$ video (e.g. video with sound), and video $^ +$ text (e.g. video with description). As one may have noticed, the language modality appears in most datasets and domains. This echos the idea of using text as the bridge modality to be able to extrapolate and generate new unseen combinations such as audio and image bridged by text, as mentioned in Section 3.2 and Section 3.4. Due to space limit, more details on training datasets and can be found in Appendix C, model architecture details in Appendix Appendix A.1, and training details in Appendix B.
|
| 105 |
+
|
| 106 |
+
Image $^ +$ Text. We use a recently developed large-scale image caption dataset, Laion400M [42]. This image-text paired data allows us to train with tasks text image, image text, and the joint generation of image and text. For the joint generation task, we propose to train with text image+text, where the prompt text is the truncated image caption, and the output text is the original caption. Since the condition information is incomplete, the text and image diffuser will need to learn to attend with each other through the joint generation process.
|
| 107 |
+
|
| 108 |
+
Table 2: COCO-caption [32] FID scores for text-to-image generation.
|
| 109 |
+
|
| 110 |
+
<table><tr><td>Method</td><td>FID↓</td></tr><tr><td>CogView [10]</td><td>27.10</td></tr><tr><td>GLIDE [36]</td><td>12.24</td></tr><tr><td>Make-a-Scene [15]</td><td>11.84</td></tr><tr><td>LDM [41]</td><td>12.63</td></tr><tr><td>Stable Diffusion-1.4</td><td>11.21</td></tr><tr><td>Stable Diffusion-1.5</td><td>11.12</td></tr><tr><td>Versatile Diffusion [53]</td><td>11.10</td></tr><tr><td>CoDi (Ours)</td><td>11.26</td></tr></table>
|
| 111 |
+
|
| 112 |
+
Table 3: MSR-VTT text-to-video Table 4: UCF-101 text-to-video generation performance. generation performance.
|
| 113 |
+
|
| 114 |
+
<table><tr><td>Method</td><td>Zero-Shot</td><td>CLIPSIM ↑</td></tr><tr><td>GODIVA [50]</td><td>No</td><td>0.2402</td></tr><tr><td>NUWA [51]</td><td>No</td><td>0.2439</td></tr><tr><td>CogVideo [22]</td><td>Yes</td><td>0.2631</td></tr><tr><td>Make-A-Video [44]</td><td>Yes</td><td>0.3049</td></tr><tr><td>Video LDM[5]</td><td>Yes</td><td>0.2929</td></tr><tr><td>CoDi(Ours)</td><td>Yes</td><td>0.2890</td></tr></table>
|
| 115 |
+
|
| 116 |
+
<table><tr><td>Method</td><td>IS(1)</td><td>FVD (↑)</td></tr><tr><td>Cog Video (Chinese)</td><td>23.55</td><td>751.34</td></tr><tr><td>CogVideo (English)</td><td>25.27</td><td>701.59</td></tr><tr><td>Make-A-Video</td><td>33.00</td><td>367.23</td></tr><tr><td>Video LDM</td><td>33.45</td><td>550.61</td></tr><tr><td>CoDi(Ours)</td><td>32.88</td><td>596.34</td></tr></table>
|
| 117 |
+
|
| 118 |
+
Table 5: The comparison between our audio diffuser and baseline TTA generation models. Evaluation is conducted on AudioCaps test set. AS, AC, FSD, BBC, and SDN stand for AudioSet, AudioCaps, Freesound, BBC Sound Effect, and Soundnet.
|
| 119 |
+
|
| 120 |
+
<table><tr><td>Model</td><td>Datasets</td><td>FD↓</td><td>IS个</td><td>KL←</td><td>FAD↓</td><td>OVL ↑</td><td>REL个</td></tr><tr><td>Ground truth</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>83.61</td><td>80.11</td></tr><tr><td>DiffSound</td><td>AS+AC</td><td>47.68</td><td>4.01</td><td>2.52</td><td>7.75</td><td>45.00</td><td>43.83</td></tr><tr><td>AudioGen</td><td>AS +AC+8others</td><td>=</td><td>-</td><td>2.09</td><td>3.13</td><td>-</td><td>=</td></tr><tr><td>AudioLDM-L-Full</td><td>AS+AC+FSD+BBC</td><td>23.31</td><td>8.13</td><td>1.59</td><td>1.96</td><td>65.91</td><td>65.97</td></tr><tr><td>CoDi(Ours)</td><td>AS+AC+FSD+BBC+SDN</td><td>22.90</td><td>8.77</td><td>1.40</td><td>1.80</td><td>66.87</td><td>67.60</td></tr></table>
|
| 121 |
+
|
| 122 |
+
Table 6: COCO image captioning scores comparison.
|
| 123 |
+
|
| 124 |
+
<table><tr><td>Model</td><td>B@4</td><td>METEOR</td><td>CIDEr</td></tr><tr><td colspan="4">Autoregressive Model</td></tr><tr><td>Oscar [31]</td><td>36.58</td><td>30.4</td><td>124.12</td></tr><tr><td>ClipCap [35]</td><td>32.15</td><td>27.1</td><td>108.35</td></tr><tr><td>OFA [49]</td><td>44.9</td><td>32.5</td><td>154.9</td></tr><tr><td>BLIP2 [30]</td><td>43.7</td><td>-</td><td>145.8</td></tr><tr><td colspan="4">Diffusion Model</td></tr><tr><td>DDCap [59]</td><td>35.0</td><td>28.2</td><td>117.8</td></tr><tr><td>SCD-Net [34]</td><td>39.4</td><td>29.2</td><td>131.6</td></tr><tr><td>CoDi (Ours)</td><td>40.2</td><td>31.0</td><td>149.9</td></tr></table>
|
| 125 |
+
|
| 126 |
+
Table 7: AudioCaps audio captioning scores comparison.
|
| 127 |
+
|
| 128 |
+
<table><tr><td>Model</td><td>SPIDEr</td><td>CIDEr</td><td>SPICE</td></tr><tr><td>AudioCaps [24]</td><td>0.369</td><td>0.593</td><td>0.144</td></tr><tr><td>BART-Finetune [17]</td><td>0.465</td><td>0.753</td><td>0.176</td></tr><tr><td>VALOR[7]</td><td></td><td>0.741</td><td></td></tr><tr><td>AL-MixGen [25]</td><td>0.466</td><td>0.755</td><td>0.177</td></tr><tr><td>CoDi (Ours)</td><td>0.480</td><td>0.789</td><td>0.182</td></tr></table>
|
| 129 |
+
|
| 130 |
+
Table 8: MSRVTT video captioning scores comparison.
|
| 131 |
+
|
| 132 |
+
<table><tr><td>Model</td><td>B@4</td><td>METEOR</td><td>CIDEr</td></tr><tr><td>ORG-TRL[58]</td><td>43.6</td><td>28.8</td><td>50.9</td></tr><tr><td>MV-GPT[43]</td><td>48.9</td><td>38.7</td><td>60.0</td></tr><tr><td>GIT[48]</td><td>54.8</td><td>33.1</td><td>75.9</td></tr><tr><td>mPLUG-2 [52]</td><td>57.8</td><td>34.9</td><td>80.3</td></tr><tr><td>CoDi(Ours)</td><td>52.1</td><td>32.5</td><td>74.4</td></tr></table>
|
| 133 |
+
|
| 134 |
+
Audio $^ +$ Text. We curated a new dataset, Freesound 500K, by crawling 500K audio samples together with tags and descriptions from the Freesound website. We also use AudioSet [42] with 2 million human-labeled 10-second sound clips from YouTube videos and AudioCaps [24] with 46K audiotext pairs derived from the AudioSet dataset. Audio samples are clipped into 10-second segments for training purposes. The paired audio $^ +$ text data enables us to train text audio, audio text, text audio $^ +$ text generation, and audio-text contrastive learning. Similar to image $^ +$ text joint generation, in text audio $^ +$ text, text prompt is the truncated text, and the output is the original text.
|
| 135 |
+
|
| 136 |
+
Video. We use the following diverse and high-quality video datasets to train video generation and video prompt encoder. WebVid [4], a large-scale dataset of web videos together with descriptions; HD-Villa-100M [54] with high resolution YouTube videos of at least 720P. We perform text video and video-text contrastive learning task with WebVid. We use HD-Villa-100M for image video generation where the middle frame is the input image.
|
| 137 |
+
|
| 138 |
+
Audiovisual. Web videos are a natural aligned audio-video data resource. However, many existing datasets, e.g., ACAV100M [28], feature heavily on videos of human speech rather than natural sounds. Therefore, we leverage sound-oriented datasets AudioSet and SoundNet [3] for joint audio-video generation. For image audio $^ +$ video, we use the middle frame of the target video as the input prompt image. We also use the middle frame as the prompt input to train the model to generate the audio, i.e., image audio.
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
Figure 4: Generation with multiple input modality conditions. Top to bottom: text+audio image, text+audio video, video+audio text.
|
| 142 |
+
|
| 143 |
+
# 5 Evaluation Results
|
| 144 |
+
|
| 145 |
+
In this section, we will evaluate the model generation quality in different settings including single modality generation, multi-condition generation, and multi-output joint generation. We provide both quantitative benchmarking on evaluation datasets as well as qualitative visualization demonstrations.
|
| 146 |
+
|
| 147 |
+
# 5.1 Single Modality Generation Results
|
| 148 |
+
|
| 149 |
+
We first show example demo in Fig. 3, where we present various single to single modality generation. Then, we evaluate the synthesis quality of the unimodal generation on text, image, video, and audio. CoDi achieves SOTA on audio captions and audio generation, as shown in Table 7 and Table 5. Notably for the first time in the field, CoDi, a diffusion-base model, exhibits comparable performance on image captioning with autoregressive transformer-based SOTA (Table 6). CoDi is the first diffusion-model based for video captioning Table 8. On image and video generation, CoDi performs competitively with state-of-the-art (Tables 2 to 4). This gives us strong starting points for multi-condition and multi-output generation that will be presented next in Section 5.2 and Section 5.3.
|
| 150 |
+
|
| 151 |
+
We demonstrate in Section 3.2 that CoDi is capable of integrating representation from different modalities in the generation. Thus, we first show multi-condition generation demo as shown in Fig. 4.
|
| 152 |
+
|
| 153 |
+
# 5.2 Multi-Condition Generation Results
|
| 154 |
+
|
| 155 |
+
For quantitative evaluation, we focus on multiple inputs to image synthesis output since the evaluation metric for this case (FID) does not require specific modality inputs like text. We test with several input combinations including text $^ +$ image, text $^ +$ audio, image $^ +$ audio, text $^ +$ video, as well as three inputs text $^ +$ audio $^ +$ image. We test on the validation set of AudioCaps [24] since all four modalities are present in this dataset. The prompt image input is the middle frame of the video. As shown in
|
| 156 |
+
|
| 157 |
+
Table 9: CoDi is capable of generating high quality output (image in this case) from various combinations of prompt modalities.
|
| 158 |
+
|
| 159 |
+
<table><tr><td>Inputs</td><td>FID↓</td></tr><tr><td> Single-modality Prompt</td><td></td></tr><tr><td>Text</td><td>14.2</td></tr><tr><td>Audio</td><td>14.3</td></tr><tr><td> Dual-modality Prompt</td><td></td></tr><tr><td>Text+Audio</td><td>14.9</td></tr></table>
|
| 160 |
+
|
| 161 |
+
Table 10: MSR-VTT text-to-video generation performance.
|
| 162 |
+
|
| 163 |
+
<table><tr><td>Inputs</td><td>CLIPSIM个</td></tr><tr><td> Single-modality Prompt</td><td></td></tr><tr><td>Text</td><td>0.2890</td></tr><tr><td> Dual-modality Prompt</td><td></td></tr><tr><td>Text+Audio</td><td>0.2912</td></tr><tr><td>Text+Image</td><td>0.2891</td></tr><tr><td>Text+Audio+Image</td><td>0.2923</td></tr></table>
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
Figure 5: Joint generation of multiple output modalities by CoDi. From top to bottom: text video+audio, tex image+text+audio, text+audio+image video+audio.
|
| 167 |
+
|
| 168 |
+
Table 9, CoDi achieves high image generation quality given assorted groups of input modalities. We also test with several input combinations with video as output including text, text $^ +$ audio, image $^ +$ image, as well as text $^ +$ audio $^ +$ image. We also test on MSRVTT [24] since all four modalities are present in this dataset. Similarly, the prompt image input is the middle frame of the video. As shown in Table 10, CoDi achieves high video and ground truth text similarity given assorted groups of input modalities. Again our model does not need to train on multi-condition generation like text $^ +$ audio or text $^ +$ image. Through bridging alignment and composable multimodal conditioning as proposed in Section 3.2, our model trained on single condition can zero-shot infer on multiple conditions.
|
| 169 |
+
|
| 170 |
+
# 5.3 Multi-Output Joint Generation Results
|
| 171 |
+
|
| 172 |
+
For joint multimodal generation, we first demonstrate high-quality multimodal output joint generation demo as shown in Fig. 5. For quantitative evaluation, there is no existing evaluation metric since we are the first model that can simultaneously generate across all 4 modalities. Therefore, we propose the following metric SIM that quantifies the coherence and consistency between the two generated modalities by cosine similarity of embeddings:
|
| 173 |
+
|
| 174 |
+
$$
|
| 175 |
+
\operatorname { S I M } ( A , B ) = \cos { ( C _ { A } ( A ) , C _ { B } ( B ) ) }
|
| 176 |
+
$$
|
| 177 |
+
|
| 178 |
+
Table 11: Similarity scores between generated modalities. The number on the left of $" / "$ represents the similarity score of independent generation, and the right it represents the case of joint generation. Jointly generated outputs consistently show stronger coherence.
|
| 179 |
+
|
| 180 |
+
<table><tr><td>Inputs</td><td>SIM-IT</td><td>SIM-AT</td><td>SIM-VT</td><td>SIM-VA</td></tr><tr><td colspan="5"> Two Joint Outputs</td></tr><tr><td>Audio → Image+Text</td><td>0.251 / 0.260</td><td></td><td></td><td></td></tr><tr><td>Image→Audio+Text</td><td>■</td><td>0.244 / 0.256</td><td></td><td></td></tr><tr><td>Text →Video+Audio</td><td></td><td></td><td></td><td>0.240 / 0.255</td></tr><tr><td>Audio →Video+Text</td><td></td><td></td><td>0.256 / 0.261</td><td></td></tr><tr><td colspan="5"> Three Joint Outputs</td></tr><tr><td>Text-→ Video+Image+Audio 0.256/0.270 0.240/0.257</td><td></td><td></td><td></td><td>0.240 / 0.257</td></tr><tr><td colspan="5"> Multi-Inputs-Outputs</td></tr><tr><td>Text+Image -→ Video+Audio</td><td></td><td></td><td></td><td>0.247 / 0.259</td></tr></table>
|
| 181 |
+
|
| 182 |
+
where $A$ , $B$ are the generated modalities, and $C _ { A }$ and $C _ { B }$ are aligned encoders that project $A$ and $B$ to the same space. We use the prompt encoder as described in Section 3.2. This metric aims to compute the cosine similarity of the embedding of two modalities using contrastive learned prompt encoders. Thus, the higher the metric, the more aligned and similar the generated modalities are.
|
| 183 |
+
|
| 184 |
+
To demonstrate the effectiveness of joint generation, assume the prompt modality is $P$ , we compare $\mathrm { S I M } ( A , B )$ of $A$ and $B$ generated separately vs. jointly, i.e., $\{ P \ { \overset { - } { \to } } \ A , \ P \ { \overset { - } { \to } } \ B \}$ vs. $\{ P $ $A + B \}$ . The benchmark is the validation set of AudioCaps [24]. We test on the following settings, audio image+text, image audio+text, and text video+audio, image video+audio. audio video+text, audio text+video+image, text video+image+audio, where the image prompt is the middle frame of the video clip. As shown in Table 11, joint generation (similarity shown on the right side of $" / "$ ) consistently outperforms independent generation (on the left side of $" / "$ ).
|
| 185 |
+
|
| 186 |
+
# 6 Conclusion
|
| 187 |
+
|
| 188 |
+
In this paper, we present Composable Diffusion (CoDi), a groundbreaking model in multimodal generation that is capable of processing and simultaneously generating modalities across text, image, video, and audio. Our approach enables the synergistic generation of high-quality and coherent outputs spanning various modalities, from assorted combinations of input modalities. Through extensive experiments, we demonstrate CoDi’s remarkable capabilities in flexibly generating single or multiple modalities from a wide range of inputs. Our work marks a significant step towards more engaging and holistic human-computer interactions, establishing a solid foundation for future investigations in generative artificial intelligence.
|
| 189 |
+
|
| 190 |
+
Limitations & Broader Impacts. See Appendix D for the discussion.
|
| 191 |
+
|
| 192 |
+
# Acknowledgement
|
| 193 |
+
|
| 194 |
+
We would like to thank Bei Liu for HD-VILA-100M data support. We also thank Shi Dong, Mahmoud Khademi, Junheng Hao, Yuwei Fang, Yichong Xu and Azure Cognitive Services Research team members for their feedback.
|
| 195 |
+
|
| 196 |
+
# References
|
| 197 |
+
|
| 198 |
+
[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716–23736, 2022. 3
|
| 199 |
+
[2] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 5, 15, 16
|
| 200 |
+
|
| 201 |
+
[3] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. Advances in neural information processing systems, 29, 2016. 6, 7
|
| 202 |
+
|
| 203 |
+
[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728–1738, 2021. 6, 7, 16
|
| 204 |
+
|
| 205 |
+
[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. arXiv preprint arXiv:2304.08818, 2023. 7
|
| 206 |
+
|
| 207 |
+
[6] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 2
|
| 208 |
+
|
| 209 |
+
[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor: Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345, 2023. 7
|
| 210 |
+
|
| 211 |
+
[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931–1942. PMLR, 2021. 3
|
| 212 |
+
|
| 213 |
+
[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5
|
| 214 |
+
|
| 215 |
+
[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. arXiv preprint arXiv:2105.13290, 2021. 7
|
| 216 |
+
|
| 217 |
+
[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687–10696, 2021. 3
|
| 218 |
+
|
| 219 |
+
[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap: Learning audio concepts from natural language supervision. arXiv preprint arXiv:2206.04769, 2022. 3
|
| 220 |
+
|
| 221 |
+
[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2, 5
|
| 222 |
+
|
| 223 |
+
[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3
|
| 224 |
+
|
| 225 |
+
[15] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pages 89–106. Springer, 2022. 7 [16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776–780. IEEE, 2017. 6
|
| 226 |
+
|
| 227 |
+
[17] Félix Gontier, Romain Serizel, and Christophe Cerisara. Automated audio captioning by fine-tuning bart with audioset tags. In Detection and Classification of Acoustic Scenes and Events-DCASE 2021, 2021. 7 [18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 15
|
| 228 |
+
|
| 229 |
+
[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3
|
| 230 |
+
|
| 231 |
+
[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020. 3, 4
|
| 232 |
+
|
| 233 |
+
[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3, 15
|
| 234 |
+
|
| 235 |
+
[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 7
|
| 236 |
+
|
| 237 |
+
[23] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. arXiv preprint arXiv:2301.12661, 2023. 2, 3
|
| 238 |
+
|
| 239 |
+
[24] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions for audios in the wild. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 119–132, 2019. 6, 7, 8, 9, 10
|
| 240 |
+
|
| 241 |
+
[25] Eungbeom Kim, Jinhee Kim, Yoori Oh, Kyungsu Kim, Minju Park, Jaeheon Sim, Jinwoo Lee, and Kyogu Lee. Improving audio-language learning with mixgen and multi-level test-time augmentation. arXiv preprint arXiv:2210.17143, 2022. 7
|
| 242 |
+
|
| 243 |
+
[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 15
|
| 244 |
+
|
| 245 |
+
[27] Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis. Advances in Neural Information Processing Systems, 33:17022–17033, 2020. 5
|
| 246 |
+
|
| 247 |
+
[28] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song. Acav $1 0 0 \mathrm { m }$ : Automatic curation of large-scale datasets for audio-visual video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10274–10284, 2021. 7 [29] Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiujun Li, Yizhe Zhang, and Jianfeng Gao. Optimus: Organizing sentences via pre-trained modeling of a latent space. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4678–4699, 2020. 5
|
| 248 |
+
|
| 249 |
+
[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 7
|
| 250 |
+
|
| 251 |
+
[31] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXX 16, pages 121–137. Springer, 2020. 7
|
| 252 |
+
|
| 253 |
+
[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740–755. Springer, 2014. 7
|
| 254 |
+
|
| 255 |
+
[33] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint arXiv:2301.12503, 2023. 2, 3, 5
|
| 256 |
+
|
| 257 |
+
[34] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semanticconditional diffusion networks for image captioning. arXiv preprint arXiv:2212.03099, 2022. 7
|
| 258 |
+
|
| 259 |
+
[35] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3, 7
|
| 260 |
+
|
| 261 |
+
[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 7
|
| 262 |
+
|
| 263 |
+
[37] Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng. Automatic prompt optimization with" gradient descent" and beam search. arXiv preprint arXiv:2305.03495, 2023. 2
|
| 264 |
+
|
| 265 |
+
[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR, 2021. 3, 4
|
| 266 |
+
|
| 267 |
+
[39] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 5
|
| 268 |
+
|
| 269 |
+
[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3
|
| 270 |
+
|
| 271 |
+
[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 2, 3, 4, 5, 7
|
| 272 |
+
|
| 273 |
+
[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 6, 7
|
| 274 |
+
|
| 275 |
+
[43] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959–17968, 2022. 7
|
| 276 |
+
|
| 277 |
+
[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 7
|
| 278 |
+
|
| 279 |
+
[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256–2265. PMLR, 2015. 2, 4
|
| 280 |
+
|
| 281 |
+
[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3
|
| 282 |
+
|
| 283 |
+
[47] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3
|
| 284 |
+
|
| 285 |
+
[48] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 7
|
| 286 |
+
|
| 287 |
+
[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-tosequence learning framework. arXiv preprint arXiv:2202.03052, 2022. 7
|
| 288 |
+
|
| 289 |
+
[50] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 7
|
| 290 |
+
|
| 291 |
+
[51] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pre-training for neural visual world creation. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVI, pages 720–736. Springer, 2022. 7
|
| 292 |
+
|
| 293 |
+
[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 7
|
| 294 |
+
|
| 295 |
+
[53] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 5, 7
|
| 296 |
+
|
| 297 |
+
[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036–5045, 2022. 6, 7, 16
|
| 298 |
+
|
| 299 |
+
[55] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian, Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework. arXiv preprint arXiv:2205.01818, 2022. 2, 3
|
| 300 |
+
|
| 301 |
+
[56] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375–16387, 2022. 3
|
| 302 |
+
[57] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634–23651, 2021. 3
|
| 303 |
+
[58] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph with teacher-recommended learning for video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13278–13288, 2020. 7
|
| 304 |
+
[59] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang, Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint arXiv:2211.11694, 2022. 7
|
| 305 |
+
|
| 306 |
+
# A Model Architecture and Configuration
|
| 307 |
+
|
| 308 |
+
# A.1 Overview
|
| 309 |
+
|
| 310 |
+
In this section, we provide more details on the model architecture as shown in Table 12, where each modality specific diffuser is based on UNet architecture with different variations detailed in the table. Another notable difference is the video architecture where we add temporal attention and temporal shift as discussed in Section 3.3 and we will discuss its detail in the next section.
|
| 311 |
+
|
| 312 |
+
Table 12: Hyperparameters for our diffusion models. Note the video and image generation uses the same diffuser.
|
| 313 |
+
|
| 314 |
+
<table><tr><td>Modality</td><td>Video (Image) LDM</td><td>Audio LDM</td><td>Text LDM</td></tr><tr><td colspan="4">Hyperparameter</td></tr><tr><td>Architecture</td><td>LDM</td><td>LDM</td><td>LDM</td></tr><tr><td>z-shape</td><td>4× #frames × 64× 64</td><td>8× 256×16</td><td>768×1×1</td></tr><tr><td>Channels</td><td>320</td><td>320</td><td>320</td></tr><tr><td>Depth</td><td>4</td><td>2</td><td>2</td></tr><tr><td>Channel multiplier</td><td>1,2,4,4</td><td>1,2,4,4</td><td>1,2,4,4</td></tr><tr><td>Attention resolutions</td><td>64,32,16</td><td>64,32,16</td><td>64,32,16</td></tr><tr><td>Head channels</td><td>32</td><td>32</td><td>32</td></tr><tr><td>Number of heads</td><td>8</td><td>8</td><td>8</td></tr><tr><td>CA embed dim</td><td>768</td><td>768</td><td>768</td></tr><tr><td>CA resolutions</td><td>64,32,16</td><td>64,32,16</td><td>64,32,16</td></tr><tr><td>Autoencoders</td><td>AutoKL</td><td>AudioLDM</td><td>Optimus</td></tr><tr><td>Weight initialization</td><td>Stable Diffusion-1.4</td><td>-</td><td>Versatile Diffusion</td></tr><tr><td>Parameterization</td><td>E</td><td>E</td><td>E</td></tr><tr><td>Learning rate</td><td>2e-5</td><td>5e-6</td><td>5e-5</td></tr><tr><td>Total batch size</td><td>256</td><td>1024</td><td>1024</td></tr><tr><td colspan="4">Diffusion Setup</td></tr><tr><td>Diffusion steps</td><td>1000</td><td>1000</td><td>1000</td></tr><tr><td>Noise schedule</td><td>Linear</td><td>Linear</td><td>Linear</td></tr><tr><td>β</td><td>0.00085</td><td>0.00085</td><td>0.00085</td></tr><tr><td>阳</td><td>0.0120</td><td>0.0120</td><td>0.0120</td></tr><tr><td colspan="4">Sampling Parameters</td></tr><tr><td>Sampler</td><td>DDIM</td><td>DDIM</td><td>DDIM</td></tr><tr><td>Steps</td><td>50</td><td>50</td><td>50</td></tr><tr><td>n</td><td>1.0</td><td>1.0</td><td>1.0</td></tr><tr><td>Guidance scale</td><td>2.0</td><td>7.5</td><td>2.0</td></tr></table>
|
| 315 |
+
|
| 316 |
+
# A.2 Video LDM Architecture
|
| 317 |
+
|
| 318 |
+
Except for the base image UNet architecture, we also add temporal attention and temporal shift [2] before each residual block. Following VDM [21], the temporal attention is a transformer attention module where we flatten the height and width dimension to batch size dimension and the self-attention is performed on the time dimension. The temporal shift is illustrated in Fig. 6 where we first split channels into $k$ chunks. Then, we shift the channel dimension numbered 0 to $k - 1$ by temporal dimension from 0 to $k - 1$ times respectively. Eventually, we concatenate the shifted chunks by the hidden dimension. Note that we use $k = 3$ in the illustration for simplicity but $k = 8$ in our implementation. We then add a convolution layer before the temporal shift module. Finally, we use residual connection [18] and add the output to the input before the convolution layer. The complete video UNet layer is shown in Fig. 7.
|
| 319 |
+
|
| 320 |
+
# B Model Training
|
| 321 |
+
|
| 322 |
+
Prompt Encoders Training. As discussed in Section 3.2, we use bridging alignment to perform contrastive learning between all prompt encoders. We use Adam [26] optimizer with learning rate 1e-4 and weight decay 1e-4.
|
| 323 |
+
|
| 324 |
+
|
| 325 |
+
Figure 6: Temporal shift [2] illustration. $C , H .$ , $W$ represent channel, height, width, respectively. The vertical line represents time steps from $t - 1 , t$ , and $t + 1$ . The grey blocks denote “padding tensors”.
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
Figure 7: Video UNet layer architecture details including normalization & activation, 2D temporal attention, followed by temporal shift and 1D spatial convolution.
|
| 329 |
+
|
| 330 |
+
Diffusion Model Training. We train diffusion model with training objectives and hyperparameters detailed in Table 1 and Table 12. For video LDM, we adopt a more specific training curriculum. We adopt curriculum learning on frame resolution and frames-per-second (FPS). First, the diffuser is trained on the WebVid dataset of a 256-frame resolution, with the training objective being textconditioned video generation. The training clips are sampled from 2-second video chunks with 4 FPS. Second, the model is further trained on HDVILLA and ACAV datasets, with a 512-frame resolution and 8 FPS, and the training objective is image-conditioned video generation (the image is a randomly sampled frame of the clip). Each training clip contains 16 frames sampled from a 2-second video chunk with 8 FPS.
|
| 331 |
+
|
| 332 |
+
Joint Generation Training. As discussed in Section 3.2, we train joint generation by aligning environment encoders and optimize cross-attention layers only in the diffusion models. We use Adam optimizer with learning rate 1e-5 and weight decay 1e-4.
|
| 333 |
+
|
| 334 |
+
# C Training Datasets
|
| 335 |
+
|
| 336 |
+
In this section, we introduce more details about the video and audiovisual training datasets.
|
| 337 |
+
|
| 338 |
+
Video. WebVid [4] is a large-scale dataset of web videos with diverse content, spanning over 40 categories such as sports, cooking, and travel. It contains over 1.2 million video clips (all without sound) that are all at least 30 seconds in duration with video descriptions. We perform text video and video-text contrastive learning task with this dataset. HD-Villa-100M [54] is a large-scale video dataset with over 100 million video clips sourced from YouTube. The dataset covers a wide range of video categories and includes high-quality videos with a resolution of at least 720P. Since it lacks curated video description and we use the middle frame as image input to perform image video generation.
|
| 339 |
+
|
| 340 |
+
Audiovisual. SoundNet originally contains over two million sounds and spans a wide range of categories including music, animal sounds, natural sounds, and environmental sounds. We collected all currently accessible 1M videos.
|
| 341 |
+
|
| 342 |
+
# D Limitations & Broader Impacts
|
| 343 |
+
|
| 344 |
+
While the paper primarily focuses on the technical advancements and potential applications of CoDi, we also consider potential negative social impacts that could arise from the development and deployment of such technology. These impacts can include:
|
| 345 |
+
|
| 346 |
+
Deepfakes and Misinformation. As part of a common issue for generative AI models, the ability of CoDi to generate realistic and synchronized multimodal outputs also raises concerns about the creation and dissemination of deepfakes. Malicious actors could exploit this technology to create highly convincing fake content, such as fabricated videos or audio clips, which can be used for misinformation, fraud, or other harmful purposes.
|
| 347 |
+
|
| 348 |
+
Bias and Stereotyping. If the training data used for CoDi is biased or contains stereotypes, the generated multimodal outputs may also reflect these.
|
| 349 |
+
|
| 350 |
+
# E License
|
| 351 |
+
|
| 352 |
+
We will publicly release our code and checkpoints. We cite licenses from the individual dataset or package we use from the community and provide the following links for references.
