Explicador de conceptos·
How does extractive question answering work in legal AI?
La noticia detrás de este explicador:Leer la noticia en NewsPals
Recent benchmark work on Indian intellectual property law highlights a practical problem in legal AI: it is not enough for a system to sound plausible. In legal work, the system must often point to the exact words that support an answer.
Why this matters now
Most professionals have seen AI systems produce fluent summaries of contracts, regulations, or statutes. Fluency is useful, but it can be risky when the answer needs legal grounding. A vague response such as “the law generally allows this” may be unacceptable if a lawyer, compliance team, or product reviewer needs to know which phrase in which provision creates that conclusion.
Extractive question answering is important because it changes the evaluation target. Instead of rewarding a model for generating a polished explanation, it tests whether the model can identify the precise span of text that answers a question. That span might be a definition, condition, exception, time period, threshold, or list of covered rights.
This is especially valuable in legal domains where wording matters. Similar provisions can differ by a single qualifier. Exceptions can reverse the practical answer. Cross references can make a nearby sentence misleading unless the right clause is selected. For professional users, span grounding is a safeguard against confident but unsupported legal prose.
How it works
Extractive question answering is a task where the system receives a question and a source passage, then returns a contiguous text span from that passage as the answer. Unlike generative answering, it does not invent a new sentence. The output is selected from the legal text itself.
- Legal question
- Source passage
- Span selection
- Grounded answer
The system selects the exact answer text from the provided legal passage.
A typical system has two related jobs. First, it must find the relevant context, such as the right section of a statute or clause of a contract. Second, it must select the exact answer span within that context. These are different skills. A retriever may bring back the correct section, while the question answering model still chooses the wrong sentence. Conversely, a span selector may work well only when the right passage has already been supplied.
Training data usually consists of question, context, and answer span triples. The answer span is marked by its start and end location in the text. During learning, the model becomes better at aligning question wording with legal phrasing, including definitions, provisos, enumerated conditions, and exceptions.
Evaluation often uses overlap metrics, because the selected span may be partially correct. But legal use also requires qualitative review. A span that is close but omits “unless,” “subject to,” or “notwithstanding” can change the legal meaning.
Real-world applications
In legal research, extractive QA can help users ask targeted questions across statutes, regulations, policy manuals, or case summaries, then inspect the exact supporting text. This makes review faster without hiding the authority behind a generated paragraph.
In compliance and risk operations, teams can use span-grounded QA to check whether a requirement, exemption, reporting duty, or eligibility condition appears in a policy corpus. The answer remains auditable because the system returns the source wording.
In contract analysis, extractive QA can locate termination rights, renewal periods, notice requirements, indemnity triggers, liability caps, and governing law clauses. The same pattern also applies beyond law: insurance policies, procurement rules, clinical protocols, and technical standards all benefit from precise answer spans.
The key product lesson is that extractive QA should not be treated as a complete lawyer substitute. It is better understood as a precision layer: a way to reduce search friction, improve traceability, and support human review.
Where to go deeper
To build skill in this area, study the difference between retrieval, extractive QA, and generative QA. Retrieval finds candidate passages. Extractive QA selects answer text. Generative QA writes a new answer, ideally grounded in retrieved evidence.
Next, learn how benchmark design affects trust. Good legal QA benchmarks avoid leakage between training and test material, include varied question types, and measure whether systems can handle definitions, exceptions, and cross referenced clauses.
Finally, practice reading model outputs like a reviewer. Ask: Is the cited passage the right authority? Is the selected span complete? Does it include necessary qualifiers? For legal AI, these questions are often more important than whether the answer sounds fluent.



