Microblog

MMO game Sea of Thieves is amazing so far. Two days in and so far we found treasure, smashed a few skeletons, sailed with an allied fleet of two galleons, and set off some fireworks.

Reminder that preprocessing user-supplied data is never trivial.

The package contained 2 boxes of Story cubes, a present from HMI’s NLP group. ✨

It’s my last week at the university, and this mysterious package suddenly showed up in my home office today... 👀

Cute! We need more cat pics. ;) Now, for throwback Tuesday, this photo was taken 10 years ago. Look at that babyface!

I am deeply impressed by this long-read by @betterthemask. It definitely has impacted the way I view NLP/NLG and games, including my own research work. Can't wait to explore indie game Sable now...

One of these 10 cuties is on its way to our home, thanks to @mrngm

A blogpost about biases in recommendation AI led me to this useful list with actions you can take to get away from addictive tech platforms and recommender algorithms: humanetech.com/take-control.
So far I really like the News Feed Eradicator browser add-on.

Will do. There will be a livestream on the day itself for people that want to watch the defense remotely!

Huzzah! Yesterday we confirmed the last member of my thesis committee, so now there's a team of 5 experts (NLG! NLP! Games! PCG! Storytelling!) ready to review my 250-page dissertation on flavor text generation for video games. My defense is planned for 25th of March, 2022. 🎈

Nee, maar het is een variatie op github.com/jd7h/spotifail, misschien dat je daar iets mee kan knutselen. ;)

Co-worker Tim showed me this beautiful thesis by @TineArts, with a watercolor illustration by Dutch artist Evelien Jagtman.

2 years ago today! My fav is still "Harry Potter and the Problem That Couldn't Possibly Be DNS" -- although "Harry Potter and the harry potter and the harry potter and the harry potter" is also really relevant when working with GPT-2...

judithvanstegeren.com/blog/2019/harr…

De Spotify zoek-machine is wel een beetje "dom": als jullie schrijven "Ayreon – Day Seven: Hope (Come Back To Me)" dan snapt Spotify niet dat dat hetzelfde is als "Ayreon - Day Seven: Hope". En het heeft ook moeite met specifieke remixes, versies of uitvoeringen.

Een Python-script dat jullie keuzelijst inleest vanaf de website, en daarna dmv de Spotify API (1) een nieuwe playlist aanmaakt, (2) alle nummers uit de keuzelijst (titel + artiestennaam) opzoekt in de Spotify database en (3) de gevonden nummers toevoegt aan de playlist.

Hey leuke mensen van @ndrgwrdrdldjs, @Snob_2000 en @PinguinRadio, ik heb (geautomatiseerd) aan Spotify gevraagd om een playlist te bakken van de Snob 2000 keuzelijst voor 2021. Voila: open.spotify.com/playlist/7jjI4…

This haiku generator also generates "auto praise" for generated artifacts. :') Did we just find an example of "framing" in the wild? poem-generator.org.uk/haiku/

I love this query-answer example. We need more tweets like this -- it's a tangible demonstration of how these models go wrong in interpreting inputs. Maybe that's the best remedy against overhyped AI: crowdsourced "fuzzing" of opaque models.

Grading mood.

New business idea: company that does pentesting/fuzzing on production-grade deep learning models. Surely this already exists?

Generative Adversarial Network Discovers Rare Zebra Cat, Shocks Researchers.

TIL "AI" stands for "Angered Italian" real-italian-coffee.herokuapp.com/brain (That would explain the scientific success of our joint projects, @lorgatti)

How did I miss this game!? It's Alba: a Wildlife Adventure (low poly game with photography) but with dogs!?

But where do you get the age (category) data from? Twitter does not share age or date of birth of users via their API, AFAIK.

Dissertation writing mood.

