Prompt engineering deliberately designs and iteratively refines natural-language instructions and examples so a language model produces useful, task-specific outputs. In knowledge-graph construction, it formulates entity and relation extraction as text completion, often avoiding a separately engineered NLP pipeline. section 5.3.1 section 5.3.4
A reliable prompting workflow combines domain analysis, explicit output constraints, representative examples, and controlled evaluation:
- Explore the data and anticipate domain-specific difficulties, such as names later appearing as initials, surnames, or pronouns. section 5.3.1
- Describe the task precisely, and enumerate the entity and relation classes that matter. section 5.3.3
- Add complex, representative input-output examples containing difficult linguistic forms and all important relation types. section 5.3.3
- Iterate on wording, class names, model choice, and parameters; lower temperature favors deterministic output, while higher temperature increases variation and hallucination risk. section 5.3.1 section 5.3.3
- Inspect outputs for inconsistent schemas, overly granular relations, hallucinations, and unresolved entities before generating or querying a knowledge graph. section 5.3 section 5.3.2
Normalizing entity and relation extraction
An initial prompt asks for every entity and relation, but produces variants such as “specializes in,” “is measuring,” and “demonstrates.” A revised prompt names target classes and prioritizes normalized relations such as “works on,” while specifying how initials and abbreviated titles should be resolved. The resulting output expands “Prof. Chem.” to “Professor of Chemistry” and links abbreviated references to full names, though it still requires validation.
Listing 5.1 Prompt version 1; identifying entities and relations
prompt_segments = dict() prompt_segments['task'] = #1 """You are an expert on constructing Knowledge Graphs from texts using named entity recognition and relation extraction. Given a prompt, identify as many entities and relations among them as possible and output a list of relations in the format [ENTITY 1, ENTITY 1 TYPE, RELATION, ENTITY 2, ENTITY 2 TYPE]. The relations are directed, so the order matters.""" prompt_segments['example'] = "J.R.Smith (Prof. Phys.) is employed by MIT and mentioned another scientist, Mary Hodge, who works on cyclotron research." #2 prompt_segments['example_output'] = #3 """["J. R. Smith", "person", "has title", "Professor of Physics", "title"] ["J. R. Smith", "person", "works for", "MIT", "organization"] ["J. R. Smith", "person", "talked about", "Mary", "person"] ["Mary", "person", "works on", "cyclotron", "occupation"]"""
Use the listing to inspect a concrete prompt assembled from a task description, an example, and an expected output format. section 5.3.2
Prompt engineering is not a guarantee of correctness. If prompting remains insufficient for a complex task, expected outputs can seed a human-annotated training set for fine-tuning. Later systems may also use chain-of-thought or scratchpad prompting to encourage intermediate reasoning before a final answer. section 5.3.1 section 14.7