|
| 353 |
+
|
| 354 |
+
LAION-400M: Creative Common CC-BY 4.0
|
| 355 |
+
|
| 356 |
+
AudioSet: Creative Common CC-BY 4.0
|
| 357 |
+
|
| 358 |
+
AudioCaps: MIT
|
| 359 |
+
|
| 360 |
+
Freesound: Creative Commons
|
| 361 |
+
|
| 362 |
+
BBC Sound Effect: The BBC’s Content Licence
|
| 363 |
+
|
| 364 |
+
SoundNet: MIT
|
| 365 |
+
|
| 366 |
+
Webvid10M: Webvid
|
| 367 |
+
|
| 368 |
+
HD-Villa-100M: Research Use of Data Agreement v1.0
|
| 369 |
+
|
| 370 |
+
PyTorch: BSD-style
|
| 371 |
+
|
| 372 |
+
Huggingface Transformers: Apache
|
| 373 |
+
|
| 374 |
+
Torchvision: BSD 3-Clause
|
| 375 |
+
|
| 376 |
+
Torchaudio: BSD 2-Clause
|
parse/dev/2EDqbSCnmF/2EDqbSCnmF_content_list.json
ADDED
|
@@ -0,0 +1,2070 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Any-to-Any Generation via Composable Diffusion ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
194,
|
| 8 |
+
122,
|
| 9 |
+
803,
|
| 10 |
+
147
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Zineng Tang1∗ ",
|
| 17 |
+
"text_level": 1,
|
| 18 |
+
"bbox": [
|
| 19 |
+
184,
|
| 20 |
+
200,
|
| 21 |
+
285,
|
| 22 |
+
215
|
| 23 |
+
],
|
| 24 |
+
"page_idx": 0
|
| 25 |
+
},
|
| 26 |
+
{
|
| 27 |
+
"type": "text",
|
| 28 |
+
"text": "Mohit Bansal1† ",
|
| 29 |
+
"text_level": 1,
|
| 30 |
+
"bbox": [
|
| 31 |
+
705,
|
| 32 |
+
200,
|
| 33 |
+
813,
|
| 34 |
+
215
|
| 35 |
+
],
|
| 36 |
+
"page_idx": 0
|
| 37 |
+
},
|
| 38 |
+
{
|
| 39 |
+
"type": "text",
|
| 40 |
+
"text": "Ziyi Yang2† Chenguang ${ \\bf Z } { \\bf h } { \\bf u } ^ { 2 \\ddagger }$ Michael Zeng2 1University of North Carolina at Chapel Hill 2Microsoft Azure Cognitive Services Research https://codi-gen.github.io ",
|
| 41 |
+
"bbox": [
|
| 42 |
+
310,
|
| 43 |
+
199,
|
| 44 |
+
676,
|
| 45 |
+
258
|
| 46 |
+
],
|
| 47 |
+
"page_idx": 0
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"type": "text",
|
| 51 |
+
"text": "Abstract ",
|
| 52 |
+
"text_level": 1,
|
| 53 |
+
"bbox": [
|
| 54 |
+
462,
|
| 55 |
+
294,
|
| 56 |
+
535,
|
| 57 |
+
310
|
| 58 |
+
],
|
| 59 |
+
"page_idx": 0
|
| 60 |
+
},
|
| 61 |
+
{
|
| 62 |
+
"type": "text",
|
| 63 |
+
"text": "We present Composable Diffusion (CoDi), a novel generative model capable of generating any combination of output modalities, such as language, image, video, or audio, from any combination of input modalities. Unlike existing generative AI systems, CoDi can generate multiple modalities in parallel and its input is not limited to a subset of modalities like text or image. Despite the absence of training datasets for many combinations of modalities, we propose to align modalities in both the input and output space. This allows CoDi to freely condition on any input combination and generate any group of modalities, even if they are not present in the training data. CoDi employs a novel composable generation strategy which involves building a shared multimodal space by bridging alignment in the diffusion process, enabling the synchronized generation of intertwined modalities, such as temporally aligned video and audio. Highly customizable and flexible, CoDi achieves strong joint-modality generation quality, and outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis. The project page with demonstrations and code is at https://codi-gen.github.io/ ",
|
| 64 |
+
"bbox": [
|
| 65 |
+
232,
|
| 66 |
+
329,
|
| 67 |
+
766,
|
| 68 |
+
535
|
| 69 |
+
],
|
| 70 |
+
"page_idx": 0
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"type": "image",
|
| 74 |
+
"img_path": "images/f6d7a60ecba1b89ddec9c2b8b8bfea37a1e8ff89e67fccf04a7b8dcb4427834d.jpg",
|
| 75 |
+
"image_caption": [
|
| 76 |
+
"Figure 1: CoDi can generate various (joint) combinations of output modalities from diverse (joint) sets of inputs: video, image, audio, and text (example combinations depicted by the colored arrows). "
|
| 77 |
+
],
|
| 78 |
+
"image_footnote": [],
|
| 79 |
+
"bbox": [
|
| 80 |
+
176,
|
| 81 |
+
565,
|
| 82 |
+
820,
|
| 83 |
+
773
|
| 84 |
+
],
|
| 85 |
+
"page_idx": 0
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"type": "text",
|
| 89 |
+
"text": "1 Introduction ",
|
| 90 |
+
"text_level": 1,
|
| 91 |
+
"bbox": [
|
| 92 |
+
174,
|
| 93 |
+
89,
|
| 94 |
+
310,
|
| 95 |
+
106
|
| 96 |
+
],
|
| 97 |
+
"page_idx": 1
|
| 98 |
+
},
|
| 99 |
+
{
|
| 100 |
+
"type": "text",
|
| 101 |
+
"text": "Recent years have seen the rise of powerful cross-modal models that can generate one modality from another, e.g. text-to-text [6, 37], text-to-image [13, 19, 22, 41, 44], or text-to-audio [23, 33]. However, these models are restricted in their real-world applicability where multiple modalities coexist and interact. While one can chain together modality-specific generative models in a multi-step generation setting, the generation power of each step remains inherently limited, and a serial, multistep process can be cumbersome and slow. Moreover, independently generated unimodal streams will not be consistent and aligned when stitched together in a post-processing way (e.g., synchronized video and audio). The development of a comprehensive and versatile model that can generate any combination of modalities from any set of input conditions has been eagerly anticipated, as it would more accurately capture the multimodal nature of the world and human comprehension, seamlessly consolidate information from a wide range of sources, and enable strong immersion in human-AI interactions (for example, by generating coherent video, audio, and text description at the same time). ",
|
| 102 |
+
"bbox": [
|
| 103 |
+
174,
|
| 104 |
+
122,
|
| 105 |
+
825,
|
| 106 |
+
289
|
| 107 |
+
],
|
| 108 |
+
"page_idx": 1
|
| 109 |
+
},
|
| 110 |
+
{
|
| 111 |
+
"type": "text",
|
| 112 |
+
"text": "In pursuit of this goal, we propose Composable Diffusion, or CoDi, the first model capable of simultaneously processing and generating arbitrary combinations of modalities as shown in Fig. 1. Training a model to take any mixture of input modalities and flexibly generate any mixture of outputs presents significant computational and data requirements, as the number of combinations for the input and output modalities scales exponentially. Also aligned training data for many groups of modalities is scarce or even non-existent, making it infeasible to train with all possible input-output combinations. To address this challenge, we propose to align multiple modalities in both the input conditioning (Section 3.2) and generation diffusion step (Section 3.4). Furthermore, a proposed “Bridging Alignment” strategy for contrastive learning (Section 3.2) allows us to efficiently model the exponential number of input-output combinations with a linear number of training objectives. ",
|
| 113 |
+
"bbox": [
|
| 114 |
+
174,
|
| 115 |
+
295,
|
| 116 |
+
825,
|
| 117 |
+
433
|
| 118 |
+
],
|
| 119 |
+
"page_idx": 1
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"type": "text",
|
| 123 |
+
"text": "Building a model with any-to-any generation capacity with exceptional generation quality requires comprehensive model design and training on diverse data resources. Therefore, we build CoDi in an integrative way. First, we train a latent diffusion model (LDM) for each modality, e.g., text, image, video, and audio. These models can be trained in parallel independently, ensuring exceptional singlemodality generation quality using widely available modality-specific training data (i.e., data with one or more modalities as input and one modality as output). For conditional cross-modality generation, such as generating images using audio+language prompts, the input modalities are projected into a shared feature space (Section 3.2), and the output LDM attends to the combination of input features. This multimodal conditioning mechanism prepares the diffusion model to condition on any modality or combination of modalities without directly training for such settings. ",
|
| 124 |
+
"bbox": [
|
| 125 |
+
174,
|
| 126 |
+
439,
|
| 127 |
+
826,
|
| 128 |
+
578
|
| 129 |
+
],
|
| 130 |
+
"page_idx": 1
|
| 131 |
+
},
|
| 132 |
+
{
|
| 133 |
+
"type": "text",
|
| 134 |
+
"text": "The second stage of training enables the model to handle many-to-many generation strategies that involve simultaneously generating arbitrary combinations of output modalities. To the best of our knowledge, CoDi is the first AI model with this capability. This is achieved by adding a crossattention module to each diffuser, and an environment encoder $V$ to project the latent variable of different LDMs into a shared latent space (Section 3.4). Next, we freeze the parameters of the LDM, training only the cross-attention parameters and $V$ . Since the environment encoder of different modalities are aligned, an LDM can cross-attend with any group of co-generated modalities by interpolating the representation’s output by $V$ . This enables CoDi to seamlessly generate any group of modalities, without training on all possible generation combinations. This reduces the number of training objectives from exponential to linear. ",
|
| 135 |
+
"bbox": [
|
| 136 |
+
174,
|
| 137 |
+
584,
|
| 138 |
+
825,
|
| 139 |
+
722
|
| 140 |
+
],
|
| 141 |
+
"page_idx": 1
|
| 142 |
+
},
|
| 143 |
+
{
|
| 144 |
+
"type": "text",
|
| 145 |
+
"text": "We demonstrate the any-to-any generation capability of CoDi, including single-to-single modality generation, multi-condition generation, and the novel capacity of joint generation of multiple modalities. For example, generating synchronized video and audio given the text input prompt; or generating video given a prompt image and audio. We also provide a quantitative evaluation of CoDi using eight multimodal datasets. As the latest work from Project i-Code [55] towards Composable AI, CoDi exhibits exceptional generation quality across assorted scenarios, with synthesis quality on par or even better than single to single modality SOTA, e.g., audio generation and audio captioning. ",
|
| 146 |
+
"bbox": [
|
| 147 |
+
174,
|
| 148 |
+
729,
|
| 149 |
+
825,
|
| 150 |
+
827
|
| 151 |
+
],
|
| 152 |
+
"page_idx": 1
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"type": "text",
|
| 156 |
+
"text": "2 Related Works ",
|
| 157 |
+
"text_level": 1,
|
| 158 |
+
"bbox": [
|
| 159 |
+
176,
|
| 160 |
+
849,
|
| 161 |
+
328,
|
| 162 |
+
866
|
| 163 |
+
],
|
| 164 |
+
"page_idx": 1
|
| 165 |
+
},
|
| 166 |
+
{
|
| 167 |
+
"type": "text",
|
| 168 |
+
"text": "Diffusion models (DMs) learn the data distribution by denoising and recovering the original data. Deep Diffusion Process (DDP) [45] adopts a sequence of reversible diffusion steps to model image probability distribution. It uses a reversible encoder to map the input image to a latent space and a decoder to map the latent variables to an output image. Denoising diffusion probabilistic model (DDPM) [20] uses a cascade of diffusion processes to gradually increase the complexity of the probability density function model. At each step, the model adds noise to the input image and estimates the corresponding noise level using an autoregressive model. This allows the model to capture the dependencies between adjacent pixels and generate high-quality images. Score-based generative models (SOG) [46] use the score function to model the diffusion process. [40] generates high-fidelity images conditioned on CLIP representations of text prompts. Latent diffusion model (LDM) [41] uses a VAE to encode inputs into latent space to reduce modeling dimension and improves efficiency. The motivation is that image compression can be separated into semantic space by a diffusion model and perceptual space by an autoencoder. By incorporating temporal modeling modules and cascading model architectures, video diffusion models have been built upon image diffusers to generate temporally consistent and inherent frames[14, 19, 21, 44]. Diffusion models have also been applied to other domains, such as generating audio from text and vision prompts[23, 33]. ",
|
| 169 |
+
"bbox": [
|
| 170 |
+
174,
|
| 171 |
+
882,
|
| 172 |
+
823,
|
| 173 |
+
911
|
| 174 |
+
],
|
| 175 |
+
"page_idx": 1
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"type": "image",
|
| 179 |
+
"img_path": "images/fd34f6f54041b6f9329e8b6d950176f23a2bd0ce1819fa8d9711a4b7d903ec66.jpg",
|
| 180 |
+
"image_caption": [
|
| 181 |
+
"Figure 2: CoDi model architecture: (a) We first train individual diffusion model with aligned prompt encoder by “Bridging Alignment”; (b) Diffusion models learn to attend with each other via “Latent Alignment”; (c) CoDi achieves any-to-any generation with a linear number of training objectives. "
|
| 182 |
+
],
|
| 183 |
+
"image_footnote": [],
|
| 184 |
+
"bbox": [
|
| 185 |
+
173,
|
| 186 |
+
87,
|
| 187 |
+
821,
|
| 188 |
+
415
|
| 189 |
+
],
|
| 190 |
+
"page_idx": 2
|
| 191 |
+
},
|
| 192 |
+
{
|
| 193 |
+
"type": "text",
|
| 194 |
+
"text": "",
|
| 195 |
+
"bbox": [
|
| 196 |
+
174,
|
| 197 |
+
496,
|
| 198 |
+
825,
|
| 199 |
+
689
|
| 200 |
+
],
|
| 201 |
+
"page_idx": 2
|
| 202 |
+
},
|
| 203 |
+
{
|
| 204 |
+
"type": "text",
|
| 205 |
+
"text": "Multimodal modeling has experienced rapid advancement recently, with researchers striving to build uniform representations of multiple modalities using a single model to achieve more comprehensive cross-modal understanding. Vision transformers [11], featuring diverse model architectures and training techniques, have been applied to various downstream tasks such as vision Q&A and image captioning. Multimodal encoders have also proven successful in vision-language [1, 8, 57], videoaudio [47] and video-speech-language [55, 56] domains. Aligning data from different modalities is an active research area [12, 38], with promising applications in cross-modality retrieval and building uniform multimodal representations [33, 35, 41]. ",
|
| 206 |
+
"bbox": [
|
| 207 |
+
174,
|
| 208 |
+
695,
|
| 209 |
+
825,
|
| 210 |
+
806
|
| 211 |
+
],
|
| 212 |
+
"page_idx": 2
|
| 213 |
+
},
|
| 214 |
+
{
|
| 215 |
+
"type": "text",
|
| 216 |
+
"text": "3 Methodology ",
|
| 217 |
+
"text_level": 1,
|
| 218 |
+
"bbox": [
|
| 219 |
+
174,
|
| 220 |
+
825,
|
| 221 |
+
313,
|
| 222 |
+
843
|
| 223 |
+
],
|
| 224 |
+
"page_idx": 2
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"type": "text",
|
| 228 |
+
"text": "3.1 Preliminary: Latent Diffusion Model ",
|
| 229 |
+
"text_level": 1,
|
| 230 |
+
"bbox": [
|
| 231 |
+
174,
|
| 232 |
+
857,
|
| 233 |
+
470,
|
| 234 |
+
872
|
| 235 |
+
],
|
| 236 |
+
"page_idx": 2
|
| 237 |
+
},
|
| 238 |
+
{
|
| 239 |
+
"type": "text",
|
| 240 |
+
"text": "Diffusion models (DM) represent a class of generative models that learn data distributions $p ( { \\pmb x } )$ by simulating the diffusion of information over time. During training, random noise is iteratively added to $_ { \\textbf { \\em x } }$ , while the model learns to denoise the examples. For inference, the model denoises data points sampled from simple distributions such as Gaussian. Latent diffusion models (LDM) [41] learn the distribution of the latent variable $_ { z }$ corresponding to $_ { \\textbf { \\em x } }$ , significantly reducing computational cost by decreasing the data dimension. ",
|
| 241 |
+
"bbox": [
|
| 242 |
+
174,
|
| 243 |
+
883,
|
| 244 |
+
821,
|
| 245 |
+
911
|
| 246 |
+
],
|
| 247 |
+
"page_idx": 2
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"type": "text",
|
| 251 |
+
"text": "",
|
| 252 |
+
"bbox": [
|
| 253 |
+
174,
|
| 254 |
+
90,
|
| 255 |
+
825,
|
| 256 |
+
147
|
| 257 |
+
],
|
| 258 |
+
"page_idx": 3
|
| 259 |
+
},
|
| 260 |
+
{
|
| 261 |
+
"type": "text",
|
| 262 |
+
"text": "In LDM, an autoencoder is first trained to reconstruct $_ { \\textbf { \\em x } }$ , i.e., $\\hat { \\pmb { x } } = D ( E ( \\pmb { x } ) )$ , where $E$ and $D$ denote the encoder and decoder, respectively. The latent variable $z = E ( { \\pmb x } )$ is iteratively diffused over time steps $t$ based on a variance schedule $\\beta _ { 1 } , \\ldots , \\beta _ { T }$ , i.e., $q ( z _ { t } | z _ { t - 1 } ) = \\mathcal { N } ( z _ { t } ; \\sqrt { 1 - \\beta _ { t } } z _ { t - 1 } , \\beta _ { t } I )$ [20, 45]. ",
|
| 263 |
+
"bbox": [
|
| 264 |
+
173,
|
| 265 |
+
152,
|
| 266 |
+
825,
|
| 267 |
+
209
|
| 268 |
+
],
|
| 269 |
+
"page_idx": 3
|
| 270 |
+
},
|
| 271 |
+
{
|
| 272 |
+
"type": "text",
|
| 273 |
+
"text": "The forward process allows the random sampling of ${ \\boldsymbol { z } } _ { t }$ at any timestep in a closed form [20, 45]: $\\boldsymbol { z } _ { t } = \\alpha _ { t } \\boldsymbol { z } + \\sigma _ { t } \\boldsymbol { \\epsilon }$ , where $\\epsilon \\sim \\mathcal { N } ( 0 , I )$ , $\\alpha _ { t } : = 1 - \\beta _ { t }$ and $\\begin{array} { r } { \\sigma _ { t } : = \\dot { 1 } - \\prod _ { s = 1 } ^ { t } \\dot { \\alpha } _ { s } } \\end{array}$ . The diffuser learns how to denoise from $\\left\\{ { z } _ { t } \\right\\}$ to recover $_ z$ . Following the reparameterization method proposed in [20], the denoising training objective can be expressed as [41]: ",
|
| 274 |
+
"bbox": [
|
| 275 |
+
173,
|
| 276 |
+
215,
|
| 277 |
+
825,
|
| 278 |
+
273
|
| 279 |
+
],
|
| 280 |
+
"page_idx": 3
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"type": "equation",
|
| 284 |
+
"img_path": "images/c0b6a42a7dddc81d5d528ed8a349c84f1a8e8e52d97b6eb0c26878eaf69606c7.jpg",
|
| 285 |
+
"text": "$$\n\\begin{array} { r } { \\mathcal { L } _ { D } = \\mathbb { E } _ { z , \\epsilon , t } \\Vert \\epsilon - \\epsilon _ { \\theta } ( z _ { t } , t , C ( \\pmb { y } ) ) \\Vert _ { 2 } ^ { 2 } . } \\end{array}\n$$",
|
| 286 |
+
"text_format": "latex",
|
| 287 |
+
"bbox": [
|
| 288 |
+
377,
|
| 289 |
+
279,
|
| 290 |
+
620,
|
| 291 |
+
297
|
| 292 |
+
],
|
| 293 |
+
"page_idx": 3
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"type": "text",
|
| 297 |
+
"text": "In data generation, the denoising process can be realized through reparameterized Gaussian sampling: ",
|
| 298 |
+
"bbox": [
|
| 299 |
+
174,
|
| 300 |
+
304,
|
| 301 |
+
826,
|
| 302 |
+
320
|
| 303 |
+
],
|
| 304 |
+
"page_idx": 3
|
| 305 |
+
},
|
| 306 |
+
{
|
| 307 |
+
"type": "equation",
|
| 308 |
+
"img_path": "images/cd64023bfb25f4aca1f1f00e845397403302ab368fd5956a291fe57286f62be3.jpg",
|
| 309 |
+
"text": "$$\np ( z _ { t - 1 } | z _ { t } ) = \\mathcal { N } \\left( z _ { t - 1 } ; \\frac { 1 } { \\sqrt { \\alpha _ { t } } } \\left( z _ { t } - \\frac { \\beta _ { t } } { \\sqrt { \\sigma _ { t } } } \\epsilon _ { \\theta } \\right) , \\beta _ { t } I \\right) .\n$$",
|
| 310 |
+
"text_format": "latex",
|
| 311 |
+
"bbox": [
|
| 312 |
+
316,
|
| 313 |
+
329,
|
| 314 |
+
679,
|
| 315 |
+
364
|
| 316 |
+
],
|
| 317 |
+
"page_idx": 3
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"type": "text",
|
| 321 |
+
"text": "In $\\mathcal { L } _ { D }$ , the diffusion time step $t \\sim \\mathcal { U } [ 1 , T ]$ ; $\\epsilon _ { \\theta }$ is a denoising model with UNet backbone parameterized by $\\theta ; { \\boldsymbol { y } }$ represents the conditional variable that can be used to control generation; $C$ is the prompt encoder. The conditioning mechanism is implemented by first featurizing $\\textbf { { y } }$ into $C ( \\boldsymbol { y } )$ , then the UNet $\\epsilon _ { \\theta }$ conditions on $C ( \\boldsymbol { y } )$ via cross-attention, as described in [41]. Distinct from previous works, our model can condition on any combinations of modalities of text, image, video and audio. Details are presented in the following section. ",
|
| 322 |
+
"bbox": [
|
| 323 |
+
173,
|
| 324 |
+
368,
|
| 325 |
+
826,
|
| 326 |
+
452
|
| 327 |
+
],
|
| 328 |
+
"page_idx": 3
|
| 329 |
+
},
|
| 330 |
+
{
|
| 331 |
+
"type": "text",
|
| 332 |
+
"text": "3.2 Composable Multimodal Conditioning ",
|
| 333 |
+
"text_level": 1,
|
| 334 |
+
"bbox": [
|
| 335 |
+
176,
|
| 336 |
+
469,
|
| 337 |
+
480,
|
| 338 |
+
484
|
| 339 |
+
],
|
| 340 |
+
"page_idx": 3
|
| 341 |
+
},
|
| 342 |
+
{
|
| 343 |
+
"type": "text",
|
| 344 |
+
"text": "To enable our model to condition on any combination of input/prompt modalities, we align the prompt encoder of text, image, video and audio (denoted by $C _ { t }$ , $C _ { i }$ , $C _ { v }$ , and $C _ { a }$ , respectively) to project the input from any modality into the same space. Multimodal conditioning can then be conveniently achieved by interpolating the representations of each modality $m$ : $\\begin{array} { r } { C ( x _ { t } , \\bar { x _ { i } } , x _ { v } , x _ { a } ) = \\sum _ { m } \\alpha _ { m } C ( \\bar { m } ) } \\end{array}$ for $m \\in \\ b { x } _ { t } , \\ b { x } _ { i } , \\ b { x } _ { v } , \\ b { x } _ { a }$ , with $\\textstyle \\sum _ { m } \\alpha _ { m } = 1$ . Through simple weighted interpolation of aligned embeddings, we enable models trained with single-conditioning (i.e., with only one input) to perform zero-shot multi-conditioning (i.e., with multiple inputs). This process is illustrated in Fig. 2 (a)(2). ",
|
| 345 |
+
"bbox": [
|
| 346 |
+
173,
|
| 347 |
+
494,
|
| 348 |
+
825,
|
| 349 |
+
593
|
| 350 |
+
],
|
| 351 |
+
"page_idx": 3
|
| 352 |
+
},
|
| 353 |
+
{
|
| 354 |
+
"type": "text",
|
| 355 |
+
"text": "Optimizing all four prompt encoders simultaneously in a combinatorial manner is computationally heavy, with $\\mathcal { O } ( n ^ { 2 } )$ pairs. Additionally, for certain dual modalities, well-aligned paired datasets are limited or unavailable e.g., image-audio pairs. To address this challenge, we propose a simple and effective technique called \"Bridging Alignment\" to efficiently align conditional encoders. As shown in Fig. 2 (a)(1), we choose the text modality as the \"bridging\" modality due to its ubiquitous presence in paired data, such as text-image, text-video, and text-audio pairs. We begin with a pretrained text-image paired encoder, i.e., CLIP [38]. We then train audio and video prompt encoders on audio-text and video-text paired datasets using contrastive learning, with text and image encoder weights frozen. ",
|
| 356 |
+
"bbox": [
|
| 357 |
+
173,
|
| 358 |
+
598,
|
| 359 |
+
825,
|
| 360 |
+
723
|
| 361 |
+
],
|
| 362 |
+
"page_idx": 3
|
| 363 |
+
},
|
| 364 |
+
{
|
| 365 |
+
"type": "text",
|
| 366 |
+
"text": "In this way, all four modalities are aligned in the feature space. As shown in Section 5.2, CoDi can effectively leverage and combine the complementary information present in any combination of modalities to generate more accurate and comprehensive outputs. The high generation quality remains unaffected with respect to the number of prompt modalities. As we will discuss in subsequent sections, we continue to apply Bridging Alignment to align the latent space of LDMs with different modalities to achieve joint multimodal generation. ",
|
| 367 |
+
"bbox": [
|
| 368 |
+
174,
|
| 369 |
+
729,
|
| 370 |
+
825,
|
| 371 |
+
813
|
| 372 |
+
],
|
| 373 |
+
"page_idx": 3
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"type": "text",
|
| 377 |
+
"text": "3.3 Composable Diffusion ",
|
| 378 |
+
"text_level": 1,
|
| 379 |
+
"bbox": [
|
| 380 |
+
174,
|
| 381 |
+
829,
|
| 382 |
+
367,
|
| 383 |
+
844
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 3
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "text",
|
| 389 |
+
"text": "Training an end-to-end anything-to-anything model requires extensive learning on various data resources. The model also needs to maintain generation quality for all synthesis flows. To address these challenges, CoDi is designed to be composable and integrative, allowing individual modalityspecific models to be built independently and then smoothly integrated later. Specifically, we start by independently training image, video, audio, and text LDMs. These diffusion models then efficiently learn to attend across modalities for joint multimodal generation (Section 3.4) by a novel mechanism named “latent alignment”. ",
|
| 390 |
+
"bbox": [
|
| 391 |
+
174,
|
| 392 |
+
856,
|
| 393 |
+
825,
|
| 394 |
+
911
|
| 395 |
+
],
|
| 396 |
+
"page_idx": 3
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"type": "text",
|
| 400 |
+
"text": "",
|
| 401 |
+
"bbox": [
|
| 402 |
+
176,
|
| 403 |
+
90,
|
| 404 |
+
823,
|
| 405 |
+
132
|
| 406 |
+
],
|
| 407 |
+
"page_idx": 4
|
| 408 |
+
},
|
| 409 |
+
{
|
| 410 |
+
"type": "text",
|
| 411 |
+
"text": "Image Diffusion Model. The image LDM follows the same structure as Stable Diffusion 1.5 [41] and is initialized with the same weights. Reusing the weights transfers the knowledge and exceptional generation fidelity of Stable Diffusion trained on large-scale high-quality image datasets to CoDi. ",
|
| 412 |
+
"bbox": [
|
| 413 |
+
174,
|
| 414 |
+
146,
|
| 415 |
+
825,
|
| 416 |
+
189
|
| 417 |
+
],
|
| 418 |
+
"page_idx": 4
|
| 419 |
+
},
|
| 420 |
+
{
|
| 421 |
+
"type": "text",
|
| 422 |
+
"text": "Video Diffusion Model. To model the temporal properties of videos and simultaneously maintain vision generation quality, we construct the video diffuser by extending the image diffuser with temporal modules. Specifically, we insert pseudo-temporal attention before the residual block [13]. However, we argue that pseudo-temporal attention only enables video frames to globally attend to each other by flattening the pixels (height, width dimension) to batch dimension, resulting in a lack of cross-frame interaction between local pixels. We argue that this results in the common temporal-inconsistency issue in video generation that locations, shapes, colors, etc. of objects can be inconsistent across generated frames. To address this problem, we propose adapting the latent shift method [2] that performs temporal-spatial shifts on latent features in accordance with temporal attention. We divide the video by the hidden dimension into $k = 8$ chunks, and for each chunk $i = 0$ to 7, we shift the temporal dimension forward by $i$ positions. Further details will be provided in the appendix. ",
|
| 423 |
+
"bbox": [
|
| 424 |
+
173,
|
| 425 |
+
203,
|
| 426 |
+
825,
|
| 427 |
+
368
|
| 428 |
+
],
|
| 429 |
+
"page_idx": 4
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"type": "text",
|
| 433 |
+
"text": "Audio Diffusion Model. To enable flexible cross-modality attention in joint generation, the audio diffuser is designed to have a similar architecture to vision diffusers, where the mel-spectrogram can be naturally viewed as an image with 1 channel. We use a VAE encoder to encode the melspectrogram of audio to a compressed latent space. In audio synthesis, a VAE decoder maps the latent variable to the mel-spectrogram, and a vocoder generates the audio sample from the mel-spectrogram. We employ the audio VAE from [33] and the vocoder from [27]. ",
|
| 434 |
+
"bbox": [
|
| 435 |
+
174,
|
| 436 |
+
382,
|
| 437 |
+
825,
|
| 438 |
+
465
|
| 439 |
+
],
|
| 440 |
+
"page_idx": 4
|
| 441 |
+
},
|
