--- title: "Learning to Paraphrase for Question Answering — Dong 2017" description: "Question answering systems often fail when the same information need is phrased differently, for example \"created\" versus \"started\" for the founder…" authors: ["Li Dong", "Jonathan Mallinson", "Siva Reddy", "Mirella Lapata"] year: 2017 venue: "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing" source: "https://aclanthology.org/D17-1091/" doi: "10.18653/v1/d17-1091" 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/en/papers/dong-2017-learning-paraphrase-question-answering/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/dong-2017-learning-paraphrase-question-answering/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2017](https://md.research.elamur.ai/free/en/papers/2017/page/2/) / Learning to Paraphrase for Question Answering # Learning to Paraphrase for Question Answering **Authors:** [Li Dong](https://md.research.elamur.ai/free/en/authors/li-dong/), [Jonathan Mallinson](https://md.research.elamur.ai/free/en/authors/jonathan-mallinson/), [Siva Reddy](https://md.research.elamur.ai/free/en/authors/siva-reddy/), [Mirella Lapata](https://md.research.elamur.ai/free/en/authors/mirella-lapata/) · **Year:** 2017 · **Venue:** Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/D17-1091/) · [PDF](https://aclanthology.org/D17-1091.pdf) · [DOI 10.18653/v1/d17-1091](https://doi.org/10.18653/v1/d17-1091) · [OpenAlex](https://openalex.org/W2964053384) **Citations:** 191 citations (openalex-snapshot · 2026-10-14) · [94](#cited-by) in our library, of the 140673 papers whose bibliographies we read ## Summary Question answering systems often fail when the same information need is phrased differently, for example "created" versus "started" for the founder relation. Earlier work that rewrote questions usually scored paraphrases with a separate model, which was not tuned to the QA task. Dong et al. propose a framework in which the answer probability is a sum over the original question and its candidate paraphrases. Each term is weighted by a neural scoring model that estimates how likely that rewrite is to produce the right answer. Candidate paraphrases come from three sources: PPDB lexical/phrasal rules, NMT back-translation through German, and templates mined from WikiAnswers. The scorer and the QA model are trained jointly, using only question-answer pairs as supervision. The framework does not depend on a particular generator or QA model. The authors test it on QA over Freebase (GraphQuestions, WebQuestions) and on answer sentence selection (WikiQA). With fairly simple QA models, it gives consistent gains, sets a new best result on GraphQuestions, and is competitive on the other two datasets. ## Key points - The model, PARA4QA, decomposes p(a|q) as a sum over q and its paraphrase set H_q. Each term multiplies a QA model term p_ψ(a|q′) by a learned paraphrase weight p_θ(q′|q). - The three paraphrase generators are: - **PPDB:** bidirectionally entailing rules from PPDB. - **NMT:** EN-DE and DE-EN NMT models, where several German pivots are decoded jointly instead of relying on one translation. - **Rules:** templates mined from WikiAnswers clusters. About 45% of question pairs in these clusters are only related, not true paraphrases. - On WikiQA with a BiLSTM base model: - Without word-matching features, PARA4QA scores 0.6759 MAP / 0.6918 MRR, up from 0.6456 / 0.6608. - With +CNT word-matching features, it scores 0.6978 MAP / 0.7131 MRR, up from 0.6722 / 0.6877. - Ablations show the most useful generator depends on the dataset: - Removing NMT hurts most on GraphQuestions and WikiQA. - Removing the rule-based generator hurts most on WebQuestions, where rule coverage is higher. - Compared with SimpleGraph, average F1 rises from 20.9 to 27.4 on simple questions and from 12.2 to 16.0 on complex ones. ## Interesting ideas - **Learn paraphrase weights from the downstream task.** Treat rewrites as latent inputs and let the end-task loss decide how much each one counts, instead of filtering them with a separate quality model. This could carry over to any input-sensitive task, such as entailment or search over Ukrainian queries. - **Keep the original input in the candidate set.** Because the untouched question competes alongside its rewrites, a bad paraphrase cannot replace a good original. This is a cheap safeguard whenever automatic rewriting is noisy. - **Decode from several pivots at once.