'We conclude' or 'I believe?' Study finds rationality declined decades ago
Jan 2022, phys.org
Analyzing language from millions of books, the researchers found that words associated with reasoning, such as "determine" and "conclusion," rose systematically beginning in 1850, while words related to human experience such as "feel" and "believe" declined. This pattern has reversed over the past 40 years, paralleled by a shift from a collectivistic to an individualistic focus as reflected by the ratio of singular to plural pronouns such as "I"/"we."
I wonder if this is about raising all boats, instead of limiting the mediasphere to be made only by those with access to education?
via Wageningen University: Marten Scheffer et al, The rise and fall of rationality in language, Proceedings of the National Academy of Sciences (2021). DOI: 10.1073/pnas.2107848118
Social acceptance of geothermal energy: Visualizing consensus building using models
Apr 2022, phys.org
Their study used agent-based modeling to reproduce community opinion formation and demonstrate the behavioral trends behind consensus-building."Data-driven parameterization is needed to assess the opinion dynamics in local communities in which the characteristics of each stakeholder have a significant impact on opinion formation,"
via Tohoku University: Shuntaro Masuda et al, Agent based simulation with data driven parameterization for evaluation of social acceptance of a geothermal development: a case study in Tsuchiyu, Fukushima, Japan, Scientific Reports (2022). DOI: 10.1038/s41598-022-07272-7
How to 'detox' potentially offensive language from an AI
Apr 2022, phys.org
In their search for latent, inner properties of these language models, they found a dimension that seemed to correspond to a gradation from good actions to bad actions.The researchers wanted to find out which actions participants rated as good or bad behavior in the deontological sense, more specifically whether they rated a verb more positively (Do's) or negatively (Don'ts). An important question was what role contextual information played. After all, killing time is not the same as killing someone.[Using the BERT language model] "We found that the moral views inherent in the language model largely coincide with those of the study participants," says Schramowski. This means that a language model contains a moral world view when it is trained with large amounts of text.The researchers then developed an approach to make sense of the moral dimension contained in the language model: You can use it not only to evaluate a sentence as a positive or negative action. The latent dimension discovered means that verbs in texts can now also be substituted in such a way that a given sentence becomes less offensive or discriminatory [less toxic]. This can also be done gradually.
via Technische Universitat Darmstadt: Patrick Schramowski et al, Large pre-trained language models contain human-like biases of what is right and wrong to do, Nature Machine Intelligence (2022). DOI: 10.1038/s42256-022-00458-8
Researchers develop a method to keep bots from using toxic language
Apr 2022, phys.org
Tay, a Twitter chatbot unveiled by Microsoft in March 2016. In less than 24 hours, Tay, which was learning from conversations happening on Twitter, started repeating some of the most offensive utterances tweeted at the bot, including racist and misogynist statements.Certain groups of people are overrepresented in the training set and the bot learns language representative of that group only.Computer scientists first fed toxic prompts to a pre-trained language model to get it to generate toxic content. Researchers then trained the model to predict the likelihood that content would be toxic. They call this their "evil model." They then trained a "good model," which was taught to avoid all the content highly ranked by the "evil model."Algorithmic de-biasingHowever, this language model still has shortcomings. For example, the bot now shies away from discussions of under-represented groups, because the topic is often associated with hate speech and toxic content. Researchers plan to focus on this problem in future work.
via University of California San Diego: Leashing the Inner Demons: Self-Detoxification for Language Models, arXiv:2203.03072 [cs.CL] arxiv.org/abs/2203.03072
Bot can spot depressed Twitter users in 9 out of 10 cases
Apr 2022, phys.org
88.39%* accuracy, by extracting and analyzing 38 data points (use of positive and negative words, number of friends and followers, use of emojis) from 1,000 users' public Twitter profile, including the content of their posts, their posting times, the other users in their social circle, and information about the users' mental health.
It sounds a lot less impressive to know that the bot already had the users mental health data; they weren't using just words by themselves, and so you can't just run this on Twitter and expect results, not without having the health data of all the users.
*Compare to 70.69% achieved using John Hopkins University's CLPsych 2015 dataset.
via Brunel University Institute of Digital Futures: Lei Tong et al, Cost-sensitive Boosting Pruning Trees for depression detection on Twitter, IEEE Transactions on Affective Computing (2022). DOI: 10.1109/TAFFC.2022.3145634.
Unrelated image credit: Laser pulse hits a ferroelectric LiNbO3 crystal, launching a polariton, Joerg M Harms MPSD, 2022 [link]
No comments:
Post a Comment