| 442 |
+
{
|
| 443 |
+
"type": "text",
|
| 444 |
+
"text": "Text Diffusion Model. The VAE of the text LDM is OPTIMUS [29], and its encoder and decoder are [9] and GPT-2 [39], respectively. For the denoising UNet, unlike the one in image diffusion, the 2D convolution in residual blocks is replaced with 1D convolution [53]. ",
|
| 445 |
+
"bbox": [
|
| 446 |
+
174,
|
| 447 |
+
479,
|
| 448 |
+
823,
|
| 449 |
+
522
|
| 450 |
+
],
|
| 451 |
+
"page_idx": 4
|
| 452 |
+
},
|
| 453 |
+
{
|
| 454 |
+
"type": "text",
|
| 455 |
+
"text": "3.4 Joint Multimodal Generation by Latent Alignment ",
|
| 456 |
+
"text_level": 1,
|
| 457 |
+
"bbox": [
|
| 458 |
+
174,
|
| 459 |
+
536,
|
| 460 |
+
565,
|
| 461 |
+
551
|
| 462 |
+
],
|
| 463 |
+
"page_idx": 4
|
| 464 |
+
},
|
| 465 |
+
{
|
| 466 |
+
"type": "text",
|
| 467 |
+
"text": "The final step is to enable cross-attention between diffusion flows in joint generation, i.e., generating two or more modalities simultaneously. This is achieved by adding cross-modal attention sublayers to the UNet $\\epsilon _ { \\theta }$ (Fig. 2 (b)(2)). Specifically, consider a diffusion model of modality $A$ that cross-attends with another modality $B$ . Let the latent variables of modalities $m _ { A }$ and $m _ { B }$ at diffusion step $t$ be denoted as $ { \\boldsymbol { z } } _ { t } ^ { A }$ and $\\hat { z _ { t } ^ { B } }$ , respectively. The proposed “Latent Alignment” technique is such that a modality-specific environment encoder $V _ { B }$ first projects $ { \\boldsymbol { z } } _ { t } ^ { B }$ into a shared latent space for different modalities. Then, in each layer of the UNet for modality $A$ , a cross-attention sublayer attends to $V _ { B } \\big ( z _ { t } ^ { B } \\big )$ . For the diffusion model of modality $A$ , the training objective in Eq. (1) now becomes: ",
|
| 468 |
+
"bbox": [
|
| 469 |
+
173,
|
| 470 |
+
561,
|
| 471 |
+
825,
|
| 472 |
+
674
|
| 473 |
+
],
|
| 474 |
+
"page_idx": 4
|
| 475 |
+
},
|
| 476 |
+
{
|
| 477 |
+
"type": "equation",
|
| 478 |
+
"img_path": "images/543f8049db9dbb90f598c7dea6ef4aa9f7c34c5bc297bb830ab29bb93064c182.jpg",
|
| 479 |
+
"text": "$$\n\\mathcal { L } _ { C r o s s } ^ { A } = \\mathbb { E } _ { z , \\epsilon , t } \\Vert \\epsilon - \\epsilon _ { \\theta _ { c } } ( z _ { t } ^ { A } , V _ { B } ( z _ { t } ^ { B } ) , t , C ( \\pmb { y } ) ) \\Vert _ { 2 } ^ { 2 } ,\n$$",
|
| 480 |
+
"text_format": "latex",
|
| 481 |
+
"bbox": [
|
| 482 |
+
328,
|
| 483 |
+
690,
|
| 484 |
+
668,
|
| 485 |
+
710
|
| 486 |
+
],
|
| 487 |
+
"page_idx": 4
|
| 488 |
+
},
|
| 489 |
+
{
|
| 490 |
+
"type": "text",
|
| 491 |
+
"text": "where $\\theta _ { c }$ denotes the weights of cross-attention modules in the UNet. ",
|
| 492 |
+
"bbox": [
|
| 493 |
+
169,
|
| 494 |
+
717,
|
| 495 |
+
630,
|
| 496 |
+
731
|
| 497 |
+
],
|
| 498 |
+
"page_idx": 4
|
| 499 |
+
},
|
| 500 |
+
{
|
| 501 |
+
"type": "text",
|
| 502 |
+
"text": "The training objective of $A + B$ joint generation is $\\mathcal { L } _ { C r o s s } ^ { A } + \\mathcal { L } _ { C r o s s } ^ { B }$ . $V ( \\cdot )$ of different modalities are trained to be aligned with contrastive learning. Since $z _ { t } ^ { A }$ and $z _ { t } ^ { B }$ at any time step can be sampled with closed form in the diffusion process Section 3.1, one can conveniently train the contrastive learning together with $\\mathcal { L } _ { C r o s s }$ . The purpose of $V$ is to achieve the generation of any combination of modalities (in polynomial) by training on a linear number of joint-generation tasks. For example, if we have trained the joint generation of modalities $A , B$ , and $B$ , $C$ independently, then we have $V _ { A } ( z _ { t } ^ { A } )$ , $V _ { B } \\big ( z _ { t } ^ { B } \\big )$ , and $V _ { C } ( z _ { t } ^ { C } )$ aligned. Therefore, CoDi can seamlessly achieve joint generation of modalities $A$ and $C$ without any additional training. Moreover, such design automatically effortlessly enables joint generation of modalities $A$ , $B$ , and $C$ concurrently. Specifically, UNet of $A$ can cross-attend with the interpolation of $V _ { B } \\big ( z _ { t } ^ { B } \\big )$ , and $V _ { C } ( z _ { t } ^ { C } )$ , although CoDi has not been trained with such task. ",
|
| 503 |
+
"bbox": [
|
| 504 |
+
173,
|
| 505 |
+
736,
|
| 506 |
+
825,
|
| 507 |
+
877
|
| 508 |
+
],
|
| 509 |
+
"page_idx": 4
|
| 510 |
+
},
|
| 511 |
+
{
|
| 512 |
+
"type": "text",
|
| 513 |
+
"text": "As shown in Fig. 2(b)(3), we follow similar designs to the \"Bridging Alignment\" in training joint generation: (1) We first train the cross-attention weights in the image and text diffusers, as well as their environment encoders $V$ , on text-image paired data. (2) We freeze the weights of the text diffuser and train the environment encoder and cross-attention weights of the audio diffuser on text-audio paired data. (3) Finally we freeze the audio diffuser and its environment encoder, and train the joint generation of the video modality on audio-video paired data. As demonstrated in Section 5.3, although only trained on three paired joint generation tasks (i.e, Text $^ +$ Audio, Text+Image, and Video+Audio), CoDi is capable of generating assorted combinations of modalities simultaneously that are unseen in training, e.g., joint image-text-audio generation in Fig. 5. ",
|
| 514 |
+
"bbox": [
|
| 515 |
+
173,
|
| 516 |
+
882,
|
| 517 |
+
821,
|
| 518 |
+
911
|
| 519 |
+
],
|
| 520 |
+
"page_idx": 4
|
| 521 |
+
},
|
| 522 |
+
{
|
| 523 |
+
"type": "table",
|
| 524 |
+
"img_path": "images/a1227e1b3cb165e5b75d57d2176ab2a10da01b03a47cc44ccf5819cbf35e61b5.jpg",
|
| 525 |
+
"table_caption": [
|
| 526 |
+
"Table 1: Training tasks (CT stands for “contrastive learning” to align prompt encoders) and datasets with corresponding statistics. \\* denotes the number of accessible examples in the original datasets. "
|
| 527 |
+
],
|
| 528 |
+
"table_footnote": [],
|
| 529 |
+
"table_body": "<table><tr><td>Categories</td><td>Tasks</td><td>Datasets</td><td># of samples</td><td>Domain</td></tr><tr><td>Image + Text</td><td>Image-→Text,Text-→Image Text-→Image+Text</td><td>Laion400M [42]</td><td>400M</td><td>Open</td></tr><tr><td>Audio + Text</td><td>Text→Audio,Audio-→Text, Text-→Audio+Text,Audio-Text CT</td><td>AudioSet [16] AudioCaps [24] Freesound 500K BBC Sound Effect</td><td>900K* 46K 2.5M 30K</td><td>YouTube YouTube Public audio samples Authentic natural sound</td></tr><tr><td>Audiovisual</td><td>Image→Audio,Image→Video+Audio</td><td>AudioSet SoundNet [3]</td><td>900K* 1.0M*</td><td>YouTube Flickr, natural sound</td></tr><tr><td>Video</td><td>Text-→Video,Image→Video, Video-Text CT</td><td>Webvid10M[4] HD-Villa-100M [54]</td><td>10.7M 100M</td><td>Short videos YouTube</td></tr></table>",
|
| 530 |
+
"bbox": [
|
| 531 |
+
174,
|
| 532 |
+
122,
|
| 533 |
+
821,
|
| 534 |
+
276
|
| 535 |
+
],
|
| 536 |
+
"page_idx": 5
|
| 537 |
+
},
|
| 538 |
+
{
|
| 539 |
+
"type": "image",
|
| 540 |
+
"img_path": "images/493f24169a89b5450efae4d85c3805ebc13a132fcf5e28a3a488cbd564530264.jpg",
|
| 541 |
+
"image_caption": [
|
| 542 |
+
"Figure 3: Single-to-single modality generation. Clockwise from top left: text image, image text, image video, audio image. "
|
| 543 |
+
],
|
| 544 |
+
"image_footnote": [],
|
| 545 |
+
"bbox": [
|
| 546 |
+
174,
|
| 547 |
+
291,
|
| 548 |
+
821,
|
| 549 |
+
493
|
| 550 |
+
],
|
| 551 |
+
"page_idx": 5
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"type": "text",
|
| 555 |
+
"text": "",
|
| 556 |
+
"bbox": [
|
| 557 |
+
174,
|
| 558 |
+
561,
|
| 559 |
+
825,
|
| 560 |
+
660
|
| 561 |
+
],
|
| 562 |
+
"page_idx": 5
|
| 563 |
+
},
|
| 564 |
+
{
|
| 565 |
+
"type": "text",
|
| 566 |
+
"text": "4 Experiments ",
|
| 567 |
+
"text_level": 1,
|
| 568 |
+
"bbox": [
|
| 569 |
+
174,
|
| 570 |
+
680,
|
| 571 |
+
312,
|
| 572 |
+
696
|
| 573 |
+
],
|
| 574 |
+
"page_idx": 5
|
| 575 |
+
},
|
| 576 |
+
{
|
| 577 |
+
"type": "text",
|
| 578 |
+
"text": "4.1 Training Objectives and Datasets ",
|
| 579 |
+
"text_level": 1,
|
| 580 |
+
"bbox": [
|
| 581 |
+
176,
|
| 582 |
+
712,
|
| 583 |
+
442,
|
| 584 |
+
727
|
| 585 |
+
],
|
| 586 |
+
"page_idx": 5
|
| 587 |
+
},
|
| 588 |
+
{
|
| 589 |
+
"type": "text",
|
| 590 |
+
"text": "We list training tasks of CoDi in Table 1, including single modality synthesis, joint multimodal generation, and contrastive learning to align prompt encoders. Table 1 provides an overview of the datasets, tasks, number of samples, and domain. Datasets are from the following domains: image $^ +$ text (e.g. image with caption), audio $^ +$ text (e.g. audio with description), audio $^ +$ video (e.g. video with sound), and video $^ +$ text (e.g. video with description). As one may have noticed, the language modality appears in most datasets and domains. This echos the idea of using text as the bridge modality to be able to extrapolate and generate new unseen combinations such as audio and image bridged by text, as mentioned in Section 3.2 and Section 3.4. Due to space limit, more details on training datasets and can be found in Appendix C, model architecture details in Appendix Appendix A.1, and training details in Appendix B. ",
|
| 591 |
+
"bbox": [
|
| 592 |
+
173,
|
| 593 |
+
738,
|
| 594 |
+
825,
|
| 595 |
+
877
|
| 596 |
+
],
|
| 597 |
+
"page_idx": 5
|
| 598 |
+
},
|
| 599 |
+
{
|
| 600 |
+
"type": "text",
|
| 601 |
+
"text": "Image $^ +$ Text. We use a recently developed large-scale image caption dataset, Laion400M [42]. This image-text paired data allows us to train with tasks text image, image text, and the joint generation of image and text. For the joint generation task, we propose to train with text image+text, where the prompt text is the truncated image caption, and the output text is the original caption. Since the condition information is incomplete, the text and image diffuser will need to learn to attend with each other through the joint generation process. ",
|
| 602 |
+
"bbox": [
|
| 603 |
+
173,
|
| 604 |
+
883,
|
| 605 |
+
823,
|
| 606 |
+
911
|
| 607 |
+
],
|
| 608 |
+
"page_idx": 5
|
| 609 |
+
},
|
| 610 |
+
{
|
| 611 |
+
"type": "table",
|
| 612 |
+
"img_path": "images/f0713a5e3868134fdb21d162b330910771c9fe92eb4e6a85f24933ece1e1781b.jpg",
|
| 613 |
+
"table_caption": [
|
| 614 |
+
"Table 2: COCO-caption [32] FID scores for text-to-image generation. "
|
| 615 |
+
],
|
| 616 |
+
"table_footnote": [],
|
| 617 |
+
"table_body": "<table><tr><td>Method</td><td>FID↓</td></tr><tr><td>CogView [10]</td><td>27.10</td></tr><tr><td>GLIDE [36]</td><td>12.24</td></tr><tr><td>Make-a-Scene [15]</td><td>11.84</td></tr><tr><td>LDM [41]</td><td>12.63</td></tr><tr><td>Stable Diffusion-1.4</td><td>11.21</td></tr><tr><td>Stable Diffusion-1.5</td><td>11.12</td></tr><tr><td>Versatile Diffusion [53]</td><td>11.10</td></tr><tr><td>CoDi (Ours)</td><td>11.26</td></tr></table>",
|
| 618 |
+
"bbox": [
|
| 619 |
+
178,
|
| 620 |
+
137,
|
| 621 |
+
370,
|
| 622 |
+
257
|
| 623 |
+
],
|
| 624 |
+
"page_idx": 6
|
| 625 |
+
},
|
| 626 |
+
{
|
| 627 |
+
"type": "table",
|
| 628 |
+
"img_path": "images/092e9968c4819972025aaf57bfda705e16f8eed1cfa742fa9dda77afc438a0d5.jpg",
|
| 629 |
+
"table_caption": [
|
| 630 |
+
"Table 3: MSR-VTT text-to-video Table 4: UCF-101 text-to-video generation performance. generation performance. "
|
| 631 |
+
],
|
| 632 |
+
"table_footnote": [],
|
| 633 |
+
"table_body": "<table><tr><td>Method</td><td>Zero-Shot</td><td>CLIPSIM ↑</td></tr><tr><td>GODIVA [50]</td><td>No</td><td>0.2402</td></tr><tr><td>NUWA [51]</td><td>No</td><td>0.2439</td></tr><tr><td>CogVideo [22]</td><td>Yes</td><td>0.2631</td></tr><tr><td>Make-A-Video [44]</td><td>Yes</td><td>0.3049</td></tr><tr><td>Video LDM[5]</td><td>Yes</td><td>0.2929</td></tr><tr><td>CoDi(Ours)</td><td>Yes</td><td>0.2890</td></tr></table>",
|
| 634 |
+
"bbox": [
|
| 635 |
+
390,
|
| 636 |
+
152,
|
| 637 |
+
609,
|
| 638 |
+
231
|
| 639 |
+
],
|
| 640 |
+
"page_idx": 6
|
| 641 |
+
},
|
| 642 |
+
{
|
| 643 |
+
"type": "table",
|
| 644 |
+
"img_path": "images/89d31c9fff04f8286fa349bee00035aa85492d934c65c9e6e455e2a373e24edf.jpg",
|
| 645 |
+
"table_caption": [],
|
| 646 |
+
"table_footnote": [],
|
| 647 |
+
"table_body": "<table><tr><td>Method</td><td>IS(1)</td><td>FVD (↑)</td></tr><tr><td>Cog Video (Chinese)</td><td>23.55</td><td>751.34</td></tr><tr><td>CogVideo (English)</td><td>25.27</td><td>701.59</td></tr><tr><td>Make-A-Video</td><td>33.00</td><td>367.23</td></tr><tr><td>Video LDM</td><td>33.45</td><td>550.61</td></tr><tr><td>CoDi(Ours)</td><td>32.88</td><td>596.34</td></tr></table>",
|
| 648 |
+
"bbox": [
|
| 649 |
+
617,
|
| 650 |
+
154,
|
| 651 |
+
821,
|
| 652 |
+
229
|
| 653 |
+
],
|
| 654 |
+
"page_idx": 6
|
| 655 |
+
},
|
| 656 |
+
{
|
| 657 |
+
"type": "table",
|
| 658 |
+
"img_path": "images/6cdd7e474acf79309e07b53bb583c906924dd31228adbed791c23115f83c02ab.jpg",
|
| 659 |
+
"table_caption": [
|
| 660 |
+
"Table 5: The comparison between our audio diffuser and baseline TTA generation models. Evaluation is conducted on AudioCaps test set. AS, AC, FSD, BBC, and SDN stand for AudioSet, AudioCaps, Freesound, BBC Sound Effect, and Soundnet. "
|
| 661 |
+
],
|
| 662 |
+
"table_footnote": [],
|
| 663 |
+
"table_body": "<table><tr><td>Model</td><td>Datasets</td><td>FD↓</td><td>IS个</td><td>KL←</td><td>FAD↓</td><td>OVL ↑</td><td>REL个</td></tr><tr><td>Ground truth</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>83.61</td><td>80.11</td></tr><tr><td>DiffSound</td><td>AS+AC</td><td>47.68</td><td>4.01</td><td>2.52</td><td>7.75</td><td>45.00</td><td>43.83</td></tr><tr><td>AudioGen</td><td>AS +AC+8others</td><td>=</td><td>-</td><td>2.09</td><td>3.13</td><td>-</td><td>=</td></tr><tr><td>AudioLDM-L-Full</td><td>AS+AC+FSD+BBC</td><td>23.31</td><td>8.13</td><td>1.59</td><td>1.96</td><td>65.91</td><td>65.97</td></tr><tr><td>CoDi(Ours)</td><td>AS+AC+FSD+BBC+SDN</td><td>22.90</td><td>8.77</td><td>1.40</td><td>1.80</td><td>66.87</td><td>67.60</td></tr></table>",
|
| 664 |
+
"bbox": [
|
| 665 |
+
205,
|
| 666 |
+
316,
|
| 667 |
+
790,
|
| 668 |
+
401
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 6
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "table",
|
| 674 |
+
"img_path": "images/3b08dd3a07ff80c69c148aa0491f160ff27946a1af6b4a57632f75d5ff7af62b.jpg",
|
| 675 |
+
"table_caption": [
|
| 676 |
+
"Table 6: COCO image captioning scores comparison. "
|
| 677 |
+
],
|
| 678 |
+
"table_footnote": [],
|
| 679 |
+
"table_body": "<table><tr><td>Model</td><td>B@4</td><td>METEOR</td><td>CIDEr</td></tr><tr><td colspan=\"4\">Autoregressive Model</td></tr><tr><td>Oscar [31]</td><td>36.58</td><td>30.4</td><td>124.12</td></tr><tr><td>ClipCap [35]</td><td>32.15</td><td>27.1</td><td>108.35</td></tr><tr><td>OFA [49]</td><td>44.9</td><td>32.5</td><td>154.9</td></tr><tr><td>BLIP2 [30]</td><td>43.7</td><td>-</td><td>145.8</td></tr><tr><td colspan=\"4\">Diffusion Model</td></tr><tr><td>DDCap [59]</td><td>35.0</td><td>28.2</td><td>117.8</td></tr><tr><td>SCD-Net [34]</td><td>39.4</td><td>29.2</td><td>131.6</td></tr><tr><td>CoDi (Ours)</td><td>40.2</td><td>31.0</td><td>149.9</td></tr></table>",
|
| 680 |
+
"bbox": [
|
| 681 |
+
176,
|
| 682 |
+
449,
|
| 683 |
+
377,
|
| 684 |
+
554
|
| 685 |
+
],
|
| 686 |
+
"page_idx": 6
|
| 687 |
+
},
|
| 688 |
+
{
|
| 689 |
+
"type": "table",
|
| 690 |
+
"img_path": "images/d48a4e4d621172d01f3a4f8bc34da6f60c2863153ae2635e211633ead5c366e1.jpg",
|
| 691 |
+
"table_caption": [
|
| 692 |
+
"Table 7: AudioCaps audio captioning scores comparison. "
|
| 693 |
+
],
|
| 694 |
+
"table_footnote": [],
|
| 695 |
+
"table_body": "<table><tr><td>Model</td><td>SPIDEr</td><td>CIDEr</td><td>SPICE</td></tr><tr><td>AudioCaps [24]</td><td>0.369</td><td>0.593</td><td>0.144</td></tr><tr><td>BART-Finetune [17]</td><td>0.465</td><td>0.753</td><td>0.176</td></tr><tr><td>VALOR[7]</td><td></td><td>0.741</td><td></td></tr><tr><td>AL-MixGen [25]</td><td>0.466</td><td>0.755</td><td>0.177</td></tr><tr><td>CoDi (Ours)</td><td>0.480</td><td>0.789</td><td>0.182</td></tr></table>",
|
| 696 |
+
"bbox": [
|
| 697 |
+
390,
|
| 698 |
+
472,
|
| 699 |
+
598,
|
| 700 |
+
534
|
| 701 |
+
],
|
| 702 |
+
"page_idx": 6
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"type": "table",
|
| 706 |
+
"img_path": "images/555831ba6bc0c59839fc917171599939764c30018cb4c379929213d7ca95eaf8.jpg",
|
| 707 |
+
"table_caption": [
|
| 708 |
+
"Table 8: MSRVTT video captioning scores comparison. "
|
| 709 |
+
],
|
| 710 |
+
"table_footnote": [],
|
| 711 |
+
"table_body": "<table><tr><td>Model</td><td>B@4</td><td>METEOR</td><td>CIDEr</td></tr><tr><td>ORG-TRL[58]</td><td>43.6</td><td>28.8</td><td>50.9</td></tr><tr><td>MV-GPT[43]</td><td>48.9</td><td>38.7</td><td>60.0</td></tr><tr><td>GIT[48]</td><td>54.8</td><td>33.1</td><td>75.9</td></tr><tr><td>mPLUG-2 [52]</td><td>57.8</td><td>34.9</td><td>80.3</td></tr><tr><td>CoDi(Ours)</td><td>52.1</td><td>32.5</td><td>74.4</td></tr></table>",
|
| 712 |
+
"bbox": [
|
| 713 |
+
611,
|
| 714 |
+
470,
|
| 715 |
+
808,
|
| 716 |
+
534
|
| 717 |
+
],
|
| 718 |
+
"page_idx": 6
|
| 719 |
+
},
|
| 720 |
+
{
|
| 721 |
+
"type": "text",
|
| 722 |
+
"text": "",
|
| 723 |
+
"bbox": [
|
| 724 |
+
174,
|
| 725 |
+
587,
|
| 726 |
+
825,
|
| 727 |
+
642
|
| 728 |
+
],
|
| 729 |
+
"page_idx": 6
|
| 730 |
+
},
|
| 731 |
+
{
|
| 732 |
+
"type": "text",
|
| 733 |
+
"text": "Audio $^ +$ Text. We curated a new dataset, Freesound 500K, by crawling 500K audio samples together with tags and descriptions from the Freesound website. We also use AudioSet [42] with 2 million human-labeled 10-second sound clips from YouTube videos and AudioCaps [24] with 46K audiotext pairs derived from the AudioSet dataset. Audio samples are clipped into 10-second segments for training purposes. The paired audio $^ +$ text data enables us to train text audio, audio text, text audio $^ +$ text generation, and audio-text contrastive learning. Similar to image $^ +$ text joint generation, in text audio $^ +$ text, text prompt is the truncated text, and the output is the original text. ",
|
| 734 |
+
"bbox": [
|
| 735 |
+
173,
|
| 736 |
+
648,
|
| 737 |
+
825,
|
| 738 |
+
746
|
| 739 |
+
],
|
| 740 |
+
"page_idx": 6
|
| 741 |
+
},
|
| 742 |
+
{
|
| 743 |
+
"type": "text",
|
| 744 |
+
"text": "Video. We use the following diverse and high-quality video datasets to train video generation and video prompt encoder. WebVid [4], a large-scale dataset of web videos together with descriptions; HD-Villa-100M [54] with high resolution YouTube videos of at least 720P. We perform text video and video-text contrastive learning task with WebVid. We use HD-Villa-100M for image video generation where the middle frame is the input image. ",
|
| 745 |
+
"bbox": [
|
| 746 |
+
174,
|
| 747 |
+
752,
|
| 748 |
+
825,
|
| 749 |
+
821
|
| 750 |
+
],
|
| 751 |
+
"page_idx": 6
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"type": "text",
|
| 755 |
+
"text": "Audiovisual. Web videos are a natural aligned audio-video data resource. However, many existing datasets, e.g., ACAV100M [28], feature heavily on videos of human speech rather than natural sounds. Therefore, we leverage sound-oriented datasets AudioSet and SoundNet [3] for joint audio-video generation. For image audio $^ +$ video, we use the middle frame of the target video as the input prompt image. We also use the middle frame as the prompt input to train the model to generate the audio, i.e., image audio. ",
|
| 756 |
+
"bbox": [
|
| 757 |
+
174,
|
| 758 |
+
828,
|
| 759 |
+
825,
|
| 760 |
+
911
|
| 761 |
+
],
|
| 762 |
+
"page_idx": 6
|
| 763 |
+
},
|
| 764 |
+
{
|
| 765 |
+
"type": "image",
|
| 766 |
+
"img_path": "images/dc63eb6c668a08077a3c76c77a471ccad92d8b97c381a63765856efb970b42b0.jpg",
|
| 767 |
+
"image_caption": [
|
| 768 |
+
"Figure 4: Generation with multiple input modality conditions. Top to bottom: text+audio image, text+audio video, video+audio text. "
|
| 769 |
+
],
|
| 770 |
+
"image_footnote": [],
|
| 771 |
+
"bbox": [
|
| 772 |
+
174,
|
| 773 |
+
89,
|
| 774 |
+
823,
|
| 775 |
+
468
|
| 776 |
+
],
|
| 777 |
+
"page_idx": 7
|
| 778 |
+
},
|
| 779 |
+
{
|
| 780 |
+
"type": "text",
|
| 781 |
+
"text": "5 Evaluation Results ",
|
| 782 |
+
"text_level": 1,
|
| 783 |
+
"bbox": [
|
| 784 |
+
174,
|
| 785 |
+
531,
|
| 786 |
+
362,
|
| 787 |
+
549
|
| 788 |
+
],
|
| 789 |
+
"page_idx": 7
|
| 790 |
+
},
|
| 791 |
+
{
|
| 792 |
+
"type": "text",
|
| 793 |
+
"text": "In this section, we will evaluate the model generation quality in different settings including single modality generation, multi-condition generation, and multi-output joint generation. We provide both quantitative benchmarking on evaluation datasets as well as qualitative visualization demonstrations. ",
|
| 794 |
+
"bbox": [
|
| 795 |
+
176,
|
| 796 |
+
564,
|
| 797 |
+
825,
|
| 798 |
+
607
|
| 799 |
+
],
|
| 800 |
+
"page_idx": 7
|
| 801 |
+
},
|
| 802 |
+
{
|
| 803 |
+
"type": "text",
|
| 804 |
+
"text": "5.1 Single Modality Generation Results ",
|
| 805 |
+
"text_level": 1,
|
| 806 |
+
"bbox": [
|
| 807 |
+
176,
|
| 808 |
+
625,
|
| 809 |
+
460,
|
| 810 |
+
640
|
| 811 |
+
],
|
| 812 |
+
"page_idx": 7
|
| 813 |
+
},
|
| 814 |
+
{
|
| 815 |
+
"type": "text",
|
| 816 |
+
"text": "We first show example demo in Fig. 3, where we present various single to single modality generation. Then, we evaluate the synthesis quality of the unimodal generation on text, image, video, and audio. CoDi achieves SOTA on audio captions and audio generation, as shown in Table 7 and Table 5. Notably for the first time in the field, CoDi, a diffusion-base model, exhibits comparable performance on image captioning with autoregressive transformer-based SOTA (Table 6). CoDi is the first diffusion-model based for video captioning Table 8. On image and video generation, CoDi performs competitively with state-of-the-art (Tables 2 to 4). This gives us strong starting points for multi-condition and multi-output generation that will be presented next in Section 5.2 and Section 5.3. ",
|
| 817 |
+
"bbox": [
|
| 818 |
+
173,
|
| 819 |
+
651,
|
| 820 |
+
825,
|
| 821 |
+
762
|
| 822 |
+
],
|
| 823 |
+
"page_idx": 7
|
| 824 |
+
},
|
| 825 |
+
{
|
| 826 |
+
"type": "text",
|
| 827 |
+
"text": "We demonstrate in Section 3.2 that CoDi is capable of integrating representation from different modalities in the generation. Thus, we first show multi-condition generation demo as shown in Fig. 4. ",
|
| 828 |
+
"bbox": [
|
| 829 |
+
173,
|
| 830 |
+
768,
|
| 831 |
+
823,
|
| 832 |
+
796
|
| 833 |
+
],
|
| 834 |
+
"page_idx": 7
|
| 835 |
+
},
|
| 836 |
+
{
|
| 837 |
+
"type": "text",
|
| 838 |
+
"text": "5.2 Multi-Condition Generation Results ",
|
| 839 |
+
"text_level": 1,
|
| 840 |
+
"bbox": [
|
| 841 |
+
176,
|
| 842 |
+
814,
|
| 843 |
+
464,
|
| 844 |
+
830
|
| 845 |
+
],
|
| 846 |
+
"page_idx": 7
|
| 847 |
+
},
|
| 848 |
+
{
|
| 849 |
+
"type": "text",
|
| 850 |
+
"text": "For quantitative evaluation, we focus on multiple inputs to image synthesis output since the evaluation metric for this case (FID) does not require specific modality inputs like text. We test with several input combinations including text $^ +$ image, text $^ +$ audio, image $^ +$ audio, text $^ +$ video, as well as three inputs text $^ +$ audio $^ +$ image. We test on the validation set of AudioCaps [24] since all four modalities are present in this dataset. The prompt image input is the middle frame of the video. As shown in ",
|
| 851 |
+
"bbox": [
|
| 852 |
+
174,
|
| 853 |
+
842,
|
| 854 |
+
823,
|
| 855 |
+
911
|
| 856 |
+
],
|
| 857 |
+
"page_idx": 7
|
| 858 |
+
},
|
| 859 |
+
{
|
| 860 |
+
"type": "table",
|
| 861 |
+
"img_path": "images/e4534d7ebcfba76448956e498aa36d45e0b4f106882850134d22e24e5043a3c9.jpg",
|
| 862 |
+
"table_caption": [
|
| 863 |
+
"Table 9: CoDi is capable of generating high quality output (image in this case) from various combinations of prompt modalities. "
|
| 864 |
+
],
|
| 865 |
+
"table_footnote": [],
|
| 866 |
+
"table_body": "<table><tr><td>Inputs</td><td>FID↓</td></tr><tr><td> Single-modality Prompt</td><td></td></tr><tr><td>Text</td><td>14.2</td></tr><tr><td>Audio</td><td>14.3</td></tr><tr><td> Dual-modality Prompt</td><td></td></tr><tr><td>Text+Audio</td><td>14.9</td></tr></table>",
|
| 867 |
+
"bbox": [
|
| 868 |
+
236,
|
| 869 |
+
137,
|
| 870 |
+
441,
|
| 871 |
+
227
|
| 872 |
+
],
|
| 873 |
+
"page_idx": 8
|
| 874 |
+
},
|
| 875 |
+
{
|
| 876 |
+
"type": "table",
|
| 877 |
+
"img_path": "images/5dac823453343eca2eec8a73d3324d6d6bf7776e7f743d995f77fa1083127e85.jpg",
|
| 878 |
+
"table_caption": [
|
| 879 |
+
"Table 10: MSR-VTT text-to-video generation performance. "
|
| 880 |
+
],
|
| 881 |
+
"table_footnote": [],
|
| 882 |
+
"table_body": "<table><tr><td>Inputs</td><td>CLIPSIM个</td></tr><tr><td> Single-modality Prompt</td><td></td></tr><tr><td>Text</td><td>0.2890</td></tr><tr><td> Dual-modality Prompt</td><td></td></tr><tr><td>Text+Audio</td><td>0.2912</td></tr><tr><td>Text+Image</td><td>0.2891</td></tr><tr><td>Text+Audio+Image</td><td>0.2923</td></tr></table>",
|
| 883 |
+
"bbox": [
|
| 884 |
+
562,
|
| 885 |
+
122,
|
| 886 |
+
805,
|
| 887 |
+
227
|
| 888 |
+
],
|
| 889 |
+
"page_idx": 8
|
| 890 |
+
},
|
| 891 |
+
{
|
| 892 |
+
"type": "image",
|
| 893 |
+
"img_path": "images/570523fa62e48124b6e7ee0927d72eb815a36ea17ca483b39355fc5b79c4dd8d.jpg",
|
| 894 |
+
"image_caption": [
|
| 895 |
+
"Figure 5: Joint generation of multiple output modalities by CoDi. From top to bottom: text video+audio, tex image+text+audio, text+audio+image video+audio. "
|
| 896 |
+
],
|
| 897 |
+
"image_footnote": [],
|
| 898 |
+
"bbox": [
|
| 899 |
+
178,
|
| 900 |
+
253,
|
| 901 |
+
818,
|
| 902 |
+
597
|
| 903 |
+
],
|
| 904 |
+
"page_idx": 8
|
| 905 |
+
},
|
| 906 |
+
{
|
| 907 |
+
"type": "text",
|
| 908 |
+
"text": "Table 9, CoDi achieves high image generation quality given assorted groups of input modalities. We also test with several input combinations with video as output including text, text $^ +$ audio, image $^ +$ image, as well as text $^ +$ audio $^ +$ image. We also test on MSRVTT [24] since all four modalities are present in this dataset. Similarly, the prompt image input is the middle frame of the video. As shown in Table 10, CoDi achieves high video and ground truth text similarity given assorted groups of input modalities. Again our model does not need to train on multi-condition generation like text $^ +$ audio or text $^ +$ image. Through bridging alignment and composable multimodal conditioning as proposed in Section 3.2, our model trained on single condition can zero-shot infer on multiple conditions. ",
|
| 909 |
+
"bbox": [
|
| 910 |
+
173,
|
| 911 |
+
662,
|
| 912 |
+
825,
|
| 913 |
+
775
|
| 914 |
+
],
|
| 915 |
+
"page_idx": 8
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"type": "text",
|
| 919 |
+