Quote from "Developing a methodology for corpus-based computer game studies" by Carrillo Masso (2009). bit.ly/3DKXnt4

Crap, I should REALLY have read this in my first year. Missed this glorious chance. ;)

"One of the biggest mistakes in the area of game analysis has been the fact that researchers have consistently drawn conclusions and made ‘analyses’ of games they have not played. In order to avoid this mistake, playing Diablo and WoW was planned for as part of the methodology."

I'm reading a 36 page journal paper about "morphological and onomastic characteristics of toponyms in The Elder Scrolls". Translation: linguistic properties of fictional place-names, such as the Dwemer ruins and ancestor tombs in Morrowind. I love my job. bit.ly/3DMcLp1

Other side-effects include obsessively searching for local butcher shops, translating obscure vegetable names from English to Dutch and silently moving large amounts of money from 'savings' to 'groceries'.

I was reading Nigel Slater's Kitchen Diaries, but after ten pages of lavish October feasts I couldn't take it any more. Now I'm eating toasted sourdough bread with olive oil and a few chunks of goat cheese. This book is dangerous!

Now I really want to read this book (1978). What other hidden gems does it contain?

Another classic video game that has influenced procedural content generation: Ham(m)urabi (1968)
acriticalhit.com/sumerian-game-…

This Guardian article (2003) describes the procedural planets in Elite (1984), a classic example in procedural content generation. theguardian.com/books/2003/oct…

@joynessthebrave @SketchesbyBoze @Iconawrites I watched Howl’s Moving Castle and Lark rise to Candleford recently, and I’m now reading Piranesi. Thank you for all the great tips you put out on Twitter!

Apparently, Twitter is now only for people who (1) can recognize chimneys in cropped photos (2) own a phone (3) consent to obligatory, unannounced and unpaid crowd-worker tasks by large social media platforms.

I know you meant the /book/ is available in all good bookshops and online... but I would also love to buy the tiny Lego robot for my office <3

Look at that cute tromp l'oeil zine cover!

Tom Gauld's drawings in full colour 😍

I'm writing my conclusion now, and I feel the best part of my thesis is the ideas for future work. Not sure if this is a very good or very bad sign.

While I was searching for screenshots of Torchlight 2, I found this database of game interface screenshots. Pretty cool, and a nice source of inspiration for UI nerds. :) interfaceingame.com

I read children's books and graphic novels!

Dissertation writing mood.

Diving into the SW:KOTOR game internals again, as I'm reviewing my work on game corpora for my thesis. Turns out that in KOTOR, characters are only allowed to flirt if they are of the opposite gender as the flirtee, and have high charisma.

Dissertation writing mood.

So @annetropy, I know your professional opinions on cats and dogs, but how do you feel about birds?

Yay, #fdg21 has started! It's great to see all the usual suspects at the Discord server. I've spent this morning pre-recording my talk... and hoping that the sounds of leafblowers, road works and neighbourhood kids are not too prominent in the background.

And of course you can ask questions via Twitter as well.

Thanks! It's prerecorded so you can watch it after your vacation as well, instead of working during your days off. ;) Worklife balance is important!

Playing around with image generation using BigSleep, for the haiku poem about a jumping frog by Matsuo Basho.

The old pond
A frog jumped in,
Kerplunk!

But there is hope. My co-author's name is now spelled properly, with an "ś"!

The new ACM TAPS workflow is not working very well so far... Helpdesk unresponsive, TAPS dashboard stuck at 'helpdesk will contact you within 24 hours' for 3 weeks, and now that I've finally received a fixed HTML proof, I can't mark it as accepted it in the online TAPS dashboard.

Fun blogpost about generating haiku using GPT-2. brianweet.com/2019/06/16/wri…

I'm writing a thesis section about the various problems and limitations of GPT-2 today. This report by OpenAI is a worthwhile starting point. arxiv.org/abs/1908.09203

Here among my trappings and belongings, I am home~
youtu.be/yVmtG9_55F0?t=…

Midnight googling for the origin of hacker cocktail 'Tschunk'. I found the first (?) recipe for Tschunk on a Gulaschprogrammiernacht (GPN) blog from 2006. chre.kocht.gulasch.myblog.de/chre.kocht.gul…

Where are all these wholesome games coming from all of a sudden? I love them, and there are not enough hours in the day.

