When to Look Up a Word vs. Infer It (Japanese)

Knowing when to look up a word vs. infer it is the word-by-word judgment that separates vocabulary-building reading from reading that stalls on every line. The honest answer to "should I look up every word reading Japanese?" is that there is no single rule. It depends on your reading mode, whether the word blocks comprehension, and how cheap the lookup is.

Overview

A lookup and a guess trade off against each other. A lookup gives you the exact dictionary meaning but interrupts reading. Inference preserves flow but can leave you with a vague or wrong meaning.¹²

Because each move works under different conditions, the skill is not picking a side. It is choosing per word. The sections below give a trade-off frame, a reading-mode rule, a single-word heuristic, and the middle ground a hover dictionary opens up.

The trade-off: flow vs. precision

Every unknown word forces a small decision with two real costs: stopping, and guessing wrong. Naming both makes the word-by-word choice deliberate rather than reflexive.

What a lookup costs and buys

A lookup interrupts the reading task, and interruptions degrade reading comprehension. In a controlled study, interruptions during reading significantly impaired comprehension of passages that required connecting and synthesizing information across the text. They did not impair simple recognition of isolated facts.³

The mechanism is working memory. Comprehension depends on briefly holding information active while you read. A mid-sentence stop to look up a word competes for exactly the resources the sentence needs, which is what "I lost the thread of the sentence" describes.³

The flow cost is real but not uniform across readers

In the same study, a 15-second processing buffer before an interruption removed its negative effect. The disruption was largest for readers with lower working-memory capacity. The cost of a lookup is a mechanism, not a fixed penalty, and it lands harder mid-sentence than at a natural pause.³

Against that cost, a self-initiated lookup improves retention. When advanced second-language readers consulted a bilingual dictionary for an unknown word, they retained those words as well as or better than words explained in margin glosses, and better than words in a no-information control.¹

The broader pattern is that supplying the meaning by any means beats leaving the word unresolved. Both the dictionary condition and the gloss condition produced more incidental vocabulary learning, meaning learning that happens without direct study. Both did better than the control in which no meaning was available.¹

A lookup also gives precision. It returns the dictionary sense, whereas a guess returns only an approximation. That precision difference is the core of the trade-off.²

What inference costs and buys

Inference keeps the eyes moving. It avoids the working-memory interruption a lookup imposes, so it protects comprehension of the surrounding text.³

Its cost is unreliability when coverage is low. Guessing from context fails when too many surrounding words are unknown. The reliability of inference rises with lexical coverage, the share of words you already know, so it is feasible mainly when the reader already knows the bulk of the words around the gap.²⁴

The contribution is real but limited. Readers who already knew 90% of the words in a text and then inferred from context raised effective coverage by roughly 5 percentage points, approaching the 95% region. Inference adds coverage at the margin. It does not create comprehension from a low base.⁴

In that same study, measured comprehension differed little across 90%, 95%, and 98% coverage of the same texts (about 81%, 82%, and 81%). The author links this to active inferencing during reading. Inference can partly compensate for missing words, but it works on top of high coverage, not in place of it.⁴

A confident wrong guess corrects itself less often than a flagged gap

A guess can be confidently wrong. A wrong meaning the reader keeps re-deriving on each encounter is worse than a known gap, because nothing prompts a correction. This is the silent-propagation risk that makes inference dangerous on load-bearing words.²

Why neither extreme works

Looking up every word maximizes the interruption cost. By removing the need to guess, it also prevents you from developing the inferencing ability that high-coverage reading rewards.³⁴

Guessing every word fails precisely where coverage is low or context is thin. Those are the conditions under which inference is least reliable. Unresolved key words then degrade comprehension, and wrong guesses propagate.²⁴

Lookup wins on precision and retention. Inference wins on flow. Because each is reliable only under certain conditions, the rational policy is to choose per word rather than apply one blanket rule.¹⁴³

Decide by reading mode

The first lever is not the word but the session. The same unknown word calls for a different response depending on whether you are reading to extract every detail or reading to cover volume.

