Talk:Abstract Wikipedia/Related and previous work/Natural language generation

From Meta, a Wikimedia project coordination wiki

Propose a link[edit]

Please add interesting links to the main page.

If you're not sure whether we might be interested, please add the link here. Any additional comments would be welcome. Please use the language you are most comfortable with.

Very interesting links[edit]

  • Book Review (Yue Zhang, 2020) of Deep Learning Approaches to Text Production (Narayan and Gardent, 2020) "Text Production" is NLG, so the book's topic is "neural NLG". If you thought you might be interested in the book, the review will help you make up your mind. A preview of the book is available here.
  • A Survey of Evaluation Metrics Used for NLG Systems (Ananya B. Sai, Akash Kumar Mohankumar, and Mitesh M. Khapra, 2020) "Over the past few years, many evaluation metrics have been proposed: some task agnostic and others task specific. In this survey ... we propose that ... the metrics can be categorised as context-free and context-dependent metrics. Within each of these categories there are trained and untrained metrics which rely on word based, character based or embedding based information to evaluate a hypothesis. This arrangement ... shows that there is still a need for developing task-specific context-dependent metrics as most of the current metrics are context-free. This is a major gap in existing works as in many NLG tasks context plays a very important role and ignoring it is not prudent." As well as surveying the field of metrics, this paper gives an accessible overview of the NLG field today (and abstraction with credit is permitted). NLG subfields referenced are:
    • Machine Translation (MT)
    • Abstractive Summarization (AS)
    • Free-form Question Answering (QA)
    • Question Generation (QG)
    • Data to Text Generation (D2T)
    • Dialogue Generation (DG)
    • Image Captioning (IC)
Each of these is described in less than half a page. Effectively excluded from the survey are "spelling and grammar correction, automatic paraphrase generation, video captioning, simplification of complex texts, automatic code generation, humour generation, etc". There is a useful one-page table of example inputs and outputs of the different activities at page 6. ("Table 1. Examples inputs and generated outputs for various Natural Language Generation tasks.") Unfortunately, apostrophes appear to have become corrupted. In Section 3 (p.9), as well as describing how evaluations have been set up, different criteria are named and defined for the different types of system (MT, AS etc). Section 4 (p.13) provides the authors' taxonomy of automated evaluation metrics. A one-page chart is at page 15. (Sections 5 to 8 are for the statisticians; there were far too many sigmas for me to cope with. Oh, and a λ... but it's only a "weighting factor" --GrounderUK (talk) 23:02, 31 August 2020 (UTC)) Section 5 (p.16) considers context-free metrics, Section 6 (p.31) considers context-dependent metrics, Section 7 (p.37) considers studies critical of automated evaluation metrics (but with no statistical analysis) and Section 8 (p.39) considers studies evaluating evaluation metrics. Section 9 (p.43) makes substantive recommendations for future research and Section 10 is a fairly brief conclusion, some of which is quoted above (after the link). 179 papers are referenced.[reply]
  • Toward Givenness Hierarchy Theoretic Natural Language Generation (Pal & Williams, 2020) is a five-pager on the generation of anaphora by robots by inferring the likely givenness hierarchy of the interlocutor. "we formulate cognitive status modeling as a Bayesian filtering problem..." [section 3] "Using this formalism, our goal is to recursively estimate, for a given object, the probability distribution over cognitive statuses for object o at time t..." [We would be less concerned with time, not being a robot.] "The GetStatus(O) function (Algorithm 1) takes an object O and returns its most likely cognitive status. If no CSF [cognitive status filter; the Bayesian given in the paper] exists for O ... “UID” is returned; otherwise the most probable cognitive status for O (as determined by the distribution maintained by O’s CSF) is returned."

