Gains:
- Ability to extract topic extract from actual content and map primary and secondary topics from approved topic list
- Being able to accept the main class of classification suggestions such as Dewey and verify the sub-levels with the official table.
- Ability to understand the judgmental and representational nature of classification and check suggestions against outdated and exclusionary terms.
Determining what a resource is "about" is one of the most judgmental tasks of librarianship. This is called topic analysis: reading the content of a document and matching it with the correct subject headings and the correct classification number. If a user can find related books side by side on the shelf or gets accurate results when searching for a topic in the catalog, there is a good topic analysis behind it. In this unit, you will learn how to use artificial intelligence in topic recommendation, classification and automatic indexing; but you will learn why the human should make the final subject decision.
Let's define a few concepts. The classification system is a system that numbers sources according to their subject; the most common are the Dewey Decimal Classification (DDC) and the Library of Congress Classification (LCC). For example, in Dewey, 500 indicates natural sciences, 900 indicates history and geography. Indexing is the task of assigning key terms to a document that make it findable. Automatic indexing is when the software suggests this job. AI produces suggestions in all three tasks, but the expert has the final say in each one.
Step by step: topic analysis with artificial intelligence
1. Give the gist of the document to the AI. Full text, abstract, table of contents or introduction. Topic analysis is based on context; If you give AI just the title, you'll get a superficial and misleading recommendation.
2. Request a content summary from the AI. First, “what is this document mainly about?” ask. This both helps you understand quickly and shows whether the AI has grasped the subject correctly.
3. Match with controlled topics. Give the AI your list of approved topics and say “suggest only the most appropriate topics from this list.” Do not allow it to produce free terminology.
4. Suggest classification number. You can ask the AI "which main class does this document fall into in Dewey?" but be sure to verify the exact digits of the number with the official classification chart. AI can produce reasonable but inaccurate numbers.
5. Make the decision and justify it. Review the AI's suggestions, select them in light of the institution's policy and user base, and reject them when necessary. The final subject decision and classification bears your signature.
Tip: A document can be given multiple topics. From AI “what is the primary topic of this document, what are its secondary topics?” Ask separately. Choosing the primary subject correctly ensures that the resource goes into the right place on the shelf; secondary topics give additional access points in the catalogue.
Judicial nature of classification
Classification appears neutral, but it is not. Whether to put a document under "women's studies" or "sociology", or even under which geography you will classify a historical event, are decisions that involve perspective. Classification systems carry the assumptions of particular periods and cultures; Some terms become obsolete or exclusionary over time. AI inherits these biases verbatim from the data it is trained on and may even reinforce them.
Therefore, topic analysis is a matter of judgment that cannot be left to AI. AI indicates which titles are “likely”; But the decision as to which title represents the source most accurately and fairly belongs to the expert who knows the values of the institution and its user base.
Caution: Make sure that a topic suggested by AI does not misrepresent or exclude the source. Compare AI-generated terms with the institution's current and inclusive vocabulary, especially on sensitive topics such as identity, ethnicity, religion and gender.
three mini cases
Case 1 — Mass topic proposal. A library would assign topics to a digital collection of 800 articles. The expert gave the summary of each article to the AI and asked for a match from the approved title list. AI recommendations correctly captured an average of 2 out of 3 eligible titles per article; the expert added a third and weeded out the wrong suggestions. The job was finished approximately 40% faster than if done entirely by hand.
Case 2 — Incorrect classification number. An expert asked for a Dewey number from AI for a medical history book. YZ suggested "610 (Medicine)"; However, the book was mainly about the history of medicine and the correct place was "610.9 (History of Medicine)". The expert looked at the official Dewey table and corrected it; AI found the main class but missed the sub-break.
Case 3 — Obsolete term caught. AI proposed an outdated and now considered exclusionary subject term for a sociology book; because this old term was common in the training data. The expert selected the correct term from the institution's current inclusive vocabulary. This was a concrete example that AI carries the bias of the past.
Depth and consistency: how specific should it be?