"text": "5.3 Multi-Output Joint Generation Results ",
|
| 920 |
+
"text_level": 1,
|
| 921 |
+
"bbox": [
|
| 922 |
+
174,
|
| 923 |
+
791,
|
| 924 |
+
485,
|
| 925 |
+
808
|
| 926 |
+
],
|
| 927 |
+
"page_idx": 8
|
| 928 |
+
},
|
| 929 |
+
{
|
| 930 |
+
"type": "text",
|
| 931 |
+
"text": "For joint multimodal generation, we first demonstrate high-quality multimodal output joint generation demo as shown in Fig. 5. For quantitative evaluation, there is no existing evaluation metric since we are the first model that can simultaneously generate across all 4 modalities. Therefore, we propose the following metric SIM that quantifies the coherence and consistency between the two generated modalities by cosine similarity of embeddings: ",
|
| 932 |
+
"bbox": [
|
| 933 |
+
174,
|
| 934 |
+
818,
|
| 935 |
+
825,
|
| 936 |
+
887
|
| 937 |
+
],
|
| 938 |
+
"page_idx": 8
|
| 939 |
+
},
|
| 940 |
+
{
|
| 941 |
+
"type": "equation",
|
| 942 |
+
"img_path": "images/442991e83372788db1166c87e1f118b81b1449278a900571ed564c3772d81cd4.jpg",
|
| 943 |
+
"text": "$$\n\\operatorname { S I M } ( A , B ) = \\cos { ( C _ { A } ( A ) , C _ { B } ( B ) ) }\n$$",
|
| 944 |
+
"text_format": "latex",
|
| 945 |
+
"bbox": [
|
| 946 |
+
379,
|
| 947 |
+
895,
|
| 948 |
+
619,
|
| 949 |
+
912
|
| 950 |
+
],
|
| 951 |
+
"page_idx": 8
|
| 952 |
+
},
|
| 953 |
+
{
|
| 954 |
+
"type": "table",
|
| 955 |
+
"img_path": "images/7c6e18d97f95273627bd5cdb16fd3a210522268a86a54bd9275114265361cbb0.jpg",
|
| 956 |
+
"table_caption": [
|
| 957 |
+
"Table 11: Similarity scores between generated modalities. The number on the left of $\" / \"$ represents the similarity score of independent generation, and the right it represents the case of joint generation. Jointly generated outputs consistently show stronger coherence. "
|
| 958 |
+
],
|
| 959 |
+
"table_footnote": [],
|
| 960 |
+
"table_body": "<table><tr><td>Inputs</td><td>SIM-IT</td><td>SIM-AT</td><td>SIM-VT</td><td>SIM-VA</td></tr><tr><td colspan=\"5\"> Two Joint Outputs</td></tr><tr><td>Audio → Image+Text</td><td>0.251 / 0.260</td><td></td><td></td><td></td></tr><tr><td>Image→Audio+Text</td><td>■</td><td>0.244 / 0.256</td><td></td><td></td></tr><tr><td>Text →Video+Audio</td><td></td><td></td><td></td><td>0.240 / 0.255</td></tr><tr><td>Audio →Video+Text</td><td></td><td></td><td>0.256 / 0.261</td><td></td></tr><tr><td colspan=\"5\"> Three Joint Outputs</td></tr><tr><td>Text-→ Video+Image+Audio 0.256/0.270 0.240/0.257</td><td></td><td></td><td></td><td>0.240 / 0.257</td></tr><tr><td colspan=\"5\"> Multi-Inputs-Outputs</td></tr><tr><td>Text+Image -→ Video+Audio</td><td></td><td></td><td></td><td>0.247 / 0.259</td></tr></table>",
|
| 961 |
+
"bbox": [
|
| 962 |
+
178,
|
| 963 |
+
136,
|
| 964 |
+
816,
|
| 965 |
+
305
|
| 966 |
+
],
|
| 967 |
+
"page_idx": 9
|
| 968 |
+
},
|
| 969 |
+
{
|
| 970 |
+
"type": "text",
|
| 971 |
+
"text": "where $A$ , $B$ are the generated modalities, and $C _ { A }$ and $C _ { B }$ are aligned encoders that project $A$ and $B$ to the same space. We use the prompt encoder as described in Section 3.2. This metric aims to compute the cosine similarity of the embedding of two modalities using contrastive learned prompt encoders. Thus, the higher the metric, the more aligned and similar the generated modalities are. ",
|
| 972 |
+
"bbox": [
|
| 973 |
+
174,
|
| 974 |
+
332,
|
| 975 |
+
825,
|
| 976 |
+
388
|
| 977 |
+
],
|
| 978 |
+
"page_idx": 9
|
| 979 |
+
},
|
| 980 |
+
{
|
| 981 |
+
"type": "text",
|
| 982 |
+
"text": "To demonstrate the effectiveness of joint generation, assume the prompt modality is $P$ , we compare $\\mathrm { S I M } ( A , B )$ of $A$ and $B$ generated separately vs. jointly, i.e., $\\{ P \\ { \\overset { - } { \\to } } \\ A , \\ P \\ { \\overset { - } { \\to } } \\ B \\}$ vs. $\\{ P $ $A + B \\}$ . The benchmark is the validation set of AudioCaps [24]. We test on the following settings, audio image+text, image audio+text, and text video+audio, image video+audio. audio video+text, audio text+video+image, text video+image+audio, where the image prompt is the middle frame of the video clip. As shown in Table 11, joint generation (similarity shown on the right side of $\" / \"$ ) consistently outperforms independent generation (on the left side of $\" / \"$ ). ",
|
| 983 |
+
"bbox": [
|
| 984 |
+
174,
|
| 985 |
+
395,
|
| 986 |
+
825,
|
| 987 |
+
492
|
| 988 |
+
],
|
| 989 |
+
"page_idx": 9
|
| 990 |
+
},
|
| 991 |
+
{
|
| 992 |
+
"type": "text",
|
| 993 |
+
"text": "6 Conclusion ",
|
| 994 |
+
"text_level": 1,
|
| 995 |
+
"bbox": [
|
| 996 |
+
174,
|
| 997 |
+
513,
|
| 998 |
+
299,
|
| 999 |
+
530
|
| 1000 |
+
],
|
| 1001 |
+
"page_idx": 9
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"type": "text",
|
| 1005 |
+
"text": "In this paper, we present Composable Diffusion (CoDi), a groundbreaking model in multimodal generation that is capable of processing and simultaneously generating modalities across text, image, video, and audio. Our approach enables the synergistic generation of high-quality and coherent outputs spanning various modalities, from assorted combinations of input modalities. Through extensive experiments, we demonstrate CoDi’s remarkable capabilities in flexibly generating single or multiple modalities from a wide range of inputs. Our work marks a significant step towards more engaging and holistic human-computer interactions, establishing a solid foundation for future investigations in generative artificial intelligence. ",
|
| 1006 |
+
"bbox": [
|
| 1007 |
+
174,
|
| 1008 |
+
546,
|
| 1009 |
+
825,
|
| 1010 |
+
657
|
| 1011 |
+
],
|
| 1012 |
+
"page_idx": 9
|
| 1013 |
+
},
|
| 1014 |
+
{
|
| 1015 |
+
"type": "text",
|
| 1016 |
+
"text": "Limitations & Broader Impacts. See Appendix D for the discussion. ",
|
| 1017 |
+
"bbox": [
|
| 1018 |
+
174,
|
| 1019 |
+
664,
|
| 1020 |
+
635,
|
| 1021 |
+
679
|
| 1022 |
+
],
|
| 1023 |
+
"page_idx": 9
|
| 1024 |
+
},
|
| 1025 |
+
{
|
| 1026 |
+
"type": "text",
|
| 1027 |
+
"text": "Acknowledgement ",
|
| 1028 |
+
"text_level": 1,
|
| 1029 |
+
"bbox": [
|
| 1030 |
+
176,
|
| 1031 |
+
700,
|
| 1032 |
+
330,
|
| 1033 |
+
717
|
| 1034 |
+
],
|
| 1035 |
+
"page_idx": 9
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"type": "text",
|
| 1039 |
+
"text": "We would like to thank Bei Liu for HD-VILA-100M data support. We also thank Shi Dong, Mahmoud Khademi, Junheng Hao, Yuwei Fang, Yichong Xu and Azure Cognitive Services Research team members for their feedback. ",
|
| 1040 |
+
"bbox": [
|
| 1041 |
+
176,
|
| 1042 |
+
733,
|
| 1043 |
+
825,
|
| 1044 |
+
775
|
| 1045 |
+
],
|
| 1046 |
+
"page_idx": 9
|
| 1047 |
+
},
|
| 1048 |
+
{
|
| 1049 |
+
"type": "text",
|
| 1050 |
+
"text": "References ",
|
| 1051 |
+
"text_level": 1,
|
| 1052 |
+
"bbox": [
|
| 1053 |
+
174,
|
| 1054 |
+
796,
|
| 1055 |
+
266,
|
| 1056 |
+
811
|
| 1057 |
+
],
|
| 1058 |
+
"page_idx": 9
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"type": "text",
|
| 1062 |
+
"text": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716–23736, 2022. 3 \n[2] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 5, 15, 16 ",
|
| 1063 |
+
"bbox": [
|
| 1064 |
+
178,
|
| 1065 |
+
819,
|
| 1066 |
+
826,
|
| 1067 |
+
911
|
| 1068 |
+
],
|
| 1069 |
+
"page_idx": 9
|
| 1070 |
+
},
|
| 1071 |
+
{
|
| 1072 |
+
"type": "text",
|
| 1073 |
+
"text": "[3] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. Advances in neural information processing systems, 29, 2016. 6, 7 ",
|
| 1074 |
+
"bbox": [
|
| 1075 |
+
178,
|
| 1076 |
+
92,
|
| 1077 |
+
823,
|
| 1078 |
+
118
|
| 1079 |
+
],
|
| 1080 |
+
"page_idx": 10
|
| 1081 |
+
},
|
| 1082 |
+
{
|
| 1083 |
+
"type": "text",
|
| 1084 |
+
"text": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728–1738, 2021. 6, 7, 16 ",
|
| 1085 |
+
"bbox": [
|
| 1086 |
+
179,
|
| 1087 |
+
128,
|
| 1088 |
+
820,
|
| 1089 |
+
167
|
| 1090 |
+
],
|
| 1091 |
+
"page_idx": 10
|
| 1092 |
+
},
|
| 1093 |
+
{
|
| 1094 |
+
"type": "text",
|
| 1095 |
+
"text": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. arXiv preprint arXiv:2304.08818, 2023. 7 ",
|
| 1096 |
+
"bbox": [
|
| 1097 |
+
178,
|
| 1098 |
+
176,
|
| 1099 |
+
821,
|
| 1100 |
+
215
|
| 1101 |
+
],
|
| 1102 |
+
"page_idx": 10
|
| 1103 |
+
},
|
| 1104 |
+
{
|
| 1105 |
+
"type": "text",
|
| 1106 |
+
"text": "[6] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 2 ",
|
| 1107 |
+
"bbox": [
|
| 1108 |
+
178,
|
| 1109 |
+
226,
|
| 1110 |
+
823,
|
| 1111 |
+
265
|
| 1112 |
+
],
|
| 1113 |
+
"page_idx": 10
|
| 1114 |
+
},
|
| 1115 |
+
{
|
| 1116 |
+
"type": "text",
|
| 1117 |
+
"text": "[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor: Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345, 2023. 7 ",
|
| 1118 |
+
"bbox": [
|
| 1119 |
+
178,
|
| 1120 |
+
275,
|
| 1121 |
+
823,
|
| 1122 |
+
314
|
| 1123 |
+
],
|
| 1124 |
+
"page_idx": 10
|
| 1125 |
+
},
|
| 1126 |
+
{
|
| 1127 |
+
"type": "text",
|
| 1128 |
+
"text": "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931–1942. PMLR, 2021. 3 ",
|
| 1129 |
+
"bbox": [
|
| 1130 |
+
178,
|
| 1131 |
+
323,
|
| 1132 |
+
821,
|
| 1133 |
+
351
|
| 1134 |
+
],
|
| 1135 |
+
"page_idx": 10
|
| 1136 |
+
},
|
| 1137 |
+
{
|
| 1138 |
+
"type": "text",
|
| 1139 |
+
"text": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 ",
|
| 1140 |
+
"bbox": [
|
| 1141 |
+
178,
|
| 1142 |
+
359,
|
| 1143 |
+
823,
|
| 1144 |
+
387
|
| 1145 |
+
],
|
| 1146 |
+
"page_idx": 10
|
| 1147 |
+
},
|
| 1148 |
+
{
|
| 1149 |
+
"type": "text",
|
| 1150 |
+
"text": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. arXiv preprint arXiv:2105.13290, 2021. 7 ",
|
| 1151 |
+
"bbox": [
|
| 1152 |
+
173,
|
| 1153 |
+
395,
|
| 1154 |
+
823,
|
| 1155 |
+
435
|
| 1156 |
+
],
|
| 1157 |
+
"page_idx": 10
|
| 1158 |
+
},
|
| 1159 |
+
{
|
| 1160 |
+
"type": "text",
|
| 1161 |
+
"text": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687–10696, 2021. 3 ",
|
| 1162 |
+
"bbox": [
|
| 1163 |
+
173,
|
| 1164 |
+
445,
|
| 1165 |
+
821,
|
| 1166 |
+
497
|
| 1167 |
+
],
|
| 1168 |
+
"page_idx": 10
|
| 1169 |
+
},
|
| 1170 |
+
{
|
| 1171 |
+
"type": "text",
|
| 1172 |
+
"text": "[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap: Learning audio concepts from natural language supervision. arXiv preprint arXiv:2206.04769, 2022. 3 ",
|
| 1173 |
+
"bbox": [
|
| 1174 |
+
173,
|
| 1175 |
+
506,
|
| 1176 |
+
823,
|
| 1177 |
+
534
|
| 1178 |
+
],
|
| 1179 |
+
"page_idx": 10
|
| 1180 |
+
},
|
| 1181 |
+
{
|
| 1182 |
+
"type": "text",
|
| 1183 |
+
"text": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2, 5 ",
|
| 1184 |
+
"bbox": [
|
| 1185 |
+
173,
|
| 1186 |
+
542,
|
| 1187 |
+
825,
|
| 1188 |
+
582
|
| 1189 |
+
],
|
| 1190 |
+
"page_idx": 10
|
| 1191 |
+
},
|
| 1192 |
+
{
|
| 1193 |
+
"type": "text",
|
| 1194 |
+
"text": "[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3 ",
|
| 1195 |
+
"bbox": [
|
| 1196 |
+
171,
|
| 1197 |
+
592,
|
| 1198 |
+
825,
|
| 1199 |
+
631
|
| 1200 |
+
],
|
| 1201 |
+
"page_idx": 10
|
| 1202 |
+
},
|
| 1203 |
+
{
|
| 1204 |
+
"type": "text",
|
| 1205 |
+
"text": "[15] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pages 89–106. Springer, 2022. 7 [16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776–780. IEEE, 2017. 6 ",
|
| 1206 |
+
"bbox": [
|
| 1207 |
+
171,
|
| 1208 |
+
640,
|
| 1209 |
+
823,
|
| 1210 |
+
680
|
| 1211 |
+
],
|
| 1212 |
+
"page_idx": 10
|
| 1213 |
+
},
|
| 1214 |
+
{
|
| 1215 |
+
"type": "text",
|
| 1216 |
+
"text": "",
|
| 1217 |
+
"bbox": [
|
| 1218 |
+
173,
|
| 1219 |
+
689,
|
| 1220 |
+
828,
|
| 1221 |
+
741
|
| 1222 |
+
],
|
| 1223 |
+
"page_idx": 10
|
| 1224 |
+
},
|
| 1225 |
+
{
|
| 1226 |
+
"type": "text",
|
| 1227 |
+
"text": "[17] Félix Gontier, Romain Serizel, and Christophe Cerisara. Automated audio captioning by fine-tuning bart with audioset tags. In Detection and Classification of Acoustic Scenes and Events-DCASE 2021, 2021. 7 [18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 15 ",
|
| 1228 |
+
"bbox": [
|
| 1229 |
+
171,
|
| 1230 |
+
751,
|
| 1231 |
+
823,
|
| 1232 |
+
777
|
| 1233 |
+
],
|
| 1234 |
+
"page_idx": 10
|
| 1235 |
+
},
|
| 1236 |
+
{
|
| 1237 |
+
"type": "text",
|
| 1238 |
+
"text": "",
|
| 1239 |
+
"bbox": [
|
| 1240 |
+
173,
|
| 1241 |
+
786,
|
| 1242 |
+
825,
|
| 1243 |
+
825
|
| 1244 |
+
],
|
| 1245 |
+
"page_idx": 10
|
| 1246 |
+
},
|
| 1247 |
+
{
|
| 1248 |
+
"type": "text",
|
| 1249 |
+
"text": "[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3 ",
|
| 1250 |
+
"bbox": [
|
| 1251 |
+
173,
|
| 1252 |
+
835,
|
| 1253 |
+
823,
|
| 1254 |
+
876
|
| 1255 |
+
],
|
| 1256 |
+
"page_idx": 10
|
| 1257 |
+
},
|
| 1258 |
+
{
|
| 1259 |
+
"type": "text",
|
| 1260 |
+
"text": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020. 3, 4 ",
|
| 1261 |
+
"bbox": [
|
| 1262 |
+
169,
|
| 1263 |
+
885,
|
| 1264 |
+
825,
|
| 1265 |
+
911
|
| 1266 |
+
],
|
| 1267 |
+
"page_idx": 10
|
| 1268 |
+
},
|
| 1269 |
+
{
|
| 1270 |
+
"type": "text",
|
| 1271 |
+
"text": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3, 15 ",
|
| 1272 |
+
"bbox": [
|
| 1273 |
+
171,
|
| 1274 |
+
92,
|
| 1275 |
+
825,
|
| 1276 |
+
118
|
| 1277 |
+
],
|
| 1278 |
+
"page_idx": 11
|
| 1279 |
+
},
|
| 1280 |
+
{
|
| 1281 |
+
"type": "text",
|
| 1282 |
+
"text": "[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 7 ",
|
| 1283 |
+
"bbox": [
|
| 1284 |
+
173,
|
| 1285 |
+
127,
|
| 1286 |
+
821,
|
| 1287 |
+
155
|
| 1288 |
+
],
|
| 1289 |
+
"page_idx": 11
|
| 1290 |
+
},
|
| 1291 |
+
{
|
| 1292 |
+
"type": "text",
|
| 1293 |
+
"text": "[23] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. arXiv preprint arXiv:2301.12661, 2023. 2, 3 ",
|
| 1294 |
+
"bbox": [
|
| 1295 |
+
174,
|
| 1296 |
+
162,
|
| 1297 |
+
821,
|
| 1298 |
+
202
|
| 1299 |
+
],
|
| 1300 |
+
"page_idx": 11
|
| 1301 |
+
},
|
| 1302 |
+
{
|
| 1303 |
+
"type": "text",
|
| 1304 |
+
"text": "[24] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions for audios in the wild. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 119–132, 2019. 6, 7, 8, 9, 10 ",
|
| 1305 |
+
"bbox": [
|
| 1306 |
+
173,
|
| 1307 |
+
210,
|
| 1308 |
+
825,
|
| 1309 |
+
262
|
| 1310 |
+
],
|
| 1311 |
+
"page_idx": 11
|
| 1312 |
+
},
|
| 1313 |
+
{
|
| 1314 |
+
"type": "text",
|
| 1315 |
+
"text": "[25] Eungbeom Kim, Jinhee Kim, Yoori Oh, Kyungsu Kim, Minju Park, Jaeheon Sim, Jinwoo Lee, and Kyogu Lee. Improving audio-language learning with mixgen and multi-level test-time augmentation. arXiv preprint arXiv:2210.17143, 2022. 7 ",
|
| 1316 |
+
"bbox": [
|
| 1317 |
+
171,
|
| 1318 |
+
272,
|
| 1319 |
+
823,
|
| 1320 |
+
310
|
| 1321 |
+
],
|
| 1322 |
+
"page_idx": 11
|
| 1323 |
+
},
|
| 1324 |
+
{
|
| 1325 |
+
"type": "text",
|
| 1326 |
+
"text": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 15 ",
|
| 1327 |
+
"bbox": [
|
| 1328 |
+
173,
|
| 1329 |
+
319,
|
| 1330 |
+
823,
|
| 1331 |
+
345
|
| 1332 |
+
],
|
| 1333 |
+
"page_idx": 11
|
| 1334 |
+
},
|
| 1335 |
+
{
|
| 1336 |
+
"type": "text",
|
| 1337 |
+
"text": "[27] Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis. Advances in Neural Information Processing Systems, 33:17022–17033, 2020. 5 ",
|
| 1338 |
+
"bbox": [
|
| 1339 |
+
174,
|
| 1340 |
+
356,
|
| 1341 |
+
825,
|
| 1342 |
+
393
|
| 1343 |
+
],
|
| 1344 |
+
"page_idx": 11
|
| 1345 |
+
},
|
| 1346 |
+
{
|
| 1347 |
+
"type": "text",
|
| 1348 |
+
"text": "[28] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song. Acav $1 0 0 \\mathrm { m }$ : Automatic curation of large-scale datasets for audio-visual video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10274–10284, 2021. 7 [29] Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiujun Li, Yizhe Zhang, and Jianfeng Gao. Optimus: Organizing sentences via pre-trained modeling of a latent space. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4678–4699, 2020. 5 ",
|
| 1349 |
+
"bbox": [
|
| 1350 |
+
173,
|
| 1351 |
+
404,
|
| 1352 |
+
823,
|
| 1353 |
+
443
|
| 1354 |
+
],
|
| 1355 |
+
"page_idx": 11
|
| 1356 |
+
},
|
| 1357 |
+
{
|
| 1358 |
+
"type": "text",
|
| 1359 |
+
"text": "",
|
| 1360 |
+
"bbox": [
|
| 1361 |
+
173,
|
| 1362 |
+
452,
|
| 1363 |
+
825,
|
| 1364 |
+
492
|
| 1365 |
+
],
|
| 1366 |
+
"page_idx": 11
|
| 1367 |
+
},
|
| 1368 |
+
{
|
| 1369 |
+
"type": "text",
|
| 1370 |
+
"text": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 7 ",
|
| 1371 |
+
"bbox": [
|
| 1372 |
+
171,
|
| 1373 |
+
500,
|
| 1374 |
+
825,
|
| 1375 |
+
526
|
| 1376 |
+
],
|
| 1377 |
+
"page_idx": 11
|
| 1378 |
+
},
|
| 1379 |
+
{
|
| 1380 |
+
"type": "text",
|
| 1381 |
+
"text": "[31] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXX 16, pages 121–137. Springer, 2020. 7 ",
|
| 1382 |
+
"bbox": [
|
| 1383 |
+
171,
|
| 1384 |
+
535,
|
| 1385 |
+
826,
|
| 1386 |
+
587
|
| 1387 |
+
],
|
| 1388 |
+
"page_idx": 11
|
| 1389 |
+
},
|
| 1390 |
+
{
|
| 1391 |
+
"type": "text",
|
| 1392 |
+
"text": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740–755. Springer, 2014. 7 ",
|
| 1393 |
+
"bbox": [
|
| 1394 |
+
173,
|
| 1395 |
+
595,
|
| 1396 |
+
823,
|
| 1397 |
+
647
|
| 1398 |
+
],
|
| 1399 |
+
"page_idx": 11
|
| 1400 |
+
},
|
| 1401 |
+
{
|
| 1402 |
+
"type": "text",
|
| 1403 |
+
"text": "[33] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint arXiv:2301.12503, 2023. 2, 3, 5 ",
|
| 1404 |
+
"bbox": [
|
| 1405 |
+
174,
|
| 1406 |
+
656,
|
| 1407 |
+
823,
|
| 1408 |
+
695
|
| 1409 |
+
],
|
| 1410 |
+
"page_idx": 11
|
| 1411 |
+
},
|
| 1412 |
+
{
|
| 1413 |
+
"type": "text",
|
| 1414 |
+
"text": "[34] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semanticconditional diffusion networks for image captioning. arXiv preprint arXiv:2212.03099, 2022. 7 ",
|
| 1415 |
+
"bbox": [
|
| 1416 |
+
171,
|
| 1417 |
+
704,
|
| 1418 |
+
823,
|
| 1419 |
+
732
|
| 1420 |
+
],
|
| 1421 |
+
"page_idx": 11
|
| 1422 |
+
},
|
| 1423 |
+
{
|
| 1424 |
+
"type": "text",
|
| 1425 |
+
"text": "[35] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3, 7 ",
|
| 1426 |
+
"bbox": [
|
| 1427 |
+
173,
|
| 1428 |
+
741,
|
| 1429 |
+
823,
|
| 1430 |
+
767
|
| 1431 |
+
],
|
| 1432 |
+
"page_idx": 11
|
| 1433 |
+
},
|
| 1434 |
+
{
|
| 1435 |
+
"type": "text",
|
| 1436 |
+
"text": "[36] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 7 ",
|
| 1437 |
+
"bbox": [
|
| 1438 |
+
173,
|
| 1439 |
+
776,
|
| 1440 |
+
823,
|
| 1441 |
+
815
|
| 1442 |
+
],
|
| 1443 |
+
"page_idx": 11
|
| 1444 |
+
},
|
| 1445 |
+
{
|
| 1446 |
+
"type": "text",
|
| 1447 |
+
"text": "[37] Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng. Automatic prompt optimization with\" gradient descent\" and beam search. arXiv preprint arXiv:2305.03495, 2023. 2 ",
|
| 1448 |
+
"bbox": [
|
| 1449 |
+
171,
|
| 1450 |
+
824,
|
| 1451 |
+
823,
|
| 1452 |
+
851
|
| 1453 |
+
],
|
| 1454 |
+
"page_idx": 11
|
| 1455 |
+
},
|
| 1456 |
+
{
|
| 1457 |
+
"type": "text",
|
| 1458 |
+
"text": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR, 2021. 3, 4 ",
|
| 1459 |
+
"bbox": [
|
| 1460 |
+
174,
|
| 1461 |
+
859,
|
| 1462 |
+
825,
|
| 1463 |
+
911
|
| 1464 |
+
],
|
| 1465 |
+
"page_idx": 11
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"type": "text",
|
| 1469 |
+
"text": "[39] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 5 ",
|
| 1470 |
+
"bbox": [
|
| 1471 |
+
169,
|
| 1472 |
+
92,
|
| 1473 |
+
823,
|
| 1474 |
+
118
|
| 1475 |
+
],
|
| 1476 |
+
"page_idx": 12
|
| 1477 |
+
},
|
| 1478 |
+
{
|
| 1479 |
+
"type": "text",
|
| 1480 |
+
"text": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3 ",
|
| 1481 |
+
"bbox": [
|
| 1482 |
+
173,
|
| 1483 |
+
127,
|
| 1484 |
+
823,
|
| 1485 |
+
154
|
| 1486 |
+
],
|
| 1487 |
+
"page_idx": 12
|
| 1488 |
+
},
|
| 1489 |
+
{
|
| 1490 |
+
"type": "text",
|
| 1491 |
+
"text": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 2, 3, 4, 5, 7 ",
|
| 1492 |
+
"bbox": [
|
| 1493 |
+
173,
|
| 1494 |
+
162,
|
| 1495 |
+
821,
|
| 1496 |
+
202
|
| 1497 |
+
],
|
| 1498 |
+
"page_idx": 12
|
| 1499 |
+
},
|
| 1500 |
+
{
|
| 1501 |
+
"type": "text",
|
| 1502 |
+
"text": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 6, 7 ",
|
| 1503 |
+
"bbox": [
|
| 1504 |
+
173,
|
| 1505 |
+
210,
|
| 1506 |
+
825,
|
| 1507 |
+
275
|
| 1508 |
+
],
|
| 1509 |
+
"page_idx": 12
|
| 1510 |
+
},
|
| 1511 |
+
{
|
| 1512 |
+
"type": "text",
|
| 1513 |
+
"text": "[43] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959–17968, 2022. 7 ",
|
| 1514 |
+
"bbox": [
|
| 1515 |
+
171,
|
| 1516 |
+
284,
|
| 1517 |
+
823,
|
| 1518 |
+
323
|
| 1519 |
+
],
|
| 1520 |
+
"page_idx": 12
|
| 1521 |
+
},
|
| 1522 |
+
{
|
| 1523 |
+
"type": "text",
|
| 1524 |
+
"text": "[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 7 ",
|
| 1525 |
+
"bbox": [
|
| 1526 |
+
173,
|
| 1527 |
+
332,
|
| 1528 |
+
823,
|
| 1529 |
+
371
|
| 1530 |
+
],
|
| 1531 |
+
"page_idx": 12
|
| 1532 |
+
},
|
| 1533 |
+
{
|
| 1534 |
+
"type": "text",
|
| 1535 |
+
"text": "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256–2265. PMLR, 2015. 2, 4 ",
|
| 1536 |
+
"bbox": [
|
| 1537 |
+
174,
|
| 1538 |
+
380,
|
| 1539 |
+
821,
|
| 1540 |
+
419
|
| 1541 |
+
],
|
| 1542 |
+
"page_idx": 12
|
| 1543 |
+
},
|
| 1544 |
+
{
|
| 1545 |
+
"type": "text",
|
| 1546 |
+
"text": "[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3 ",
|
| 1547 |
+
"bbox": [
|
| 1548 |
+
173,
|
| 1549 |
+
428,
|
| 1550 |
+
821,
|
| 1551 |
+
467
|
| 1552 |
+
],
|
| 1553 |
+
"page_idx": 12
|
| 1554 |
+
},
|
| 1555 |
+
{
|
| 1556 |
+
"type": "text",
|
| 1557 |
+
"text": "[47] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3 ",
|
| 1558 |
+
"bbox": [
|
| 1559 |
+
171,
|
| 1560 |
+
474,
|
| 1561 |
+
823,
|
| 1562 |
+
515
|
| 1563 |
+
],
|
| 1564 |
+
"page_idx": 12
|
| 1565 |
+
},
|
| 1566 |
+
{
|
| 1567 |
+
"type": "text",
|
| 1568 |
+
"text": "[48] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 7 ",
|
| 1569 |
+
"bbox": [
|
| 1570 |
+
173,
|
| 1571 |
+
523,
|
| 1572 |
+
823,
|
| 1573 |
+
563
|
| 1574 |
+
],
|
| 1575 |
+
"page_idx": 12
|
| 1576 |
+
},
|
| 1577 |
+
{
|
| 1578 |
+
"type": "text",
|
| 1579 |
+
"text": "[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-tosequence learning framework. arXiv preprint arXiv:2202.03052, 2022. 7 ",
|
| 1580 |
+
"bbox": [
|
| 1581 |
+
171,
|
| 1582 |
+
570,
|
| 1583 |
+
823,
|
| 1584 |
+
611
|
| 1585 |
+
],
|
| 1586 |
+
"page_idx": 12
|
| 1587 |
+
},
|
| 1588 |
+
{
|
| 1589 |
+
"type": "text",
|
| 1590 |
+
"text": "[50] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 7 ",
|
| 1591 |
+
"bbox": [
|
| 1592 |
+
171,
|
| 1593 |
+
619,
|
| 1594 |
+
825,
|
| 1595 |
+
659
|
| 1596 |
+
],
|
| 1597 |
+
"page_idx": 12
|
| 1598 |
+
},
|
| 1599 |
+
{
|
| 1600 |
+
"type": "text",
|
| 1601 |
+
"text": "[51] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pre-training for neural visual world creation. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVI, pages 720–736. Springer, 2022. 7 ",
|
| 1602 |
+
"bbox": [
|
| 1603 |
+
173,
|
| 1604 |
+
667,
|
| 1605 |
+
828,
|
| 1606 |
+
718
|
| 1607 |
+
],
|
| 1608 |
+
"page_idx": 12
|
| 1609 |
+
},
|
| 1610 |
+
{
|
| 1611 |
+
"type": "text",
|
| 1612 |
+
"text": "[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 7 ",
|
| 1613 |
+
"bbox": [
|
| 1614 |
+
171,
|
| 1615 |
+
728,
|
| 1616 |
+
823,
|
| 1617 |
+
767
|
| 1618 |
+
],
|
| 1619 |
+
"page_idx": 12
|
| 1620 |
+
},
|
| 1621 |
+
{
|
| 1622 |
+
"type": "text",
|
| 1623 |
+
"text": "[53] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 5, 7 ",
|
| 1624 |
+
"bbox": [
|
| 1625 |
+
171,
|
| 1626 |
+
776,
|
| 1627 |
+
825,
|
| 1628 |
+
803
|
| 1629 |
+
],
|
| 1630 |
+
"page_idx": 12
|
| 1631 |
+
},
|
| 1632 |
+
{
|
| 1633 |
+
"type": "text",
|
| 1634 |
+
"text": "[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036–5045, 2022. 6, 7, 16 ",
|
| 1635 |
+
"bbox": [
|
| 1636 |
+
173,
|
| 1637 |
+
811,
|
| 1638 |
+
825,
|
| 1639 |
+
863
|
| 1640 |
+
],
|
| 1641 |
+
"page_idx": 12
|
| 1642 |
+
},
|
| 1643 |
+
{
|
| 1644 |
+
"type": "text",
|
| 1645 |
+
"text": "[55] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian, Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework. arXiv preprint arXiv:2205.01818, 2022. 2, 3 ",
|
| 1646 |
+
"bbox": [
|
| 1647 |
+
171,
|
| 1648 |
+
872,
|
| 1649 |
+
823,
|
| 1650 |
+
911
|
| 1651 |
+
],
|
| 1652 |
+
"page_idx": 12
|
| 1653 |
+
},
|
| 1654 |
+
{
|
| 1655 |
+
"type": "text",
|
| 1656 |
+
"text": "[56] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375–16387, 2022. 3 \n[57] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634–23651, 2021. 3 \n[58] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph with teacher-recommended learning for video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13278–13288, 2020. 7 \n[59] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang, Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint arXiv:2211.11694, 2022. 7 ",
|
| 1657 |
+
"bbox": [
|
| 1658 |
+
171,
|
| 1659 |
+
92,
|
| 1660 |
+
826,
|
| 1661 |
+
286
|
| 1662 |
+
],
|
| 1663 |
+
"page_idx": 13
|
| 1664 |
+
},
|
| 1665 |
+
{
|
| 1666 |
+
"type": "text",
|
| 1667 |
+
"text": "A Model Architecture and Configuration ",
|
| 1668 |
+
"text_level": 1,
|
| 1669 |
+
"bbox": [
|
| 1670 |
+
174,
|
| 1671 |
+
89,
|
| 1672 |
+
532,
|
| 1673 |
+
107
|
| 1674 |
+
],
|
| 1675 |
+
"page_idx": 14
|
| 1676 |
+
},
|
| 1677 |
+
{
|
| 1678 |
+
"type": "text",
|
| 1679 |
+
"text": "A.1 Overview ",
|
| 1680 |
+
"text_level": 1,
|
| 1681 |
+
"bbox": [
|
| 1682 |
+
174,
|
| 1683 |
+
122,
|
| 1684 |
+
282,
|
| 1685 |
+
137
|
| 1686 |
+
],
|
| 1687 |
+
"page_idx": 14
|
| 1688 |
+
},
|
| 1689 |
+
{
|
| 1690 |
+
"type": "text",
|
| 1691 |
+
"text": "In this section, we provide more details on the model architecture as shown in Table 12, where each modality specific diffuser is based on UNet architecture with different variations detailed in the table. Another notable difference is the video architecture where we add temporal attention and temporal shift as discussed in Section 3.3 and we will discuss its detail in the next section. ",
|
| 1692 |
+
"bbox": [
|
| 1693 |
+
174,
|
| 1694 |
+
148,
|
| 1695 |
+
825,
|
| 1696 |
+
204
|
| 1697 |
+
],
|
| 1698 |
+
"page_idx": 14
|
| 1699 |
+
},
|
| 1700 |
+
{
|
| 1701 |
+
"type": "table",
|
| 1702 |
+
"img_path": "images/0f1eb37557b3404ed3d97b4e9270dbc9d4725f2fb42c6ba33e4d45037b6bd75d.jpg",
|
| 1703 |
+
"table_caption": [
|
| 1704 |
+
"Table 12: Hyperparameters for our diffusion models. Note the video and image generation uses the same diffuser. "
|
| 1705 |
+
],
|
| 1706 |
+
"table_footnote": [],
|
| 1707 |
+
