Crash Blossom

A crash blossom is a newspaper headline that is unintentionally ambiguous because its compressed wording allows more than one valid parse, producing an unintended (and often comical) alternative reading.^[4] The canonical example, which gave the term its name, is "Violinist Linked to JAL Crash Blossoms," a 2009 headline in which "blossoms" was meant as a verb but reads as if "crash blossoms" were a single mysterious noun.^[1][2] In computational linguistics, crash blossoms are a vivid case of syntactic ambiguity: they arise because headlines drop function words such as articles, auxiliary verbs, and the copula to save space, leaving constructions where the same word can be read as either a noun or a verb. They sit alongside garden-path sentences and the telegraphic register known as headlinese as classic challenges for natural language processing (NLP) parsers.^[5]

What is a crash blossom?

A crash blossom is an ambiguously worded headline that can be parsed in more than one way. The ambiguity is structural rather than purely lexical: even when every individual word is understood, the omitted grammatical cues leave the reader (or a machine parser) unable to decide which word is the subject, which is the verb, and how the phrases attach. Merriam-Webster, which added the term to its "Words We're Watching" list, defines a crash blossom as "a headline that is ambiguous because of its wording or punctuation (or absence thereof)."^[3] Wiktionary defines it as "a sentence, often a news headline, that is subject to incorrect interpretation due to syntactic and/or lexical ambiguity."^[4]

Because the misreading usually resolves into something absurd (a violinist linked to flowering crashes, a squad that helps a dog bite someone), crash blossoms are simultaneously a copy-editing failure, a source of newsroom humor, and a compact illustration of why parsing natural language is hard.

Where does the term "crash blossom" come from?

The word "crash blossom" was coined in August 2009 on the Testy Copy Editors online forum. Mike O'Connell, an American editor based in Sapporo, Japan, encountered the headline "Violinist Linked to JAL Crash Blossoms" in the newspaper Japan Today and posted asking, in effect, "What's a crash blossom?"^[1][2] The headline referred to Diana Yukawa, a Japanese-British violinist and composer whose father, Akihisa Yukawa, was killed in the crash of Japan Air Lines Flight 123 on August 12, 1985. Yukawa was born about a month after the disaster, and the article described how her career had "blossomed" despite the tragedy. "Blossoms" was intended as a verb, but O'Connell parsed "crash blossoms" as a compound noun, producing the confusion.^[1]

Another forum member, Dan Bloom, suggested that the phrase would make an excellent label for this whole class of ambiguous headlines, and the name spread rapidly through the linguistics blogosphere. The linguist and lexicographer Ben Zimmer wrote about it on the Language Log blog on August 26, 2009.^[1] Zimmer then devoted his "On Language" column in The New York Times Magazine to the phenomenon (dated January 31, 2010), which popularized the term beyond copy desks.^[2] In that column Zimmer explained the underlying mechanism directly: headlines, he wrote, "sweep away those little words, particularly articles, auxiliary verbs and forms of 'to be,' robbing the reader of crucial context."^[2]

How do crash blossoms work?

Crash blossoms exploit the compressed register known as headlinese. Standard English headlines follow conventions that differ from ordinary prose: articles ("a," "the") are dropped, the copula ("is," "are") is omitted, and verbs are often written in the simple present tense regardless of when the event occurred. These compressions save column space but remove the syntactic cues that readers rely on to parse sentence structure correctly.

The most common mechanism behind crash blossoms is noun-verb ambiguity. Many English words function as both nouns and verbs ("crash," "fire," "hit," "race," "drop," "hold"). When such a word appears in a headline without the function words that would clarify its grammatical role, readers may assign the wrong part of speech, leading to an unintended interpretation.^[4] As Zimmer put it in his column, English has "countless nouns that can also be verbs, and vice versa," and a headline strips away the surrounding words that would normally settle which is which.^[2]

Other contributing factors include:

• Prepositional phrase (PP) attachment ambiguity: A prepositional phrase can modify either the noun or the verb in a headline, producing different meanings.
• Noun pile-ups: Headlines sometimes stack multiple nouns together without hyphens or other punctuation, making it hard to determine which nouns modify which.
• Missing hyphens: A missing hyphen between a compound modifier can split the modifier into separate words, changing the parse. For example, "dog bite victim" without a hyphen between "dog" and "bite" can be misread as a dog that bites a victim.