@maxkreminski Reminiscent of your gardening games idea at FDG 2018? eurogamer.net/articles/2021-…

Congrats! I have a master student working on coherence & stories atm (apart from my own dissertation work) so this is welcome reading material. :)

My latest paper as a haiku:

GPT-2 read
many World of Warcraft quests
to help humans write

Maybe we should just ditch the 'A PhD student can only go to a conf if they get a paper accepted there' rule?

If I could re-do my PhD and start in year 1, I would write journal papers instead of conference papers... Conferences are fun, but cramming multiple experiments in 8 pages just doesn't do them justice.

"To open the frozen door and claim the honeycomb inside, you’ll need to find an old woman to spin you a pair of mittens out of wool from undersea sheep."

@tomgauld Can we get more macro-shots of the contents of your bookcase? ;)

Guess the book from the top n words

dream, fire, art, feeling, creative, hear, yes, worksheet, passion, walk, strength, coach, session, burn, bring, feel, earn, exercise, drive, service, trust, get, wait, vision, enthusiasm

Yes, I have the same experience. :(

I have sent a message to Support, we'll see what happens. To clarify: I can still access results from previous experiments, but I'm not allowed to start new jobs -- I only receive the fatal error after setup (meh). AFAIK I never had a Figure8 subscription, so this is unexpected.

Soooooo it seems I'm in the market for a new crowd-source platform. What does everyone use for their NLG/PCG evaluation experiments nowadays?

I'm trying to continue with a crowd-sourcing experiment in Appen (previously CrowdFlower/Figure 8), but despite ample funds in my account, the platform suddenly complains that my "subscription has expired". :/

I have just received word that the technical term for this phenomenon is 'Buffett overflow' or 'stonk overflow'. :')

Never underestimate the power of first-year computer science knowledge.

Seeing procedural art on Twitter always gives me sweaty palms and makes me want to drop all my existing projects and rush off to learn the newest equivalent of Processing.

Yes, it's mentioned in his autiobiography as well. I believe Russell lived until the age of 98. He changed his views, opinions, politics and lifestyle frequently during his lifetime, and made a few big mistakes too. In the book he does not hesitate to reflect on his past.

I'm reading the autobiography of Bertrand Russell today. This is what he wrote about his experiences with writing his fellowship dissertation, which he wrote when he was 21...

If you like this topic, this excellent blogpost by @mark_riedl is also worth reading. mark-riedl.medium.com/ai-democratiza…

I don't think GPT-3 can generate Literature: Charles Dickens and Victor Hugo tried to address societal problems with their writing. Tolstoy wrote about complex human emotions + experiences. Large language models can't do that, even if you train them on every novel ever written.

I love to see GPT-3 used for creative writing, but it's becoming overhyped to the same extent that blockchain technology is overhyped. It's not a silver bullet and it has some serious drawbacks.

If the majority of people views AI-generated content as an extension of the human who used the AI, as something with common sense and intentions, it could be dangerous to society.

This view is not explicitly part of the article, I know. But the author spends only a few lines on the caveats of a large model trained on a part of the internet. The internet, which can be a bit like a public bathroom at times.

Societal enthusiasm about applications of AI and large language models: great. But the view that GPT-3 (by some high-tech magic) can create high-quality text, consistently and effortlessly, for any creative task, is just too optimistic.

Correction, the article was published in the New Yorker, not the NYTimes.

Maybe it's these lines: "But what is writing this? Who is writing this? In a sense, Coleridge wrote it. He articulated the pathways of its manufacture."
A GPT-3 generated poem generated from one line of Coleridge as a prompt, however sophisticated, is not "written by Coleridge".