"table_body": "<table><tr><td>Modality</td><td>Video (Image) LDM</td><td>Audio LDM</td><td>Text LDM</td></tr><tr><td colspan=\"4\">Hyperparameter</td></tr><tr><td>Architecture</td><td>LDM</td><td>LDM</td><td>LDM</td></tr><tr><td>z-shape</td><td>4× #frames × 64× 64</td><td>8× 256×16</td><td>768×1×1</td></tr><tr><td>Channels</td><td>320</td><td>320</td><td>320</td></tr><tr><td>Depth</td><td>4</td><td>2</td><td>2</td></tr><tr><td>Channel multiplier</td><td>1,2,4,4</td><td>1,2,4,4</td><td>1,2,4,4</td></tr><tr><td>Attention resolutions</td><td>64,32,16</td><td>64,32,16</td><td>64,32,16</td></tr><tr><td>Head channels</td><td>32</td><td>32</td><td>32</td></tr><tr><td>Number of heads</td><td>8</td><td>8</td><td>8</td></tr><tr><td>CA embed dim</td><td>768</td><td>768</td><td>768</td></tr><tr><td>CA resolutions</td><td>64,32,16</td><td>64,32,16</td><td>64,32,16</td></tr><tr><td>Autoencoders</td><td>AutoKL</td><td>AudioLDM</td><td>Optimus</td></tr><tr><td>Weight initialization</td><td>Stable Diffusion-1.4</td><td>-</td><td>Versatile Diffusion</td></tr><tr><td>Parameterization</td><td>E</td><td>E</td><td>E</td></tr><tr><td>Learning rate</td><td>2e-5</td><td>5e-6</td><td>5e-5</td></tr><tr><td>Total batch size</td><td>256</td><td>1024</td><td>1024</td></tr><tr><td colspan=\"4\">Diffusion Setup</td></tr><tr><td>Diffusion steps</td><td>1000</td><td>1000</td><td>1000</td></tr><tr><td>Noise schedule</td><td>Linear</td><td>Linear</td><td>Linear</td></tr><tr><td>β</td><td>0.00085</td><td>0.00085</td><td>0.00085</td></tr><tr><td>阳</td><td>0.0120</td><td>0.0120</td><td>0.0120</td></tr><tr><td colspan=\"4\">Sampling Parameters</td></tr><tr><td>Sampler</td><td>DDIM</td><td>DDIM</td><td>DDIM</td></tr><tr><td>Steps</td><td>50</td><td>50</td><td>50</td></tr><tr><td>n</td><td>1.0</td><td>1.0</td><td>1.0</td></tr><tr><td>Guidance scale</td><td>2.0</td><td>7.5</td><td>2.0</td></tr></table>",
|
| 1708 |
+
"bbox": [
|
| 1709 |
+
220,
|
| 1710 |
+
251,
|
| 1711 |
+
769,
|
| 1712 |
+
616
|
| 1713 |
+
],
|
| 1714 |
+
"page_idx": 14
|
| 1715 |
+
},
|
| 1716 |
+
{
|
| 1717 |
+
"type": "text",
|
| 1718 |
+
"text": "A.2 Video LDM Architecture ",
|
| 1719 |
+
"text_level": 1,
|
| 1720 |
+
"bbox": [
|
| 1721 |
+
174,
|
| 1722 |
+
648,
|
| 1723 |
+
392,
|
| 1724 |
+
662
|
| 1725 |
+
],
|
| 1726 |
+
"page_idx": 14
|
| 1727 |
+
},
|
| 1728 |
+
{
|
| 1729 |
+
"type": "text",
|
| 1730 |
+
"text": "Except for the base image UNet architecture, we also add temporal attention and temporal shift [2] before each residual block. Following VDM [21], the temporal attention is a transformer attention module where we flatten the height and width dimension to batch size dimension and the self-attention is performed on the time dimension. The temporal shift is illustrated in Fig. 6 where we first split channels into $k$ chunks. Then, we shift the channel dimension numbered 0 to $k - 1$ by temporal dimension from 0 to $k - 1$ times respectively. Eventually, we concatenate the shifted chunks by the hidden dimension. Note that we use $k = 3$ in the illustration for simplicity but $k = 8$ in our implementation. We then add a convolution layer before the temporal shift module. Finally, we use residual connection [18] and add the output to the input before the convolution layer. The complete video UNet layer is shown in Fig. 7. ",
|
| 1731 |
+
"bbox": [
|
| 1732 |
+
173,
|
| 1733 |
+
674,
|
| 1734 |
+
825,
|
| 1735 |
+
814
|
| 1736 |
+
],
|
| 1737 |
+
"page_idx": 14
|
| 1738 |
+
},
|
| 1739 |
+
{
|
| 1740 |
+
"type": "text",
|
| 1741 |
+
"text": "B Model Training ",
|
| 1742 |
+
"text_level": 1,
|
| 1743 |
+
"bbox": [
|
| 1744 |
+
174,
|
| 1745 |
+
835,
|
| 1746 |
+
339,
|
| 1747 |
+
853
|
| 1748 |
+
],
|
| 1749 |
+
"page_idx": 14
|
| 1750 |
+
},
|
| 1751 |
+
{
|
| 1752 |
+
"type": "text",
|
| 1753 |
+
"text": "Prompt Encoders Training. As discussed in Section 3.2, we use bridging alignment to perform contrastive learning between all prompt encoders. We use Adam [26] optimizer with learning rate 1e-4 and weight decay 1e-4. ",
|
| 1754 |
+
"bbox": [
|
| 1755 |
+
174,
|
| 1756 |
+
869,
|
| 1757 |
+
823,
|
| 1758 |
+
911
|
| 1759 |
+
],
|
| 1760 |
+
"page_idx": 14
|
| 1761 |
+
},
|
| 1762 |
+
{
|
| 1763 |
+
"type": "image",
|
| 1764 |
+
"img_path": "images/e11082636d4b4499d75ad6f7451413683dac74d3ebb613e652b9ea9f4c995784.jpg",
|
| 1765 |
+
"image_caption": [
|
| 1766 |
+
"Figure 6: Temporal shift [2] illustration. $C , H .$ , $W$ represent channel, height, width, respectively. The vertical line represents time steps from $t - 1 , t$ , and $t + 1$ . The grey blocks denote “padding tensors”. "
|
| 1767 |
+
],
|
| 1768 |
+
"image_footnote": [],
|
| 1769 |
+
"bbox": [
|
| 1770 |
+
222,
|
| 1771 |
+
87,
|
| 1772 |
+
772,
|
| 1773 |
+
337
|
| 1774 |
+
],
|
| 1775 |
+
"page_idx": 15
|
| 1776 |
+
},
|
| 1777 |
+
{
|
| 1778 |
+
"type": "image",
|
| 1779 |
+
"img_path": "images/4409bc09bf832575915923690ca96fa4e7ccd04266897f8300b3a97df95ceeee.jpg",
|
| 1780 |
+
"image_caption": [
|
| 1781 |
+
"Figure 7: Video UNet layer architecture details including normalization & activation, 2D temporal attention, followed by temporal shift and 1D spatial convolution. "
|
| 1782 |
+
],
|
| 1783 |
+
"image_footnote": [],
|
| 1784 |
+
"bbox": [
|
| 1785 |
+
348,
|
| 1786 |
+
388,
|
| 1787 |
+
643,
|
| 1788 |
+
500
|
| 1789 |
+
],
|
| 1790 |
+
"page_idx": 15
|
| 1791 |
+
},
|
| 1792 |
+
{
|
| 1793 |
+
"type": "text",
|
| 1794 |
+
"text": "Diffusion Model Training. We train diffusion model with training objectives and hyperparameters detailed in Table 1 and Table 12. For video LDM, we adopt a more specific training curriculum. We adopt curriculum learning on frame resolution and frames-per-second (FPS). First, the diffuser is trained on the WebVid dataset of a 256-frame resolution, with the training objective being textconditioned video generation. The training clips are sampled from 2-second video chunks with 4 FPS. Second, the model is further trained on HDVILLA and ACAV datasets, with a 512-frame resolution and 8 FPS, and the training objective is image-conditioned video generation (the image is a randomly sampled frame of the clip). Each training clip contains 16 frames sampled from a 2-second video chunk with 8 FPS. ",
|
| 1795 |
+
"bbox": [
|
| 1796 |
+
174,
|
| 1797 |
+
564,
|
| 1798 |
+
826,
|
| 1799 |
+
690
|
| 1800 |
+
],
|
| 1801 |
+
"page_idx": 15
|
| 1802 |
+
},
|
| 1803 |
+
{
|
| 1804 |
+
"type": "text",
|
| 1805 |
+
"text": "Joint Generation Training. As discussed in Section 3.2, we train joint generation by aligning environment encoders and optimize cross-attention layers only in the diffusion models. We use Adam optimizer with learning rate 1e-5 and weight decay 1e-4. ",
|
| 1806 |
+
"bbox": [
|
| 1807 |
+
174,
|
| 1808 |
+
705,
|
| 1809 |
+
823,
|
| 1810 |
+
747
|
| 1811 |
+
],
|
| 1812 |
+
"page_idx": 15
|
| 1813 |
+
},
|
| 1814 |
+
{
|
| 1815 |
+
"type": "text",
|
| 1816 |
+
"text": "C Training Datasets ",
|
| 1817 |
+
"text_level": 1,
|
| 1818 |
+
"bbox": [
|
| 1819 |
+
174,
|
| 1820 |
+
767,
|
| 1821 |
+
357,
|
| 1822 |
+
784
|
| 1823 |
+
],
|
| 1824 |
+
"page_idx": 15
|
| 1825 |
+
},
|
| 1826 |
+
{
|
| 1827 |
+
"type": "text",
|
| 1828 |
+
"text": "In this section, we introduce more details about the video and audiovisual training datasets. ",
|
| 1829 |
+
"bbox": [
|
| 1830 |
+
169,
|
| 1831 |
+
797,
|
| 1832 |
+
769,
|
| 1833 |
+
813
|
| 1834 |
+
],
|
| 1835 |
+
"page_idx": 15
|
| 1836 |
+
},
|
| 1837 |
+
{
|
| 1838 |
+
"type": "text",
|
| 1839 |
+
"text": "Video. WebVid [4] is a large-scale dataset of web videos with diverse content, spanning over 40 categories such as sports, cooking, and travel. It contains over 1.2 million video clips (all without sound) that are all at least 30 seconds in duration with video descriptions. We perform text video and video-text contrastive learning task with this dataset. HD-Villa-100M [54] is a large-scale video dataset with over 100 million video clips sourced from YouTube. The dataset covers a wide range of video categories and includes high-quality videos with a resolution of at least 720P. Since it lacks curated video description and we use the middle frame as image input to perform image video generation. ",
|
| 1840 |
+
"bbox": [
|
| 1841 |
+
174,
|
| 1842 |
+
827,
|
| 1843 |
+
825,
|
| 1844 |
+
911
|
| 1845 |
+
],
|
| 1846 |
+
"page_idx": 15
|
| 1847 |
+
},
|
| 1848 |
+
{
|
| 1849 |
+
"type": "text",
|
| 1850 |
+
"text": "",
|
| 1851 |
+
"bbox": [
|
| 1852 |
+
173,
|
| 1853 |
+
90,
|
| 1854 |
+
823,
|
| 1855 |
+
119
|
| 1856 |
+
],
|
| 1857 |
+
"page_idx": 16
|
| 1858 |
+
},
|
| 1859 |
+
{
|
| 1860 |
+
"type": "text",
|
| 1861 |
+
"text": "Audiovisual. SoundNet originally contains over two million sounds and spans a wide range of categories including music, animal sounds, natural sounds, and environmental sounds. We collected all currently accessible 1M videos. ",
|
| 1862 |
+
"bbox": [
|
| 1863 |
+
174,
|
| 1864 |
+
133,
|
| 1865 |
+
825,
|
| 1866 |
+
176
|
| 1867 |
+
],
|
| 1868 |
+
"page_idx": 16
|
| 1869 |
+
},
|
| 1870 |
+
{
|
| 1871 |
+
"type": "text",
|
| 1872 |
+
"text": "D Limitations & Broader Impacts ",
|
| 1873 |
+
"text_level": 1,
|
| 1874 |
+
"bbox": [
|
| 1875 |
+
176,
|
| 1876 |
+
195,
|
| 1877 |
+
473,
|
| 1878 |
+
213
|
| 1879 |
+
],
|
| 1880 |
+
"page_idx": 16
|
| 1881 |
+
},
|
| 1882 |
+
{
|
| 1883 |
+
"type": "text",
|
| 1884 |
+
"text": "While the paper primarily focuses on the technical advancements and potential applications of CoDi, we also consider potential negative social impacts that could arise from the development and deployment of such technology. These impacts can include: ",
|
| 1885 |
+
"bbox": [
|
| 1886 |
+
176,
|
| 1887 |
+
227,
|
| 1888 |
+
823,
|
| 1889 |
+
268
|
| 1890 |
+
],
|
| 1891 |
+
"page_idx": 16
|
| 1892 |
+
},
|
| 1893 |
+
{
|
| 1894 |
+
"type": "text",
|
| 1895 |
+
"text": "Deepfakes and Misinformation. As part of a common issue for generative AI models, the ability of CoDi to generate realistic and synchronized multimodal outputs also raises concerns about the creation and dissemination of deepfakes. Malicious actors could exploit this technology to create highly convincing fake content, such as fabricated videos or audio clips, which can be used for misinformation, fraud, or other harmful purposes. ",
|
| 1896 |
+
"bbox": [
|
| 1897 |
+
174,
|
| 1898 |
+
284,
|
| 1899 |
+
825,
|
| 1900 |
+
353
|
| 1901 |
+
],
|
| 1902 |
+
"page_idx": 16
|
| 1903 |
+
},
|
| 1904 |
+
{
|
| 1905 |
+
"type": "text",
|
| 1906 |
+
"text": "Bias and Stereotyping. If the training data used for CoDi is biased or contains stereotypes, the generated multimodal outputs may also reflect these. ",
|
| 1907 |
+
"bbox": [
|
| 1908 |
+
173,
|
| 1909 |
+
367,
|
| 1910 |
+
823,
|
| 1911 |
+
396
|
| 1912 |
+
],
|
| 1913 |
+
"page_idx": 16
|
| 1914 |
+
},
|
| 1915 |
+
{
|
| 1916 |
+
"type": "text",
|
| 1917 |
+
"text": "E License ",
|
| 1918 |
+
"text_level": 1,
|
| 1919 |
+
"bbox": [
|
| 1920 |
+
174,
|
| 1921 |
+
415,
|
| 1922 |
+
272,
|
| 1923 |
+
433
|
| 1924 |
+
],
|
| 1925 |
+
"page_idx": 16
|
| 1926 |
+
},
|
| 1927 |
+
{
|
| 1928 |
+
"type": "text",
|
| 1929 |
+
"text": "We will publicly release our code and checkpoints. We cite licenses from the individual dataset or package we use from the community and provide the following links for references. ",
|
| 1930 |
+
"bbox": [
|
| 1931 |
+
176,
|
| 1932 |
+
446,
|
| 1933 |
+
823,
|
| 1934 |
+
474
|
| 1935 |
+
],
|
| 1936 |
+
"page_idx": 16
|
| 1937 |
+
},
|
| 1938 |
+
{
|
| 1939 |
+
"type": "text",
|
| 1940 |
+
"text": "LAION-400M: Creative Common CC-BY 4.0 ",
|
| 1941 |
+
"bbox": [
|
| 1942 |
+
176,
|
| 1943 |
+
481,
|
| 1944 |
+
483,
|
| 1945 |
+
494
|
| 1946 |
+
],
|
| 1947 |
+
"page_idx": 16
|
| 1948 |
+
},
|
| 1949 |
+
{
|
| 1950 |
+
"type": "text",
|
| 1951 |
+
"text": "AudioSet: Creative Common CC-BY 4.0 ",
|
| 1952 |
+
"bbox": [
|
| 1953 |
+
176,
|
| 1954 |
+
502,
|
| 1955 |
+
449,
|
| 1956 |
+
516
|
| 1957 |
+
],
|
| 1958 |
+
"page_idx": 16
|
| 1959 |
+
},
|
| 1960 |
+
{
|
| 1961 |
+
"type": "text",
|
| 1962 |
+
"text": "AudioCaps: MIT ",
|
| 1963 |
+
"bbox": [
|
| 1964 |
+
174,
|
| 1965 |
+
522,
|
| 1966 |
+
294,
|
| 1967 |
+
536
|
| 1968 |
+
],
|
| 1969 |
+
"page_idx": 16
|
| 1970 |
+
},
|
| 1971 |
+
{
|
| 1972 |
+
"type": "text",
|
| 1973 |
+
"text": "Freesound: Creative Commons ",
|
| 1974 |
+
"bbox": [
|
| 1975 |
+
174,
|
| 1976 |
+
542,
|
| 1977 |
+
387,
|
| 1978 |
+
558
|
| 1979 |
+
],
|
| 1980 |
+
"page_idx": 16
|
| 1981 |
+
},
|
| 1982 |
+
{
|
| 1983 |
+
"type": "text",
|
| 1984 |
+
"text": "BBC Sound Effect: The BBC’s Content Licence ",
|
| 1985 |
+
"bbox": [
|
| 1986 |
+
176,
|
| 1987 |
+
563,
|
| 1988 |
+
500,
|
| 1989 |
+
578
|
| 1990 |
+
],
|
| 1991 |
+
"page_idx": 16
|
| 1992 |
+
},
|
| 1993 |
+
{
|
| 1994 |
+
"type": "text",
|
| 1995 |
+
"text": "SoundNet: MIT ",
|
| 1996 |
+
"bbox": [
|
| 1997 |
+
174,
|
| 1998 |
+
584,
|
| 1999 |
+
285,
|
| 2000 |
+
598
|
| 2001 |
+
],
|
| 2002 |
+
"page_idx": 16
|
| 2003 |
+
},
|
| 2004 |
+
{
|
| 2005 |
+
"type": "text",
|
| 2006 |
+
"text": "Webvid10M: Webvid ",
|
| 2007 |
+
"bbox": [
|
| 2008 |
+
174,
|
| 2009 |
+
604,
|
| 2010 |
+
320,
|
| 2011 |
+
619
|
| 2012 |
+
],
|
| 2013 |
+
"page_idx": 16
|
| 2014 |
+
},
|
| 2015 |
+
{
|
| 2016 |
+
"type": "text",
|
| 2017 |
+
"text": "HD-Villa-100M: Research Use of Data Agreement v1.0 ",
|
| 2018 |
+
"bbox": [
|
| 2019 |
+
178,
|
| 2020 |
+
626,
|
| 2021 |
+
544,
|
| 2022 |
+
640
|
| 2023 |
+
],
|
| 2024 |
+
"page_idx": 16
|
| 2025 |
+
},
|
| 2026 |
+
{
|
| 2027 |
+
"type": "text",
|
| 2028 |
+
"text": "PyTorch: BSD-style ",
|
| 2029 |
+
"bbox": [
|
| 2030 |
+
174,
|
| 2031 |
+
646,
|
| 2032 |
+
315,
|
| 2033 |
+
661
|
| 2034 |
+
],
|
| 2035 |
+
"page_idx": 16
|
| 2036 |
+
},
|
| 2037 |
+
{
|
| 2038 |
+
"type": "text",
|
| 2039 |
+
"text": "Huggingface Transformers: Apache ",
|
| 2040 |
+
"bbox": [
|
| 2041 |
+
174,
|
| 2042 |
+
667,
|
| 2043 |
+
424,
|
| 2044 |
+
683
|
| 2045 |
+
],
|
| 2046 |
+
"page_idx": 16
|
| 2047 |
+
},
|
| 2048 |
+
{
|
| 2049 |
+
"type": "text",
|
| 2050 |
+
"text": "Torchvision: BSD 3-Clause ",
|
| 2051 |
+
"bbox": [
|
| 2052 |
+
174,
|
| 2053 |
+
688,
|
| 2054 |
+
361,
|
| 2055 |
+
702
|
| 2056 |
+
],
|
| 2057 |
+
"page_idx": 16
|
| 2058 |
+
},
|
| 2059 |
+
{
|
| 2060 |
+
"type": "text",
|
| 2061 |
+
"text": "Torchaudio: BSD 2-Clause ",
|
| 2062 |
+
"bbox": [
|
| 2063 |
+
174,
|
| 2064 |
+
708,
|
| 2065 |
+
361,
|
| 2066 |
+
723
|
| 2067 |
+
],
|
| 2068 |
+
"page_idx": 16
|
| 2069 |
+
}
|
| 2070 |
+
]
|
parse/dev/2EDqbSCnmF/2EDqbSCnmF_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/2EDqbSCnmF/2EDqbSCnmF_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/3RBY8fKjHeu/images/159d86a4fe017221206965fa98efc6ce35e16bebec7536f231b04a5fa470830b.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/17f2d11eee9937e70f62a1993623ebccd221887d067e71919c350fa57662f4d3.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/24fe88bf92baa43778d9defa3450750bc0d2c910fa9c12c5902630d7c2316e1e.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/29586778ea39548465206a088f601c3e740302c103692a574d7a70d8188b3983.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/372acb53fc3f9ab9f178baac319f8b0cb0c4ee1ca96f374bada7397c4ec23630.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/476b14497f73d983953dbe7a34cff454d210243e28b1aa97dafb54ae6a92e42f.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/69863294723843746383e47fe99dcd32744e499d1c526f51df4121c52ff99fe8.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/69e7a0dc11e3a7ecd812dec526777f2a39e7aed63605587ef789903e7f57fb8c.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/bbd9aa6b3f541685e1ecf9dd1c4451b92904b361a6547ee2e39414769cb64de4.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/bced559cf73caae8b60315e212417c82fda960dadfbf4c1dc28fc8ea7be19aee.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/bd74dd0968c35c4dc8ffeba6cc51d9f109521ca66008a561e30a7fcccfe148a6.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/d3f8da6b396f76d54da62528b9e91defa6543c22e48371ec0f3e6d668424058f.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/d59a106b3ad7e1f46762af4cb6c79589191a386db6dd18da2381b810f0e72627.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/db6cf431ae9355646aa06c810c30e311d8db38009707b4dea4bad788085ac2bb.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/dfce202941b6d7b7a3b4e91b152da625264b3b1c43837193ab53e137e11b01f3.jpg
ADDED
|
Git LFS Details
|
parse/dev/3RBY8fKjHeu/images/e30f877426a1aa2686b70c08629a56889403c96105b017890f8ea57b7982c4a2.jpg
ADDED
|
Git LFS Details
|
parse/dev/3jooF27-0Wy/3jooF27-0Wy_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:213b7f7cc0a8c9f79f4b4679a33e747c85e204c79a1962e41d069a901f8a0c80
|
| 3 |
+
size 2090090
|
parse/dev/3jooF27-0Wy/3jooF27-0Wy_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9ab05785720cdb5cbe34ba164da4bc54815bf96dfa9a7bdb764e35ec20a9d487
|
| 3 |
+
size 1782421
|
parse/dev/3jooF27-0Wy/3jooF27-0Wy_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6bc3f99fc7c5f8768431adecae1d3c5af4ce52184cbbb1162f0c0dd0a4886084
|
| 3 |
+
size 2086544
|
parse/dev/3mRwyG5one/3mRwyG5one_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e9325cdff54c3c0d0f250f946474e19512eb6e38ac6f5a48a2dfd5f2f29b51ba
|
| 3 |
+
size 2189525
|
parse/dev/3mRwyG5one/3mRwyG5one_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2477ac6b0fb89f3e8eb5fea6ad4298820847e4fbf027b4833277f1c07f3b753a
|
| 3 |
+
size 1961956
|
parse/dev/3mRwyG5one/3mRwyG5one_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2668dfbe9d2f87838e211c3934118fdda9dd9324bfd38530d688c41d8dcbcdb1
|
| 3 |
+
size 2195239
|
parse/dev/4NpoSrT8uU-/4NpoSrT8uU-_content_list.json
ADDED
|
@@ -0,0 +1,1264 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Deep Bidirectional Language-Knowledge Graph Pretraining ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
209,
|
| 8 |
+
122,
|
| 9 |
+
790,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Michihiro Yasunaga,1 Antoine Bosselut,2 Hongyu Ren,1 Xikun Zhang1 Christopher D Manning,1 Percy Liang,1⇤ Jure Leskovec1⇤ ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
238,
|
| 19 |
+
224,
|
| 20 |
+
754,
|
| 21 |
+
255
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "1Stanford University 2EPFL ⇤Equal senior authorship {myasu,antoineb,hyren,xikunz2,manning,pliang,jure}@cs.stanford.edu ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
250,
|
| 30 |
+
256,
|
| 31 |
+
746,
|
| 32 |
+
284
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Abstract ",
|
| 39 |
+
"text_level": 1,
|
| 40 |
+
"bbox": [
|
| 41 |
+
462,
|
| 42 |
+
318,
|
| 43 |
+
535,
|
| 44 |
+
335
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "Pretraining a language model (LM) on text has been shown to help various downstream NLP tasks. Recent works show that a knowledge graph (KG) can complement text data, offering structured background knowledge that provides a useful scaffold for reasoning. However, these works are not pretrained to learn a deep fusion of the two modalities at scale, limiting the potential to acquire fully joint representations of text and KG. Here we propose DRAGON (Deep Bidirectional Language-Knowledge Graph Pretraining), a self-supervised method to pretrain a deeply joint language-knowledge foundation model from text and KG at scale. Specifically, our model takes pairs of text segments and relevant KG subgraphs as input and bidirectionally fuses information from both modalities. We pretrain this model by unifying two self-supervised reasoning tasks, masked language modeling and KG link prediction. DRAGON outperforms existing LM and $_ { \\mathrm { L M + K G } }$ models on diverse downstream tasks including question answering across general and biomedical domains, with $+ 5 \\%$ absolute gain on average. In particular, DRAGON achieves strong performance on complex reasoning about language and knowledge $( + 1 0 \\%$ on questions involving long contexts or multi-step reasoning) and low-resource QA $+ 8 \\%$ on OBQA and RiddleSense), and new state-of-the-art results on various BioNLP tasks. Our code and trained models are available at https://github.com/michiyasunaga/dragon. ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
232,
|
| 53 |
+
351,
|
| 54 |
+
766,
|
| 55 |
+
613
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "1 Introduction ",
|
| 62 |
+
"text_level": 1,
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
641,
|
| 66 |
+
310,
|
| 67 |
+
657
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Pretraining learns self-supervised representations from massive raw data to help various downstream tasks [1]. Language models (LMs) pretrained on large amounts of text data, such as BERT [2] and GPTs [3], have shown strong performance on many natural language processing (NLP) tasks. The success of these models comes from deeply interactive (contextualized) representations of input tokens learned at scale via self-supervision [2, 4]. Meanwhile, large knowledge graphs (KGs), such as Freebase [5], Wikidata [6] and ConceptNet [7], can provide complementary information to text data. KGs offer structured background knowledge by representing entities as nodes and relations between them as edges, and also offer scaffolds for structured, multi-step reasoning about entities [8, 9, 10, 11] (§3.4.1). The dual strengths of text data and KGs motivate research in pretraining deeply interactive representations of the two modalities at scale. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
672,
|
| 77 |
+
825,
|
| 78 |
+
810
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "How to effectively combine text and KGs for pretraining is an open problem and presents challenges. Given text and KG, we need both (i) a deeply bidirectional model for the two modalities to interact, and (ii) a self-supervised objective to learn joint reasoning over text and KG at scale. Several existing works [12, 13, 14, 15, 16] propose methods for self-supervised pretraining, but they fuse text and KG in a shallow or uni-directional manner. Another line of work [8, 9] proposes bidirectional models for text and KG, but these models focus on finetuning on labeled downstream tasks and do not perform self-supervised learning. Consequently, existing methods may have limited their potential to model and learn deep interactions over text and KG. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
818,
|
| 88 |
+
825,
|
| 89 |
+
900
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/1ab98d065bce88d339bfe2f9ad29f53937430d9e3f7347ee0eb7fe530fdcec25.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: Overview of our approach, DRAGON. Left: Given raw data of a text corpus and a large knowledge graph, we create aligned (text, local KG) pairs by sampling a text segment from the corpus and extracting a relevant subgraph from the KG (§2.1). As the structured knowledge in KG can ground the text and the text can provide the KG with rich context for reasoning, we aim to pretrain a language-knowledge model jointly from the text-KG pairs (DRAGON). Right: To model the interactions over text and KG, DRAGON uses a cross-modal encoder that bidirectionally exchanges information between them to produce fused text token and KG node representations (§2.2). To pretrain DRAGON jointly on text and KG, we unify two self-supervised reasoning tasks: (1) masked language modeling, which masks some tokens in the input text and then predicts them, and (2) link prediction, which holds out some edges from the input KG and then predicts them. This joint objective encourages text and KG to mutually inform each other, facilitating the model to learn joint reasoning over text and KG (§2.3). "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
178,
|
| 102 |
+
68,
|
| 103 |
+
823,
|
| 104 |
+
281
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "",
|
| 111 |
+
"bbox": [
|
| 112 |
+
174,
|
| 113 |
+
428,
|
| 114 |
+
823,
|
| 115 |
+
457
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "To address both of the above challenges and fully unify the strengths of text and KG, we propose DRAGON (Deep Bidirectional Language-Knowledge Graph Pretraining), an approach that performs deeply bidirectional, self-supervised pretraining of a language-knowledge model from text and KG. DRAGON has two core components: a cross-modal model that bidirectionally fuses text and KG, and a bidirectional self-supervised objective that learns joint reasoning over text and KG. Concretely, as in Figure 1, we take a text corpus and a KG as raw data, and create inputs for the model by sampling a text segment from the corpus and extracting a relevant subgraph from the KG via entity linking, obtaining a (text, local $K G$ ) pair. We use a cross-modal model to encode this input into fused representations, where each layer of the model encodes the text with an LM and the KG with a graph neural network (GNN), and fuses the two with a bidirectional modality interaction module (GreaseLM [9]). We pretrain this model by unifying two self-supervised reasoning tasks: (1) masked language modeling (MLM), which masks and predicts tokens in the input text, and (2) link prediction, which drops and predicts edges in the input KG. The intuition is that by combining the two tasks, MLM makes the model use the text jointly with structured knowledge in the KG to reason about masked tokens in the text (e.g., in Figure 1, using the “round brush”–“art supply” multi-hop path from the KG helps), and link prediction makes the model use the KG structure jointly with the textual context to reason about missing links in the KG (e.g., recognizing that “round brush could be used for hair” from the text helps). This joint objective thus enables text to be grounded by KG structure and KG to be contextualized by text simultaneously, producing a deeply-unified language-knowledge pretrained model where information flows bidirectionally between text and KG for reasoning. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
462,
|
| 125 |
+
825,
|
| 126 |
+
738
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "We pretrain DRAGON in two domains: a general domain, using the Book corpus and ConceptNet KG [7] (§3), and a biomedical domain, using the PubMed corpus and UMLS KG [17] (§4). We show that DRAGON improves on existing LM and $_ { \\mathrm { L M + K G } }$ models on diverse downstream tasks across domains. For the general domain, DRAGON outperforms RoBERTa [18], our base LM without KGs, on various commonsense reasoning tasks such as CSQA, OBQA, RiddleSense and HellaSwag, with $+ 8 \\%$ absolute accuracy gain on average. For the biomedical domain, DRAGON improves on the previous best LM, BioLinkBERT [19], and sets a new state of the art on BioNLP tasks such as MedQA and PubMedQA, with $+ 3 \\%$ accuracy gain. In particular, DRAGON exhibits notable improvements on QA tasks involving complex reasoning $+ 1 0 \\%$ gain on multi-step, negation, hedge, or long context reasoning) and on downstream tasks with limited training data $( + 8 \\%$ gain). These results show that our deep bidirectional self-supervision over text and KG produces significantly improved language-knowledge representations compared to existing models. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