Famous examples

Ambiguous headlines have existed for as long as newspapers themselves, but they were not called "crash blossoms" until 2009. The Columbia Journalism Review collected such headlines for decades in its column "The Lower Case" and published two anthologies: Squad Helps Dog Bite Victim (1980) and Red Tape Holds Up New Bridge (1987), edited by Gloria Cooper.^[6]

The table below lists well-known crash blossoms along with their intended and unintended readings.

Scroll sideways for more →

How is a crash blossom related to syntactic ambiguity?

Syntactic ambiguity (also called structural ambiguity, amphiboly, or amphibology) is the broader linguistic phenomenon that underlies crash blossoms. A sentence is syntactically ambiguous when its grammatical structure allows more than one valid parse tree. Unlike lexical ambiguity, which arises from a single word having multiple meanings (e.g., "bank" as a financial institution or a river bank), syntactic ambiguity comes from the relationships among words and clauses.^[5]

Syntactic ambiguity can be divided into two categories:

• Global ambiguity: The sentence remains ambiguous even after the reader reaches the end. "The woman held the baby in the green blanket" retains multiple readings (the baby is in the blanket, or the woman is wearing the blanket while holding the baby).
• Local ambiguity: The sentence is temporarily ambiguous but resolves to a single interpretation by the end. These often produce garden-path effects.

Crash blossoms are typically instances of global ambiguity, because headline writers do not provide enough additional context to resolve the ambiguity.

How are crash blossoms related to garden-path sentences?

Garden-path sentences are closely related to crash blossoms. A garden-path sentence leads the reader toward an initial interpretation that turns out to be incorrect, forcing them to reanalyze the sentence's structure. The classic example is:

"The horse raced past the barn fell."

Most readers initially parse "raced" as the main verb ("The horse raced past the barn"), but the sentence actually means "The horse [that was] raced past the barn fell." Here, "raced" is a reduced relative clause, not the main verb.

The Garden-Path Model of sentence processing, proposed by Lyn Frazier and others, holds that the parser initially builds the simplest possible syntactic structure based on two principles:^[11]

1. Minimal Attachment: The parser prefers the structure that requires the fewest syntactic nodes.
2. Late Closure: New words are attached to the phrase currently being processed rather than starting a new phrase.

When these heuristics lead to an incorrect parse, the reader must perform reanalysis, backtracking to find the correct structure. This reanalysis is cognitively costly, which is why garden-path sentences feel confusing.

Crash blossoms function like garden-path sentences in miniature. The compressed syntax of headlinese removes many of the cues that readers would normally use to parse a sentence correctly on the first pass.

Types of ambiguity in NLP

Crash blossoms sit at the intersection of several types of linguistic ambiguity that NLP systems must handle. The table below summarizes these types.

Scroll sideways for more →

Why are crash blossoms hard for NLP parsers?

Crash blossoms present a concentrated form of the ambiguity problems that NLP systems face more broadly. Parsing an ambiguous headline requires a system to handle part-of-speech tagging, syntactic parsing, semantic interpretation, and world knowledge simultaneously, and the telegraphic style of headlinese removes exactly the cues that modern parsers were trained to rely on.

