Hi Louis, all,
Louis, thanks for raising that important issue!
I have been looking in a number of related NLG systems, and one thing I noticed is a pattern of many of these projects being developed very much in isolation from each other, and also often without much concern for ongoing linguistic research. That is what I tried to capture in the research paper by stating that there is no consensus on this, and that it seems too early to commit to a specific solution.
I had given a quick look to UNL, but the project looked pretty stale to me - I could not see any activity in the last decade. Furthermore, the page didn't provide access to the source code and instead mentioned that part of the technology is under patents, which is quite a red flag for me, and I usually don't look into something like that any further, in order to honestly be able to say that I didn't get any ideas from those patents. If I am mistaken, and there is a freely usable write-up or implementation, I'd be happy to come back and read up more.
Thank you for the annotated bibliography! That is super useful.
But I did look into detail into a (small) number of other, similar systems, such as Grammatical Framework or KPML. Tiago mentioned FrameNet, and I learned a lot about that too. To get an overview of the whole field has been a rather frustrating experience, especially since the major textbook in that area - Dale & Reiter - doesn't cover these systems, nor the 2018 update to that book by Gatt & Krahmer, and it seems that research work in that area often omits these practical systems. Accordingly, when I talk with the professors and researchers in this area, also about the proposal here, they are more focussed on specific issues, and don't know that much about the concrete systems (which is understandable - the flow from research to practical systems is a more established flow in many areas). Never mind that when you get to the linguistic side of it, instead of the computer science part, there are even more competing theories, many of which are aimed toward much more encompassing goals and are about covering the whole of language and natural language understanding, which we want to be shying away from.
The goal of the paper was never meant to be a comprehensive account of the state of the art in natural language generation. That's what Dale & Reiter and Gatt & Krahmer have aimed for, and their works are hundreds of pages. I had the feeling my paper was already too long, and putting in an overview of the state of the art would have made it at least double the length.
So, given that (and other reasons, as lined out in the paper), it seems that a system which could support any of these approaches seemed a more promising way. So far, for my own prototype, I have been mostly following Grammatical Framework (because it has a very accessible book, the software is free, the community was friendly, etc.), and it worked good enough to leave me convinced that the whole thing is worth trying out. But I don't know whether that's the best approach.
As mentioned by Chris Cooley, the goal will be to create a new wiki, a library of functions, that can support any of these approaches. My dream would be - and I see that Chris had already suggested that - that experts like you and your colleagues create an overview of the state of the art that will be accessible to the community and that will allow us to make a well-informed decision when the time comes as to which path to explore first. In a parallel track, we will be creating the function wiki, and then, when the time is ripe we can bring these two strands of work together. So, would you be willing to work on that?
How does this sound for a plan?
Some further points:
This is way easier to implement, test and deliver than to implement 10
different backends with various progress in implementation, incompatibilities and runtime characteristics.
Regarding your point about evaluation environments: I agree, it would be a huge task if the WMF core team were to develop all these different environments. But that's not the plan. The goal is really that *others* will hopefully build these :) All we need to do is to make sure that's possible and encouraged and simple enough. But yeah, not the core team.
The paper presents AW as sitting on top on WL. Both are big projects.
Sitting a big project on top of another one is really risky, as it means a significant milestone must first be reached in the dependency (here WL), which would likely took some years, before even starting the work on the other project.
Yes, that's correct. That is exactly the time that allows us do the appropriate state of the art analysis. I hope it won't take us years, but that we will be faster.
AW can be realised with current tools and engineering practices.
Only if you commit to a specific implementation, which I am hesitant to do.
[English is an obstacle to programming] This strong affirmation needs to
be sourced.
https://dl.acm.org/doi/10.1145/3051457.3051464
As I spend a significant time (~10 hours) gathering references and
writing this email (which is 5 pages long in Word), I would like to be mentioned as co-author in the final paper if any idea or references presented here is used in it.
Thank you for your detailed comments, which will certainly improve the second version of the paper. I am happy to mention you in the acknowledgments. For co-authorship, I usually go for a more substantial engagement ;) If you're willing to write up the "Previous work" section along the lines you mentioned above (maybe with Tiago? Maybe with others to join?), but for a comprehensive overview of existing systems, then I am open to talk about co-authorship :)
For French, the gender of every noun entity *must* be present ... For
Chinese and Japanese, classifier information must be present for all noun, in case one must be enumerated.
That's exactly the goal of the lexicographic project on Wikidata, as was pointed out:
https://www.wikidata.org/wiki/Lexeme:L12449
You'll find plenty of Lexemes with their classifiers, forms, etc. The lexicographic project was started with the Abstract Wikipedia in mind, knowing that exactly that will be needed.
