The unread oracle bone characters

An oracle bone is a prepared ox scapula or turtle plastron with rows of hollows drilled into its back. A diviner applied a heated point to a hollow so that a crack formed on the front, read the crack, and the record was cut beside it with a knife: a preface giving the day in the sixty-day cycle and the diviner's name, the charge put to the ancestors or spirits, sometimes the king's reading of the crack, and sometimes what happened⁠[1]. Charges concern sacrifice, harvest, rain, war, hunting, illness, childbirth and the ten-day period ahead. Pieces were stored in pits at Xiaotun; pit YH127, found in 1936, held over 17,000 inscribed pieces, and the 1991 Huayuanzhuang East cache held 561 inscribed pieces from a prince's household⁠[24]⁠[1].

About 150,000 inscribed pieces are known, held in China and about fourteen other countries; roughly 90 percent are fragments⁠[13]⁠[20]. The National Library of China holds upwards of 35,000, Academia Sinica about 20,000 from the 1928 to 1937 excavations, and the British Library about 500 under shelfmark Or. 7694⁠[23]⁠[24]. The standard edition is the Jiaguwen heji (1978 to 1982), 41,956 rubbings and drawings in 13 volumes arranged by Dong Zuobin's periods, with a 1999 supplement of 13,450 pieces⁠[22]. Later excavated groups have their own volumes. Digital access is split between a subscription database of over 143,000 inscriptions from Zhonghua Book Company, the free Academia Sinica rubbings database, the Anyang Normal University platform Yinqi Wenyuan, the CHANT transcriptions at the Chinese University of Hong Kong, and Xiaoxuetang's graph-by-graph forms⁠[25]⁠[26]⁠[10]⁠[27]⁠[28].

The language is early Chinese and the script is the direct ancestor of modern Chinese characters. Its graphs work as later Chinese graphs do: some are pictures of what they mean, some are borrowed for a word that sounds alike, and most combine a sound element with a meaning element⁠[7]. Text runs in columns, usually top to bottom, and paired charges on one bone are often cut in mirror image on left and right⁠[1]. Numerals, the sixty-day cycle, the names of the kings and their consorts, the diviners, the sacrifices and the common verbs are all read, so most inscriptions can be translated in full and the rest in large part⁠[1]⁠[29].

The chronology is fixed on three legs. Wang Guowei's 1917 matching of sacrificial recipients to the Shiji king list established that the archive runs from Wu Ding to Di Xin⁠[3]. Dong Zuobin's ten criteria of 1933, since refined into diviner groups by Li Xueqin, Qiu Xigui and Lin Yun, assign most pieces to a reign or pair of reigns⁠[4]⁠[5]. Radiocarbon dating of 48 king-attributed bones in 2021 gave 1254 to 1197 BC for Wu Ding and about 1041 BC for the end, and placed the disputed Li diviner group in the Wu Ding to Zu Geng reigns rather than later⁠[6].

What is known about the unread graphs is mostly negative. Counts of distinct graphs range from about 3,800 to about 4,500 depending on how scribal variants are merged, and the share read ranges from a third to a half depending on whether disputed readings count⁠[8]⁠[9]⁠[10]⁠[11]. Most unread graphs are rare. Context often shows what kind of word a graph is, a place after the verb for hunting, a person after a verb of sacrifice, a clan sign on a bronze, without showing which word⁠[11]⁠[7].

The method that has produced every accepted reading since 1914 is the chain of forms. A graph on bone is matched to a bronze form of the Western Zhou, then to Warring States and small seal forms, and finally to a character whose sound and meaning are recorded; the reading is then checked against every context in which the graph occurs⁠[2]⁠[7]. Jiang Yubin's 2018 reading of chun, meaning stirring or unrest, followed this route through bronze inscriptions and was accepted by the Anyang museum's reviewers; Zhou Zhongbing's reading of tai, accepted in 2023, took two years of collection and drafting⁠[13]⁠[14]. The bounty has produced three awards in nine years, which is the measure of how hard the remaining graphs are⁠[10].

