TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Wikipedia-grounded chatbot “outperforms all baselines” on factual accuracy

233 点作者 akolbe将近 2 年前

26 条评论

hammock将近 2 年前
“Breaking: Trivia bot trained on the dictionary spells words better than trivia bot trained on high school English papers”
评论 #36758885 未加载
评论 #36761842 未加载
评论 #36761102 未加载
评论 #36763745 未加载
评论 #36758301 未加载
评论 #36759626 未加载
charlieo88将近 2 年前
I wish I had the time or facility to take a snapshot of wikipedia now before the imminent deluge of Chat-GPT based updates that start materially modifying wikipedia is some weird and unpredictable manner.
评论 #36761926 未加载
评论 #36760447 未加载
评论 #36763223 未加载
评论 #36760406 未加载
评论 #36763546 未加载
评论 #36762072 未加载
评论 #36761045 未加载
评论 #36762618 未加载
评论 #36760431 未加载
评论 #36762910 未加载
评论 #36763717 未加载
评论 #36760579 未加载
jefftk将近 2 年前
There are multiple studies discussed on this page; the one we&#x27;re looking at is partway down the page, under &quot;Wikipedia-based LLM chatbot &quot;outperforms all baselines&quot; regarding factual accuracy&quot;. This link will take you there if your browser supports scrolling to text fragments: <a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Wikipedia:Wikipedia_Signpost&#x2F;2023-07-17&#x2F;Recent_research#:~:text=Wikipedia%2Dbased%20LLM%20chatbot%20%22outperforms%20all%20baselines%22%20regarding%20factual%20accuracy" rel="nofollow noreferrer">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Wikipedia:Wikipedia_Signpost&#x2F;2...</a>
评论 #36759381 未加载
dr_dshiv将近 2 年前
Here is a direct link to the arxiv article: <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2305.14292" rel="nofollow noreferrer">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2305.14292</a><p><i>WikiChat: A Few-Shot LLM-Based Chatbot Grounded with Wikipedia</i>
评论 #36759471 未加载
TillE将近 2 年前
The trouble with Wikipedia is that it&#x27;s an inch deep. For any given topic, especially history, there&#x27;s a trove of information in scholarly books from good publishers, but the corresponding Wikipedia article is like three pages long.
评论 #36759011 未加载
评论 #36758420 未加载
评论 #36758789 未加载
评论 #36758376 未加载
评论 #36758415 未加载
评论 #36759966 未加载
评论 #36758565 未加载
评论 #36759363 未加载
评论 #36758482 未加载
评论 #36758704 未加载
评论 #36766191 未加载
评论 #36759291 未加载
评论 #36761313 未加载
评论 #36759414 未加载
评论 #36759367 未加载
评论 #36759102 未加载
altilunium将近 2 年前
Related : <a href="https:&#x2F;&#x2F;diff.wikimedia.org&#x2F;2023&#x2F;07&#x2F;13&#x2F;exploring-paths-for-the-future-of-free-knowledge-new-wikipedia-chatgpt-plugin-leveraging-rich-media-social-apps-and-other-experiments&#x2F;" rel="nofollow noreferrer">https:&#x2F;&#x2F;diff.wikimedia.org&#x2F;2023&#x2F;07&#x2F;13&#x2F;exploring-paths-for-th...</a>
indymike将近 2 年前
This article was more about open the value of open access when publishing research being 15% more likely to be cited on Wikipedia. The AI part was somewhat weak as it did not compare against ChatGPT.
评论 #36758169 未加载
评论 #36758235 未加载
spacephysics将近 2 年前
This is great and all, but we still run into the problem with political biases embedded in the source data [0]<p>Musk’s AI’s aim is to get to the truth, not eliminate biases retroactively. I think that’s a noble goal, politics aside.<p>I agree with him that teaching an AI to lie is a dangerous path. Currently it’s probably not exactly akin to lying, but it’s close enough to be on that path.<p>We should find a way to feed source material from all “biases” if you will, and have it produce what’s closest to reality. It’s obviously easier said than done, but I don’t think the AI Czar VP Harris aims to do this.<p>If we’re too divided or hellbent on pushing our own agenda, it’ll be a bad outcome for all.<p>Unfortunately the differences we have are at a very fundamental level that really is a question of how reality is perceived, and what we consider meaningful. The difference of if something by its nature has meaning, or if we give meaning to it culturally&#x2F;societally.<p>The former is a more “conservative” (personality wise, not political) view.<p>The later is more of, “everything that has meaning is based off the meaning we say it has, thus we can ascribe the level of meaning to that or other things as we wish”. The idea that many things are social constructs, and we can change those as we wish to craft what we’d like to see.<p>I’m probably doing a poor job of wording it, but this fundamental difference in perception is going to very quickly be at the forefront of AI ethics.<p>[0] <a href="https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Ideological_bias_on_Wikipedia#:~:text=The%20authors%20found%20that%20%22Wikipedia,about%20immigration%20trended%20toward%20Republican" rel="nofollow noreferrer">https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Ideological_bias_on_Wikipedi...</a>
