--- title: "Learning Structured Text Representations — Liu 2018" description: "The authors propose a hierarchical model. 3 references, cited by 56 papers here. Among them: 145." authors: ["Yang Liu", "Mirella Lapata"] year: 2018 venue: "Transactions of the Association for Computational Linguistics" source: "https://aclanthology.org/Q18-1005/" doi: "10.1162/tacl_a_00005" 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/liu-2018-learning-structured-text-representations/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/liu-2018-learning-structured-text-representations/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2018](https://md.research.elamur.ai/free/en/papers/2018/page/3/) / Learning Structured Text Representations # Learning Structured Text Representations **Authors:** [Yang Liu](https://md.research.elamur.ai/free/en/authors/yang-liu/) and [Mirella Lapata](https://md.research.elamur.ai/free/en/authors/mirella-lapata/) · **Year:** 2018 · **Venue:** [Transactions of the Association for Computational Linguistics](https://md.research.elamur.ai/free/en/venues/tacl/) ([Transactions of the Association for Computational Linguistics (TACL) 2018](https://md.research.elamur.ai/free/en/venues/tacl/2018/)) · **Publisher:** [The MIT Press](https://md.research.elamur.ai/free/en/publishers/the-mit-press/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/Q18-1005/) · [PDF](https://aclanthology.org/Q18-1005.pdf) · [DOI 10.1162/tacl\_a\_00005](https://doi.org/10.1162/tacl_a_00005) · [OpenAlex](https://openalex.org/W2963899396) **Citations:** 145 citations (OpenAlex, 2026-10-14) · [56](https://md.research.elamur.ai/free/en/papers/liu-2018-learning-structured-text-representations/cited-by/) in our library, of the 141,840 papers whose bibliographies we read ## Summary Liu and Lapata look for a way to make neural document encoders aware of discourse structure without annotated discourse corpora or an external parser, whose mistakes spread through pipeline systems. Earlier structured attention (Kim et al., 2017) used the inside-outside algorithm. That approach covers only projective trees and parallelizes poorly. The authors propose a hierarchical model. Bidirectional LSTMs first encode words into sentence vectors, then sentences into a document vector. Each LSTM output is split into a semantic part and a structure part. The structure part drives an attention layer whose weights are the marginal probabilities of non-projective dependency trees, computed with a variant of Kirchhoff's Matrix-Tree Theorem. Because every step is differentiable, the latent trees are learned end-to-end for the target task. The model was tested on SNLI and on document datasets (Yelp, IMDB, CZ Movies, Debates). The authors report state-of-the-art document modeling results, though the exact scores fall in the omitted part of the provided text. The induced trees are readable. The work matters because it offers discourse-aware representations without costly annotation. ## Key points - The core mechanism is structured attention based on the Matrix-Tree Theorem (Tutte, 1984). It allows crossing (non-projective) edges and runs efficiently on GPUs, despite O(n³) worst-case complexity. - By comparison, adding an inside-outside pass made training about 10 times slower in the authors' experiments. - On the SNLI test set, induced sentence trees were shallower than those of the Stanford parser, which was trained on the Penn Treebank: average height 5.78 versus 8.99. About half were projective without any explicit constraint, and many edges matched the parser's output. - Document-level trees were extracted with the Chu-Liu-Edmonds algorithm. The share of projective trees was 79.6% on Yelp, 74.9% on IMDB, 82.8% on CZ Movies and 62.4% on Debates. Most document trees reach only depth 3. - In qualitative examples, the induced links resemble "elaboration" relations, connect sentences about the same topic, and group sentences that share sentiment. One example is a Czech Movies review. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions - [University of Edinburgh](https://md.research.elamur.ai/free/en/universities/university-of-edinburgh/) — University, Edinburgh, United Kingdom · 255 works · 68,771 citations 📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 1,282 works ## Cite Cite as: Yang Liu and Mirella Lapata. 2018. Learning Structured Text Representations. Transactions of the Association for Computational Linguistics. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @article{liu2018learning, title = {{Learning Structured Text Representations}}, author = {Yang Liu and Mirella Lapata}, journal = {Transactions of the Association for Computational Linguistics}, year = {2018}, doi = {10.1162/tacl\_a\_00005}, url = {https://doi.org/10.1162/tacl_a_00005} } ``` - [← previous A Hybrid Approach to Automatic Corpus Generation for Chinese Spelling Check](https://md.research.elamur.ai/free/en/papers/wang-2018-hybrid-approach-automatic-corpus-generation/) - [next → Dual Conditional Cross-Entropy Filtering of Noisy Parallel Corpora](https://md.research.elamur.ai/free/en/papers/dowmunt-2018-dual-conditional-cross-entropy-filtering/) - [↑ back to papers of 2018](https://md.research.elamur.ai/free/en/papers/2018/page/3/) ## References 3 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. [Neural machine translation by jointly learning to align and translate](https://md.research.elamur.ai/free/en/papers/bahdanau-2014-neural-machine-translation-jointly-learning/). In Proceedings of the ICLR Conference. 2. James K. Baker. 1979. Trainable grammars for speech recognition. The Journal of the Acoustical Society of America 65(S1):S132–S132. [find](https://scholar.google.com/scholar?q=James+K%2E+Baker%2E+1979%2E+Trainable+grammars+for+speech+recognition%2E+The+Journal+of+the+Acoustical+Society+of+America+65%28S1%29%3AS132%E2%80%93S132%2E) 3. Regina Barzilay and Mirella Lapata. 2008. [Modeling local coherence: An entity-based approach](https://md.research.elamur.ai/free/en/papers/barzilay-2008-modeling-local-coherence-entity-based/). Computational Linguistics 34(1):1–34. \#\# Conclusions *Updated 24 October 2026*