Wikipedia and Wikidata related[edit]

  • Kaffee, Lucie-Aimée; Vougiouklis, Pavlos; Simperl, Elena. "Using Natural Language Generation to Bootstrap Missing Wikipedia Articles: A Human-centric Perspective" (PDF). Semantic Web: Interoperability, Usability, Applicability ( Q15817015).  Arabic and Esperanto; ML. This is the most immediately relevant paper I've come across. The study used a machine-learning approach, but in other respects it explores the Abstract Wikipedia context. It prototypes an extension of ArticlePlaceholder with an introductory sentence generated from Wikidata triples. This prototype is presented to a number of experienced Wikipedia editors and their reactions are explored. This includes whether and how they edit or replace the introductory sentence, as well as structured interviews after the "confrontation". "The evaluation, which includes an automatic, a judgement-based, and a task-based component, shows that the [introductory] sentences score well in terms of perceived fluency and appropriateness for Wikipedia, and can help editors bootstrap new articles. It also hints at several potential implications of using NLG solutions in Wikipedia at large, including content quality, trust in technology, and algorithmic transparency." Of concern to the Abstract Wikipedia project is the fact that it gets no mention.
  • Language Models as Knowledge Bases? (Petroni, Rocktäschel, Lewis, Bakhtin, Wu, Miller and Riedel, 2019) ML; English.
  • Drawing Questions from Wikidata (Geng, 2016) "...ratings suggest that our application can generate questions that can compete with manually created ones. However, there are still many that are deemed irrelevant. Our results indicate that Wikidata still has a lot of incomplete and imprecise data. ...we experienced that an update in Wikidata’s data set does impact our quiz and makes it more precise. Furthermore, our quiz application can be used to detect incomplete Wikidata items in an entertaining manner." This builds on work "introducing Wikidata quiz" described in:
  • Drawing Questions from Wikidata (Bissig, 2015) "We construct a graph by querying multiple Wikidata items originating from any chosen topic. The structure of the resulting graph is used to generate relevant questions with answer options ... participants found good questions and they saw improvements in the algorithm over time. We found that the incomplete status of Wikidata negatively impacts the quality of the generated graph. We found limits in the types of questions that are suitable for generating from knowledge bases..."

NLG in practice[edit]

Haapanen, Lauri; Leppänen, Leo (2020-10-07). "Recycling a genre for news automation: The production of Valtteri the Election Bot". AILA Review (Q15749404) 33: 67–85. ISSN 1461-0213. doi:10.1075/aila.00030.haa. Retrieved 2020-11-09.  Swedish, Finnish, English; data to news; pipeline. As well as having tri-lingual and deterministic text generation (aligned to the pipeline approach), the user experience in this case study allowed the vantage point of the news story to be selected, including geographic localization and party-political perspective. Sadly, readers rated the generated texts less highly than similar stories produced by human journalists; the "most frequent complaints were about language errors, obtrusive repetition, and “dry” language, and the most common words in the negative feedback were words like boring, confusing, monotone and incoherent. On the positive side, the computer-written stories were generally praised for being based on facts and for being clear and to-the-point."

Leppänen, Leo; Munezero, Myriam; Granroth-Wilding, Mark; Toivonen, Hannu (Sep 2017). "Data-Driven News Generation for Automated Journalism". Proceedings of the 10th International Conference on Natural Language Generation. Santiago de Compostela, Spain: Association for Computational Linguistics. pp. 188–197. doi:10.18653/v1/W17-3528. Retrieved 2020-11-09.  The authors "explore the field and challenges associated with building a journalistic natural language generation system [...and...] present a set of requirements that should guide system design, including transparency, accuracy, modifiability and transferability." The outcome is "a data-driven architecture for automated journalism that is largely domain and language independent."
Leppänen, Leo; Munezero, Myriam; Sirén-Heikel, Stefanie; Granroth-Wilding, Mark; Toivonen, Hannu (2017). "Finding and expressing news from structured data" (PDF). Proceedings of the 21st International Academic Mindtrek Conference on - AcademicMindtrek '17. the 21st International Academic Mindtrek Conference. Tampere, Finland: ACM Press. pp. 174–183. ISBN 978-1-4503-5426-4. doi:10.1145/3131085.3131112. Retrieved 2020-11-09.  "through automation of the news generation process, including the generation of textual news articles, a large amount of news can be expressed in digestible formats to audiences, at varying local levels, and in multiple languages. In addition, automation allows the audience to tailor or personalize the news they want to read."