评论 #36759071 未加载
评论 #36760540 未加载
评论 #36763499 未加载
quotemstr将近 2 年前
Accuracy according to whom? Wikipedia is a battleground for ideologues. You can&#x27;t trust anything even remotely controversial present there.
评论 #36759667 未加载
评论 #36764089 未加载
sebzim4500将近 2 年前
Can someone update the link to <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2305.14292" rel="nofollow noreferrer">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2305.14292</a> ?<p>The headline refers to only a small portion of the linked page.
sam0x17将近 2 年前
You know you could probably get really far just training an LLM on wikipedia and all linked citations, and nothing else<p>The whole problem of wikipedia only being an &quot;inch deep&quot; on any given topic is basically solved if the LLM also has to read every cited work in full<p>And maybe citation counts could affect exposure to that work during training
评论 #36759996 未加载
评论 #36759765 未加载
评论 #36760041 未加载
评论 #36759220 未加载
评论 #36760126 未加载
edgyquant将近 2 年前
How does one even go about testing such a thing? Comparing it to Wikipedia articles? Even if it is factual does it spew the interpretations present in most Wikipedia articles?
评论 #36758425 未加载
评论 #36758268 未加载
sandworm101将近 2 年前
Lol, I wonder how many of their fact checkers silently used Wikipedia to verify the facts outputted by the AI.
ouraf将近 2 年前
I&#x27;ve been seeing a bit of buzz about &quot;grounded&quot; models. Other than setting temperature to Zero and adding better context matching (with a vector database and embeddings, for example), what other changes in methodology or implementation makes it better?
jakearmitage将近 2 年前
How did they manage to get it to stop hallucinating? I can&#x27;t prevent my llama-index based chatbot from making up absurd things from my own documents, even though I&#x27;ve been trying to restrict it to that specific area of knowledge.
ogou将近 2 年前
&quot;this first draft is probably not solid enough to be cited in Wikipedia&quot;
Julesman将近 2 年前
We can talk about alignment of LLMs. We can also talk about alignment of people who write Wikipedia. To imagine there is no bias is foolish and dangerous. More accurate isn&#x27;t truth. More accurate for whom?
评论 #36759331 未加载
siva7将近 2 年前
I would be very careful assuming that knowledge from Wikipedia is correct. Take medical articles, many of them are debatable at best and misleading at worst from a professional POV.
ec109685将近 2 年前
“All baselines” is doing a lot of heavy lifting in that sentence.
tcbawo将近 2 年前
I hope somebody makes a game of Trivial Pursuit with generated questions sourced from Wikipedia.
jokoon将近 2 年前
People on the internet still often criticizes wikipedia when I link to it, I don&#x27;t understand why.<p>It&#x27;s true that it&#x27;s not good enough for academic work (is it?), but it&#x27;s largely enough for everything else.
guestbest将近 2 年前
That value was calculated and verified using Wikipedia?
c0brac0bra将近 2 年前
But did they measure the truthiness of those facts?
freitzkriesler2将近 2 年前
[citation needed]<p>And no you can&#x27;t cite Wikipedia ;)
评论 #36758508 未加载
Outright0133将近 2 年前
&quot;Wikipedia Is Badly Biased&quot;: <a href="https:&#x2F;&#x2F;larrysanger.org&#x2F;2020&#x2F;05&#x2F;wikipedia-is-badly-biased&#x2F;" rel="nofollow noreferrer">https:&#x2F;&#x2F;larrysanger.org&#x2F;2020&#x2F;05&#x2F;wikipedia-is-badly-biased&#x2F;</a><p>By cofounder Larry Sanger
评论 #36761363 未加载
评论 #36761328 未加载
评论 #36762603 未加载
评论 #36762870 未加载
评论 #36761243 未加载
评论 #36761228 未加载
评论 #36765551 未加载
评论 #36765588 未加载
EGreg将近 2 年前
And how is factual accuracy determined? Using the exact same sources as Wikipedia, right?