pyrealb at the GEM'24 Data-to-text Task: Symbolic English Text Generation from RDF Triples
What it is about
The paper describes a symbolic English generator for the GEM'24 data-to-text task, where a set of RDF triples (at most seven) must become text. The pipeline is simple. Triples are grouped by subject and sorted by a hand-set priority per predicate, so that the text tells a story (birth before career, creation before activities). Each group becomes one sentence; groups of more than three triples are split. Each predicate is verbalised by one of a set of hand-written templates, written as calls to the PYREALB realizer, which handles agreement, passive, pronouns and punctuation. There are 250 templates for the frequent predicates, and a default template (the predicate name split at case changes) covers about 5% of cases. The author reports that the system is fast on a laptop, that a new predicate takes about a minute to write, and that the preliminary scores are close to those of the other, mostly neural, participants. The paper gives no exact scores in the text read here.
Ideas
- I1.
world/src/knowledge.rs: keep one verbalisation table per relation, keyed by aBTreeMapof relation name to a priority, a subject-type flag and a list of templates. Choose the first template in table order, never at random, so that the output is deterministic. Measure how many school facts are realised from a template and how many fall back to the default. - I2. Default-template fallback: split a camel-case relation name (
servedAsChiefOf…) into words, as the paper does. Count how often this fallback is used on the school corpus, and rank the unseen relations by frequency, so the author knows which templates to write next. - I3.
world/src/narr.rs: a priority per relation, applied when facts about the same subject are ordered into a story. Test it with a negative control: a shuffled order of the same facts must score worse on a reading-order check. - I4. Aggregation rule: merge facts with a shared subject into one sentence, and split groups above a size limit. Measure the length of the output sentences on the school texts before and after.
For Mova
The verbalisation table belongs with the world: world/src/knowledge.rs (lessons, facts) and world/src/narr.rs. The sentence building itself is closest to en/src/lin.rs and the templates in 22. Phrases: noun, adjective, adverb and prepositional phrases. Measures: the template-hit rate and the default-fallback rate on the school lessons, and the determinism check (run twice, identical output).
Cite
@inproceedings{lapalme2024pyrealb,
title = {{pyrealb at the GEM'24 Data-to-text Task: Symbolic English Text Generation from RDF Triples}},
author = {Lapalme, Guy},
booktitle = {Proceedings of the 17th International Natural Language Generation Conference: Generation Challenges},
year = {2024},
doi = {10.18653/v1/2024.inlg-genchal.5},
url = {https://doi.org/10.18653/v1/2024.inlg-genchal.5}
}References
- Guy Lapalme. 2020. RDFjsRealB: a symbolic approach for generating text from RDF triples. In WebNLG 2020 : 3rd Workshop on Natural Language Generation from the Semantic Web, pages 144–153, Dublin, Ireland (virtual). SIGGEN. find
- Guy Lapalme. 2022. The jsRealB text realizer: Organization and use cases. (arXiv:2012.15425). ↗ our card
- Guy Lapalme. 2023. Data-to-text bilingual generation. Technical Report arXiv:2311.14808. ↗
- Simon Mille, João Sedoc, Yixin Liu, Elizabeth Clark, Agnes Axelsson, Miruna-Adriana Clinciu, Yufang Hou, Saad Mahamood, Ishmael Obonyo, and Lining Zhang. 2024. The 2024 GEM shared task on multilingual data-to-text generation and summarization: Overview and preliminary results. In Proceedings of the 17th International Conference on Natural Language Generation: Generation Challenges, Tokyo, Japan. Association for Computational Linguistics. find
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/lapalme-2024-pyrealb-gem-24-data-text/