Part-of-speech tagging

Part-of-speech (POS) tagging is the task of assigning grammatical categories (noun, verb, adjective, etc.) to each word in a sentence. On well-formed prose, this is largely a solved problem: state-of-the-art taggers "routinely achieve token-level accuracies of over 97% on news body text."^[16] However, the same models underperform on headlines, because, as Benton, Li, and Malioutov note, "the register of English news headlines, 'headlinese', is very different from the register of long-form text, causing POS tagging models to underperform on headlines."^[16]

The reason is precisely the structure that creates crash blossoms. In the headline "Eye Drops Off Shelf," a POS tagger must decide whether "drops" is a noun (eye drops, the product) or a verb (something drops off a shelf), and whether "eye" is an adjective modifying "drops" or the subject of the verb "drops." Without articles or auxiliary verbs, these decisions become much harder. Research on cross-register POS tagging has shown that models trained primarily on well-formed sentences (such as newswire text from the Penn Treebank) can struggle when applied to headlines, tweets, and other non-standard text.^[7][16]

Syntactic parsing

Syntactic parsers construct parse trees that represent the grammatical structure of a sentence. When a sentence is ambiguous, a parser may produce multiple candidate trees. Traditional approaches use probabilistic context-free grammars (PCFGs) and algorithms such as the Cocke-Kasami-Younger (CKY) algorithm to find the most probable parse. More recent neural parsers use deep learning architectures to score candidate structures.

Headlines have nonetheless received "surprisingly little attention from the NLP syntactic parsing community," and Benton and colleagues released the first news-headline corpus annotated with Universal Dependencies syntactic trees specifically to measure how well state-of-the-art dependency parsers handle them.^[17] They found that models trained only on gold-annotated long-form text parse headlines poorly, and that bootstrapping additional training data from headline-and-lead-sentence pairs improved accuracy, though the gains were moderated by constructions idiosyncratic to each news outlet.^[17]

The prepositional phrase (PP) attachment problem is one of the most studied forms of syntactic ambiguity. In the sentence "I saw the man with the telescope," the PP "with the telescope" can attach to the verb phrase (I used the telescope to see) or the noun phrase (the man has the telescope). The Penn Treebank annotates PP attachment ambiguities using a system called "pseudo-attachment," where annotators can flag globally ambiguous structures even when context does not resolve them.^[7]

Word sense disambiguation

Word sense disambiguation (WSD) is the task of determining which meaning of a polysemous word is intended in a given context.^[12] WSD approaches fall into several categories:

• Supervised methods: Treat WSD as a classification problem, training models on datasets where word instances have been manually annotated with correct senses. Common algorithms include support vector machines, decision trees, and neural networks.
• Unsupervised methods: Rely on distributional patterns in large text corpora, clustering word instances by context similarity without labeled training data.
• Knowledge-based methods: Exploit structured lexical resources such as WordNet or BabelNet, using graph-based algorithms to select the most appropriate sense.
• Contextualized embeddings: Models such as ELMo, BERT, and GPT generate different vector representations for the same word depending on context, providing an implicit form of sense disambiguation.^[9][8]

In the context of crash blossoms, WSD is necessary to resolve lexical ambiguity (e.g., "holds up" meaning "delays" vs. "supports"), but it is not sufficient on its own; structural disambiguation is also required.

How do modern NLP models handle ambiguity?

Contextualized word representations

Before 2018, most NLP systems used static word embeddings such as Word2Vec or GloVe, which assign a single vector to each word regardless of context. These representations cannot distinguish between different senses of a polysemous word. Contextualized models changed this.

ELMo (Embeddings from Language Models), introduced by Peters et al. in 2018, generates word representations that vary depending on the surrounding sentence, using a bidirectional LSTM architecture.^[9] BERT (Bidirectional Encoder Representations from Transformers), introduced by Devlin et al. later that same year, uses the self-attention mechanism to process all words in a sentence simultaneously, allowing each word's representation to be influenced by every other word.^[8] These approaches improved performance on tasks that require disambiguating word meaning in context.

Transformer attention and syntactic structure

The self-attention mechanism in transformer models allows each token to attend to every other token in the input. Research has shown that different attention heads in BERT specialize in capturing different types of linguistic information. Some heads track syntactic dependencies (such as subject-verb agreement), while others capture positional relationships or semantic similarity. This multi-faceted attention can help models resolve syntactic ambiguities by weighing structural and semantic evidence simultaneously.