I can't articulate why, but this NYtimes article gave me a foreboding feeling. "For writers who don’t like writing—which, in my experience, is nearly all of us—Sudowrite may well be a salvation."
Sudowrite is built on top of GPT-3.
newyorker.com/culture/cultur…

I'm reading TVTropes this Saturday morning, and I learned a new word! tvtropes.org/pmwiki/pmwiki.…

I wish I'd known the difference between narrative, story, discourse and plot when I started out three years ago. Most research papers don't contain the basics, so it's very useful for a research field to have newcomer-friendly, well-written survey papers. :)

If you're interested in narrative generation, be sure to read the survey on computational narratives by Kybartas and Bidarra (2017). It has definitions of core terms and provides a nice framework for story gen systems based on two axes 'plot' and 'space'. repository.tudelft.nl/islandora/obje…

I played 4 demos in 3 hours: Game Director Story (project manager & HR simulation), The Invisible Hand (stock market mystery), Forgotten Fields, and Nuts (90s biology research assistant sim with video camera's and pop art graphics).

It's Sunday, and LudoNarraCon is on! Hold my cappuccino while I download a subset of the 34 playable demo's. :3

So games research Twitter mentioned Blaseball at some point, and now #LudoNarraCon is sponsoring a "blaseball season"... I think it's time to find out what Blaseball is, and see what the fuss is all about.

I remember at COG 2019 (an academic games conferences I attended pre-COVID), some of the games that were showcased at #LudoNarraCon were the ones that everyone kept recommending to each other!

The exhibitors are mostly indie devs, and the games are unlike anything you normally see: their stories explore interesting themes (privacy, diversity, politics, societal issues) and the games often have really interesting art styles as well.

I cannot stress enough how much I love this event. If you like narrative-driven games, be sure to join this weekend. It's online and freely accessible. I attended last year by playing a few of the demos and ended up buying multiple games because I loved the demos so much.

LudoNarraCon, an online celebration of narrative games, starts this Friday! It runs from April 23 to April 26 with all kinds of cool activities on Steam. Looking forward to all the demos! More info here: ludonarracon.com

I am pretty sure this can be built with some basic NLP magic. Sounds like some good thesis procrastination...

New blogpost: Networking tips for your first conference (aka the blogpost I wish I had in 2016)
Thanks to @WEareTROOPERS for the great conf experience. <3
judithvanstegeren.com/blog/2021/netw…

This post has nice visualisations of encoder-decoder architectures and attention mechanisms:
jalammar.github.io/visualizing-ne…

Haha, one of my tweets ended up in a fun blogpost by @IrisVanRooij: irisvanrooijcogsci.com/2018/12/21/can…

Humor in the footnotes of the original GPT-2 paper by OpenAI, in a section about evaluating GPT-2 on a question answering dataset. GPT-2 scored pretty bad with an accuracy of only 4%, which is 5 times better than its predecessor GPT. The (human) first author scored 17%. #nlproc

A quote from Common Stocks and Uncommon Profits (1958, p. 203), in which Philip A. Fisher talks about economic forecasting.

"At present (1958), able men are attempting to harness electronic computers to establish “input-output” series of sufficient intricacy that perhaps at some future date it may be possible to know with a fair degree of precision what the coming business trends will be."

What's an FSI enterprise?

If you know about companies or people that do something like this, let me know!

I'm thinking along these lines:
- use text processing (NLP) to analyze ESG criteria in public companies, e.g. gender equality like @equileap does
- use text generation to generate better & faster security reports
- use NLP to process security reports for faster monitoring
- ???

I've been thinking a lot about my post-PhD career lately. My contract ends in December, after which I'll work @Datakami_nl fulltime... but what kind of projects should I contribute to and why? I'd love to do something meaningful with my skills within infosec, NLP and/or finance.

