--- title: "Breaking NLP — Mahler 2017" description: "The authors, a team from The Ohio State University, took part as \"breakers\" in the \"Build It Break It\" sentiment analysis competition held at EMNLP in…" authors: ["Taylor Mahler", "Willy Cheung", "Micha Elsner", "David King", "Marie-Catherine de Marneffe", "Cory Shain", "Symon Stevens-Guille", "Michael White"] year: 2017 venue: "Proceedings of the First Workshop on Building Linguistically Generalizable NLP Systems" source: "https://aclanthology.org/W17-5405/" doi: "10.18653/v1/w17-5405" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/papers/mahler-2017-breaking-nlp-morphosyntax-semantics-pragmatics/" --- [Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/mahler-2017-breaking-nlp-morphosyntax-semantics-pragmatics/) [home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2017](https://md.research.elamur.ai/free/papers/2017/) / Breaking NLP: Using Morphosyntax, Semantics, Pragmatics and World Knowledge to Fool Sentiment Analysis Systems # Breaking NLP: Using Morphosyntax, Semantics, Pragmatics and World Knowledge to Fool Sentiment Analysis Systems **Authors:** Taylor Mahler, Willy Cheung, [Micha Elsner](https://md.research.elamur.ai/free/authors/micha-elsner/), David King, [Marie-Catherine de Marneffe](https://md.research.elamur.ai/free/authors/marie-catherine-de-marneffe/), Cory Shain, Symon Stevens-Guille, [Michael White](https://md.research.elamur.ai/free/authors/michael-white/) · **Year:** 2017 · **Venue:** Proceedings of the First Workshop on Building Linguistically Generalizable NLP Systems · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/W17-5405/) · [PDF](https://aclanthology.org/W17-5405.pdf) · [DOI 10.18653/v1/w17-5405](https://doi.org/10.18653/v1/w17-5405) · [OpenAlex](https://openalex.org/W2793353517) **Citations:** 19 citations (openalex-snapshot · 2026-10-14) · [15](#cited-by) in our library, of the 140673 papers whose bibliographies we read ## Summary The authors, a team from The Ohio State University, took part as "breakers" in the "Build It Break It" sentiment analysis competition held at EMNLP in 2017. Their goal was to probe how well "builder" sentiment classifiers handle linguistic structure. They took sentences from the blind test set of movie-review snippets and made small edits to each one, producing minimal pairs where a system should get exactly one member wrong. The edits followed four linguistically motivated strategies: - morphosyntactic changes, such as negation, comparatives and prefixes like *un-* or *anti-*; - semantic changes, such as *too*/*enough*, modals, embedding under operators, antonyms and swapping the film's genre; - pragmatic tricks, such as scare quotes, Gricean implicatures and reordering clauses; - edits whose sentiment depends on world knowledge. The submitted pairs fooled the builder systems most of the time. Semantic and pragmatic edits caused the most trouble. This suggests that even modern sentiment models stay brittle when faced with tiny, adversarial-style edits. The authors argue that such curated example sets can guide the development of more robust systems. ## Key points - The team submitted 99 minimal pairs, which broke systems at an average rate of 72.6%. Their weighted F1 of 28.67 placed them second among four breaker teams. - 72 of the 99 pairs broke more than half of the six builder systems. All 12 same-sentiment pairs were among the less effective cases. - Putting a single praise word in scare quotes (e.g., "terrific") broke every builder system in both pairs that used this trick. - Lexical-semantic and world-knowledge edits proved less challenging than expected. Failed pairs often kept negative keywords that gave the answer away. - The authors ran no significance tests and note several limitations. The source data had only 521 examples, which made rare constructions hard to find. They also could not empirically verify how natural their edited sentences were. