What Do You Do With a Word That Appears Once? Guess, Carefully

The meaning of a word is normally worked out by comparing the contexts it appears in, since a word used fifty times across varied passages reveals what it can and cannot be applied to. A word appearing once supplies a single context and nothing to compare it with, so the evidence available is one sentence. If that sentence is itself obscure, damaged or poetic, the word may be effectively undecipherable. The problem is acute for ancient texts, where the surviving body of writing in a language may be small and where no native speaker exists to ask.

Several lines of evidence are combined, none decisive on its own:

• •The immediate context, and what the sentence requires the word to mean
• •Parallel passages elsewhere saying something similar in other words
• •Related words in the same language sharing a root
• •Cognates in related languages, where the family is known
• •Early translations, which record how somebody once understood it
• •Later commentary, which may preserve a lost tradition

The Hebrew Bible contains a large number and they cause real difficulty, because the surviving corpus of ancient Hebrew is small and consists largely of that one body of text. A word appearing once in it may appear nowhere else in the language at all, so there is no comparison available within Hebrew and scholars turn to Akkadian, Ugaritic, Arabic and Aramaic for related forms. The resulting uncertainty is visible in any comparison of translations, where a verse may be rendered quite differently in different versions, and a footnote saying the meaning of the Hebrew is uncertain frequently marks exactly this situation.

Several such words have become well known in their own right because so much has been written about them. Shakespeare's works contain a number of words appearing nowhere else in his writing and, in some cases, nowhere else in English before him, which supports the long argument about how much of the language he coined and how much simply survives only in his texts. Homer contains words whose meaning was already uncertain to ancient Greek commentators, who preserved competing explanations that scholars still weigh. And the practice of counting such words across an author's work is used in authorship studies, since the rate at which a writer uses words they use nowhere else is a measurable stylistic fingerprint.

The same problem has a modern form in language technology and it is a persistent one. A system trained on text learns about a word from its occurrences, so words appearing once supply almost nothing and words appearing zero times supply nothing at all, which is the out of vocabulary problem. Since word frequencies follow a distribution in which a large proportion of distinct words occur only once in any corpus, that is a substantial share of vocabulary rather than an edge case. Modern systems address it by breaking words into smaller pieces, so that an unseen word is handled as a sequence of familiar fragments, which handles morphology well and names badly.

A word occurring once gives one context and nothing to compare it against, so its meaning must be inferred from what the sentence requires, from related words, from cognate languages and from early translations. Ancient Hebrew has many, because the surviving corpus is essentially one book, which is why translations of certain verses differ sharply. Language systems face the same problem and split unseen words into fragments.