** Back-translation that combines multiple pivot translations at each decoding step is less fragile than trusting a single translation. This is useful for paraphrase or data augmentation in lower-resource languages. - **Compare against simpler ways of using paraphrases.** The baselines cover uniform averaging (AVGPARA), plain data augmentation, and a separately trained scorer (SEPPARA). This makes it clear that the joint weighting itself causes the gain, which is a good template for evaluating any augmentation method. - **Analyse what kind of structure gets rewritten.** Manually tagging rewrites by category (question word, question focus, relation verb or NP, constraints/aggregation) shows where paraphrasing helps. This kind of analysis could guide which generators to build for a new domain. *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: [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/) · [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) Topics: [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) ## Cited by Cited by 94 papers of our library cite it (of the 140673 whose bibliographies we read); 6 of them have a page here, 88 not yet. - [Controllable Paraphrase Generation with a Syntactic Exemplar](https://md.research.elamur.ai/free/en/papers/chen-2019-controllable-paraphrase-generation-syntactic-exemplar/) - [Adversarial Example Generation with Syntactically Controlled Paraphrase Networks](https://md.research.elamur.ai/free/en/papers/iyyer-2018-adversarial-example-generation-syntactically-controlled/) - [Paraphrase Generation with Deep Reinforcement Learning](https://md.research.elamur.ai/free/en/papers/li-2018-paraphrase-generation-deep-reinforcement-learning/) - [Style Transfer Through Back-Translation](https://md.research.elamur.ai/free/en/papers/prabhumoye-2018-style-transfer-through-back-translation/) - [Paraphrase Generation as Zero-Shot Multilingual Translation: Disentangling Semantic Similarity from Lexical and Syntactic Diversity](https://md.research.elamur.ai/free/en/papers/thompson-2020-paraphrase-generation-as-zero-shot/) - [QANet: Combining Local Convolution with Global Self-Attention for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/yu-2018-qanet-combining-local-convolution-global/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{dong2017learning, title = {{Learning to Paraphrase for Question Answering}}, author = {Dong, Li and Mallinson, Jonathan and Reddy, Siva and Lapata, Mirella}, booktitle = {Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing}, year = {2017}, doi = {10.18653/v1/d17-1091}, url = {https://doi.org/10.18653/v1/d17-1091} } ``` - [← previous Learning to Generate Reviews and Discovering Sentiment](https://md.research.elamur.ai/free/en/papers/radford-2017-learning-generate-reviews-discovering-sentiment/) - [next → Learning to Parse and Translate Improves Neural Machine Translation](https://md.research.elamur.ai/free/en/papers/eriguchi-2017-learning-parse-translate-improves-neural/) - [↑ back to papers of 2017](https://md.research.elamur.ai/free/en/papers/2017/page/2/) ## References 2 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Table 7: We group G RAPH Q UESTIONS into simple and complex questions and report model performance in each split. Best results in each group are shown in bold. The values in brackets are absolute improvements of average F1 scores.* Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Table+7%3A+We+group+G+RAPH+Q+UESTIONS+into+simple+and+complex+questions+and+report+model+performance+in+each+split%2E+Best+results+in+each+group+are+shown+in+bold%2E+The+values+in+brackets+are+absolute+impr) [our card](https://md.research.elamur.ai/free/en/papers/bahdanau-2014-neural-machine-translation-jointly-learning/) 2. Colin Bannard and Chris Callison-Burch. 2005. Paraphrasing with bilingual parallel corpora. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL’05), pages 597–604, Ann Arbor, Michigan. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Colin+Bannard+and+Chris+Callison%2DBurch%2E+2005%2E+Paraphrasing+with+bilingual+parallel+corpora%2E+In+Proceedings+of+the+43rd+Annual+Meeting+of+the+Association+for+Computational+Linguistics+%28ACL%E2%80%9905%29%2C+pages+5) [our card](https://md.research.elamur.ai/free/en/papers/bannard-2005-paraphrasing-bilingual-parallel-corpora/)