Glossary
Slot filling
Slot filling finds a value for a specified field of a specified entity from available evidence.
Knowledge fusion
Slot filling asks for a value in a predefined relation or attribute slot, such as an organization's registration number. The subject and requested field are known; the answer may be a value, several values where the field permits them, or no supported answer. The task differs from open-ended extraction because the output shape and target relation are fixed before evidence is read.
A practical slot-filling task searches candidate sources, checks whether mentions refer to the intended entity and reads the local context around a proposed value. The field definition matters: a headquarters address and a registered address are different slots. Dates matter too, because an old value may be correct historically and wrong for the requested as-of date.
An organization defending a filled field needs more than the final string. It needs the source span, the interpretation of the field and an explicit state when evidence is absent or conflicting. Information extraction is often used to propose candidates; entity disambiguation ensures they belong to the right subject; truth discovery helps when credible candidates disagree.
Example
For Halden Therapeutics AG (fictional), a request for registered address does not accept a laboratory mailing address as the same slot.
Task definition and its origins
In slot filling the system is told which entity and which attribute to look for, and must return the value or report that none is found. The best-known formulation is the slot-filling task in the TAC knowledge base population evaluations, in which systems received a query entity, such as a person or an organisation, together with a list of slots, such as place of birth or parents, and returned fillers with the document passages that justified them. The task has an important feature that is easy to overlook: a correct answer may be that there is no filler, and a system is judged on that decision as well as on the values it returns.
A more recent line recasts slot filling as reading comprehension. Levy and colleagues turn each relation into a natural-language question about an entity and use a question-answering model to find the answer in a sentence, which allows relations unseen in training to be handled by writing a new question. Work on unanswerable questions in SQuAD 2.0 made the same point from the evaluation side: a system that always returns a span, even when the passage does not contain the answer, will look better on answerable cases and be worse in use.
Handling absence, multiplicity and time
Three properties of slots deserve explicit design. Absence: a missing value may mean the attribute does not apply, the source is silent, or the extractor failed, and these are three different states with different consequences, as set out in null semantics. Multiplicity: a slot such as subsidiaries or inventors can hold many values, so the task is to return a set and to be scored on that set. Time: a slot such as chief executive holds different values at different dates, so the filler needs a period, and the valid time of the statement must be kept with it.
For each filler, keep the passage that supports it. A filler without evidence cannot be audited; a filler whose passage does not contain the stated value is a hallucination of the extraction step. Where two passages give different fillers, the disagreement is handled as a data conflict and not by taking the first.
A worked example
The query is the chief financial officer of Halden Therapeutics AG (fictional) as of 30 June. One document names a person as chief financial officer in a biography section; another, dated earlier that month, announces that this person will step down at the end of the quarter and that a successor takes over on 1 July. The correct fill for 30 June is the first person, with the second document as supporting evidence for the end date. The correct fill for 2 July would be the successor. A system that returns only the name most often attached to the title, without regard to dates, will be right for one date and wrong for the other. The abstention case is also real: if no document in the collection names a holder of the post, the correct output is no value with a reason, not the nearest plausible name.
Related terms and boundaries
Slot filling is a constrained form of information extraction used to populate a knowledge base. It relies on entity disambiguation to be sure that the query entity is the entity the passage describes, and on abstention as the system-level behaviour when no filler is supported. Its conflicts feed truth discovery.
Checks a reviewer can run
Ask what the system returns when no value is found, and how that is distinguished from a failure. Test a slot that changes over time with two different dates. Check that multivalued slots return sets. Confirm that each filler carries its supporting passage and that the passage contains the value.
Common misreadings
- Returning the first matching string without context.
- Mixing similarly named fields.
- Treating no retrieved answer as proof of absence.
Appears in
- LLM extraction vs rule-based extraction
Language models interpret varied prose; explicit rules suit stable, mechanically defined fields and formats.
References
- Surdeanu (2013), Overview of the TAC2013 Knowledge Base Population Evaluation: English Slot Filling and Temporal Slot Filling, TAC.
- NIST, Text Analysis Conference (TAC), Knowledge Base Population evaluations.
- Levy et al. (2017), Zero-Shot Relation Extraction via Reading Comprehension, CoNLL.
- Rajpurkar et al. (2018), Know What You Don't Know: Unanswerable Questions for SQuAD, ACL.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Slot filling" bear on a decision of yours?
This entry sits under knowledge fusion. If a decision in your organization turns on it, send the entity concerned, the sources you fuse to describe it and who may read which of them. We reply with what a first record would cover.
Or write to [email protected] in your own words.