Yet, the use of any existing formalism is dismissed, which mean all the
situations I illustrated in this email will need to be dealt with in an ad hoc fashion.
No, not at all it doesn't have to be ad-hoc, that's exactly what we can start working on now, long before we get to the point that we need to make that ad-hoc decision. I hope you'll join us to figure out the best way!
Thanks to Charles, Amir, Tiago, Christopher, Arthur, and Adam for your beautiful answers, who raised a number of great replies much better than I ever could. And thanks to Louis for starting this more than interesting thread! Let's continue in this vein!
Cheers, Denny
On Sun, Jul 5, 2020 at 9:49 PM Adam Sobieski adamsobieski@hotmail.com wrote:
Brainstorming: resembling what the document object model (DOM) [1] is for XML and attributed trees, perhaps we could create and specify an object model for sets of attributed predicate calculus expressions.
With an attributed predicate calculus object model (e.g. “APCOM”) for sets of attributed predicate calculus expressions:
{
r1.@a1(o1(icl>domain1).@a2, o2(icl>domain2).@a3).@a4
r2.@a5(o3(icl>domain3).@a6, o4(icl>domain4).@a7).@a8
r3.@a9(o5(icl>domain5).@a10, o6(icl>domain6).@a11, o7(icl>domain7).@a12).@a13
}.@a14
developers could more conveniently utilize sets of attributed predicate calculus expressions from JavaScript and Lua.
Drawing from XML, we can consider that objects, relations, attributes could be, instead of plain text strings, uniform resource identifiers (URI’s). “r1” could be a URI, “a1” could be a URI, “o1” could be a URI, and so forth.
We can also consider that the attributes in a model could have values:
{
r1.[@a1=v1](o1(icl>domain1).[@a2=v2], o2(icl>domain2).[@a3=v3]).[@a4=v4]
r2.[@a5=v5](o3(icl>domain3).[@a6=v6], o4(icl>domain4).[@a7=v7]).[@a8=v8]
r3.[@a9=v9](o5(icl>domain5).[@a10=v10], o6(icl>domain6).[@a11=v11], o7(icl>domain7).[@a12=v12]).[@a13=v13]
}.[@a14=v14]
We can consider creating a scripting API (e.g. “APCOM”) for a semantic model to convenience developers. We can also consider adding attribute-value pairs to a semantic model.
Best regards,
Adam
[1] https://en.wikipedia.org/wiki/Document_Object_Model
*From: *Tiago Timponi Torrent tiago.torrent@ufjf.edu.br *Sent: *Sunday, July 5, 2020 9:06 PM *To: *General public mailing list for the discussion of Abstract Wikipedia (aka Wikilambda) abstract-wikipedia@lists.wikimedia.org *Subject: *Re: [Abstract-wikipedia] NLP issues severely overlooked (Amir E. Aharoni)
That’s a good idea, but I think you would need more than that. Take FrameNet, for example, but now departing from verbs instead of nouns. FrameNet has a very detailed model for dealing with verbs, their semantic arguments and the way they surface in morphosyntax. Nonetheless, to apply such a model in a text comprehension and/or generation task, you need more than that. You need to know prototypical fillers for the positions, which, in turn, are associated to other frames and, therefore, participate in other clusters of the network of frames. Moreover, you’d want those prototypical fillers to function as departing points for analogical extensions in the model, since not every sentence is a prototypical combination of words. In other words, the collection of attributes and relations you refer to should be defined in a way that they can be analogically extended to other collections not originally assigned to the item you’re looking at.
Cheers
Tiago
Em dom, 5 de jul de 2020 às 20:03, Arthur Smith arthurpsmith@gmail.com escreveu:
Yes, thank you for the UNL background, that is extremely helpful. I've been reading some of the articles Louis provided as references, and it seems to me from just this perhaps naive point of view, that a lot of the complexity is associated with disambiguation of meaning - for nouns I think Wikidata items (and their relations to lexeme senses) solve that problem, but we are still missing I think a lot of the detail needed to do the same with adjectives and verbs (at least). So there is definitely some room for finding better ways to model - but maybe Wikidata could be expanded to handle the adjective/verb cases too. In general the concept of a single meaning associated with a Wikidata item as its identifier and a collection of attributes and relationships attached to that item is a powerful one that could resolve many such issues.
Arthur
On Sun, Jul 5, 2020 at 6:55 PM Adam Sobieski adamsobieski@hotmail.com wrote:
Louis,
Thank you for the information about the Universal Networking Language [1] and the World Atlas of Language Structures [2].