746,
|
| 136 |
+
825,
|
| 137 |
+
911
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "1.1 Related work ",
|
| 144 |
+
"text_level": 1,
|
| 145 |
+
"bbox": [
|
| 146 |
+
174,
|
| 147 |
+
92,
|
| 148 |
+
305,
|
| 149 |
+
106
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 2
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "Knowledge-augmented LM pretraining. Knowledge integration is active research for improving LMs. One line of works is retrieval-augmented LMs [20, 21, 22], which retrieve relevant text from a corpus and integrate it into LMs as additional knowledge. Orthogonal to these works, we focus on using knowledge bases as background knowledge, to ground reasoning about entities and facts. ",
|
| 156 |
+
"bbox": [
|
| 157 |
+
174,
|
| 158 |
+
116,
|
| 159 |
+
825,
|
| 160 |
+
171
|
| 161 |
+
],
|
| 162 |
+
"page_idx": 2
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"type": "text",
|
| 166 |
+
"text": "Closest to our work are works that integrate knowledge bases in LM pretraining. One line of research aims to add entity features to LMs [12, 23, 24]; Some works use the KG entity information or structure to create additional training signals [13, 25, 14, 26, 27, 28]; Several works add KG triplet information directly to the LM input [29, 16, 15, 30, 31]. While these methods have achieved substantial progress, they typically propagate information between text and KG in a shallow or uni-directional (e.g., KG to text) manner, which might limit the potential to perform fully joint reasoning over the two modalities. To improve on the above works, we propose to bidirectionally interact text and KG via a deep cross-modal model and joint self-supervision, so that text and KG are grounded and contextualized by each other. We find that this improves model performance on various reasoning tasks (§3). Another distinction is that existing works in this space typically focus on adding entity- or triplet-level knowledge from KGs to LMs, and focus on solving entity/relation classification tasks. Our work significantly expands this scope in that we use larger KG subgraphs (200 nodes) as input to enable richer contextualization between KG and text, and we achieve performance improvements on a broader set of NLP tasks including QA, reasoning and text classification tasks. ",
|
| 167 |
+
"bbox": [
|
| 168 |
+
174,
|
| 169 |
+
179,
|
| 170 |
+
825,
|
| 171 |
+
372
|
| 172 |
+
],
|
| 173 |
+
"page_idx": 2
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"type": "text",
|
| 177 |
+
"text": "KG-augmented question answering. Various works designed KG-augmented reasoning models for question answering [32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42]. In particular, recent works such as QA-GNN [8] and GreaseLM [9] suggest that a KG can scaffold reasoning about entities with its graph structure, and help for complex question answering (e.g., negation, multi-hop reasoning). These works typically focus on training or finetuning models on particular QA datasets. In contrast, we generalize this and integrate KG-augmented reasoning into general-purpose pretraining. Our motivation is that self-supervised pretraining allows the model to learn from larger and more diverse data, helping to learn richer interactions between text and KGs and to acquire more diverse reasoning abilities beyond specific QA tasks. We find that our proposed pretraining approach (DRAGON) offers significant boosts over the baseline QA models (e.g. GreaseLM) on diverse downstream tasks (§3). This opens a new research avenue in scaling up various carefully-designed QA models to pretraining. ",
|
| 178 |
+
"bbox": [
|
| 179 |
+
174,
|
| 180 |
+
378,
|
| 181 |
+
825,
|
| 182 |
+
530
|
| 183 |
+
],
|
| 184 |
+
"page_idx": 2
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"type": "text",
|
| 188 |
+
"text": "KG representation learning. Our link prediction task used in pretraining is motivated by research in KG representation learning. Link prediction is a fundamental task in KGs [43, 44], and various works study methods to learn KG entity and relation embeddings for link prediction, such as TransE [45], DistMult [46] and RotatE [47]. Several works additionally use textual data or pretrained LMs to help learn KG embeddings and link prediction [48, 49, 50, 51, 52, 53]. While these works focus on the KG-side representations, we extend the scope and use the KG-side objective (link prediction) jointly with a text-side objective (language modeling) to train a mutually-interactive text-KG model. ",
|
| 189 |
+
"bbox": [
|
| 190 |
+
174,
|
| 191 |
+
536,
|
| 192 |
+
825,
|
| 193 |
+
633
|
| 194 |
+
],
|
| 195 |
+
"page_idx": 2
|
| 196 |
+
},
|
| 197 |
+
{
|
| 198 |
+
"type": "text",
|
| 199 |
+
"text": "2 Deep Bidirectional Language-Knowledge Graph Pretraining (DRAGON) ",
|
| 200 |
+
"text_level": 1,
|
| 201 |
+
"bbox": [
|
| 202 |
+
171,
|
| 203 |
+
651,
|
| 204 |
+
805,
|
| 205 |
+
670
|
| 206 |
+
],
|
| 207 |
+
"page_idx": 2
|
| 208 |
+
},
|
| 209 |
+
{
|
| 210 |
+
"type": "text",
|
| 211 |
+
"text": "We propose DRAGON, an approach that performs deeply bidirectional, self-supervised pretraining of a language-knowledge model from text and KG. Specifically, as illustrated in Figure 1, we take a text corpus and a large knowledge graph as raw data, and create input instances for the model by sampling coarsely-aligned (text segment, local KG) pairs (§2.1). To learn mutual interactions over text and KG, DRAGON consists of a cross-modal encoder (GreaseLM) that fuses the input text-KG pair bidirectionally $( \\ S 2 . 2 )$ , and a pretraining objective that performs bidirectional self-supervision on the text-KG input $( \\ S 2 . 3 )$ . Our pretraining objective unifies masked language modeling (MLM) and KG link prediction (LinkPred) to make text and KG mutually inform each other and learn joint reasoning over them. Finally, we describe how we finetune the pretrained DRAGON model for downstream tasks (§2.4). While each individual piece of our approach (GreaseLM, MLM, LinkPred) is not new in itself, we are the first to bring them together effectively and demonstrate that the resulting model has strong empirical results (§3, §4). ",
|
| 212 |
+
"bbox": [
|
| 213 |
+
174,
|
| 214 |
+
683,
|
| 215 |
+
825,
|
| 216 |
+
849
|
| 217 |
+
],
|
| 218 |
+
"page_idx": 2
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"type": "text",
|
| 222 |
+
"text": "Definitions. We define a text corpus $\\mathcal { W }$ as a set of text segments $\\mathcal { W } = \\{ W \\}$ , and each text segment $W$ as a sequence of tokens (words), $W = ( w _ { 1 } , . . . , w _ { I } )$ . We define a knowledge graph (KG) as a multi-relational graph $\\mathcal { G } = ( \\nu , \\mathcal { E } )$ , where $\\nu$ is the set of entity nodes in the KG and $\\mathcal { E } \\subseteq \\mathcal { V } \\times \\mathcal { R } \\times \\mathcal { V }$ is the set of edges (triplets) that connect nodes in $\\nu$ , with $\\mathcal { R }$ being the set of relation types $\\{ r \\}$ ",
|
| 223 |
+
"bbox": [
|
| 224 |
+
176,
|
| 225 |
+
856,
|
| 226 |
+
825,
|
| 227 |
+
911
|
| 228 |
+
],
|
| 229 |
+
"page_idx": 2
|
| 230 |
+
},
|
| 231 |
+
{
|
| 232 |
+
"type": "text",
|
| 233 |
+
"text": "Each triplet $( h , r , t )$ in a KG can represent a knowledge fact such as (Paris, in, France). As a raw KG is often large, with millions of nodes, a subgraph of the raw KG (local KG) is considered: $G = ( V , E )$ where $\\bar { V } = \\{ v _ { 1 } , . . . , v _ { J } \\} \\subseteq \\mathcal { V }$ and $E \\subseteq { \\mathcal { E } }$ . We define a language-knowledge model to be a composition of two functions, $f _ { \\mathrm { h e a d } } ( f _ { \\mathrm { e n c } } ( X ) )$ , where the encoder $f _ { \\mathrm { e n c } }$ takes in an input $X =$ (text segment $W$ , local ${ \\bf K G } \\breve { G }$ ), and produces a contextualized vector representation for each text token, $( \\mathbf { H } _ { 1 } , . . . , \\mathbf { H } _ { I } )$ , and for each KG node, $( \\mathbf { V } _ { 1 } , . . . , \\mathbf { V } _ { J } )$ . A language model is a special case of a language-knowledge model with no KG ( $J = 0$ ). The head $f _ { \\mathrm { h e a d } }$ uses these representations to perform self-supervised tasks in the pretraining step and to perform downstream tasks in the finetuning step. ",
|
| 234 |
+
"bbox": [
|
| 235 |
+
173,
|
| 236 |
+
90,
|
| 237 |
+
825,
|
| 238 |
+
203
|
| 239 |
+
],
|
| 240 |
+
"page_idx": 3
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"type": "text",
|
| 244 |
+
"text": "2.1 Input representation ",
|
| 245 |
+
"text_level": 1,
|
| 246 |
+
"bbox": [
|
| 247 |
+
174,
|
| 248 |
+
219,
|
| 249 |
+
356,
|
| 250 |
+
234
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 3
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "text",
|
| 256 |
+
"text": "Given a text corpus $\\mathcal { W }$ and a large knowledge graph $\\mathcal { G }$ , we create input instances for the model by preparing (text segment $W$ , local KG $G$ ) pairs. We want each pair’s text and KG to be (roughly) semantically aligned so that the text and KG can mutually inform each other and facilitate the model to learn interactive reasoning between the two modalities. Specifically, for each text segment $W$ from $\\mathcal { W }$ , we extract a relevant local KG $G$ for it from $\\mathcal { G }$ via the following KG retrieval process. ",
|
| 257 |
+
"bbox": [
|
| 258 |
+
174,
|
| 259 |
+
244,
|
| 260 |
+
825,
|
| 261 |
+
314
|
| 262 |
+
],
|
| 263 |
+
"page_idx": 3
|
| 264 |
+
},
|
| 265 |
+
{
|
| 266 |
+
"type": "text",
|
| 267 |
+
"text": "KG retrieval. Given a text segment $W$ , we link entity mentions in $W$ to entity nodes in $\\mathcal { G }$ to get an initial set of nodes $V _ { \\mathrm { e l } }$ . We then add their 2-hop bridge nodes from $\\mathcal { G }$ to get the total retrieved nodes $V \\subseteq \\mathcal { V }$ . Lastly, we add all edges that span these nodes in $\\mathcal { G }$ to get $E \\subseteq { \\mathcal { E } }$ , which yields the final local KG, $G = ( V , E )$ , as well as our final input instance $X = ( W , G )$ . Appendix B.1 provides more details on KG retrieval. Henceforth, we use “KG” to refer to this local KG $G$ unless noted otherwise. ",
|
| 268 |
+
"bbox": [
|
| 269 |
+
174,
|
| 270 |
+
320,
|
| 271 |
+
825,
|
| 272 |
+
390
|
| 273 |
+
],
|
| 274 |
+
"page_idx": 3
|
| 275 |
+
},
|
| 276 |
+
{
|
| 277 |
+
"type": "text",
|
| 278 |
+
"text": "Modality interaction token/node. For each resulting (text, KG) pair, we further add a special token (interaction token) $w _ { \\mathrm { i n t } }$ to the text and a special node (interaction node) $v _ { \\mathrm { i n t } }$ to the KG, which will serve as an information pooling point for each modality as well as an interface for modality interaction in our cross-modal encoder (§2.2). Specifically, we prepend $w _ { \\mathrm { i n t } }$ to the original text $W { = } ( w _ { 1 } , . . . , w _ { I } )$ , and connect $v _ { \\mathrm { i n t } }$ to the entity-linked nodes in the original KG, $V _ { \\mathrm { e l } } \\subseteq V = \\{ v _ { 1 } , . . . , v _ { J } \\}$ , using a new relation type $r _ { \\mathrm { e l } }$ . The interaction token and node can also be used to produce a pooled representation of the whole input, e.g., when finetuning for classification tasks (§2.4). ",
|
| 279 |
+
"bbox": [
|
| 280 |
+
173,
|
| 281 |
+
396,
|
| 282 |
+
825,
|
| 283 |
+
494
|
| 284 |
+
],
|
| 285 |
+
"page_idx": 3
|
| 286 |
+
},
|
| 287 |
+
{
|
| 288 |
+
"type": "text",
|
| 289 |
+
"text": "2.2 Cross-modal encoder ",
|
| 290 |
+
"text_level": 1,
|
| 291 |
+
"bbox": [
|
| 292 |
+
174,
|
| 293 |
+
511,
|
| 294 |
+
361,
|
| 295 |
+
525
|
| 296 |
+
],
|
| 297 |
+
"page_idx": 3
|
| 298 |
+
},
|
| 299 |
+
{
|
| 300 |
+
"type": "text",
|
| 301 |
+
"text": "To model mutual interactions over the text and KG, we use a bidirectional sequence-graph encoder for $f _ { \\mathrm { e n c } }$ which takes in the text tokens and KG nodes and exchanges information across them for multiple layers to produce a fused representation of each token and node (Figure 1 right): ",
|
| 302 |
+
"bbox": [
|
| 303 |
+
173,
|
| 304 |
+
536,
|
| 305 |
+
825,
|
| 306 |
+
579
|
| 307 |
+
],
|
| 308 |
+
"page_idx": 3
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"type": "equation",
|
| 312 |
+
"img_path": "images/ca8e44835026d135daa70f8efa0f564cd344b939b41272cb002ae77a839e2f3e.jpg",
|
| 313 |
+
"text": "$$\n\\mathbf { \\boldsymbol { \\mathbf { V } } } _ { 1 } , . . . , \\mathbf { \\boldsymbol { \\mathbf { V } } } _ { J } ) = f _ { \\mathrm { e n c } } \\big ( \\big ( \\boldsymbol { w } _ { \\mathrm { i n t } } , \\boldsymbol { w } _ { 1 } , . . . , \\boldsymbol { w } _ { I } \\big ) , \\big ( \\boldsymbol { v } _ { \\mathrm { i n t } } , \\boldsymbol { v } _ { 1 } , . . . , \\boldsymbol { v } _ { J } \\big )\n$$",
|
| 314 |
+
"text_format": "latex",
|
| 315 |
+
"bbox": [
|
| 316 |
+
406,
|
| 317 |
+
584,
|
| 318 |
+
735,
|
| 319 |
+
599
|
| 320 |
+
],
|
| 321 |
+
"page_idx": 3
|
| 322 |
+
},
|
| 323 |
+
{
|
| 324 |
+
"type": "text",
|
| 325 |
+
"text": "While we may use any deep bidirectional sequence-graph encoder for $f _ { \\mathrm { e n c } }$ , for controlled comparison with existing works, we adopt the existing top-performing sequence-graph architecture, GreaseLM [9], which combines Transformers [54] and graph neural networks (GNNs) to fuse text-KG inputs. ",
|
| 326 |
+
"bbox": [
|
| 327 |
+
173,
|
| 328 |
+
604,
|
| 329 |
+
826,
|
| 330 |
+
647
|
| 331 |
+
],
|
| 332 |
+
"page_idx": 3
|
| 333 |
+
},
|
| 334 |
+
{
|
| 335 |
+
"type": "text",
|
| 336 |
+
"text": "Specifically, GreaseLM first uses $N$ layers of Transformer language model (LM) layers to map the input text into initial token representations, and uses KG node embeddings to map the input KG nodes into initial node representations, ",
|
| 337 |
+
"bbox": [
|
| 338 |
+
174,
|
| 339 |
+
652,
|
| 340 |
+
821,
|
| 341 |
+
694
|
| 342 |
+
],
|
| 343 |
+
"page_idx": 3
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"type": "equation",
|
| 347 |
+
"img_path": "images/c145087203c3dbb871ca978f5574dc44f62fea1936247a88ea5e3ef7e4709a24.jpg",
|
| 348 |
+
"text": "$$\n\\begin{array} { r l } & { ( \\mathbf { H } _ { \\mathrm { i n t } } ^ { ( 0 ) } , \\mathbf { H } _ { 1 } ^ { ( 0 ) } , . . . , \\mathbf { H } _ { I } ^ { ( 0 ) } ) = \\mathrm { L M } \\mathrm { - } \\mathrm { L a y e r s } ( w _ { \\mathrm { i n t } } , w _ { 1 } . . . , w _ { I } ) , } \\\\ & { ( \\mathbf { V } _ { \\mathrm { i n t } } ^ { ( 0 ) } , \\mathbf { V } _ { 1 } ^ { ( 0 ) } , . . . , \\mathbf { V } _ { J } ^ { ( 0 ) } ) = \\mathrm { N o d e - } \\mathrm { E m b e d d i n g } ( v _ { \\mathrm { i n t } } , v _ { 1 } , . . . , v _ { J } ) . } \\end{array}\n$$",
|
| 349 |
+
"text_format": "latex",
|
| 350 |
+
"bbox": [
|
| 351 |
+
330,
|
| 352 |
+
696,
|
| 353 |
+
669,
|
| 354 |
+
736
|
| 355 |
+
],
|
| 356 |
+
"page_idx": 3
|
| 357 |
+
},
|
| 358 |
+
{
|
| 359 |
+
"type": "text",
|
| 360 |
+
"text": "Then it uses $M$ layers of text-KG fusion layers to encode these token/node representations jointly into the final token/node representations, ",
|
| 361 |
+
"bbox": [
|
| 362 |
+
174,
|
| 363 |
+
738,
|
| 364 |
+
823,
|
| 365 |
+
767
|
| 366 |
+
],
|
| 367 |
+
"page_idx": 3
|
| 368 |
+
},
|
| 369 |
+
{
|
| 370 |
+
"type": "equation",
|
| 371 |
+
"img_path": "images/73e238db0b2b8cfca2e2f9f7806820d8cfcfcb90133f9ba15aaae71a17b51eee.jpg",
|
| 372 |
+
"text": "$$\n( \\mathbf { H } _ { \\mathrm { i n t } } , . . . , \\mathbf { H } _ { I } ) , ( { \\mathbf { V } } _ { \\mathrm { i n t } } , . . . , \\mathbf { V } _ { J } ) = \\mathrm { F u s i o n } . \\mathrm { L a y e r s } ( ( \\mathbf { H } _ { \\mathrm { i n t } } ^ { ( 0 ) } , . . . , \\mathbf { H } _ { I } ^ { ( 0 ) } ) , ( { \\mathbf { V } } _ { \\mathrm { i n t } } ^ { ( 0 ) } , . . . , \\mathbf { V } _ { J } ^ { ( 0 ) } ) ) ,\n$$",
|
| 373 |
+
"text_format": "latex",
|
| 374 |
+
"bbox": [
|
| 375 |
+
264,
|
| 376 |
+
768,
|
| 377 |
+
732,
|
| 378 |
+
787
|
| 379 |
+
],
|
| 380 |
+
"page_idx": 3
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"type": "text",
|
| 384 |
+
"text": "where each of the fusion layers $( \\ell { = } 1 , . . . , M )$ performs the following: ",
|
| 385 |
+
"bbox": [
|
| 386 |
+
171,
|
| 387 |
+
791,
|
| 388 |
+
638,
|
| 389 |
+
806
|
| 390 |
+
],
|
| 391 |
+
"page_idx": 3
|
| 392 |
+
},
|
| 393 |
+
{
|
| 394 |
+
"type": "equation",
|
| 395 |
+
"img_path": "images/fcc8f823b7020d66644ac9868e4f75619bb61071d083025d2e2747ed17d43c9a.jpg",
|
| 396 |
+
"text": "$$\n\\begin{array} { r l } & { ( \\widetilde { \\mathbf { H } } _ { \\mathrm { i n t } } ^ { ( \\ell ) } , \\mathbf { H } _ { 1 } ^ { ( \\ell ) } , . . . , \\mathbf { H } _ { I } ^ { ( \\ell ) } ) = \\mathrm { L M } \\mathrm { L a y e r } ( \\mathbf { H } _ { \\mathrm { i n t } } ^ { ( \\ell - 1 ) } , \\mathbf { H } _ { 1 } ^ { ( \\ell - 1 ) } , . . . , \\mathbf { H } _ { I } ^ { ( \\ell - 1 ) } ) , } \\\\ & { ( \\widetilde { \\mathbf { V } } _ { \\mathrm { i n t } } ^ { ( \\ell ) } , \\mathbf { V } _ { 1 } ^ { ( \\ell ) } , . . . , \\mathbf { V } _ { J } ^ { ( \\ell ) } ) = \\mathrm { G N N } \\mathrm { - L a y e r } ( \\mathbf { V } _ { \\mathrm { i n t } } ^ { ( \\ell - 1 ) } , \\mathbf { V } _ { 1 } ^ { ( \\ell - 1 ) } , . . . , \\mathbf { V } _ { J } ^ { ( \\ell - 1 ) } ) , } \\\\ & { \\qquad [ \\mathbf { H } _ { \\mathrm { i n t } } ^ { ( \\ell ) } ; \\mathbf { V } _ { \\mathrm { i n t } } ^ { ( \\ell ) } ] = \\mathrm { M I n t } ( [ \\widetilde { \\mathbf { H } } _ { \\mathrm { i n t } } ^ { ( \\ell ) } ; \\widetilde { \\mathbf { V } } _ { \\mathrm { i n t } } ^ { ( \\ell ) } ] ) . } \\end{array}\n$$",
|
| 397 |
+
"text_format": "latex",
|
| 398 |
+
"bbox": [
|
| 399 |
+
310,
|
| 400 |
+
808,
|
| 401 |
+
686,
|
| 402 |
+
867
|
| 403 |
+
],
|
| 404 |
+
"page_idx": 3
|
| 405 |
+
},
|
| 406 |
+
{
|
| 407 |
+
"type": "text",
|
| 408 |
+
"text": "Here GNN induces graph structure-aware representations of KG nodes, $[ \\cdot ; \\cdot ]$ does concatenation, and MInt (modality interaction module) exchanges information between the interaction token (text side) and interaction node (KG side) via an MLP. For more details on GreaseLM, we refer readers to [9]. ",
|
| 409 |
+
"bbox": [
|
| 410 |
+
178,
|
| 411 |
+
869,
|
| 412 |
+
823,
|
| 413 |
+
911
|
| 414 |
+
],
|
| 415 |
+
"page_idx": 3
|
| 416 |
+
},
|
| 417 |
+
{
|
| 418 |
+
"type": "text",
|
| 419 |
+
"text": "2.3 Pretraining objective ",
|
| 420 |
+
"text_level": 1,
|
| 421 |
+
"bbox": [
|
| 422 |
+
174,
|
| 423 |
+
90,
|
| 424 |
+
359,
|
| 425 |
+
106
|
| 426 |
+
],
|
| 427 |
+
"page_idx": 4
|
| 428 |
+
},
|
| 429 |
+
{
|
| 430 |
+
"type": "text",
|
| 431 |
+
"text": "We aim to pretrain the DRAGON model so that it learns joint reasoning over text and a KG. To ensure that the text and KG mutually inform each other and the model learns bidirectional information flow, we unify two self-supervised reasoning tasks: masked language modeling and KG link prediction. ",
|
| 432 |
+
"bbox": [
|
| 433 |
+
174,
|
| 434 |
+
117,
|
| 435 |
+
825,
|
| 436 |
+
160
|
| 437 |
+
],
|
| 438 |
+
"page_idx": 4
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"type": "text",
|
| 442 |
+
"text": "Masked language modeling (MLM). MLM is a common pretraining task used for language models (e.g., BERT [2], RoBERTa [18]), which masks some tokens in the input text and predicts them. This task makes the model use non-masked context to reason about masked tokens, and in particular, as our approach takes a joint text-KG pair as input, we expect that MLM can encourage the model to learn to use the text jointly with structured knowledge in the KG to reason about masks in the text (e.g., in the example of Figure 1, besides the textual context, recognizing the “round brush”–“art supply” path from the KG can help together to predict the masked tokens “art supplies”). ",
|
| 443 |
+
"bbox": [
|
| 444 |
+
173,
|
| 445 |
+
165,
|
| 446 |
+
825,
|
| 447 |
+
263
|
| 448 |
+
],
|
| 449 |
+
"page_idx": 4
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"type": "text",
|
| 453 |
+
"text": "Concretely, to perform the MLM task, we mask a subset of tokens in the input text, $M \\subseteq W$ , with a special token [MASK], and let the task head $f _ { \\mathrm { h e a d } }$ be a linear layer that takes the contextualized token vectors $\\left\\{ \\mathbf { H } _ { i } \\right\\}$ from the encoder to predict the original tokens. The objective is a cross-entropy loss: ",
|
| 454 |
+
"bbox": [
|
| 455 |
+
174,
|
| 456 |
+
268,
|
| 457 |
+
825,
|
| 458 |
+
310
|
| 459 |
+
],
|
| 460 |
+
"page_idx": 4
|
| 461 |
+
},
|
| 462 |
+
{
|
| 463 |
+
"type": "equation",
|
| 464 |
+
"img_path": "images/55cb351675c00cb9b00d348283caee185fd72b722eb4a959d5b01d8dbd27a7c6.jpg",
|
| 465 |
+
"text": "$$\n\\mathcal { L } _ { \\mathrm { M L M } } = - \\sum _ { i \\in M } \\log p ( w _ { i } \\mid \\mathbf { H } _ { i } ) .\n$$",
|
| 466 |
+
"text_format": "latex",
|
| 467 |
+
"bbox": [
|
| 468 |
+
398,
|
| 469 |
+
314,
|
| 470 |
+
598,
|
| 471 |
+
347
|
| 472 |
+
],
|
| 473 |
+
"page_idx": 4
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"type": "text",
|
| 477 |
+
"text": "Link prediction (LinkPred). While the MLM task predicts for the text side, link prediction holds out some edges and predicts them for the input KG. Link prediction is a fundamental task in KGs [47] and makes the model use the structure of KGs to perform reasoning (e.g., using a compositional path “X’s mother’s husband is $\\mathbf { Y } ^ { \\ast }$ to deduce a missing link “X’s father is $\\mathbf { Y } ^ { \\prime \\prime }$ ). In particular, as our approach takes a joint text-KG pair as input, we expect that link prediction can encourage the model to learn to use the KG structure jointly with the textual context to reason about missing links in the KG (e.g., in Figure 1, besides the KG structure, recognizing that “round brush could be used for hair” from the text can help together to predict the held-out edge (round_brush, at, hair)). ",
|
| 478 |
+
"bbox": [
|
| 479 |
+
173,
|
| 480 |
+
362,
|
| 481 |
+
825,
|
| 482 |
+
476
|
| 483 |
+
],
|
| 484 |
+
"page_idx": 4
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"type": "text",
|
| 488 |
+
"text": "Concretely, to perform the link prediction task, we hold out a subset of edge triplets from the input KG, $S = \\mathcal { \\bar { \\{ } ( h , r , t ) \\} } \\subseteq E$ . For the task head $f _ { \\mathrm { h e a d } }$ , we adopt a KG representation learning framework, which maps each entity node $h$ or $t$ ) and relation $( r )$ in the KG to a vector, $\\mathbf { h } , \\mathbf { t } , \\mathbf { r }$ , and defines a scoring function $\\phi _ { r } ( \\mathbf { h } , \\mathbf { t } )$ to model positive/negative triplets. Specifically, we let $\\mathbf { h } = \\mathbf { V } _ { h }$ , $\\mathbf { t } = \\mathbf { V } _ { t }$ , $\\mathbf { r } = \\mathbf { R } _ { r }$ , with $\\{ \\mathbf { V } _ { j } \\}$ being the contextualized node vectors from the encoder, and $\\mathbf { R } = \\{ \\mathbf { r } _ { 1 } , . . . , \\mathbf { r } _ { | \\mathcal { R } | } \\}$ being learnable relation embeddings. We consider a KG triplet scoring function $\\phi _ { r } ( \\mathbf { h } , \\mathbf { t } )$ such as ",
|
| 489 |
+
"bbox": [
|
| 490 |
+
173,
|
| 491 |
+
479,
|
| 492 |
+
826,
|
| 493 |
+
565
|
| 494 |
+
],
|
| 495 |
+
"page_idx": 4
|
| 496 |
+
},
|
| 497 |
+
{
|
| 498 |
+
"type": "equation",
|
| 499 |
+
"img_path": "images/477b7c488dfe218200d45405a1fa41e7d7071ccd814ae711b7607dd3cab53469.jpg",
|
| 500 |
+
"text": "$$\n\\mathrm { D i s t M u l t } \\ [ 4 6 ] \\cdot \\langle \\mathbf { h } , \\mathbf { r } , \\mathbf { t } \\rangle , \\quad \\mathrm { T r a n s E } \\ [ 4 5 ] \\colon - \\| \\mathbf { h } + \\mathbf { r } - \\mathbf { t } \\| , \\quad \\mathrm { R o t a t E } \\ [ 4 7 ] \\colon - \\| \\mathbf { h } \\odot \\mathbf { r } - \\mathbf { t } \\| ,\n$$",
|
| 501 |
+
"text_format": "latex",
|
| 502 |
+
"bbox": [
|
| 503 |
+
220,
|
| 504 |
+
571,
|
| 505 |
+
774,
|
| 506 |
+
589
|
| 507 |
+
],
|
| 508 |
+
"page_idx": 4
|
| 509 |
+
},
|
| 510 |
+
{
|
| 511 |
+
"type": "text",
|
| 512 |
+
"text": "where $\\langle \\cdot , \\cdot , \\cdot \\rangle$ denotes the trilinear dot product and $\\odot$ the Hadamard product. A higher $\\phi$ indicates a higher chance of $( h , r , t )$ being a positive triplet (edge) instead of negative (no edge). We analyze the choices of scoring functions in $\\ S 3 . 4 . 3$ . For training, we optimize the objective: ",
|
| 513 |
+
"bbox": [
|
| 514 |
+
174,
|
| 515 |
+
594,
|
| 516 |
+
825,
|
| 517 |
+
637
|
| 518 |
+
],
|
| 519 |
+
"page_idx": 4
|
| 520 |
+
},
|
| 521 |
+
{
|
| 522 |
+
"type": "equation",
|
| 523 |
+
"img_path": "images/b1aa64da3ade84b3f83d5fae6f0130a1f73809a7372df9f7e19dd05e4318d12a.jpg",
|
| 524 |
+
"text": "$$\n\\mathcal { L } _ { \\mathrm { L i n k P r e d } } = \\sum _ { ( h , r , t ) \\in S } \\left( - \\log \\sigma ( \\phi _ { r } ( \\mathbf { h } , \\mathbf { t } ) + \\gamma ) + \\frac { 1 } { n } \\sum _ { ( h ^ { \\prime } , r , t ^ { \\prime } ) } \\log \\sigma ( \\phi _ { r } ( \\mathbf { h } ^ { \\prime } , \\mathbf { t } ^ { \\prime } ) + \\gamma ) \\right) ,\n$$",
|
| 525 |
+
"text_format": "latex",
|
| 526 |
+
"bbox": [
|
| 527 |
+
240,
|
| 528 |
+
641,
|
| 529 |
+
754,
|
| 530 |
+
688
|
| 531 |
+
],
|
| 532 |
+
"page_idx": 4
|
| 533 |
+
},
|
| 534 |
+
{
|
| 535 |
+
"type": "text",
|
| 536 |
+
"text": "where $( h ^ { \\prime } , r , t ^ { \\prime } )$ are $n$ negative samples corresponding to the positive triplet $( h , r , t )$ , $\\gamma$ is the margin, and $\\sigma$ is the sigmoid function. The intuition of this objective is to make the model predict triplets of the held-out edges $S$ as positive and other random triplets as negative. ",
|
| 537 |
+
"bbox": [
|
| 538 |
+
174,
|
| 539 |
+
693,
|
| 540 |
+
825,
|
| 541 |
+
736
|
| 542 |
+
],
|
| 543 |
+
"page_idx": 4
|
| 544 |
+
},
|
| 545 |
+
{
|
| 546 |
+
"type": "text",
|
| 547 |
+
"text": "Joint training. To pretrain DRAGON, we optimize the MLM and LinkPred objectives jointly: $\\begin{array} { r } { \\mathcal { L } = \\mathcal { L } _ { \\mathrm { M L M } } + \\mathcal { L } _ { \\mathrm { L i n k P r e d } } } \\end{array}$ . This joint objective unifies the effects of MLM and LinkPred, which encourage the model to simultaneously ground text with KG structure and contextualize KG with text, facilitating bidirectional information flow between text and KGs for reasoning. We show in $\\ S 3 . 4 . 3$ that the joint objective yields a more performant model than using one of the objectives alone. ",