BTW this is a short, 4-page publication in Dutch, in a magazine for economists, which highlights the method and main findings. Publishing a peer-reviewed academic publication is still in progress, according to the interview with the 2 authors in FD: fd.nl/weekend/137720…

This project is a glorious combination of some of my personal interests: data mining, text mining, NLP, sentiment analysis, investing, stock market sentiment, and finance.

Fun! Researchers of @DNB_NL performed sentiment analysis on @FD_Nieuws articles to get macro-economic sentiment for the Netherlands, which is faster than waiting for reports by Statistics Netherlands (CBS). You can read the pub here (via the first author): jasperdewinter.github.io/pp/files/DeWin…

Fun! Researchers of @DNB_NL performed sentiment analysis on @FD_Nieuws articles to get macro-economic sentiment for the Netherlands, which is faster than waiting for reports by Statistics Netherlands (CBS). You can read the pub here (via the first author): jasperdewinter.github.io/pp/files/DeWin…

Yes, but established domain-specific lexicons don't change overnight, esp. in research. If you reuse the old lexicon for new research, it is easier to compare your results. And industry generally does not publish their lexicons.

Just substitute each 'bad' word with a synonym that is not in the list, or rewrite sentences that trigger a negative rating in the sentiment analyzer.

To be honest, I expect sentiment analysis for finance is done with sentiment word lists (which is more transparant than black box sentiment analysis with deep learning), and then you can skew the analysis result by simple word substitution.

"Companies go beyond machine readability and manage the sentiment and tone of their disclosures to induce algorithmic readers to draw favorable conclusions about the content. For example, companies avoid words that are listed as negative in the directions given to algorithms."

nber.org/digest-202012/…

And again I turn to the Internet Archive to find classic-but-hard-to-find-papers. This time, "Building applied natural language generation systems" (1997) by @EhudReiter and Dale. Link: web.archive.org/web/2018050812…

Nope, I'm not familiar with it.

Recently I read research that shows that once researchers started publishing about sentiment analysis on annual reports (and investors started trading based on automated text analysis) companies slightly adapted their language to push the algorithms to more favorable conclusions.

Last Saturday, Dutch newspaper FD reported about sentiment analysis research on news articles to automatically assess the economy. Now one of their journalists has written a news article titled 'Dear Algorithm' addressed to the sentiment analyzer. :D
fd.nl/opinie/1378588…

Wat als je geen Facebookaccount hebt? Kan ik deze belangrijke info ook op een andere plek dan een Amerikaans advertentieplatform vinden?

Just found out that one of the academic lexicon's I've been working with was published under a Creative Commons, No Derivatives license, which seems a bit silly to me. Why release an academic dataset and not allow other researchers to build upon it?! creativecommons.org/2020/04/21/aca…

I'm writing the background section on NLG now, and I've gone back to the basics, such as template-based NLG. It's fun to read this paper (1995) by @EhudReiter, in which he discusses 'real' NLG vs templates: arxiv.org/pdf/cmp-lg/950…. Related blogpost: ehudreiter.com/2016/12/18/nlg…

explore argue discuss explain assume state pose observe introduce investigate focus research question criticize compare contrast adapt improve build develop propose

Schrödinger's final year PhD student: when you send off a chapter draft to your supervisor, and feel both proud and horrified at the same time when you re-read the words you've written so far.

(NL) Hihi, ik heb een nieuw woord geleerd: "gewichtigewartaalmachine" nrc.nl/nieuws/2021/02…

He also describes the various roles AI can fulfill in the context of games: "Mechanic, Alter/Ego, Observer, Protector, Player, Creator," and (speculatively) "God".

It nice to read papers that are different from my usual fare. In this paper, Michael Straeubig comments on the public discourse on video games and AI, while referencing a whole bunch of interesting people and projects: eludamos.org/index.php/elud…

It's about ReGEN, a narrative/quest generator with graph rewriting. The authors also describe metrics for analyzing branching narrative graphs, and compare their outputs w/ SQUEGE, Skyrim's Radiant generator, and main quests of Skyrim and The Witcher.