Semantic Modeling
Do you opine that adding attributes to objects, relations and expressions enhances expressiveness for various features of natural language?
r.@a1.@a2(o1(icl>domain1).@a3.@a4, o2(icl>domain2).@a5.@a6).@a7.@a8
I wonder whether there exist mappings or workarounds with which to obtain such expressiveness for models such as Wikidata’s.
Scripting Environments for Natural Language Generation
Supposing that Wikilambda could be JavaScript / WebAssembly based, and observing that Lua / WebAssembly solutions exist, we can note that scripting engines such as V8 are easy to use and to add global objects and API to. Resembling how Web browsers provide scripting environments and API for functions, we can envision providing scripting environments and API for natural language generation functions.
I wonder what you might think about scripting environments and API for natural language generation scenarios?
Best regards,
Adam
[1] https://en.wikipedia.org/wiki/Universal_Networking_Language
*From: *Louis Lecailliez louis.lecailliez@outlook.fr *Sent: *Saturday, July 4, 2020 2:10 PM *To: *abstract-wikipedia@lists.wikimedia.org *Subject: *Re: [Abstract-wikipedia] NLP issues severely overlooked (Amir E. Aharoni)
Hi Amir,
I understand the process is different that usual research. In fact I've seen Wikipedia grown from an unknown website to the biggest encyclopedia it is now. I use it daily in multiple languages and love it. I know what crowd sourcing could achieve.
It's also possible that the mere *finding* of these stumbling blocks by
such a big, diverse, open, and active community, will itself be a contribution to the scientific knowledge around this subject.
I disagree here. It would be contribution to scientic knowledge if and only if it wasn't discovered before. My email was precisely about that: capitalizing on the knowledge that has already been discovered, to avoid making the same mistake them again. It would not matter for a small project, but this one is really ambitious. We are speaking of 40 years of work by a horde of talented and very knowledgeable people, so this isn't to be dismissed easily.
This thing is, my previous email was a bit abstract, because it were a review of the paper, not of the project itself. I should have made more examples to illustrate where the problem lies.
Let's start with a simple example, in English, with corresponding Wikidata entities in-between parenthesis. I'm also using pseudo-turtle notation with made up relationships.
France (Q142) is a country (Q6256).
<Q142> <rel_is> <Q6256> .
Creating the English sentence is straightforward with the naive approach presented in the paper.
What is the French equivalent?
La France est un pays.
More information is required in the abstract representation: the text generator needs to know about the gender of both nouns (France and pays). So we need to extend the model as such:
<Q142> <rel_gender> <Q1775415> .
<Q6256> <rel_gender> <Q499327> .
Fine! Now what about Chinese?
法國是一個國家。
What we have in the middle of the sentence is a classifier (個). The model needs the following update:
<Q499327> <rel_use_classifier> <Q63153> .
To handle these 3 languages, the model has already 3 additional triples just for accounting for linguistic facts occuring in these languages. Wikipedia exists in more than 300 languages, and the world has about 6000 of them, each of them having particularities that must be taken into account. Fortunately they recoup themselves in-between languages. Nonetheless the World Atlas Language Structures ( https://wals.info/chapter/s1) count 144 distinct language features. Some are related to speech, but this means there is probably something like a hundred features that must be taken into account in the data model to produce valid natural language sentence.
Note that in the Chinese example, there is also a number (一, one) showing up. This is a phenomenon that must be taken into account; and it's not always appearing when using 是 (to be). How complex the "lambda" system will be just to deal with this issue? Hint: very much. It also needs to be compatible with dozen of other phenomena.
Then each of those features require extensive and complete data. For French, the gender of every noun entity *must* be present, otherwise there is half a chance of producing a wrong sentence each time a noun entity is encountered. For Chinese and Japanese, classifier information must be present for all noun, in case one must be enumerated. Where does the project will get the data from? (we are speaking of millions of item, most not referenced in existing dictionaries) How will this be encoded? Those are real questions that must be answered.
Suppose now we have done the work for "renderers" in these three languages. They both use the more or less similar A X B structure where X is a verb meaning "to be".
What would be the Japanese equivalent?
The more natural structure would be like:
フランスは国(だ)。
What is a play here is a topicalization (Q63105) of France, followed by a predicate (it's a country). This is very different from the previous structure, which, not surprisingly enough, needs it's own representation. To make situation more difficult, the previous (A be B) structure can also exists in Japanese, but would lead to a totally different sentence if used.