Computation has been tested on three tasks. For recognition of known graphs, datasets such as OBC306 (309,551 scanned images, 306 classes) and HUST-OBC (140,053 images; 1,588 read classes and 9,411 unread classes) support classifiers that reach high accuracy on read graphs⁠[15]⁠[16]⁠[9]. For decipherment, EVOBC provides 229,170 images along six stages from bone to clerical script so that a model can learn how forms change; a baseline predicts the modern character of a held-out read graph with 16.7 percent top-1 and 55.8 percent top-20 accuracy⁠[17]. The OBSD diffusion model generates a candidate modern glyph directly from a bone graph image and was judged the best paper at ACL 2024, but its output for unread graphs is a visual hint rather than a reading, and a 2026 review finds such hints of limited use to palaeographers⁠[18]⁠[11]. OBI-Bench tested 23 large multimodal models on 5,523 images and found the best of them at about the level of untrained people on deciphering and below them on fine perception⁠[19].

For rejoining, Zhang and colleagues' 2020 system reached 98.39 percent top-10 accuracy on held-out joins, the Microsoft and Capital Normal University system Diviner matches fragment outlines across publications, and the Tencent and Anyang Normal University platform has found about 500 duplicate rubbings⁠[21]⁠[20]⁠[10]. These recover context rather than readings, but context is what the chain-of-forms method needs.

Five positions are live about the unread graphs, and they are not exclusive. The first is that most are proper names: people, places, clans, victims and spirits whose graphs were never used for a common word and so left no descendant to be traced. On this view they can be identified as names and dated but never read as words, and the undeciphered count will never fall far⁠[11]. The second is that many are common words that died before the bronze age of writing, so that no later form exists; these can be read only from internal evidence, by showing from every context what the word must mean, which the museum's reviewers have so far accepted only when a form chain also exists⁠[7]⁠[14].

The third is that a large share of the unread classes are variants, miscuts or damaged forms of graphs already read. HUST-OBC lists 9,411 unread classes against 1,588 read ones, far more than the 2,000 to 3,000 unread graphs scholars cite, which shows how much depends on merging variants⁠[16]⁠[9]. If so, the true problem is smaller than the headline figures and a careful variant study would shrink it. The fourth is that many unread graphs do have later descendants obscured by changes of form, and that systematic chains such as EVOBC's six stages, together with the slow stream of bounty awards, will keep yielding readings⁠[17]⁠[10]. The fifth is that generative and multimodal models will propose readings that experts then confirm; no accepted reading has yet come this way, and the benchmarks place current models at the level of untrained people⁠[18]⁠[19]⁠[11].

All five share a dependence on context, and context is what fragments lack. The rejoining work is therefore relevant to every position⁠[21].

The central question is not whether the script can be read but how many of the unread graphs can ever be read, and each position predicts something testable. A full variant study of the unread classes in HUST-OBC and Yinqi Wenyuan, merging allographs by position, period and diviner group, would show how many distinct unread words there are; if the count falls well below 2,000 the variant position gains⁠[16]⁠[10]. A classification of every unread graph by syntactic slot, using the CHANT transcriptions, would show what share can only be names; if it is most of them, the names position is confirmed and the bounty's ceiling is known⁠[27]⁠[11].

For the computational position the test is a held-out one. A model trained on graphs read before a cut-off date should be scored on the graphs accepted since 2018 by the Anyang reviewers, with the generated glyph or ranked candidates compared with the accepted reading; EVOBC's own baseline of 16.7 percent top-1 on read graphs is the number to beat, and a model that places an accepted new reading in its top 20 on a graph it never saw would be the first real evidence⁠[17]⁠[18]⁠[14]. Any proposed reading must then do what Jiang Yubin's did: fit every attested context and connect to a later form or a documented word⁠[13]⁠[7].

New finds remain possible. Yinxu is still excavated, the 1991 Huayuanzhuang cache and the Zhouyuan bones of the 1970s and 2003 were unexpected, and more rejoins among the 150,000 fragments would supply the contexts that rare graphs need⁠[24]⁠[21].