Intensive (precision) mode: look up freely

Intensive reading is close reading of shorter, denser text. You read for full, precise comprehension and analysis rather than for pleasure or volume.⁵

When the goal is full comprehension of a hard text, resolving unknown words is the task. A lookup is not a digression here. Its precision-and-retention payoff is exactly what the mode wants.¹

Extensive (flow) mode: infer by default, look up sparingly

Extensive reading is high-volume reading of easy material. In Day and Bamford's principles, the material sits well within the reader's competence. The purpose is pleasure or general understanding rather than full comprehension, and reading speed is meant to be faster, not slower.⁵

The mode discourages the dictionary by design. The material is easy enough that constant lookups are unnecessary, and stopping to look words up works against the volume-and-speed purpose.⁵

Inferring first is viable here because extensive material is, by definition, high-coverage for the reader. That is the condition under which inference is most reliable. This is why inference can be the default in extensive mode, but not in intensive mode on hard text.⁴⁶

Inferring is not merely a flow-preserving fallback. In one reading experiment, readers retained words they worked out from an informative context at least as well as words resolved by a retrieval prompt, a cue to recall the meaning. The guess itself is a genuine learning event, not only a way to keep moving.⁷

How lookups interact with reading speed

A lookup is an interruption, and interruptions during reading slow processing and impair comprehension of connected text. Repeated stops therefore lower the sustained reading rate that extensive reading is built to develop.³

Faster reading speed is an explicit aim of extensive reading, and frequent dictionary stops work against that aim. That tension is one reason the mode limits lookups rather than banning them or allowing them freely.⁵

A heuristic for the single word

Within a session, each unknown word still needs its own verdict. Four questions, asked in order, resolve almost every case: look up, infer, or skip.

The decision has this short tree shape.

Does this word block comprehension?

The first question is whether not knowing the word breaks the meaning of the sentence or paragraph. If the gist survives without it, inference or skipping is defensible. If the gist collapses, the word is load-bearing and warrants a lookup.⁶⁴

This follows from the coverage logic: comprehension can tolerate some unknown words, so only the words that actually carry meaning need resolving.⁴

The 98% coverage figure is a strong rule of thumb, with a year on it

Hu and Nation (2000) found that adequate unassisted comprehension needed about 98% lexical coverage, roughly one unknown word in 50. Below that, comprehension fell off. A later replication (104 Sri Lankan adult learners, reported 2023) did not fully reproduce the 98% figure, so treat it as a rule of thumb rather than a constant. The "does it block meaning?" test does not depend on the exact number.⁶

In practice, a few unknown words per page are normal and tolerable. A reader can afford to resolve only the ones that block meaning.⁶

Is it recurring or one-off?

Frequency earns the stop. Hulstijn and colleagues found that repeated occurrence boosted incidental learning more when meaning information was available than when it was not. A word you keep meeting both repays the lookup cost and is more likely to stick once resolved.¹

A high-frequency unknown word recurs across texts, so a single lookup pays off over many future encounters. A rare word in throwaway text may never reappear and returns little on the investment.²

How confident is your inference?

Inference reliability tracks how much of the surrounding text you know. A guess from rich, high-coverage context is far more trustworthy than one from thin context. Calibrate confidence to context strength, not to how plausible the guess feels.⁴

Because a confidently wrong guess can propagate silently, a low-confidence guess on a word that also blocks comprehension most needs confirmation. A high-confidence guess on a non-blocking word can be noted and passed.²

The unknown-word situation, and the advice to confirm it, is stated plainly in Japanese itself.

読よんでる時ときに知しらない単語たんごが出でてきたら、辞書じしょでちゃんとした意味いみを調しらべてね。⁸
"If you come across a word you don't know while reading, look up its proper meaning in the dictionary."

When confidence is low, whether to resolve the word really does turn on the surrounding context.

それは文脈ぶんみゃくによる。⁸
"It depends on the context."