This is still one of my favorite procedural content generation papers: "Analysis of ReGEN as a graph-rewriting system for quest generation" by @QKybartas and Verbrugge (2014) gram.cs.mcgill.ca/papers/kybarta… #procgen

With great power comes great responsibility. ;)

Jason Boog describes how he uses GPT-2 for long-form creative writing. towardsdatascience.com/how-to-fine-tu…

Mood: "I made some painful edits"

BTW this is the paper "Towards Qualitative Procedural Generation" by Mark R. Johnson, 2016, Computational Creativity in Games Workshop.
Google Scholar doesn't have it, and the workshop website is down, but you can find the paper in the Internet Archive: web.archive.org/web/2017060819…

This is a hard to find but very fun to read paper by Mark R. Johnson about the procedural culture (religion, language, beliefs/politics and aesthetics) of experimental rogue-like game Ultima Regio Regum. #procgen web.archive.org/web/2017060819…

As soundtrack I recommend Asynje, Kaunan, Forndom and Sowulo.

Instead of hunting monsters and farming metals, I run a ferry service from Our Village Island to Troll Island With Copper Mines and get paid in raw ore. This is the life.

Valheim is the biggest gaming crack I've come across in a long time. It has everything: sailing (Zelda TWW), building (Lego), exploration (Morrowind/Skyrim), farming (Stardew Valley) and coop monster slaying.

Why do you write like you're running out of time~

A new logo, the first prototype, lots of cool collabs... December and January were good months for Datakami. :)

Your fellow researchers aren't your competition, that's stupid. They're your teachers (and students!), your mentors, your hype team, your cheerleaders, your support and guidance. I wouldn't be where I am without other researchers and I acknowledge that wholeheartedly.

Dissertation writing mood.

Researchers annotated a dataset of video game development post-mortems of @gamasutra!
Pre-print: arxiv.org/abs/2001.00491
Dataset: github.com/game-dev-datab…
Now I want to poke around in this dataset, instead of writing my dissertation intro...

My ANLP students (team #2) trained BERT on the Star Wars: KOTOR game dialogue dataset to create a sentiment analyzer for Star Wars. They created 2 short surveys to evaluate the sentiment labels mined from the dataset:
forms.gle/RyUbSAQNDWPu2P…
forms.gle/jaDwLsQcqMDbEB… #nlproc

Also... did I accidentally cause a parent NPC to be released from parental authority and get separated from their child, because I unknowingly chatted with their meddlesome neighbour!? WHAT JUST HAPPENED

OK I'll admit that it also has its flaws: the game is not optimized for performance; sometimes the animal people are a bit scary; the writing is generally kid-friendly but sometimes unexpectedly switches to dark Terry-Pratchett-esk irony. :')

Eastshade is the vacation simulator that I didn't know I needed! It has gorgeous environments and walking and camping and painting and it is WHOLESOME.

My students have built a DistillBert-based text generator for RPGs Star Wars: KOTOR and Torchlight 2, and they are now looking for people who want to help evaluate their model w/ a survey. Participants are asked to rank sentences on coherence: survey.sogosurvey.com/Survey1.aspx?k… #nlproc

I miss you, dear FDG and COG conference friends! To continue my favorite coffee-break conversation online: my favorite narrative-heavy games of 2020 were Alba and Assemble with Care by @ustwogames, Pillars of Eternity and The Long Dark. What were yours?

This year my online cookbook Mosterdgeel.nl will be ten years old! I'm thinking of making a sticker pack of food illlustrations to celebrate it...

Today in "fun academic papers that fit my PhD topic AND video game preferences": content.sciendo.com/view/journals/…

TIL there's a new contemporary translation of Beowulf and I can't wait to read it. The author has translated the first word 'Hwaet!' with 'Bro!'
aworkinglibrary.com/reading/headle… via @aworkinglibrary