The paper states that Figure 1 and 2 are examples that will be more complex in real life. Yet, the use of any existing formalism is dismissed, which mean all the situations I illustrated in this email will need to be dealt with in an ad hoc fashion. Moreover, changing formalism (be it ad hoc or not) will require to change every piece of code/data using it. This will happen everytime a language with unsupported feature(s) is added to the project. It's not hard to see how this will waste a huge amount of time and goodwill from involved people. The very code focussed tone of the paper, the english-centric approach used in the examples and the lack of references shows that the complexity of the task on the NLP front is not sufficiently conceptualized.
Best Regards,
Louis Lecailliez
*De :* Abstract-Wikipedia abstract-wikipedia-bounces@lists.wikimedia.org de la part de abstract-wikipedia-request@lists.wikimedia.org < abstract-wikipedia-request@lists.wikimedia.org> *Envoyé :* samedi 4 juillet 2020 15:06 *À :* abstract-wikipedia@lists.wikimedia.org < abstract-wikipedia@lists.wikimedia.org> *Objet :* Abstract-Wikipedia Digest, Vol 1, Issue 6
Send Abstract-Wikipedia mailing list submissions to abstract-wikipedia@lists.wikimedia.org
To subscribe or unsubscribe via the World Wide Web, visit https://lists.wikimedia.org/mailman/listinfo/abstract-wikipedia or, via email, send a message with subject or body 'help' to abstract-wikipedia-request@lists.wikimedia.org
You can reach the person managing the list at abstract-wikipedia-owner@lists.wikimedia.org
When replying, please edit your Subject line so it is more specific than "Re: Contents of Abstract-Wikipedia digest..."
Today's Topics:
- Re: NLP issues severely overlooked (Charles Matthews)
- Use case: generation of short description (Jakob Voß)
- Re: NLP issues severely overlooked (Amir E. Aharoni)
Message: 1 Date: Sat, 4 Jul 2020 14:05:09 +0100 (BST) From: Charles Matthews charles.r.matthews@ntlworld.com To: "General public mailing list for the discussion of Abstract Wikipedia (aka Wikilambda)" < abstract-wikipedia@lists.wikimedia.org> Subject: Re: [Abstract-wikipedia] NLP issues severely overlooked Message-ID: 2126327926.39940.1593867909152@mail2.virginmedia.com Content-Type: text/plain; charset="utf-8"
It is interesting to be on a list where one can hear about software issues, and then computational linguistic problems. I'm not an expert in either area.
I do have 17 years of varied Wikimedia experience (and I use my real name there).
On 04 July 2020 at 12:25 Louis Lecailliez louis.lecailliez@outlook.fr
wrote:
<snip>
Nothing precise is said about linguistic resources in the AW paperexcept for "These function finally can call the lexicographic knowlegde stored in Wikidata.", which is not very convincing: first because Wiktionary projects themselves severely lacks content and structure for those who has some content at all, secondly since specialized NLP ressources are missing there too (note: I'm not familiar with Wikidata so I could be wrong, however I never saw it cited for the kind of NLP resources I'm talking about).
I can comment about this. Besides Wiktionary, there is the "lexeme" namespace of Wikidata. It is a relatively new part of Wikidata, dealing with verbal forms.
To finish on a positive note, I would like to highlight the points I
really like in the paper. First, its collaborative and open nature, like all Wikimedia projects, gives him a much higher chance of success than its predecessors.
It is worth saying, for context, that there is a certain style or philosophy coming from the wiki side: more precisely, from the wikis before Wikipedia. There is the slogan "what is the simplest thing that would actually work?" You might argue, plausibly, that Wikipedia at nearly 20 years old, shows that there is a bit more to engineering than that.
On the other hand, looking at Wikidata at seven years old, there is some point to the comment. It has a rather simple approach to linked structured data, compared to the Semantic Web environment. (Really, just write a very large piece of JSON and try to cope with it!) But the number of binary relations used (8K, if you count the "external links" handling) is now quite large, and has grown organically. The data modelling is in a sense primitive, sometimes non-existent. But the range of content handled really is encyclopedic. And in an area like scientific bibliography, at a scale of tens of millions of entities, the advantages of not much ontological fussiness begin to be seen.
Wikidata started as an index of all Wikipedia articles, and is now five times the size needed for that: a very enriched "index".
I suppose the NLP required to code up, for example, 50K chemistry articles about molecules, might be a problem that could be solved, leaving aside the general problems for the moment.
In any case, there is a certain approach, neither academic nor commercial, that comes with Wikimedia and its communities, and it will be interesting to see how the issues are addressed.
Charles Matthews (in Cambridge UK)