Plain-text vs. furigana vs. hover-enabled

The decision rule shifts with how expensive a lookup is. On paper, a lookup means leaving the page for a dictionary. That large interruption raises the bar for stopping. With a digital pop-up dictionary, the meaning appears when you hover over the word, which lowers that bar.³⁹

Furigana resolves the reading, not the meaning

Furigana supplies the pronunciation of a kanji word but not its meaning. A furigana'd unknown word is still a meaning gap. Treat it as the read-but-not-understood case covered under "Good to know," not as a resolved word.²

The Yomitan-hover middle ground

A hover dictionary sits between paper and a full dictionary stop. It changes the arithmetic of the word-by-word decision without erasing the trade-off.

How a near-zero-cost lookup changes the math

Yomitan is a browser pop-up dictionary extension and the open-source successor to Yomichan. It shows a word's definition and reading when the reader hovers or points at it, without leaving the page.⁹

Since the interruption cost of a lookup is what lowers comprehension and flow, a hover that returns a definition in place reduces that cost toward a brief glance. The "is it worth stopping?" threshold drops, and a reader can justify checking more words than on paper.³⁹

That reasoning is mechanistic, not measured. No peer-reviewed efficacy study is cited here for Yomitan specifically. The durable evidence is that self-initiated lookups aid retention in general, so attach no learning-gain number to the tool itself.¹

When to still infer even with a hover available

Inference reliability is a trainable, coverage-linked ability. Tolerance of ambiguity, the ability to keep reading without resolving every unknown, is also a learner trait that fluent reading relies on. A reader who hovers on every word stops exercising inference. The recommended discipline is guess-first-then-verify: attempt the meaning, then hover to confirm.⁴¹⁰¹¹

The trade-off does not fully vanish under a hover. Even a cheap hover is a micro-interruption that returns a single dictionary sense rather than contextual nuance. Hovering every word pushes the reader toward word-by-word reading, which works against flow.³

Feeding lookups into review

A lookup that clears the "blocks comprehension and/or recurs" bar is exactly the candidate for a spaced-repetition card. The retention benefit of the original lookup is real but partial. Review is what converts a one-time resolution into durable knowledge.¹

The standard workflow is to mine those words into your own deck, turning the sentences you actually read into cards you actually review.

A pre-built deck if you would rather not build one from lookups

Building cards from your own lookups is the default. When you would rather not assemble a deck from scratch, J-Compass recommends Amenokori. Its vocabulary and grammar decks are scheduled with the Free Spaced Repetition Scheduler (FSRS), mapped by level (N5 to N1, the Japanese-Language Proficiency Test scale), and pre-built for the day you want them. This covers the high-frequency core without hand-curation. Keep it secondary to mining your own reading, which produces cards tied to context you have already met.

Yomitan-specific setup and card creation are out of scope here. This article covers the reading-mode decision, not the installation.⁹

Good to know

Tolerance of ambiguity is a skill you can lose

Tolerance of ambiguity, the capacity to keep reading comfortably without resolving every unknown, is identified in the good-language-learner literature as a trait of successful second-language learners.¹¹

It affects how learners use strategies. Learners with higher tolerance of ambiguity use learning strategies more selectively and flexibly, while lower-tolerance learners use them more rigidly. A moderate level is generally favored.¹⁰

Because inference is coverage-linked and improves with practice, reading where every unknown is immediately resolved removes the chances to practice and lets the guessing skill decay. Protect it deliberately by inferring first in extensive mode.⁴¹⁰

Word frequency tells you what is worth the stop

High-frequency unknown words recur across many texts, so a lookup pays off over many future encounters. With meaning information available, the word is more likely to be learned incidentally when you meet it again.¹²

Rare words in casual reading rarely return the investment. Letting frequency decide which stops to make keeps the total interruption budget aimed at words that will pay it back.²

Passive recognition is enough for most reading

The realistic goal for most looked-up words is recognition on the next encounter, not active production. Incidental learning from reading and from lookups builds receptive knowledge, or recognition ability, first.¹

Expecting every looked-up word to become productive vocabulary sets the bar higher than reading alone can deliver. Recognizing the word next time is the win to aim for.²

Kanji you can read but cannot define

A word whose reading is known, or is supplied by furigana, but whose meaning is not, is a meaning gap rather than a reading gap. Because the reading is already secured, the sound form can trigger recognition or aid a context guess. Such a word is often a good inference candidate, and a lookup is optional unless it blocks comprehension.²

This is the word-by-word "does it block comprehension?" test applied to the read-but-not-understood case.