Transformer Grammars, introduced by Sartran et al. in 2022, combine the scalability of transformers with recursive syntactic compositions, using a special attention mask and deterministic tree transformation to build syntactic structure explicitly into the model.^[13]

Performance of large language models on ambiguity

Recent research has tested how large language models (LLMs) handle various forms of ambiguity:

• Garden-path sentences: A 2024 study (arXiv 2405.16042) examined whether GPT-2, LLaMA-2, Flan-T5, and RoBERTa process garden-path sentences similarly to humans. The researchers found that these models exhibit some of the same initial misinterpretation patterns as human readers, but their reanalysis behavior differs.^[14]
• Scope ambiguity: Research published in Transactions of the Association for Computational Linguistics (2024) found that GPT-4 can recognize scope ambiguity but struggles to formalize it correctly, often misapplying quantifier scopes in generated lambda calculus representations.^[15]
• Structural ambiguity across languages: A study on GPT's multilingual capabilities found that the model tends to rely on English-centric attachment preferences when parsing syntactically ambiguous constructions in Korean and Japanese, rather than capturing language-specific patterns.

These findings indicate that while LLMs have made significant progress in handling ambiguity, they are not yet fully reliable, particularly on the kinds of compressed, context-poor constructions found in crash blossoms and headlinese.^[10]

Crash blossoms in computational humor and creativity

The unintended humor in crash blossoms has attracted interest from researchers working on computational humor and natural language generation. Detecting whether a headline is a crash blossom requires a system to identify that multiple parses exist and that one of them produces a semantically incongruous reading. This overlaps with work on automatic humor detection, where models must identify incongruity, surprise, and semantic contrast.

From a natural language generation perspective, avoiding crash blossoms in automatically generated headlines is a practical concern. News headline generation systems, which use sequence-to-sequence models or transformer-based architectures, must produce headlines that are both concise and unambiguous. Evaluating generated headlines for potential crash blossom readings is an open problem in NLG research.

Cognitive science perspective

Crash blossoms also provide a window into human sentence processing. Psycholinguistic research uses ambiguous sentences, including garden-path constructions, to study how the brain parses language in real time.

Several models of human sentence processing are relevant:

• Reanalysis Model: Processing difficulty arises when readers realize their initial parse was wrong and must restructure the sentence. Eye-tracking studies confirm that readers experience increased fixation durations at the point where disambiguation occurs.
• Competition-Based Model: Multiple syntactic analyses compete for activation. When two analyses receive roughly equal support, processing becomes more difficult.
• Unrestricted Race Model: Multiple possible structures "race" to completion, and the one with the strongest support from available information wins. This model incorporates the "good-enough" account of language comprehension, which proposes that readers often settle for incomplete or approximate interpretations rather than fully resolving every ambiguity.

Research has also found differences between how children and adults process syntactic ambiguity. Children are slower to commit to an initial syntactic interpretation and have more difficulty with reanalysis, partly because of limited working memory capacity. Adults with higher working memory spans can maintain multiple interpretations simultaneously but may actually experience greater interference from competing analyses.

Historical context

Ambiguous writing has been a source of both confusion and literary effect for centuries. Two notable historical examples predate the modern concept of crash blossoms:

• Shakespeare's Henry VI: The line "The duke yet lives that Henry shall depose" is deliberately ambiguous. It could mean that Henry will depose the duke, or that the duke will depose Henry. Shakespeare uses this ambiguity to create dramatic irony.
• Marlowe's Edward II: The Latin sentence "Eduardum occidere nolite timere bonum est" changes meaning depending on where a comma is placed. Read one way, it means "Do not be afraid to kill Edward; it is good." Read another way, it means "Do not kill Edward; it is good to be afraid."

These examples demonstrate that syntactic ambiguity is not merely a modern newspaper problem but a longstanding feature of natural language that has been exploited for literary and political purposes.

How can editors avoid crash blossoms?

