Machine Reading Comprehension (MRC)
Also known as: MRC, question answering over passages, extractive question answering, Makine Okuma Anlama (MRC)
Machine reading comprehension (MRC), popularised by the SQuAD benchmark of Rajpurkar, Zhang, Lopyrev and Liang (2016), is a natural-language-processing task in which a model reads a given passage and answers multiple-choice or open-ended questions about it. It turns a passage plus a question into a machine-generated answer, supporting information retrieval, educational technology, and querying research databases.
Key highlights
- Grounds every answer in a supplied passage, so responses are tied to a verifiable source rather than free recall.
- Handles both multiple-choice and open-ended questions within a single pipeline.
- Builds on pretrained transformer models, so strong comprehension is available without training from scratch.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use MRC when you have text passages and want a machine to answer questions grounded in them — for example knowledge access, tutoring tools, or interrogating a research database. It requires that a context passage is available and that each question is clearly formulated; without a supporting passage or with a vague question, the task cannot produce a reliable answer. A modest collection of passage-question pairs is enough to start; larger labelled sets support stronger evaluation.
Strengths & limitations
- Grounds every answer in a supplied passage, so responses are tied to a verifiable source rather than free recall.
- Handles both multiple-choice and open-ended questions within a single pipeline.
- Builds on pretrained transformer models, so strong comprehension is available without training from scratch.
- Answers are only as good as the passage; if the text does not contain the answer, the model cannot supply it.
- Vague or ambiguous questions degrade accuracy, since the model has no way to clarify intent.
- Questions that require combining evidence from multiple passages (multi-hop reasoning) remain hard.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is the difference between multiple-choice and extractive MRC?
In multiple-choice MRC the model picks the correct option from a fixed set, while in extractive MRC it identifies and returns the exact span of the passage that answers the question. Open-ended MRC instead generates a short free-form answer. All three read the same supplied passage; they differ only in how the answer is delivered.
Can the model answer if the passage does not contain the answer?
No — MRC answers with respect to the given context. If the passage does not contain or support the answer, the task cannot produce a reliable one, which is why a relevant, sufficient passage is a prerequisite.
Why does question phrasing matter so much?
The model aligns the question against the passage to locate the answer, so a clearly formulated, unambiguous question is essential. Vague or compound questions split the model's focus and lead to partial or incorrect answers.
What makes multi-hop questions harder?
Multi-hop questions require combining evidence from several places — often several documents — before an answer can be formed. Single-passage models that excel at locating one span struggle here, which is why dedicated multi-hop datasets such as HotpotQA were created.
Sources
- 1.Rajpurkar, P., Zhang, J., Lopyrev, K. & Liang, P. (2016). SQuAD: 100,000+ Questions for Machine Comprehension of Text. EMNLP, 2383-2392.
- 2.Yang, Z. et al. (2018). HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering. EMNLP.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Machine Reading Comprehension. ScholarGate. https://scholargate.app/text-mining/neural-machine-reading