|
| 548 |
+
"bbox": [
|
| 549 |
+
173,
|
| 550 |
+
741,
|
| 551 |
+
825,
|
| 552 |
+
811
|
| 553 |
+
],
|
| 554 |
+
"page_idx": 4
|
| 555 |
+
},
|
| 556 |
+
{
|
| 557 |
+
"type": "text",
|
| 558 |
+
"text": "2.4 Finetuning ",
|
| 559 |
+
"text_level": 1,
|
| 560 |
+
"bbox": [
|
| 561 |
+
174,
|
| 562 |
+
829,
|
| 563 |
+
289,
|
| 564 |
+
844
|
| 565 |
+
],
|
| 566 |
+
"page_idx": 4
|
| 567 |
+
},
|
| 568 |
+
{
|
| 569 |
+
"type": "text",
|
| 570 |
+
"text": "Lastly, we describe how we finetune DRAGON for downstream tasks such as text classification and multiple-choice QA (MCQA). Given an input text $W$ (e.g., concatenation of a question and an answer choice in the case of MCQA), we follow the same steps as $\\ S 2 . 1$ and $\\ S 2 . 2$ to retrieve a relevant local KG $G$ and encode them jointly into contextualized token/node vectors, $( \\mathbf { H } _ { \\mathrm { i n t } } , \\mathbf { H } _ { 1 } , \\ldots$ , ${ \\mathbf { H } } _ { I }$ ), $( \\mathbf { V } _ { \\mathrm { i n t } } , \\mathbf { V } _ { 1 } , . . . , \\mathbf { V } _ { J } )$ . We then compute a pooled representation of the whole input as $\\mathbf { X } =$ $\\mathbf { M L P } ( \\mathbf { H } _ { \\mathrm { i n t } } , \\mathbf { V } _ { \\mathrm { i n t } } , \\mathbf { G } )$ , where $\\mathbf { G }$ denotes attention-based pooling of $\\{ \\mathbf { V } _ { j } \\mid v _ { j } \\in \\{ v _ { 1 } , . . . , \\bar { v } _ { J } \\} \\}$ using $\\mathbf { H } _ { \\mathrm { i n t } }$ as a query. Finally, the pooled representation $\\mathbf { X }$ is used to perform the downstream task, in the same way as how the [CLS] representation is used in LMs such as BERT and RoBERTa. ",
|
| 571 |
+
"bbox": [
|
| 572 |
+
174,
|
| 573 |
+
854,
|
| 574 |
+
825,
|
| 575 |
+
911
|
| 576 |
+
],
|
| 577 |
+
"page_idx": 4
|
| 578 |
+
},
|
| 579 |
+
{
|
| 580 |
+
"type": "text",
|
| 581 |
+
"text": "",
|
| 582 |
+
"bbox": [
|
| 583 |
+
174,
|
| 584 |
+
90,
|
| 585 |
+
825,
|
| 586 |
+
147
|
| 587 |
+
],
|
| 588 |
+
"page_idx": 5
|
| 589 |
+
},
|
| 590 |
+
{
|
| 591 |
+
"type": "text",
|
| 592 |
+
"text": "The difference from GreaseLM is that while GreaseLM only performs finetuning as described in this section (hence, it is an LM finetuned with KGs), DRAGON performs self-supervised pretraining as described in $\\ S 2 . 3$ (hence, it can be viewed as an LM pretrained $^ +$ finetuned with KGs). ",
|
| 593 |
+
"bbox": [
|
| 594 |
+
174,
|
| 595 |
+
152,
|
| 596 |
+
825,
|
| 597 |
+
195
|
| 598 |
+
],
|
| 599 |
+
"page_idx": 5
|
| 600 |
+
},
|
| 601 |
+
{
|
| 602 |
+
"type": "text",
|
| 603 |
+
"text": "3 Experiments: General domain ",
|
| 604 |
+
"text_level": 1,
|
| 605 |
+
"bbox": [
|
| 606 |
+
176,
|
| 607 |
+
214,
|
| 608 |
+
459,
|
| 609 |
+
232
|
| 610 |
+
],
|
| 611 |
+
"page_idx": 5
|
| 612 |
+
},
|
| 613 |
+
{
|
| 614 |
+
"type": "text",
|
| 615 |
+
"text": "We experiment with the proposed approach DRAGON in a general domain first. We pretrain DRAGON using the Book corpus and ConceptNet KG (§3.1), and evaluate on diverse downstream tasks (§3.2). We show that DRAGON significantly improves on existing models (§3.4). We extensively analyze the effect of DRAGON’s key design choices such as self-supervision and use of KGs (§3.4.1, 3.4.2, 3.4.3). We also experiment in the biomedical domain in $\\ S 4$ . ",
|
| 616 |
+
"bbox": [
|
| 617 |
+
174,
|
| 618 |
+
246,
|
| 619 |
+
826,
|
| 620 |
+
315
|
| 621 |
+
],
|
| 622 |
+
"page_idx": 5
|
| 623 |
+
},
|
| 624 |
+
{
|
| 625 |
+
"type": "text",
|
| 626 |
+
"text": "3.1 Pretraining setup ",
|
| 627 |
+
"text_level": 1,
|
| 628 |
+
"bbox": [
|
| 629 |
+
174,
|
| 630 |
+
332,
|
| 631 |
+
334,
|
| 632 |
+
347
|
| 633 |
+
],
|
| 634 |
+
"page_idx": 5
|
| 635 |
+
},
|
| 636 |
+
{
|
| 637 |
+
"type": "text",
|
| 638 |
+
"text": "Data. For the text data, we use documents involving commonsense, BookCorpus [55]. BookCorpus has 6GB of text from online books and is widely used in LM pretraining (e.g., BERT, RoBERTa). For the KG data, we use ConceptNet [7], a general-domain knowledge graph designed to capture background commonsense knowledge. It has 800K nodes and 2M edges in total. To create a training instance, we sample a text segment of length up to 512 tokens from the text corpus, then retrieve a relevant KG subgraph of size up to 200 nodes (details in Appendix B.1), by which we obtain an aligned (text, local KG) pair. ",
|
| 639 |
+
"bbox": [
|
| 640 |
+
173,
|
| 641 |
+
357,
|
| 642 |
+
825,
|
| 643 |
+
454
|
| 644 |
+
],
|
| 645 |
+
"page_idx": 5
|
| 646 |
+
},
|
| 647 |
+
{
|
| 648 |
+
"type": "text",
|
| 649 |
+
"text": "Implementation. For our encoder (§2.2), we use the exact same architecture as GreaseLM [9] (19 LM layers followed by 5 text-KG fusion layers; 360M parameters in total). As done by [9], we initialize parameters in the LM component with the RoBERTa-Large release [18] and initialize the KG node embeddings with pre-computed ConceptNet entity embeddings (details in Appendix B.2). For the link prediction objective (§2.3, Equation 10), we use DistMult [46] for KG triplet scoring, with a negative exampling of 128 triplets and a margin of $\\gamma = 0$ . To pretrain the model, we perform MLM with a token masking rate of $15 \\%$ and link prediction with an edge drop rate of $15 \\%$ . We pretrain for 20,000 steps with a batch size of 8,192 and a learning rate of 2e-5 for parameters in the LM component and 3e-4 for the others. Training took 7 days on eight A100 GPUs using FP16. Additional details on the hyperparameters can be found in Appendix B.3. ",
|
| 650 |
+
"bbox": [
|
| 651 |
+
174,
|
| 652 |
+
460,
|
| 653 |
+
826,
|
| 654 |
+
599
|
| 655 |
+
],
|
| 656 |
+
"page_idx": 5
|
| 657 |
+
},
|
| 658 |
+
{
|
| 659 |
+
"type": "text",
|
| 660 |
+
"text": "3.2 Downstream evaluation tasks ",
|
| 661 |
+
"text_level": 1,
|
| 662 |
+
"bbox": [
|
| 663 |
+
176,
|
| 664 |
+
616,
|
| 665 |
+
416,
|
| 666 |
+
630
|
| 667 |
+
],
|
| 668 |
+
"page_idx": 5
|
| 669 |
+
},
|
| 670 |
+
{
|
| 671 |
+
"type": "text",
|
| 672 |
+
"text": "We finetune and evaluate DRAGON on nine diverse commonsense reasoning benchmarks: CommonsenseQA (CSQA) [56], OpenbookQA (OBQA) [57], RiddleSense (Riddle) [58], AI2 Reasoning Challenge–Challenge Set (ARC) [59], CosmosQA [60], HellaSwag [61], Physical Interaction QA (PIQA) [62], Social Interaction QA (SIQA) [63], and Abductive Natural Language Inference (aNLI) [64]. For CSQA, we follow the in-house data splits used by prior works [32]. For OBQA, we follow the original setting where the models only use the question as input and do not use the extra science facts. Appendix B.4 provides the full details on these tasks and data splits. Hyperparameters used for finetuning can be found in Appendix B.3. ",
|
| 673 |
+
"bbox": [
|
| 674 |
+
174,
|
| 675 |
+
641,
|
| 676 |
+
825,
|
| 677 |
+
752
|
| 678 |
+
],
|
| 679 |
+
"page_idx": 5
|
| 680 |
+
},
|
| 681 |
+
{
|
| 682 |
+
"type": "text",
|
| 683 |
+
"text": "3.3 Baselines ",
|
| 684 |
+
"text_level": 1,
|
| 685 |
+
"bbox": [
|
| 686 |
+
174,
|
| 687 |
+
768,
|
| 688 |
+
277,
|
| 689 |
+
782
|
| 690 |
+
],
|
| 691 |
+
"page_idx": 5
|
| 692 |
+
},
|
| 693 |
+
{
|
| 694 |
+
"type": "text",
|
| 695 |
+
"text": "LM. To study the effect of using KGs, we compare DRAGON with the vanilla language model, RoBERTa [18]. As we initialize DRAGON’s parameters using the RoBERTa-Large release (§3.1), for fair comparison, we let the baseline be such that we take the RoBERTa-Large release and continue pretraining it with the vanilla MLM objective on the same text data for the same number of steps as DRAGON. Hence, the only difference is that DRAGON uses KGs during pretraining while RoBERTa does not. We then perform standard LM finetuning of RoBERTa on downstream tasks. ",
|
| 696 |
+
"bbox": [
|
| 697 |
+
174,
|
| 698 |
+
792,
|
| 699 |
+
825,
|
| 700 |
+
877
|
| 701 |
+
],
|
| 702 |
+
"page_idx": 5
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"type": "text",
|
| 706 |
+
"text": "LM finetuned with KG. We also compare with existing KG-augmented QA models, QAGNN [8] and GreaseLM [9], which finetune a vanilla LM (i.e. RoBERTa-Large) with a KG on downstream tasks, but do not pretrain with a KG. GreaseLM is the existing top-performing model in this paradigm. As we use the same encoder architecture as GreaseLM for DRAGON, the only difference from GreaseLM is that DRAGON performs self-supervised pretraining while GreaseLM does not. ",
|
| 707 |
+
"bbox": [
|
| 708 |
+
173,
|
| 709 |
+
882,
|
| 710 |
+
823,
|
| 711 |
+
911
|
| 712 |
+
],
|
| 713 |
+
"page_idx": 5
|
| 714 |
+
},
|
| 715 |
+
{
|
| 716 |
+
"type": "table",
|
| 717 |
+
"img_path": "images/5909285334a3b68ad8f6a34757237ad849525b13556b63bbdadb9c523e7a7bae.jpg",
|
| 718 |
+
"table_caption": [
|
| 719 |
+
"Table 1: Accuracy on downstream commonsense reasoning tasks. DRAGON consistently outperforms the existing LM (RoBERTa) and KG-augmented QA models (QAGNN, GreaseLM) on all tasks. The gain is especially significant on tasks that have small training data (OBQA, Riddle, ARC) and tasks that require complex reasoning (CosmosQA, HellaSwag). "
|
| 720 |
+
],
|
| 721 |
+
"table_footnote": [],
|
| 722 |
+
"table_body": "<table><tr><td></td><td>CSQA</td><td>OBQA</td><td>Riddle</td><td>ARC</td><td>CosmosQA</td><td>HellaSwag</td><td>PIQA</td><td>SIQA</td><td>aNLI</td></tr><tr><td>RoBERTa [18]</td><td>68.7</td><td>64.9</td><td>60.7</td><td>43.0</td><td>80.5</td><td>82.3</td><td>79.4</td><td>75.9</td><td>82.7</td></tr><tr><td>QAGNN [8]</td><td>73.4</td><td>67.8</td><td>67.0</td><td>44.4</td><td>80.7</td><td>82.6</td><td>79.6</td><td>75.7</td><td>83.0</td></tr><tr><td>GreaseLM[9]</td><td>74.2</td><td>66.9</td><td>67.2</td><td>44.7</td><td>80.6</td><td>82.8</td><td>79.6</td><td>75.5</td><td>83.3</td></tr><tr><td>DRAGON (Ours)</td><td>76.0</td><td>72.0</td><td>71.3</td><td>48.6</td><td>82.3</td><td>85.2</td><td>81.1</td><td>76.8</td><td>84.0</td></tr></table>",
|
| 723 |
+
"bbox": [
|
| 724 |
+
223,
|
| 725 |
+
66,
|
| 726 |
+
772,
|
| 727 |
+
141
|
| 728 |
+
],
|
| 729 |
+
"page_idx": 6
|
| 730 |
+
},
|
| 731 |
+
{
|
| 732 |
+
"type": "table",
|
| 733 |
+
"img_path": "images/99e915ae49df0845b8fd89a15b7b420b916364eaf3803ad853338eeab2d6f1fe.jpg",
|
| 734 |
+
"table_caption": [
|
| 735 |
+
"Table 2: Accuracy of DRAGON on $C S Q A + O B Q A$ dev sets for questions involving complex reasoning such as negation terms, conjunction terms, hedge terms, prepositional phrases, and more entity mentions. DRAGON consistently outperforms the existing LM (RoBERTa) and KG-augmented QA models (QAGNN, GreaseLM) in these complex reasoning settings. "
|
| 736 |
+
],
|
| 737 |
+
"table_footnote": [],
|
| 738 |
+
"table_body": "<table><tr><td></td><td>Negation</td><td>Conjunction</td><td>Hedge</td><td colspan=\"4\">#Prepositional Phrases</td><td>#Entities</td></tr><tr><td></td><td></td><td></td><td></td><td>0</td><td>1</td><td>2</td><td>3</td><td>>10</td></tr><tr><td>RoBERTa</td><td>61.7</td><td>70.9</td><td>68.6</td><td>67.6</td><td>71.0</td><td>71.1</td><td>73.1</td><td>74.5</td></tr><tr><td>QAGNN</td><td>65.1</td><td>74.5</td><td>74.2</td><td>72.1</td><td>71.6</td><td>75.6</td><td>71.3</td><td>78.6</td></tr><tr><td>GreaseLM</td><td>65.1</td><td>74.9</td><td>76.6</td><td>75.6</td><td>73.8</td><td>74.7</td><td>73.6</td><td>79.4</td></tr><tr><td>DRAGON (Ours)</td><td>75.2</td><td>79.6</td><td>77.5</td><td>79.1</td><td>78.2</td><td>77.8</td><td>80.9</td><td>83.5</td></tr></table>",
|
| 739 |
+
"bbox": [
|
| 740 |
+
277,
|
| 741 |
+
193,
|
| 742 |
+
715,
|
| 743 |
+
267
|
| 744 |
+
],
|
| 745 |
+
"page_idx": 6
|
| 746 |
+
},
|
| 747 |
+
{
|
| 748 |
+
"type": "text",
|
| 749 |
+
"text": "",
|
| 750 |
+
"bbox": [
|
| 751 |
+
176,
|
| 752 |
+
327,
|
| 753 |
+
821,
|
| 754 |
+
369
|
| 755 |
+
],
|
| 756 |
+
"page_idx": 6
|
| 757 |
+
},
|
| 758 |
+
{
|
| 759 |
+
"type": "text",
|
| 760 |
+
"text": "3.4 Results ",
|
| 761 |
+
"text_level": 1,
|
| 762 |
+
"bbox": [
|
| 763 |
+
174,
|
| 764 |
+
387,
|
| 765 |
+
263,
|
| 766 |
+
402
|
| 767 |
+
],
|
| 768 |
+
"page_idx": 6
|
| 769 |
+
},
|
| 770 |
+
{
|
| 771 |
+
"type": "text",
|
| 772 |
+
"text": "Table 1 shows performance on the 9 downstream commonsense reasoning tasks. Across all tasks, DRAGON consistently outperforms the existing LM (RoBERTa) and KG-augmented QA models (QAGNN, GreaseLM), e.g., $+ 7 \\%$ absolute accuracy boost over RoBERTa and $+ 5 \\%$ over GreaseLM on OBQA. These accuracy boosts indicate the advantage of DRAGON over RoBERTa (KG reasoning) and over GreaseLM (pretraining). The gain is especially significant on datasets that have small training data such as ARC, Riddle and $O B Q A$ , and datasets that require complex reasoning such as CosmosQA and HellaSwag, which we analyze in more detail in the following sections. ",
|
| 773 |
+
"bbox": [
|
| 774 |
+
174,
|
| 775 |
+
414,
|
| 776 |
+
825,
|
| 777 |
+
511
|
| 778 |
+
],
|
| 779 |
+
"page_idx": 6
|
| 780 |
+
},
|
| 781 |
+
{
|
| 782 |
+
"type": "text",
|
| 783 |
+
"text": "3.4.1 Analysis: Effect of knowledge graph ",
|
| 784 |
+
"text_level": 1,
|
| 785 |
+
"bbox": [
|
| 786 |
+
174,
|
| 787 |
+
527,
|
| 788 |
+
477,
|
| 789 |
+
542
|
| 790 |
+
],
|
| 791 |
+
"page_idx": 6
|
| 792 |
+
},
|
| 793 |
+
{
|
| 794 |
+
"type": "text",
|
| 795 |
+
"text": "The first key contribution of DRAGON (w.r.t. existing LM pretraining methods) is that we incorporate KGs. We find that this significantly improves the model’s performance for robust and complex reasoning, such as resolving multi-step reasoning and negation, as we discuss below. ",
|
| 796 |
+
"bbox": [
|
| 797 |
+
176,
|
| 798 |
+
551,
|
| 799 |
+
823,
|
| 800 |
+
594
|
| 801 |
+
],
|
| 802 |
+
"page_idx": 6
|
| 803 |
+
},
|
| 804 |
+
{
|
| 805 |
+
"type": "text",
|
| 806 |
+
"text": "Quantitative analysis. In Table 2, we study downstream task performance of DRAGON on questions involving complex reasoning. Building on [8, 9], we consider several proxies to categorize complex questions: (i) presence of negation (e.g. no, never), (ii) presence of conjunction (e.g. and, but), (iii) presence of hedge (e.g. sometimes, maybe), (iv) number of prepositional phrases, and (v) number of entity mentions. Having negation or conjunction indicates logical multi-step reasoning, having more prepositional phrases or entity mentions indicates involving more reasoning steps or constraints, and having hedge terms indicates involving complex textual nuance. DRAGON significantly outperforms the baseline LM (RoBERTa) across all these categories (e.g., $+ 1 4 \\%$ accuracy for negation), which confirms that our joint language-knowledge pretraining boosts reasoning performance. DRAGON also consistently outperforms the existing KG-augmented QA models (QAGNN, GreaseLM). We find that QAGNN and GreaseLM only improve moderately on RoBERTa for some categories like conjunction or many prepositional phrases $( = 2 , 3 )$ , but DRAGON provides substantial boosts. This suggests that through self-supervised pretraining with larger and diverse data, DRAGON has learned more general-purpose reasoning abilities than the finetuning-only models like GreaseLM. ",
|
| 807 |
+
"bbox": [
|
| 808 |
+
174,
|
| 809 |
+
601,
|
| 810 |
+
825,
|
| 811 |
+
794
|
| 812 |
+
],
|
| 813 |
+
"page_idx": 6
|
| 814 |
+
},
|
| 815 |
+
{
|
| 816 |
+
"type": "text",
|
| 817 |
+
"text": "Qualitative analysis. Using the CSQA dataset, we further conducted case studies on the behavior of DRAGON’s KG reasoning component, where we visualize how graph attention weights change given different question variations (Figure 2). We find that DRAGON exhibits abilities to extrapolate and perform robust reasoning. For instance, DRAGON adjusts the entity attention weights and final predictions accordingly when we add conjunction or negation about entities (A1, A2) or when we add extra context to an original question $\\mathbf { B } 1 \\mathbf { B } 2$ ), but existing models, RoBERTa and GreaseLM, struggle to predict the correct answers. As these questions are more complex than ones typically seen in the CSQA training set, our insight is that while vanilla LMs (RoBERTa) and finetuning (GreaseLM) ",
|
| 818 |
+
"bbox": [
|
| 819 |
+
174,
|
| 820 |
+
800,
|
| 821 |
+
825,
|
| 822 |
+
911
|
| 823 |
+
],
|
| 824 |
+
"page_idx": 6
|
| 825 |
+
},
|
| 826 |
+
{
|
| 827 |
+
"type": "image",
|
| 828 |
+
"img_path": "images/5e7691298a65a715095daa92e7f26b765402511f7fcb77ed8033aa253aae32eb.jpg",
|
| 829 |
+
"image_caption": [
|
| 830 |
+
"Figure 2: Analysis of DRAGON’s graph reasoning, where we visualize how graph attention weights and final predictions change given question variations. Darker and thicker edges indicate higher attention weights. DRAGON exhibits abilities to extrapolate and perform robust reasoning. DRAGON adjusts the entity attention weights and final predictions accordingly when conjunction or negation is given about entities (A1, A2) or when extra context is added to an original question $( \\mathbf { B } 1 \\to \\mathbf { B } 2$ ), but existing models, RoBERTa and GreaseLM, struggle to predict the correct answers. A1: DRAGON’s final GNN layer shows strong attention to “school” but weak attention to “trip”, likely because the question states “and store one”—hence, the chair is not used for a trip. A2: DRAGON shows strong attention to “trip” and “beach”, likely because the question now states “but not store one”—hence, the chair is used for a trip. $\\mathbf { B } \\mathbf { 1 } \\to \\mathbf { B } 2$ : DRAGON’s final GNN layer shows strong attention to “movie” in the original question (B1), but after adding the extra context “don’t enjoy pre-record” (B2), DRAGON shows strong attention to “live” and “concert”, leading to making the correctly adjusted prediction “concert hall”. One interpretation of these findings is that DRAGON leverages the KG’s graph structure as a scaffold for performing complex reasoning. This insight is related to recent works that provide LMs with scratch space for intermediate reasoning [8, 65, 66]. "
|
| 831 |
+
],
|
| 832 |
+
"image_footnote": [],
|
| 833 |
+
"bbox": [
|
| 834 |
+
166,
|
| 835 |
+
83,
|
| 836 |
+
828,
|
| 837 |
+
176
|
| 838 |
+
],
|
| 839 |
+
"page_idx": 7
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"type": "table",
|
| 843 |
+
"img_path": "images/b725c49343aa1f5ecb1edc102060f5e850bfd92fe96721a45f93b014a4d6ea5d.jpg",
|
| 844 |
+
"table_caption": [
|
| 845 |
+
"Table 3: Performance in low-resource setting where $10 \\%$ of finetuning data is used. DRAGON attains large gains, suggesting its benefit for downstream data efficiency. "
|
| 846 |
+
],
|
| 847 |
+
"table_footnote": [],
|
| 848 |
+
"table_body": "<table><tr><td>Method</td><td>CosmosQA (10% train)</td><td>PIQA (10% train)</td></tr><tr><td>RoBERTa</td><td>72.2</td><td>66.4</td></tr><tr><td>GreaseLM</td><td>73.0</td><td>67.0</td></tr><tr><td>DRAGON (Ours)</td><td>77.9</td><td>72.3</td></tr></table>",
|
| 849 |
+
"bbox": [
|
| 850 |
+
183,
|
| 851 |
+
337,
|
| 852 |
+
464,
|
| 853 |
+
390
|
| 854 |
+
],
|
| 855 |
+
"page_idx": 7
|
| 856 |
+
},
|
| 857 |
+
{
|
| 858 |
+
"type": "table",
|
| 859 |
+
"img_path": "images/4587842c339ee262b11c3a2a7b40a55dddec271b97dea5262fada4dec5b4ee89.jpg",
|
| 860 |
+
"table_caption": [
|
| 861 |
+
"Table 4: Downstream performance when model capacity—number of text-KG fusion layers—is increased (“-Ex”). Increased capacity does not help for the finetuning-only model (GreaseLM), but helps when pretrained (DRAGON), suggesting the promise of DRAGON to be further scaled up. "
|
| 862 |
+
],
|
| 863 |
+
"table_footnote": [],
|
| 864 |
+
"table_body": "<table><tr><td>Method</td><td>CSQA</td><td>OBQA</td></tr><tr><td>GreaseLM</td><td>74.2</td><td>66.9</td></tr><tr><td>GreaseLM-Ex</td><td>73.9</td><td>66.2</td></tr><tr><td>DRAGON (Ours)</td><td>76.0</td><td>72.0</td></tr><tr><td>DRAGON-Ex (Ours)</td><td>76.3</td><td>72.8</td></tr></table>",
|
| 865 |
+
"bbox": [
|
| 866 |
+
227,
|
| 867 |
+
440,
|
| 868 |
+
418,
|
| 869 |
+
502
|
| 870 |
+
],
|
| 871 |
+
"page_idx": 7
|
| 872 |
+
},
|
| 873 |
+
{
|
| 874 |
+
"type": "table",
|
| 875 |
+
"img_path": "images/cee75e9f3d826577aa92b91e208aecbf41d7ef3467f5d7feeb58d117046bb155.jpg",
|
| 876 |
+
"table_caption": [
|
| 877 |
+
"Table 5: Ablation study of DRAGON. Using joint pretraining objective $\\mathbf { M L M + }$ LinkPred (§2.3) outperforms using one of them only. All variants of LinkPred scoring models (DistMult, TransE, RotatE) outperform the baseline without LinkPred (“MLM only”), suggesting that DRAGON can be combined with various KG representation learning models. Cross-modal model with bidirectional modality interaction (§2.2) outperforms combining text and KG representations only at the end. Finally, using KG as graph outperforms converting KG as sentences, suggesting the benefit of graph structure for reasoning. "
|
| 878 |
+
],
|
| 879 |
+
"table_footnote": [],
|
| 880 |
+
"table_body": "<table><tr><td>Ablation Type</td><td>Ablation</td><td>CSQA</td><td>OBQA</td></tr><tr><td rowspan=\"3\">Pretraining objective</td><td>MLM + LinkPred (final)</td><td>76.0</td><td>72.0</td></tr><tr><td>MLM only</td><td>74.3</td><td>67.2</td></tr><tr><td>LinkPred only</td><td>73.8</td><td>66.4</td></tr><tr><td rowspan=\"3\">LinkPred head </td><td>DistMult (final)</td><td>76.0</td><td>72.0</td></tr><tr><td>TransE</td><td>75.7</td><td>71.4</td></tr><tr><td>RotatE</td><td>75.8</td><td>71.7</td></tr><tr><td rowspan=\"2\">Cross-modal model</td><td>Bidirectional interaction (final)</td><td>76.0</td><td>72.0</td></tr><tr><td>Concatenate at end</td><td>74.5</td><td>68.0</td></tr><tr><td rowspan=\"2\">KG structure</td><td>Use graph (final)</td><td>76.0</td><td>72.0</td></tr><tr><td>Convert to sentence</td><td>74.7</td><td>70.1</td></tr></table>",
|
| 881 |
+
"bbox": [
|
| 882 |
+
508,
|
| 883 |
+
338,
|
| 884 |
+
797,
|
| 885 |
+
454
|
| 886 |
+
],
|
| 887 |
+
"page_idx": 7
|
| 888 |
+
},
|
| 889 |
+
{
|
| 890 |
+
"type": "text",
|
| 891 |
+
"text": "have limitation in learning complex reasoning, KG-augmented pretraining (DRAGON) helps acquire generalizable reasoning abilities that extrapolate to harder test examples. ",
|
| 892 |
+
"bbox": [
|
| 893 |
+
173,
|
| 894 |
+
606,
|
| 895 |
+
823,
|
| 896 |
+
635
|
| 897 |
+
],
|
| 898 |
+
"page_idx": 7
|
| 899 |
+
},
|
| 900 |
+
{
|
| 901 |
+
"type": "text",
|
| 902 |
+
"text": "3.4.2 Analysis: Effect of pretraining ",
|
| 903 |
+
"text_level": 1,
|
| 904 |
+
"bbox": [
|
| 905 |
+
174,
|
| 906 |
+
651,
|
| 907 |
+
437,
|
| 908 |
+
666
|
| 909 |
+
],
|
| 910 |
+
"page_idx": 7
|
| 911 |
+
},
|
| 912 |
+
{
|
| 913 |
+
"type": "text",
|
| 914 |
+
"text": "Another key contribution of DRAGON (w.r.t. existing QA models like GreaseLM) is pretraining. Here we discuss when and why our pretraining is useful. Considering the three core factors in machine learning (data, task complexity, and model capacity), pretraining helps when the available downstream task data is smaller compared to the downstream task complexity or model capacity. Concretely, we find that DRAGON is especially helpful for the following three scenarios. ",
|
| 915 |
+
"bbox": [
|
| 916 |
+
174,
|
| 917 |
+
676,
|
| 918 |
+
825,
|
| 919 |
+
746
|
| 920 |
+
],
|
| 921 |
+
"page_idx": 7
|
| 922 |
+
},
|
| 923 |
+
{
|
| 924 |
+
"type": "text",
|
| 925 |
+
"text": "Downstream tasks with limited data. In Table 1, we find that DRAGON provides significant boosts over GreaseLM on downstream tasks with limited finetuning data available, such as ARC (3K training instances; $+ 4 \\%$ accuracy gain), Riddle (3K instances; $+ 4 \\%$ accuracy) and OBQA (5K instances; $+ 5 \\%$ accuracy). For other tasks, we also experimented with a low-resource setting where $10 \\%$ of finetuning data is used (Table 3). Here we also see that DRAGON attains significant gains over GreaseLM $+ 5 \\%$ accuracy on PIQA), suggesting the improved data-efficiency of DRAGON. ",
|
| 926 |
+
"bbox": [
|
| 927 |
+
174,
|
| 928 |
+
752,
|
| 929 |
+
825,
|
| 930 |
+
835
|
| 931 |
+
],
|
| 932 |
+
"page_idx": 7
|
| 933 |
+
},
|
| 934 |
+
{
|
| 935 |
+
"type": "text",
|
| 936 |
+
"text": "Complex downstream tasks. In Table 1, we find that DRAGON provides substantial gains over GreaseLM on downstream tasks involving more complex reasoning, such as CosmosQA and HellaSwag, where the inputs have longer context and more entities (thus bigger local KGs). For these tasks, improvements of GreaesLM over RoBERTa were small $( + 0 . 1 \\%$ on CosmosQA), but DRAGON provides substantial boosts $( + 1 . 8 \\% )$ . Our insight is that through self-supervised pretraining with larger and more diverse data, DRAGON has learned richer text-KG interactions than GreaseLM, enabling solving more complex downstream tasks. Similarly, as seen in $\\ S 3 . 4 . 1$ , DRAGON also attains large gains over GreaseLM on complex questions containing negation, conjunction and prepositional phrases (Table 2), and extrapolates to questions more complex than seen in training sets (Figure 2). ",
|
| 937 |
+
"bbox": [
|
| 938 |
+
174,
|
| 939 |
+
842,
|
| 940 |
+
823,
|
| 941 |
+
911
|
| 942 |
+
],
|
| 943 |
+
"page_idx": 7
|
| 944 |
+
},
|
| 945 |
+
{
|
| 946 |
+
"type": "text",
|
| 947 |
+
"text": "",
|
| 948 |
+
"bbox": [
|
| 949 |
+
174,
|
| 950 |
+
90,
|
| 951 |
+
825,
|
| 952 |
+
147
|
| 953 |
+
],
|
| 954 |
+
"page_idx": 8
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"type": "text",
|
| 958 |
+
"text": "Increased model capacity. In Table 4, we study downstream performance when the model capacity is increased—the number of text-KG fusion layers is increased from 5 to 7—for both GreaseLM and DRAGON. We find that increased capacity does not help for the finetuning-only model (GreaseLM) as was also reported in the original GreaseLM paper, but it helps when pretrained (DRAGON). This result reveals that increased model capacity can actually be beneficial when combined with pretraining, and suggests the promise of DRAGON to be further scaled up. ",
|
| 959 |
+
"bbox": [
|
| 960 |
+
174,
|
| 961 |
+
154,
|
| 962 |
+
825,
|
| 963 |
+
237
|
| 964 |
+
],
|
| 965 |
+
"page_idx": 8
|
| 966 |
+
},
|
| 967 |
+
{
|
| 968 |
+
"type": "text",
|
| 969 |
+
"text": "3.4.3 Analysis: Design choices of DRAGON ",
|
| 970 |
+
"text_level": 1,
|
| 971 |
+
"bbox": [
|
| 972 |
+
176,
|
| 973 |
+
250,
|
| 974 |
+
480,
|
| 975 |
+
265
|
| 976 |
+
],
|
| 977 |
+
"page_idx": 8
|
| 978 |
+
},
|
| 979 |
+
{
|
| 980 |
+
"type": "text",
|
| 981 |
+
"text": "Pretraining objective (Table 5 top). The first important design choice of DRAGON is the joint pretraining objective: ${ \\bf M L M + }$ LinkPred (§2.3). Using the joint objective outperforms using MLM or LinkPred alone ( $+ 5 \\%$ accuracy on OBQA). This suggests that having the bidirectional self-supervised tasks on text and KG facilitates the model to fuse the two modalities for reasoning. ",