Editors and style guides recommend several strategies to prevent crash blossoms:

1. Use hyphens for compound modifiers: Writing "dog-bite victim" instead of "dog bite victim" eliminates the ambiguity in "Squad Helps Dog Bite Victim."
2. Choose unambiguous verbs: Replacing "holds up" with "delays" in "Red Tape Holds Up New Bridge" removes the double meaning.
3. Add function words where space permits: Including an article ("a," "the") or auxiliary verb ("is," "are") can clarify which word is the verb and which is the noun.
4. Read the headline in isolation: Editors should test whether a headline makes sense without the accompanying article, since many readers scan headlines without reading further.
5. Check for noun-verb ambiguity: Any word in a headline that could function as either a noun or a verb should be flagged and reviewed.

Explain like I'm 5 (ELI5)

Imagine you read a short sentence like "Kids Make Nutritious Snacks." You could think it means kids are cooking healthy snacks in the kitchen. But you could also read it as if the kids themselves are the snacks, which is silly and kind of funny. A "crash blossom" is the name for a headline like this one that accidentally says two things at once because the words are squished together without enough helper words to make the meaning clear. Computers that try to read and understand language have the same problem: when words are missing from a sentence, it is harder to figure out what the sentence is really saying.

See also

• Natural language processing
• Syntactic ambiguity
• Word sense disambiguation
• Part-of-speech tagging
• Garden-path sentence
• BERT
• Transformer
• Parsing

References

1. Zimmer, Ben. "Crash Blossoms." *Language Log*, August 26, 2009. https://languagelog.ldc.upenn.edu/nll/?p=1693 ↩
2. Zimmer, Ben. "On Language: Crash Blossoms." *The New York Times Magazine*, January 31, 2010. ↩
3. "Crash Blossom: Words We're Watching." *Merriam-Webster*. https://www.merriam-webster.com/wordplay/crash-blossom-words-were-watching ↩
4. "Crash blossom." *Wiktionary*. https://en.wiktionary.org/wiki/crash_blossom ↩
5. "Syntactic ambiguity." *Wikipedia*. https://en.wikipedia.org/wiki/Syntactic_ambiguity ↩
6. Cooper, Gloria, ed. *Squad Helps Dog Bite Victim, and Other Flubs from the Nation's Press*. Columbia Journalism Review, 1980. ↩
7. Marcus, Mitchell P., Beatrice Santorini, and Mary Ann Marcinkiewicz. "Building a Large Annotated Corpus of English: The Penn Treebank." *Computational Linguistics* 19, no. 2 (1993): 313-330. ↩
8. Devlin, Jacob, et al. "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding." *Proceedings of NAACL-HLT*, 2019. ↩
9. Peters, Matthew E., et al. "Deep Contextualized Word Representations." *Proceedings of NAACL-HLT*, 2018. ↩
10. Bender, Emily M., and Alexander Koller. "Climbing Towards NLU: On Meaning, Form, and Understanding in the Age of Data." *Proceedings of ACL*, 2020. ↩
11. Frazier, Lyn, and Janet Dean Fodor. "The Sausage Machine: A New Two-Stage Parsing Model." *Cognition* 6, no. 4 (1978): 291-325. ↩
12. Navigli, Roberto. "Word Sense Disambiguation: A Survey." *ACM Computing Surveys* 41, no. 2 (2009): 1-69. ↩
13. Sartran, Laurent, et al. "Transformer Grammars: Augmenting Transformer Language Models with Syntactic Inductive Biases at Scale." *Transactions of the Association for Computational Linguistics* 10 (2022): 1423-1439. ↩
14. Nour, Youssef, et al. "Incremental Comprehension of Garden-Path Sentences by Large Language Models." *arXiv preprint* 2405.16042, 2024. ↩
15. Mandelkern, Matthew, et al. "Scope Ambiguities in Large Language Models." *Transactions of the Association for Computational Linguistics* 12 (2024). ↩
16. Benton, Adrian, Hanyang Li, and Igor Malioutov. "Cross-Register Projection for Headline Part of Speech Tagging." *Proceedings of EMNLP*, 2021. https://aclanthology.org/2021.emnlp-main.521/ ↩
17. Benton, Adrian, Tianze Shi, Ozan Irsoy, and Igor Malioutov. "Weakly Supervised Headline Dependency Parsing." *Findings of EMNLP*, 2022. https://arxiv.org/abs/2301.10371 ↩