Search engines don't always help chatbots generate accurate answers
- Reference: 1686155590
- News link: https://www.theregister.co.uk/2023/06/07/search_engines_make_chatbots_inaccurate/
- Source link:
Large language models (LLMs) like GPT-3.5 – the basis for ChatGPT – are trained on text scraped from the internet up until September 2021. Companies like Google and Microsoft try to augment LLMs with search engines, giving them access to knowledge in current web pages.
As demonstrated by their respective [1]Bard and [2]Bing chatbots, Google and Microsoft still struggle to produce accurate responses to search queries – even though the correct answer may be on the internet somewhere.
[3]
"One might think connecting the search engine and ChatGPT is a perfect solution, but the reality is more challenging because of the limited accuracy of search results," Hongyin Luo, a postdoctoral associate at MIT's Computer Science & Artificial Intelligence Laboratory, told The Register .
[4]
[5]
Luo explains that search engines are keyword-based retrieval systems and do not always provide direct answers to most questions. Also, different web pages might contain unrelated, contradictory, or false information. Bing incorrectly claimed [6]Adolf Hitler was a member of the band Radiohead in one search result, for example.
Netizens [7]speculated whether the error could have been caused by a [8]page on Wikidata that mentioned Radiohead and Adolf Hitler.
[9]
If Bard and Bing are to be useful, developers will need to figure out how to make LLMs extract the most useful information from a sea of text that is noisy, confusing and inconsistent. Luo and his colleagues from MIT and the Chinese University of Hong Kong believe that models need to be fine-tuned further so they can better follow instructions on how to generate responses for web search.
[10]Eating disorder non-profit pulls chatbot for emitting 'harmful advice'
[11]Texas judge demands lawyers declare AI-generated docs
[12]Red Hat promises AI trained on 'curated' and 'domain-specific' data
[13]OpenAI calls for global watchdog focused on 'existential risk' posed by superintelligence
The team tweaked Meta's [14]LLaMA , a seven-billion-parameter LLM, fine-tuning it on a database containing 52,000 pairs of text-based instructions and corresponding responses generated by GPT-4. The researchers also constructed a separate dataset containing the top five web pages associated with each instruction, and trained the model to generate the correct response by ranking the sources on how relevant and closely aligned they were with the right response.
Luo said the fine-tuned model – nicknamed [15]SAIL-7B , which stands for search-augmented instruction learning – is better at ignoring distracting or untrustworthy search results and generates higher quality answers. The details have been [16]published [PDF] in a paper released on arXiv, and the model's [17]code is on GitHub. You can also play with a [18]demo of the system hosted on Hugging Face.
"Our model learns to find helpful information from noisy search results and generate as accurate responses as possible. As a result, our model can better summarize valuable information and generate better answers for various search queries, even when search engines cannot handle them very well," Luo said.
"Our training explicitly includes a step that clarifies if each search result is helpful or not, and the language model follows the selected helpful information. This process filters out most unreliable and unrelated search results and improves the average instruction-following performance."
[19]
Initial experiments showed that SAIL-7B outperformed GPT-3.5 and other models containing more parameters at a range of tasks. The experiments assessed their abilities to answer common sense and open-ended questions, as well as fact checking, and detecting hate speech. The models were fed web pages from Wikipedia and search results from DuckDuckGo to help them pick the right answers from a list of candidate responses. GPT-4, however, was still better than SAIL-7B.
"The challenge is that larger models have much stronger knowledge, memorizing and reasoning abilities, so our model is not as good as GPT-4 yet. However, SAIL-7B is a proof of concept with a 'small' model, and our next step is training a larger model with the strategy we have proposed," Luo told us.
Models fine-tuned with the current search-augmented instruction learning technique aren't perfect, however. The researchers noted that they cannot explain why a search result is trustworthy or not. They hope to come up with another strategy to increase accuracy and reliability in the future. ®
Get our [20]Tech Resources
[1] https://www.theregister.com/2023/02/08/alphabet_bard_mistake/
[2] https://www.theregister.com/2023/02/14/microsoft_ai_bing_error/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZID@A@UcPm9aVh1OdBSIvQAAAAs&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZID@A@UcPm9aVh1OdBSIvQAAAAs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZID@A@UcPm9aVh1OdBSIvQAAAAs&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://twitter.com/jdx/status/1658876595921190912
[7] https://news.ycombinator.com/item?id=35978058
[8] https://www.wikidata.org/w/index.php?title=Q44190&diff=prev&oldid=1882635444
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZID@A@UcPm9aVh1OdBSIvQAAAAs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://www.theregister.com/2023/05/31/ai_chatbot_eating_union/
[11] https://www.theregister.com/2023/05/31/texas_ai_law_court/
[12] https://www.theregister.com/2023/05/26/red_hat_ai/
[13] https://www.theregister.com/2023/05/24/openai_superintelligence_global_agency/
[14] https://www.theregister.com/2023/02/25/ai_in_brief/
[15] https://openlsr.org/sail-7b
[16] https://arxiv.org/pdf/2305.15225.pdf
[17] https://github.com/luohongyin/SAIL
[18] https://huggingface.co/spaces/luohy/SAIL-7B
[19] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZID@A@UcPm9aVh1OdBSIvQAAAAs&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[20] https://whitepapers.theregister.com/
I agree that most search engines now return results that are vastly inferior to those that would have come up even 10 years ago (ebay included). However, I have to say that using a - (minus) before a search word does still work on ebay; but does require there to be no space between the minus sign and the word that you wish to exclude.
