According to the official case of Kouzi, Shidian Ancient Books was jointly launched by ByteDance and Peking University. Its intelligent assistant can obtain context based on questions or selected text, call dictionary and ancient book search results, and then use a large language model to organize answers and attach links to the original texts. This article examines the process for using this material without independently verifying its search coverage or interpretation accuracy.
Where the reader is reading changes the meaning of the question
The following is ChinaOPC’s independent analysis of data reading services, and is not the undisclosed product rules of Shidian Ancient Books. When a reader selects a sentence and asks "Who does this refer to?", what they need may be the relationship between the characters in the preceding and following paragraphs; if the same sentence is entered directly, the system may not know which book it comes from. Service design can first allow users to see the currently selected content and its location, making it easier to find that the wrong paragraph has been quoted.
The question information can also include the version or version information currently being viewed. If a person with the same name or a similar title appears in other documents and the search yields more materials, it cannot be immediately determined that they are all related to the text in front of you. The assistant can present possible connections and explain which contexts still need to be compared, leaving it up to the reader to judge.
Dictionary explanations, usage in the text and related materials each have their own role
Dictionary entries provide clues to understanding the word, the current passage helps determine specific usage, and other documents may provide comparative material. When answering, you can first explain which type of information you mainly rely on, and then list separately the items that can be further consulted. If materials of different natures are lumped together into one single sentence for a definite explanation, it will be difficult for the reader to know where to look back.
If you are still unsure about character recognition or sentence segmentation, you can leave questions and be prompted to read the original pictures and text first. After a recognition result enters subsequent searches, the system may find many materials that match the wrong text; more information cannot make up for the uncertainty of the starting point. When a small team provides data services, the words to be checked can be handed over as separate items to personnel with corresponding experience.
Automatic punctuation, vernacular explanations and comparison with the original text should also make it easy for readers to identify. Explanatory text formed for the convenience of reading can be displayed separately; the original text should retain its source and positioning. In this way, when comparing different interpretations, readers can still clearly know which part is the original material and which part is subsequently compiled or generated.
The link needs to return to the specific location with the question
Attaching the title of the book or the homepage of the information website may be of limited help in continuing reading. If conditions permit, the link can be positioned to the corresponding volume, chapter, page or identifiable paragraph, and the version used will be prompted. After the reader clicks on it, he should be able to see the quoted content and the necessary context, and then judge whether it supports the previous explanation.
When comparing across versions, differences in the original text and the location of the data can be presented side by side, separating the explanation of the differences from the judgment of which interpretation to adopt. When a certain paragraph of text is not found in the search results, it can also explain the scope of this query and the expressions that have been tried; you cannot directly write a miss as not being recorded in all ancient books.
The data assistant can also support continued questioning. For example, the reader first looks up the meaning of the word, and then asks how this text is related to the previous text. The system needs to continue working on the current material; if the question has moved to another book, the user should be informed of the change in search scope to avoid mistaking the newly found fragment for still being from the original passage.
For small teams, you can start with a set of approved topic materials, first clarify the positioning, dictionary query and related reading, and then discuss expansion. The value delivered is to help readers understand and examine the material more easily, and generating explanations still needs to be reviewed for purpose. Shidian Ancient Books is a case of cooperation platform. This article does not use it to represent the main body of OPC entrepreneurship, nor does it use the platform’s results as academic verification of ChinaOPC.
Extended reading: How to check after converting scanned documents to text and PDF Q&A Why should we return to the original page, compare the different tasks of text recognition, data positioning and interpretation generation.