A common decision in topic analysis is the level of specificity of the term to assign. If you give a resource a title that's too general ("Date"), the user searching for the topic will be drowned in hundreds of irrelevant results. If you give a title that is too narrow, a user looking for a broader source will not be able to find it. The rule is to choose the term that most specifically but fully covers the source: if a source deals only with Seljuk period architecture, "Seljuk architecture" and not "Architecture" is the correct level. AI often misses this balance; It produces suggestions that are either too general or too dispersed. Therefore, it is necessary to give clear instructions to AI such as "suggest the most specific topic covered by the source, and do not add subtopics that the source does not actually touch upon." Additionally, consistent labeling of resources on the same topic over time is as important to the integrity of the catalog as specificity; Saying "Seljuk architecture" one day and "Seljuk period buildings" the next day divides the user.
Tip: When giving multiple titles to a resource, do not unnecessarily stack terms that are subsets of each other. Giving both "Ottoman history" and "History" makes the latter unnecessary; choose the most specific and only include titles that really add a different reach.
Four copyable templates
1) Topic extraction:
Write the primary topic of the document below in one sentence and the secondary topics in 3 items. Adding a topic that is not in the text, based only on the text. Document/summary: [here]
2) Matching from approved header:
Below is a document summary and list of approved subject headings. Match only the 3-5 headings from the list that are most appropriate to the document, starting with the most appropriate. Write a one-sentence justification for each choice. If there is no suitable title in the list, specify it. Summary: [here] / List: [here]
3) Classification master class suggestion:
For the following document, suggest a possible major class (3 digits) in the Dewey Decimal Classification and explain why you chose it. Mark the lower digits as "must be verified from the official table", do not give exact numbers. Document: [here]
4) Bias and currency checking:
Are there any terms that may be outdated, exclusionary, or one-sided among the topic suggestions below? If so, mark it and explain why it might be problematic. Suggestions: [here]
Weak prompt / Strong prompt
Weak prompt:
Give the topics of this book: "Commerce in the Ottoman Empire".
Working solely from the title, the AI makes up general terms without knowing the context and ignores the controlled vocabulary. Result: inconsistent, perhaps incorrect titles.
Powerful prompt:
Your role: assistant subject matter analyst. Below is the introduction and contents of the book; I also provided a list of approved topics. (1) Summarize the primary topic of the book in one sentence. (2) Match only the 3-5 most appropriate titles from the list with the rationale. (3) Don't go off the list, don't make up terms. Text: [here] /Approved list: [here]
The strong prompt limits the AI to real content and controlled vocabulary; maintains both accuracy and consistency.
Topic analysis task chart
Quest
Role of AI
man's decision
Content summary
quick draft
accuracy check
Title suggestion
Match from list
Final choice, justification
Classification number
Master class suggestion
Subdivision verification
Bias check
marking
Inclusive term decision
Common mistakes
- Just asking for a topic from the title. Topic analysis is based on context; The title is misleading.
- Accepting the classification number without verifying it. Although AI finds the main class, it misses the subbreak.
- Skipping controlled vocabulary. Free term breaks consistency.
- Ignoring prejudice. AI may carry over the outdated, exclusionary terms of the past.
- Leaving the decision entirely to AI. Representation and justice require judgment; The decision is up to the person.
In summary
In topic analysis and classification, AI is a fast assistant that extracts content summary, matches topics from the approved list, and suggests classification master class. But subject analysis is essentially a matter of judgment and representation: it is up to the expert to decide which title most accurately and fairly describes the source, the exact digits of the classification number, and the currency of the terms. Run the AI with real content and controlled vocabulary, verify classification numbers with the official table, and be alert to bias.
Application task
Prepare an introduction to a source you have and a short list of approved topics. Get AI suggestions with “Topic extraction” and “Approved topic matching” templates. Then compare the suggestions against the official classification table and the institution's current vocabulary: how many suggestions were correct, how many did you correct? Check recommendations for outdated terms with the “Bias and recency check” template.
checklist
- [ ] I had the topic analysis done from the actual content, not the title.
- [ ] I mapped the titles from the controlled vocabulary.
- [ ] I verified the lower digits of the classification number with the official table.
- [ ] I have checked suggestions against outdated and exclusionary terms.
- [ ] I made the final subject and classification decision at my own discretion.