As for the sentence in the article suggesting that developers will need to "find new techniques to make the interaction more useful" - my response would be that unless the (so-called) AI chatbots can return results that can be relied upon to be 100% accurate and up-to-date, they are of limited use to anybody. Most likely of no use at all.
As we have already seen from the results that some people have recieved to their queries (and thinking in particular of the experience of Alexander Hanff who was informed by ChatGPT that he had died four years ago: https://www.theregister.com/2023/03/02/chatgpt_considered_harmful/), I think "of no use at all" is probably the most suitable description.
LOL, whouldcouldaknowed?
Hell, search engines don't even give the rest of US accurate answers.
And the brainiacs thought it would different for them, why?
Just a thought, but maybe they should work on the input first. You know, GIGO.
*sigh* the world just gets stupider.
You mean to tell me...
...That the same eight year old search results I get from Google because someone gamed the robots.txt and paid a bit of money are being fed into AI LLMs? I am shocked. Who let this happen? Why did they let happen? How is this even possible? When is there going to be a solution to this problem? Even more important: What is the ultimate driver behind this screwed up paradigm future generations will have to endure?
Stop calling them "parameters", call them, I dunno, "nadans" or just plain old "numbers"
> Meta's LLaMA, a seven-billion-parameter LLM
Aaargh - it is a seven-billion-near-as-dammit-arbitrary-number LLM!
A seven-billion-nadan LLM! There you are, "nadan" - doesn't that even sound more exotic and intriguing than "parameter"?
A "parameter" is an input that has understood meaning - if you "know the parameters of the problem" it means that you can identify and *describe* each parameter, *explaining* how they affect the outcome. That is why we say that a function has parameters - which we give meaningful names to - and we pass over values that we know to be sensible (well, when thing are going to right).
The LLM - and any of its relations, the varieties of Neural Nets - tweaks weights between connections as it is trained. Looking at the result, this huge pile of simple numbers is totally incomprehensible: you won't gain any useful insight by asking "Why is this 7.01" - it just is. You can't say "I want to make this change to the outputs, to do so I will change that 7.01 to 7.02", there is (as yet) no way to determine whether any given number in model even *has* an effect on the outcome (it may be blocked at any level in the layers preceding or succeeding it). Unlike a Markov Chain, the layering makes these large stochastic models totally opaque. There is ongoing research to see if this situation can be improved upon, without "damaging" the "usefulness" of these models but we are not there yet.
Rather obviously, the LLM pushers call them "parameters" - and expect everyone else to follow suit - because it makes them (both LLM and peddlar) sound cleverer than they are (old advertising trick - even when people say "oh just call them parameter, what does it matter? Don't you know language changes all the time!" the mind still attaches some of the gravitas of the word to its lesser usage).
> The researchers noted that they cannot explain why a search result is trustworthy or not.
Because they are just flinging nadans not parameters! These nets have *no* explanatory power, unlike other approaches to "AI".
> They hope to come up with another strategy to increase accuracy and reliability in the future
But, note, not to increase (well, start) providing any ability to *explain* why the result is trustworthy (or not).
Web search engines, including those on trading site, universally fail to perform as well as at least one I used back in the '80s. They simply OR all the search terms together, even if terms are entered in an attempt to exclude*. At best they may use the combination of terms to partially order the results. To use the results of the search engine the chatbot would need to parse the query properly and then use the parsed query to filter the results of the search engine. It is, of course, the search engines that should be doing that parsing and filtering.
* eBay's search used to respect minus signs so that "car - Honda" would return cars of any make except Honda. Then they changed it so it would return all makes of car including Honda plus Honda motorcycles and any other Honda non-car products. Now it seems to return only Hondas and even provides check-boxes for Honda models.