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/shelves/linguistics/) Topics: [Implicature](https://md.research.elamur.ai/free/topics/implicature/) ## Cited by Cited by 15 papers of our library cite it (of the 140673 whose bibliographies we read); 6 of them have a page here, 9 not yet. - [ACES: Translation Accuracy Challenge Sets for Evaluating Machine Translation Metrics](https://md.research.elamur.ai/free/papers/amrhein-2022-aces-translation-accuracy-challenge-sets/) - [Multilingual Irony Detection with Dependency Syntax and Neural Models](https://md.research.elamur.ai/free/papers/cignarella-2020-multilingual-irony-detection-dependency-syntax/) - [Probing What Different NLP Tasks Teach Machines about Function Word Comprehension](https://md.research.elamur.ai/free/papers/kim-2019-probing-what-different-nlp-tasks/) - [The Gricean Maxims in NLP — A Survey](https://md.research.elamur.ai/free/papers/krause-2024-gricean-maxims-nlp-survey/) - [Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets](https://md.research.elamur.ai/free/papers/moghe-2025-machine-translation-meta-evaluation-through/) - [Evaluating Universal Dependency Parser Recovery of Predicate Argument Structure via CompChain Analysis](https://md.research.elamur.ai/free/papers/ud-2021-indurkhya-evaluating-universal-dependency-parser-recovery/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{mahler2017breaking, title = {{Breaking NLP: Using Morphosyntax, Semantics, Pragmatics and World Knowledge to Fool Sentiment Analysis Systems}}, author = {Mahler, Taylor and Cheung, Willy and Elsner, Micha and King, David and Marneffe, Marie-Catherine de and Shain, Cory and Stevens-Guille, Symon and White, Michael}, booktitle = {Proceedings of the First Workshop on Building Linguistically Generalizable NLP Systems}, year = {2017}, doi = {10.18653/v1/w17-5405}, url = {https://doi.org/10.18653/v1/w17-5405} } ``` - [← previous Billion-scale similarity search with GPUs](https://md.research.elamur.ai/free/papers/johnson-2017-billion-scale-similarity-search-gpus/) - [next → Challenges in Data-to-Document Generation](https://md.research.elamur.ai/free/papers/wiseman-2017-challenges-data-document-generation/) - [↑ back to papers of 2017](https://md.research.elamur.ai/free/papers/2017/) ## References 7 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Rajesh Bhatt and Shoichi Takahashi. 2011. Reduced and unreduced phrasal comparatives. Natural Language & Linguistic Theory 29(3):581–620. [find](https://scholar.google.com/scholar?q=Rajesh+Bhatt+and+Shoichi+Takahashi%2E+2011%2E+Reduced+and+unreduced+phrasal+comparatives%2E+Natural+Language+%26+Linguistic+Theory+29%283%29%3A581%E2%80%93620%2E) 2. Hope C. Dawson and Michael Phelan, editors. 2016. Language Files. The Ohio State University Press, 12th edition. [find](https://scholar.google.com/scholar?q=Hope+C%2E+Dawson+and+Michael+Phelan%2C+editors%2E+2016%2E+Language+Files%2E+The+Ohio+State+University+Press%2C+12th+edition%2E) 3. H. Paul Grice. 1975. Logic and conversation. In P Cole and J Morgan, editors, Syntax and Semantics 3: Speech Acts, Academic Press, New York, pages 64–75. [find](https://scholar.google.com/scholar?q=H%2E+Paul+Grice%2E+1975%2E+Logic+and+conversation%2E+In+P+Cole+and+J+Morgan%2C+editors%2C+Syntax+and+Semantics+3%3A+Speech+Acts%2C+Academic+Press%2C+New+York%2C+pages+64%E2%80%9375%2E) 4. Christopher Kennedy. 2002. Comparative deletion and optimality in syntax. Natural Language & Linguistic Theory 20(3):553–621. [find](https://scholar.google.com/scholar?q=Christopher+Kennedy%2E+2002%2E+Comparative+deletion+and+optimality+in+syntax%2E+Natural+Language+%26+Linguistic+Theory+20%283%29%3A553%E2%80%93621%2E) 5. Jonathan K Kummerfeld, David Hall, James R Curran, and Dan Klein. 2012. Parser showdown at the wall street corral: An empirical investigation of error types in parser output. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Lan… [find](https://scholar.google.com/scholar?q=Jonathan+K+Kummerfeld%2C+David+Hall%2C+James+R+Curran%2C+and+Dan+Klein%2E+2012%2E+Parser+showdown+at+the+wall+street+corral%3A+An+empirical+investigation+of+error+types+in+parser+output%2E+In+Proceedings+of+the+201) [our card](https://md.research.elamur.ai/free/papers/aclbulk-2012-kummerfeld-parser-showdown-wall-street-corral/) 6. Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2014. Intriguing properties of neural networks. ICLR https://doi.org/abs/1312.6199. [↗](https://doi.org/abs/1312.6199) [our card](https://md.research.elamur.ai/free/papers/szegedy-2013-intriguing-properties-neural-networks/) 7. Byron C. Wallace, Do Kook Choe, Laura Kertz, and Eugene Charniak. 2014. Humans require context to infer ironic intent (so computers probably do, too). In ACL (2). pages 512–516. [find](https://scholar.google.com/scholar?q=Byron+C%2E+Wallace%2C+Do+Kook+Choe%2C+Laura+Kertz%2C+and+Eugene+Charniak%2E+2014%2E+Humans+require+context+to+infer+ironic+intent+%28so+computers+probably+do%2C+too%29%2E+In+ACL+%282%29%2E+pages+512%E2%80%93516%2E)