|
| 982 |
+
"bbox": [
|
| 983 |
+
173,
|
| 984 |
+
273,
|
| 985 |
+
825,
|
| 986 |
+
329
|
| 987 |
+
],
|
| 988 |
+
"page_idx": 8
|
| 989 |
+
},
|
| 990 |
+
{
|
| 991 |
+
"type": "text",
|
| 992 |
+
"text": "Link prediction head choice (Table 5 middle 1). KG representation learning is an active area of research, and various KG triplet scoring models are proposed (Equation 9). We hence experimented with using different scoring models for DRAGON’s link prediction head (§2.3). We find that while DistMult has a slight edge, all variants we tried (DistMult, TransE, RotatE) are effective, outperforming the baseline without LinkPred (“MLM only”). This result suggests the generality of DRAGON and its promise to be combined with various KG representation learning techniques. ",
|
| 993 |
+
"bbox": [
|
| 994 |
+
174,
|
| 995 |
+
333,
|
| 996 |
+
825,
|
| 997 |
+
417
|
| 998 |
+
],
|
| 999 |
+
"page_idx": 8
|
| 1000 |
+
},
|
| 1001 |
+
{
|
| 1002 |
+
"type": "text",
|
| 1003 |
+
"text": "Cross-modal model (Table 5 middle 2). Another core component of DRAGON is the cross-modal encoder with bidirectional text-KG fusion layers (§2.2). We find that if we ablate them and simply concatenate text and KG representations at the end, the performance drops substantially. This result suggests that deep bidirectional fusion is crucial to model interactions over text and KG for reasoning. ",
|
| 1004 |
+
"bbox": [
|
| 1005 |
+
174,
|
| 1006 |
+
422,
|
| 1007 |
+
825,
|
| 1008 |
+
478
|
| 1009 |
+
],
|
| 1010 |
+
"page_idx": 8
|
| 1011 |
+
},
|
| 1012 |
+
{
|
| 1013 |
+
"type": "text",
|
| 1014 |
+
"text": "KG structure (Table 5 bottom). The final key design of DRAGON is that we leverage the graph structure of KGs via a sequence-graph encoder and link prediction objective. Here we experimented with an alternative pretraining method that drops the graph structure: we convert triplets in the local KG into sentences using a template [33], append them to the main text input, and perform vanilla MLM pretraining. We find that DRAGON substantially outperforms this variant $+ 2 \\%$ accuracy on OBQA), which suggests that the graph structure of KGs helps the model perform reasoning. ",
|
| 1015 |
+
"bbox": [
|
| 1016 |
+
174,
|
| 1017 |
+
482,
|
| 1018 |
+
825,
|
| 1019 |
+
565
|
| 1020 |
+
],
|
| 1021 |
+
"page_idx": 8
|
| 1022 |
+
},
|
| 1023 |
+
{
|
| 1024 |
+
"type": "text",
|
| 1025 |
+
"text": "4 Experiments: Biomedical domain ",
|
| 1026 |
+
"text_level": 1,
|
| 1027 |
+
"bbox": [
|
| 1028 |
+
174,
|
| 1029 |
+
584,
|
| 1030 |
+
486,
|
| 1031 |
+
602
|
| 1032 |
+
],
|
| 1033 |
+
"page_idx": 8
|
| 1034 |
+
},
|
| 1035 |
+
{
|
| 1036 |
+
"type": "text",
|
| 1037 |
+
"text": "Biomedicine is a domain with extensive background knowledge [67, 68, 69, 1], and experts curate various knowledge bases for it [70, 17, 71, 72]. We hypothesize that these biomedical KGs can enable deeper understanding and reasoning about biomedical text. With this motivation, we pretrain DRAGON on a biomedical corpus and KG, and evaluate on biomedical downstream tasks. ",
|
| 1038 |
+
"bbox": [
|
| 1039 |
+
174,
|
| 1040 |
+
614,
|
| 1041 |
+
825,
|
| 1042 |
+
670
|
| 1043 |
+
],
|
| 1044 |
+
"page_idx": 8
|
| 1045 |
+
},
|
| 1046 |
+
{
|
| 1047 |
+
"type": "text",
|
| 1048 |
+
"text": "Pretraining setup. For the text data, we use PubMed [73], a widely-used corpus in biomedial LM training (e.g., BioBERT [74], PubmedBERT [75]). It contains the abstracts of biomedical papers on PubMed and has 21GB of text. For the KG data, we use the Unified Medical Language System (UMLS) [17], a widely-used knowledge graph in biomedicine. It has 300K nodes and 1M edges in total. For training, we follow the same procedure as the experiment in the general domain (§3.1), except that we initialize DRAGON’s LM component with BioLinkBERT-Large [19], the state-of-theart biomedical LM, instead of RoBERTa-Large. Note that while “BioLinkBERT” has “Link” in its name, it is not about KG links but about citation links that the model was originally pretrained with. ",
|
| 1049 |
+
"bbox": [
|
| 1050 |
+
173,
|
| 1051 |
+
676,
|
| 1052 |
+
826,
|
| 1053 |
+
787
|
| 1054 |
+
],
|
| 1055 |
+
"page_idx": 8
|
| 1056 |
+
},
|
| 1057 |
+
{
|
| 1058 |
+
"type": "text",
|
| 1059 |
+
"text": "Downstream evaluation tasks. We finetune and evaluate DRAGON on three popular biomedical NLP and reasoning benchmarks: MedQA-USMLE (MedQA) [76], PubMedQA [77], and BioASQ [78]. Appendix B.4 provides details on these tasks and data splits. ",
|
| 1060 |
+
"bbox": [
|
| 1061 |
+
173,
|
| 1062 |
+
794,
|
| 1063 |
+
529,
|
| 1064 |
+
863
|
| 1065 |
+
],
|
| 1066 |
+
"page_idx": 8
|
| 1067 |
+
},
|
| 1068 |
+
{
|
| 1069 |
+
"type": "text",
|
| 1070 |
+
"text": "Baselines. We compare DRAGON with the vanilla LM (BioLinkBERT) and LMs finetuned with the KG (QAGNN and GreaseLM seeded with BioLinkBERT). ",
|
| 1071 |
+
"bbox": [
|
| 1072 |
+
173,
|
| 1073 |
+
869,
|
| 1074 |
+
529,
|
| 1075 |
+
911
|
| 1076 |
+
],
|
| 1077 |
+
"page_idx": 8
|
| 1078 |
+
},
|
| 1079 |
+
{
|
| 1080 |
+
"type": "table",
|
| 1081 |
+
"img_path": "images/a3bfdf2aaa9124cac3d69a26c94aeff6a7dc71ed6ef2f357741f68b7caac99e2.jpg",
|
| 1082 |
+
"table_caption": [],
|
| 1083 |
+
"table_footnote": [
|
| 1084 |
+
"Table 6: Accuracy on biomedical NLP tasks. DRAGON outperforms all previous biomedical LMs. "
|
| 1085 |
+
],
|
| 1086 |
+
"table_body": "<table><tr><td>Method</td><td>MedQA</td><td>PubMedQA</td><td>BioASQ</td></tr><tr><td>BioBERT[74]</td><td>36.7</td><td>60.2</td><td>84.1</td></tr><tr><td>PubmedBERT[75]</td><td>38.1</td><td>55.8</td><td>87.5</td></tr><tr><td>BioLinkBERT[19]</td><td>44.6</td><td>72.2</td><td>94.8</td></tr><tr><td>+ QAGNN</td><td>45.0</td><td>72.1</td><td>95.0</td></tr><tr><td>+ GreaseLM</td><td>45.1</td><td>72.4</td><td>94.9</td></tr><tr><td>DRAGON (Ours)</td><td>47.5</td><td>73.4</td><td>96.4</td></tr></table>",
|
| 1087 |
+
"bbox": [
|
| 1088 |
+
549,
|
| 1089 |
+
796,
|
| 1090 |
+
833,
|
| 1091 |
+
886
|
| 1092 |
+
],
|
| 1093 |
+
"page_idx": 8
|
| 1094 |
+
},
|
| 1095 |
+
{
|
| 1096 |
+
"type": "text",
|
| 1097 |
+
"text": "Results. Table 6 summarizes model performance on the downstream tasks. Across tasks, DRAGON outperforms all the existing biomedical LMs and KG-augmented QA models, e.g., $+ 3 \\%$ absolute accuracy boost over BioLinkBERT and $+ 2 \\%$ over GreaseLM on MedQA, achieving new state-of-theart performance on these tasks. This result suggests significant efficacy of KG-augmented pretraining for improving biomedical reasoning tasks. Combined with the results in the general commonsense domain (§3.4), our experiments also suggest the domain-generality of DRAGON, serving as an effective pretraining method across domains with different combinations of text, KGs and seed LMs. ",
|
| 1098 |
+
"bbox": [
|
| 1099 |
+
174,
|
| 1100 |
+
92,
|
| 1101 |
+
825,
|
| 1102 |
+
188
|
| 1103 |
+
],
|
| 1104 |
+
"page_idx": 9
|
| 1105 |
+
},
|
| 1106 |
+
{
|
| 1107 |
+
"type": "text",
|
| 1108 |
+
"text": "5 Conclusion ",
|
| 1109 |
+
"text_level": 1,
|
| 1110 |
+
"bbox": [
|
| 1111 |
+
174,
|
| 1112 |
+
207,
|
| 1113 |
+
299,
|
| 1114 |
+
224
|
| 1115 |
+
],
|
| 1116 |
+
"page_idx": 9
|
| 1117 |
+
},
|
| 1118 |
+
{
|
| 1119 |
+
"type": "text",
|
| 1120 |
+
"text": "We presented DRAGON, a self-supervised pretraining method to learn a deeply bidirectional languageknowledge model from text and knowledge graphs (KGs) at scale. In both general and biomedical domains, DRAGON outperforms existing language models and KG-augmented models on various NLP tasks, and exhibits strong performance on complex reasoning such as answering questions involving long context or multi-step reasoning. ",
|
| 1121 |
+
"bbox": [
|
| 1122 |
+
174,
|
| 1123 |
+
238,
|
| 1124 |
+
825,
|
| 1125 |
+
308
|
| 1126 |
+
],
|
| 1127 |
+
"page_idx": 9
|
| 1128 |
+
},
|
| 1129 |
+
{
|
| 1130 |
+
"type": "text",
|
| 1131 |
+
"text": "One limitation of DRAGON is that it is currently an encoder model (analogous to BERT) and does not perform language generation. An important future research would be to extend DRAGON to generation, and advance KG-enhanced language generation [28, 79]. ",
|
| 1132 |
+
"bbox": [
|
| 1133 |
+
174,
|
| 1134 |
+
314,
|
| 1135 |
+
823,
|
| 1136 |
+
356
|
| 1137 |
+
],
|
| 1138 |
+
"page_idx": 9
|
| 1139 |
+
},
|
| 1140 |
+
{
|
| 1141 |
+
"type": "text",
|
| 1142 |
+
"text": "Reproducibility ",
|
| 1143 |
+
"text_level": 1,
|
| 1144 |
+
"bbox": [
|
| 1145 |
+
174,
|
| 1146 |
+
375,
|
| 1147 |
+
305,
|
| 1148 |
+
392
|
| 1149 |
+
],
|
| 1150 |
+
"page_idx": 9
|
| 1151 |
+
},
|
| 1152 |
+
{
|
| 1153 |
+
"type": "text",
|
| 1154 |
+
"text": "Pretrained models, code and data are available at https://github.com/michiyasunaga/dragon. Experiments are available at https://worksheets.codalab.org/worksheets/0xcf9cddffff864fb382e1a2f1393c8934. ",
|
| 1155 |
+
"bbox": [
|
| 1156 |
+
173,
|
| 1157 |
+
406,
|
| 1158 |
+
828,
|
| 1159 |
+
449
|
| 1160 |
+
],
|
| 1161 |
+
"page_idx": 9
|
| 1162 |
+
},
|
| 1163 |
+
{
|
| 1164 |
+
"type": "text",
|
| 1165 |
+
"text": "Acknowledgment ",
|
| 1166 |
+
"text_level": 1,
|
| 1167 |
+
"bbox": [
|
| 1168 |
+
174,
|
| 1169 |
+
468,
|
| 1170 |
+
321,
|
| 1171 |
+
484
|
| 1172 |
+
],
|
| 1173 |
+
"page_idx": 9
|
| 1174 |
+
},
|
| 1175 |
+
{
|
| 1176 |
+
"type": "text",
|
| 1177 |
+
"text": "We thank Rok Sosic, Hamed Nilforoshan, Michael Moor, Qian Huang, members of the Stanford SNAP, P-Lambda, and NLP groups, as well as our anonymous reviewers for valuable feedback. We also gratefully acknowledge the support of HAI Google Cloud Credits 1051203844499; DARPA under Nos. HR00112190039 (TAMI), N660011924033 (MCS); ARO under Nos. W911NF-16-1-0342 (MURI), W911NF-16-1-0171 (DURIP); NSF under Nos. OAC-1835598 (CINES), OAC-1934578 (HDR), CCF-1918940 (Expeditions), IIS-2030477 (RAPID), NIH under No. R56LM013365; Stanford Data Science Initiative, Wu Tsai Neurosciences Institute, Chan Zuckerberg Biohub, Amazon, JPMorgan Chase, Docomo, Hitachi, Intel, JD.com, KDDI, Toshiba, NEC, and UnitedHealth Group. The content is solely the responsibility of the authors and does not necessarily represent the official views of the funding entities. ",
|
| 1178 |
+
"bbox": [
|
| 1179 |
+
173,
|
| 1180 |
+
498,
|
| 1181 |
+
826,
|
| 1182 |
+
637
|
| 1183 |
+
],
|
| 1184 |
+
"page_idx": 9
|
| 1185 |
+
},
|
| 1186 |
+
{
|
| 1187 |
+
"type": "text",
|
| 1188 |
+
"text": "References ",
|
| 1189 |
+
"text_level": 1,
|
| 1190 |
+
"bbox": [
|
| 1191 |
+
174,
|
| 1192 |
+
656,
|
| 1193 |
+
266,
|
| 1194 |
+
672
|
| 1195 |
+
],
|
| 1196 |
+
"page_idx": 9
|
| 1197 |
+
},
|
| 1198 |
+
{
|
| 1199 |
+
"type": "text",
|
| 1200 |
+
"text": "[1] Rishi Bommasani et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. \n[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics (NAACL), 2019. \n[3] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), 2020. \n[4] Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In North American Chapter of the Association for Computational Linguistics (NAACL), 2018. \n[5] Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase: a collaboratively created graph database for structuring human knowledge. In SIGMOD, 2008. \n[6] Denny Vrandeciˇ c and Markus Krötzsch. Wikidata: A free collaborative knowledgebase. ´ Communications of the ACM, 2014. \n[7] Robyn Speer, Joshua Chin, and Catherine Havasi. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. \n[8] Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, and Jure Leskovec. QAGNN: Reasoning with language models and knowledge graphs for question answering. In North American Chapter of the Association for Computational Linguistics (NAACL), 2021. \n[9] Xikun Zhang, Antoine Bosselut, Michihiro Yasunaga, Hongyu Ren, Percy Liang, Christopher D Manning, and Jure Leskovec. Greaselm: Graph reasoning enhanced language models for question answering. In International Conference on Learning Representations (ICLR), 2022. \n[10] Hongyu Ren, Weihua Hu, and Jure Leskovec. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In International Conference on Learning Representations (ICLR), 2020. \n[11] Hongyu Ren, Hanjun Dai, Bo Dai, Xinyun Chen, Michihiro Yasunaga, Haitian Sun, Dale Schuurmans, Jure Leskovec, and Denny Zhou. Lego: Latent execution-guided reasoning for multi-hop question answering on knowledge graphs. In International Conference on Machine Learning (ICML), 2021. \n[12] Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. Ernie: Enhanced language representation with informative entities. In Association for Computational Linguistics (ACL), 2019. \n[13] Wenhan Xiong, Jingfei Du, William Yang Wang, and Veselin Stoyanov. Pretrained encyclopedia: Weakly supervised knowledge-pretrained language model. In International Conference on Learning Representations (ICLR), 2020. \n[14] Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian Tang. Kepler: A unified model for knowledge embedding and pre-trained language representation. Transactions of the Association for Computational Linguistics (TACL), 2021. \n[15] Oshin Agarwal, Heming Ge, Siamak Shakeri, and Rami Al-Rfou. Knowledge graph based synthetic corpus generation for knowledge-enhanced language model pre-training. In North American Chapter of the Association for Computational Linguistics (NAACL), 2021. \n[16] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. \n[17] Olivier Bodenreider. The unified medical language system (UMLS): Integrating biomedical terminology. Nucleic acids research, 2004. \n[18] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. \n[19] Michihiro Yasunaga, Jure Leskovec, and Percy Liang. LinkBERT: Pretraining language models with document links. In Association for Computational Linguistics (ACL), 2022. \n[20] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Realm: Retrieval-augmented language model pre-training. In International Conference on Machine Learning (ICML), 2020. \n[21] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems (NeurIPS), 2020. \n[22] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillions of tokens. arXiv preprint arXiv:2112.04426, 2021. \n[23] Matthew E. Peters, Mark Neumann, IV RobertLLogan, Roy Schwartz, V. Joshi, Sameer Singh, and Noah A. Smith. Knowledge enhanced contextual word representations. In Empirical Methods in Natural Language Processing (EMNLP), 2019. \n[24] Corby Rosset, Chenyan Xiong, Minh Phan, Xia Song, Paul Bennett, and Saurabh Tiwary. Knowledge-aware language model pretraining. arXiv preprint arXiv:2007.00655, 2020. \n[25] Tao Shen, Yi Mao, Pengcheng He, Guodong Long, Adam Trischler, and Weizhu Chen. Exploiting structured knowledge in text via graph-guided representation learning. In Empirical Methods in Natural Language Processing (EMNLP), 2020. \n[26] Fangyu Liu, Ehsan Shareghi, Zaiqiao Meng, Marco Basaldella, and Nigel Collier. Self-alignment pretraining for biomedical entity representations. In North American Chapter of the Association for Computational Linguistics (NAACL), 2021. \n[27] Donghan Yu, Chenguang Zhu, Yiming Yang, and Michael Zeng. Jaket: Joint pre-training of knowledge graph and language understanding. In AAAI Conference on Artificial Intelligence, 2022. \n[28] Pei Ke, Haozhe Ji, Yu Ran, Xin Cui, Liwei Wang, Linfeng Song, Xiaoyan Zhu, and Minlie Huang. Jointgt: Graph-text joint representation learning for text generation from knowledge graphs. In Findings of ACL, 2021. \n[29] Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and P. Wang. K-bert: Enabling language representation with knowledge graph. In AAAI Conference on Artificial Intelligence, 2020. \n[30] Tianxiang Sun, Yunfan Shao, Xipeng Qiu, Qipeng Guo, Yaru Hu, Xuan-Jing Huang, and Zheng Zhang. Colake: Contextualized language and knowledge embedding. In International Conference on Computational Linguistics (COLING), 2020. \n[31] Bin He, Di Zhou, Jinghui Xiao, Xin Jiang, Qun Liu, Nicholas Jing Yuan, and Tong Xu. Integrating graph contextualized knowledge into pre-trained language models. In Findings of EMNLP, 2020. \n[32] Bill Yuchen Lin, Xinyue Chen, Jamin Chen, and Xiang Ren. Kagnet: Knowledge-aware graph networks for commonsense reasoning. In Empirical Methods in Natural Language Processing (EMNLP), 2019. \n[33] Yanlin Feng, Xinyue Chen, Bill Yuchen Lin, Peifeng Wang, Jun Yan, and Xiang Ren. Scalable multi-hop relational reasoning for knowledge-aware question answering. In Empirical Methods in Natural Language Processing (EMNLP), 2020. \n[34] Shangwen Lv, Daya Guo, Jingjing Xu, Duyu Tang, Nan Duan, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, and Songlin Hu. Graph-based reasoning over heterogeneous external knowledge for commonsense question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. \n[35] Kuan Wang, Yuyu Zhang, Diyi Yang, Le Song, and Tao Qin. Gnn is a counter? revisiting gnn for question answering. In International Conference on Learning Representations (ICLR), 2022. \n[36] Todor Mihaylov and Anette Frank. Knowledgeable reader: Enhancing cloze-style reading comprehension with external commonsense knowledge. In Association for Computational Linguistics (ACL), 2018. \n[37] An Yang, Quan Wang, Jing Liu, Kai Liu, Yajuan Lyu, Hua Wu, Qiaoqiao She, and Sujian Li. Enhancing pre-trained language representations with rich knowledge for machine reading comprehension. In Association for Computational Linguistics (ACL), 2019. \n[38] Haitian Sun, Bhuwan Dhingra, Manzil Zaheer, Kathryn Mazaitis, Ruslan Salakhutdinov, and William W Cohen. Open domain question answering using early fusion of knowledge bases and text. In Empirical Methods in Natural Language Processing (EMNLP), 2018. \n[39] Haitian Sun, Tania Bedrax-Weiss, and William W Cohen. Pullnet: Open domain question answering with iterative retrieval on knowledge bases and text. In Empirical Methods in Natural Language Processing (EMNLP), 2019. \n[40] Jun Yan, Mrigank Raman, Aaron Chan, Tianyu Zhang, Ryan Rossi, Handong Zhao, Sungchul Kim, Nedim Lipka, and Xiang Ren. Learning contextualized knowledge structures for commonsense reasoning. In Findings of ACL, 2021. \n[41] Yueqing Sun, Qi Shi, Le Qi, and Yu Zhang. Jointlk: Joint reasoning with language models and knowledge graphs for commonsense question answering. In North American Chapter of the Association for Computational Linguistics (NAACL), 2022. \n[42] Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. Human parity on commonsenseqa: Augmenting self-attention with external attention. In Association for Computational Linguistics (ACL), 2022. \n[43] Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. Complex embeddings for simple link prediction. In International conference on machine learning (ICML), 2016. \n[44] Seyed Mehran Kazemi and David Poole. Simple embedding for link prediction in knowledge graphs. In Advances in Neural Information Processing Systems (NeurIPS), 2018. \n[45] Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems (NeurIPS), 2013. \n[46] Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learning and inference in knowledge bases. In International Conference on Learning Representations (ICLR), 2015. \n[47] Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations (ICLR), 2019. \n[48] Sebastian Riedel, Limin Yao, Andrew McCallum, and Benjamin M Marlin. Relation extraction with matrix factorization and universal schemas. In North American Chapter of the Association for Computational Linguistics (NAACL), 2013. \n[49] Kristina Toutanova, Danqi Chen, Patrick Pantel, Hoifung Poon, Pallavi Choudhury, and Michael Gamon. Representing text for joint embedding of text and knowledge bases. In Empirical Methods in Natural Language Processing (EMNLP), 2015. \n[50] Ruobing Xie, Zhiyuan Liu, Jia Jia, Huanbo Luan, and Maosong Sun. Representation learning of knowledge graphs with entity descriptions. In Proceedings of the AAAI Conference on Artificial Intelligence, 2016. \n[51] Liang Yao, Chengsheng Mao, and Yuan Luo. Kg-bert: Bert for knowledge graph completion. arXiv preprint arXiv:1909.03193, 2019. \n[52] Bosung Kim, Taesuk Hong, Youngjoong Ko, and Jungyun Seo. Multi-task learning for knowledge graph completion with pre-trained language models. In International Conference on Computational Linguistics (COLING), 2020. \n[53] Da Li, Sen Yang, Kele Xu, Ming Yi, Yukai He, and Huaimin Wang. Multi-task pre-training language model for semantic network completion. arXiv preprint arXiv:2201.04843, 2022. \n[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems (NeurIPS), 2017. \n[55] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In International Conference on Computer Vision (ICCV), 2015. \n[56] Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In North American Chapter of the Association for Computational Linguistics (NAACL), 2019. \n[57] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Empirical Methods in Natural Language Processing (EMNLP), 2018. \n[58] Bill Yuchen Lin, Ziyi Wu, Yichi Yang, Dong-Ho Lee, and Xiang Ren. Riddlesense: Reasoning about riddle questions featuring linguistic creativity and commonsense knowledge. In Findings of ACL, 2021. \n[59] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. \n[60] Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Cosmos qa: Machine reading comprehension with contextual commonsense reasoning. In Empirical Methods in Natural Language Processing (EMNLP), 2019. \n[61] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? In Association for Computational Linguistics (ACL), 2019. \n[62] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about physical commonsense in natural language. In AAAI Conference on Artificial Intelligence, 2020. \n[63] Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. Socialiqa: Commonsense reasoning about social interactions. In Empirical Methods in Natural Language Processing (EMNLP), 2019. \n[64] Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Wen-tau Yih, and Yejin Choi. Abductive commonsense reasoning. In International Conference on Learning Representations (ICLR), 2020. \n[65] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114, 2021. \n[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022. \n[67] Elliot G Brown, Louise Wood, and Sue Wood. The medical dictionary for regulatory activities (meddra). Drug safety, 20(2):109–117, 1999. \n[68] Carolyn E Lipscomb. Medical subject headings (mesh). Bulletin of the Medical Library Association, 88(3):265, 2000. \n[69] Marinka Zitnik, Monica Agrawal, and Jure Leskovec. Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics, 34(13):i457–i466, 2018. \n[70] Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. Gene ontology: tool for the unification of biology. Nature genetics, 25(1):25–29, 2000. \n[71] David S Wishart, Yannick D Feunang, An C Guo, Elvis J Lo, Ana Marcu, Jason R Grant, Tanvir Sajed, Daniel Johnson, Carin Li, Zinat Sayeeda, et al. Drugbank 5.0: a major update to the drugbank database for 2018. Nucleic acids research, 2018. \n[72] Camilo Ruiz, Marinka Zitnik, and Jure Leskovec. Identification of disease treatment mechanisms through the multiscale interactome. Nature communications, 12(1):1–15, 2021. \n[73] PubMed. https://pubmed.ncbi.nlm.nih.gov/. \n[74] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 2020. \n[75] Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. Domain-specific language model pretraining for biomedical natural language processing. arXiv preprint arXiv:2007.15779, 2020. \n[76] Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences, 2021. \n[77] Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William W Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. In Empirical Methods in Natural Language Processing (EMNLP), 2019. \n[78] Anastasios Nentidis, Konstantinos Bougiatiotis, Anastasia Krithara, and Georgios Paliouras. Results of the seventh edition of the bioasq challenge. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2019. \n[79] Wenhao Yu, Chenguang Zhu, Zaitang Li, Zhiting Hu, Qingyun Wang, Heng Ji, and Meng Jiang. A survey of knowledge-enhanced text generation. ACM Computing Surveys (CSUR), 2022. \n[80] Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. Towards controllable biases in language generation. In the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)-Findings, long, 2020. \n[81] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, et al. Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359, 2021. \n[82] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. In Findings of EMNLP, 2020. \n[83] Ninareh Mehrabi, Pei Zhou, Fred Morstatter, Jay Pujara, Xiang Ren, and A. G. Galstyan. Lawyers are dishonest? quantifying representational harms in commonsense knowledge resources. ArXiv, abs/2103.11320, 2021. ",
|
| 1201 |
+
"bbox": [
|
| 1202 |
+
178,
|
| 1203 |
+
681,
|
| 1204 |
+
828,
|
| 1205 |
+
911
|
| 1206 |
+
],
|
| 1207 |
+
"page_idx": 9
|
| 1208 |
+
},
|
| 1209 |
+
{
|
| 1210 |
+
"type": "text",
|
| 1211 |
+
"text": "",
|
| 1212 |
+
"bbox": [
|
| 1213 |
+
171,
|
| 1214 |
+
42,
|
| 1215 |
+
828,
|
| 1216 |
+
920
|
| 1217 |
+
],
|
| 1218 |
+
"page_idx": 10
|
| 1219 |
+
},
|
| 1220 |
+
{
|
| 1221 |
+
"type": "text",
|
| 1222 |
+
"text": "",
|
| 1223 |
+
"bbox": [
|
| 1224 |
+
171,
|
| 1225 |
+
31,
|
| 1226 |
+
826,
|
| 1227 |
+
920
|
| 1228 |
+
],
|
| 1229 |
+
"page_idx": 11
|
| 1230 |
+
},
|
| 1231 |
+
{
|
| 1232 |
+
"type": "text",
|
| 1233 |
+
"text": "",
|
| 1234 |
+
"bbox": [
|
| 1235 |
+
171,
|
| 1236 |
+
69,
|
| 1237 |
+
828,
|
| 1238 |
+
912
|
| 1239 |
+
],
|
| 1240 |
+
"page_idx": 12
|
| 1241 |
+
},
|
| 1242 |
+
{
|
| 1243 |
+
"type": "text",
|
| 1244 |
+
"text": "",
|
| 1245 |
+
"bbox": [
|
| 1246 |
+
171,
|
| 1247 |
+
39,
|
| 1248 |
+
828,
|
| 1249 |
+
920
|
| 1250 |
+
],
|
| 1251 |
+
"page_idx": 13
|
| 1252 |
+
},
|
| 1253 |
+
{
|
| 1254 |
+
"type": "text",
|
| 1255 |
+
"text": "",
|
| 1256 |
+
"bbox": [
|
| 1257 |
+
171,
|
| 1258 |
+
90,
|
| 1259 |
+
828,
|
| 1260 |
+
272
|
| 1261 |
+
],
|
| 1262 |
+
"page_idx": 14
|
| 1263 |
+
}
|
| 1264 |
+
]
|
parse/dev/4NpoSrT8uU-/4NpoSrT8uU-_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/4NpoSrT8uU-/4NpoSrT8uU-_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/4WgqjmYacAf/4WgqjmYacAf_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2c694c96774ede05d382a08a88156dfaf848771340ab00694922c9009e165ece
|
| 3 |
+
size 12993333
|
parse/dev/4WgqjmYacAf/4WgqjmYacAf_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:08ce3c5f8dc5e872a6de5be29b4c3cc94380f7c9b272ab80b1cccec2f388dfdd
|
| 3 |
+
size 12784928
|
parse/dev/4WgqjmYacAf/4WgqjmYacAf_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:99fafd4770bc19a4ce1ea229cada4c0f268579316b00d95fa83093631446fe8a
|
| 3 |
+
size 13024043
|