私わたしが意味いみを知しらない言葉ことばがたくさんあります。⁸
"There are many words whose meaning I don't know."

See also

• Intensive vs. Extensive Reading in Japanese
• The i+1 Principle for Reading Japanese
• Yomitan + Anki: One-Click Card Creation
• Sentence Mining: Building Your Own Japanese Anki Deck From What You Read
• How Reading Builds Japanese Ability
• Building a Daily Japanese Reading Habit

References

Footnotes

1. Hulstijn, Jan H., Merel Hollander, and Tine Greidanus. "Incidental Vocabulary Learning by Advanced Foreign Language Students: The Influence of Marginal Glosses, Dictionary Use, and Reoccurrence of Unknown Words." The Modern Language Journal, vol. 80, no. 3, 1996, pp. 327–339. https://onlinelibrary.wiley.com/doi/10.1111/j.1540-4781.1996.tb01614.x ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

2. Nation, I. S. P. Learning Vocabulary in Another Language. 2nd ed. Cambridge University Press, 2013. Chapter 8, "Learning words from context." https://www.cambridge.org/core/books/learning-vocabulary-in-another-language/996F5DE68A6EDA0C25B4D082C4B4289A ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

3. Foroughi, Cyrus K., Nicole E. Werner, Daniela Barragán, and Deborah A. Boehm-Davis. "Interruptions Disrupt Reading Comprehension." Journal of Experimental Psychology: General, vol. 144, no. 3, 2015, pp. 704–709. https://www.apa.org/pubs/journals/features/xge-0000074.pdf (record: https://pubmed.ncbi.nlm.nih.gov/25867225/) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

4. Laufer, Batia. "Lexical Coverages, Inferencing Unknown Words and Reading Comprehension: How Are They Related?" TESOL Quarterly, vol. 54, no. 4, 2020, pp. 1076–1085. https://onlinelibrary.wiley.com/doi/abs/10.1002/tesq.3004 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

5. Day, Richard R., and Julian Bamford. "Top Ten Principles for Teaching Extensive Reading." Reading in a Foreign Language, vol. 14, no. 2, 2002, pp. 136–141. https://nflrc.hawaii.edu/rfl/item/61 ↩ ↩² ↩³ ↩⁴

6. Hu, Marcella, and I. S. P. Nation. "Unknown Vocabulary Density and Reading Comprehension." Reading in a Foreign Language, vol. 13, no. 1, 2000, pp. 403–430. https://files.eric.ed.gov/fulltext/EJ626518 (record: https://eric.ed.gov/?id=EJ626518) ↩ ↩² ↩³ ↩⁴

7. van den Broek, Gesa S. E., Eva Wesseling, Linske Huijssen, Maj Lettink, and Tamara van Gog. "Vocabulary Learning During Reading: Benefits of Contextual Inferences Versus Retrieval Opportunities." Cognitive Science, vol. 46, no. 4, 2022, e13135. https://pmc.ncbi.nlm.nih.gov/articles/PMC9285746/ ↩

8. Tatoeba Project. Open multilingual sentence corpus, Japanese–English pairs cited by numeric sentence ID. https://tatoeba.org ↩ ↩² ↩³

9. Yomitan project. Browser pop-up dictionary extension (open-source successor to Yomichan). Project repository and documentation. https://github.com/yomidevs/yomitan ↩ ↩² ↩³ ↩⁴

10. Ely, Christopher M. "Tolerance of Ambiguity and Use of Second Language Strategies." Foreign Language Annals, vol. 22, no. 5, 1989, pp. 437–445. https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1944-9720.1989.tb02766.x ↩ ↩² ↩³

11. Naiman, Neil, Maria Fröhlich, H. H. Stern, and Angela Todesco. The Good Language Learner. Research in Education Series No. 7. The Ontario Institute for Studies in Education, 1978. (Reissued: Multilingual Matters, 1996.) ↩ ↩²