Not obviously relevant links[edit]

Comments without a link[edit]

@???:

Mailing list and chat contributions[edit]

https://lists.wikimedia.org/mailman/listinfo/abstract-wikipedia
mail archives
July 2020

There have been some worth-while references cited in our various emails. The original email contributors are encouraged to include any relevant links on the main page. In the mean time, here are a few links to glean from...

https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000001.html
[https://meta.m.wikimedia.org/wiki/Wikilambda
[https://lists.wikimedia.org/mailman/listinfo/abstract-wikipedia
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000002.html
[https://www.mediawiki.org/wiki/Extension:WikiLambda
[http://pronoun.is/he
[http://pronoun.is/they/.../themself
[https://wikimediafoundation.org/
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000003.html
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000004.html
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia/July_2020_announcement
[https://news.ycombinator.com/item?id=23714875
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000005.html
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000006.html
Racket - Rhombus brainstorming
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000007.html
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000008.html
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia/Plan
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia/Plan#Task_P1.12:_JavaScript-based_implementations
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia/Plan#Task_P1.15:_Lua-based_implementations
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia/Plan#Task_O6:_Python-based_implementations
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia/Plan#Task_O7:_Implementations_in_other_languages
[https://meta.wikimedia.org/wiki/Abstract_Wikipedia/Plan#Task_O25:_Integrate_into_IDEs
[http://cohmetrix.com/
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000009.html
[https://arxiv.org/abs/2004.04733
[https://www.researchgate.net/profile/Hiroshi_Uchida2/publication/239328725_A_Gift_for_a_Millennium/links/54c6953e0cf22d626a34f224/A-Gift-for-a-Millennium.pdf
[https://pdfs.semanticscholar.org/b030/ea4662e393657b9a134c006ca5b08e8a23b3.pdf?_ga=2.109286021.1099995837.1593757540-1424212949.1593757540
[http://www.afcp-parole.org/doc/Archives_JEP/2002_XXIVe_JEP_Nancy/talnrecital/TALN/actes_taln/articles/TALN26.pdf
[https://arxiv.org/pdf/1902.08061.pdf
[http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.212.2058&rep=rep1&type=pdf
[https://www.cicling.org/2005/unl-book/Papers/003.pdf
UNL DECONVERTER FOR TAMIL (T.Dhanabalan, T.V.Geetha. 2003)
[http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.78.979&rep=rep1&type=pdf
Abstract Meaning Representation for Sembanking (2013)
Heloise — An Ariane-G5 compatible environment for developing expert MT systems online (Berment/Boitet, 2012)
Heloise — A reengineering of Ariane-G5 SLLPs for application to π-languages (Berment/Boitet, 2012) π-languages are "poorly-resourced" languages (langues peu dotées). The overlap between these two papers is significant and their references are identical.
ONLINE TRANSLATION SERVICES FOR THE LAO LANGUAGE (Berment, 2005)
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-July/000010.html
https://lists.wikimedia.org/pipermail/abstract-wikipedia/2020-August/000236.html]
Drawing Questions from Wikidata (Geng, 2016)

#wikipedia-abstract[edit]

26 July 2020
Between the Brackets ep.60 Richard Knipel
Between the Brackets ep.60 Amir Aharoni
GraalEneyj (Lucas Werkmeister)
LinGO Grammar Matrix
The LinGO Grammar Matrix Customization System (2009)
Professor Emily M. Bender
21 July 2020
Essence of Random ;)

Wikispore[edit]

@DVrandecic (WMF): I think all of this should be on an NLG Wikispore, but I can't decide whether it should be sponsored by the project or just happen to emerge.--GrounderUK (talk) 14:38, 26 July 2020 (UTC)[reply]

Organize by language, support, other dimensions?[edit]

I get the feeling that a lot of these projects focus on English NLG, but what we're probably interested in most are approaches that generalize as far as possible across many different languages. Can the listed projects be organized (perhaps in a table somewhere) by the language(s) they aim to support? Other dimensions might be the degree of activity/support, openness of project, any specific language domains they may focus on (eg. technical writing?) ArthurPSmith (talk) 20:09, 28 July 2020 (UTC)[reply]

@ArthurPSmith: It's a fair bet. Please see my suggestion to Denny above. To put this in context, the 2018 survey by Gatt & Krahmer references 214 different papers from 2010 to 2017 and 134 more from 2000 to 2009 (plus maybe another 100). A more up-to-date selection might be constrained to reference this survey. There are only just over 100 of those on Google Scholar (or "about 264", in Google gigo). The book by Reiter and Dale (2000) has 2347 citations.--GrounderUK (talk) 23:27, 28 July 2020 (UTC)[reply]