Hacker News Reader

Comment by ragazzina | Hacker News Reader

>“We want to make sure that we’re not wasting the business’s time,” Fox says. “We want to make sure throughout everything we do here, that this is good experience for the business and that they’re not getting frustrated talking to an assistant while they’re trying to run their business.”

What's the added value of sounding exactly like a real person then? Why can't it be a robotic voice following a simple, efficient algorithm?

- Hi. I'm a robot trying to book for 4 at 20 tonight for my client. Is it possible? Please say Yes or No

- Wait what?

- I'm a robot trying to book for 4 at 20 tonight for my client. Is it possible? Please say Yes or No

- No

- Is it possible at any other time tonight? If yes, when?

- Yes, at 21

- I would like to book for 4 at 21. Thank you.

We have had this kind of technology for years.

View on HN

Because people will hang up on it like they do all other robocalls, or they will talk to it like many people talk to "computers". They won't use grammar, they will speak louder, they will slow down, and they will try to keep the conversation as simple as possible fearing that the system can't handle anything else.

It needs to sound human so that humans will interact with it correctly. If it sounds human, then the person on the other end of the call might say things like "We don't have anything available on the 3rd but we do on the 5th, is that okay?"

If it sounds like a robot, people will just respond to questions with "no", "yes", or say "phone number" instead of "can i have your phone number?"

Sounding human puts the human in the drivers seat, sounding robotic means the "robot" needs to drive the conversation.

or they will talk to it like many people talk to "computers". They won't use grammar, they will speak louder, they will slow down, and they will try to keep the conversation as simple as possible fearing that the system can't handle anything else

So, pretty much how we've trained ourselves to use Google, the search engine.

Nobody types "I'd like to know what the best way is to braise pork with white wine" into Google. They type "pork recipe braise wine" because we've been trained by Google that this will produce better results.

If Google's current search worked as well as it claims Duplex will, then we'd finally be getting somewhere. But Google's taken its eye off the ball.

Right now, at least for the things I search for, searching on Google is like playing an Infocom game.

> Nobody types "I'd like to know what the best way is to braise pork with white wine" into Google

No, but lots of people say stuff like that to Google Assistant all the time, and Google is tuning their search engine to accommodate queries like that:

https://www.google.com/search?q=What's+the+best+way+to+brais...

The above query results in a card displaying step-by-step directions on how to braise pork with white wine. Maybe Google's better at this than you thought?

Name a search engine that's better than Google. Nobody type "I'd like to know what the best way is to braise pork with white wine", doesn't mean Google can't give you good search result if you typed like that, you will get just as good result if you type "pork recipe braise wine", but its fewer words and more efficient to type.

It's less of "google trained us to type like that" and more or "We opted to use the more efficient way of doing that same thing". If Google didn't give us good results if we typed "pork recipe braise wine", it would be a pretty shitty search engine in my book.

Name a search engine that's better than Google

I never stated that Google isn't the best search engine out there. I stated that I hoped the natural language learning from Duplex would filter down to the Google search engine.

But while we're on the topic, find a search query in which Google will return a list of movie reviewers in the city of Chicago. It can't. It is so focused on second-guessing search queries that it will only return lists of reviews of the film Chicago.

I had this exact very frustrating experience once trying to find the name of a woman I met who publishes a popular movie review blog in Chicago. There's even a "guild" of sorts of movie reviewers in Chicago, but Google could only surface that as high as the 10th page of results.

I know I'm not the only person frustrated with Google's desire to second-guess my searches because I heard a comedian on the radio last week who did a bit that lasted five whole minutes on Google second-guessing searches.

Kind of tangential to this, but I was really amazed at the accuracy of Google's search results when a few years back I tried to find the name of a song I'd heard at a nightclub in Hong Kong. Neither did I know the artist, nor really any part of the lyrics (because it had been very loud and crowded), so I was pleasantly suprised when Google Search took my query "guys in jean overalls singing too ra loo ra loo" and actually returned "Come on Eileen".

I feel that we have such a high standard for google search result quality that we can get frustrated by such highly specific search result edge cases. Google search is not perfect, but its the best search engine out there by a long margin for almost 20 years.

There will be a subset of users who will bemoan user privacy but expect highly specific personally tailored quality search result. You can't have both and there will always be edge cases in such a difficult domain. The fact that so many industry heavyweights can't compete with Google in search result quality, says a lot about how good Google is. Does it has problems? Yes, no one is denying it.

we can get frustrated by such highly specific search result edge cases

Every case is an edge case.

"Edge case" has become the tech industry's excuse for everything.

Every case is not an edge case. You can't be serious if you call of the following queries as edge cases -

1. Chicago

2. Chicago movie

3. Chicago movie review

4. Movies in Chicago

5. Movie reviewers

6. Movie reviewers in Chicago

7. List of movie reviewers in Chicago.

If you still feel so, you may be talking about edge case in non-technical sense.

Disc: Googler but nowhere close to search or duplex.

1. Chicago

But with Google, there is no pure search for "Chicago."

There is only a search for "Chicago" by someone at a particular location with a particular device using a particular browser with a particular search history with a particular number of other parameters that Google's pieced together that we don't even know about.

So, I'll be more specific: Every Google search is an edge case. Google spends billions making sure no two searches return identical results.

"movie reviewers located in Chicago" returns the Chicago film critics association as the second result.

For you. Right now. But when I did the search about a year ago, the results were as described above.

I'd say your issue couldn't be reaperduced.

Well, looks like your problem is fixed then.

> It's less of "google trained us to type like that" and more or "We opted to use the more efficient way of doing that same thing".

I'm a bit baffled at this, given it reads as though the past 20 years of searching hasn't been learning to adapt to what one might call search engine speak for better results. The obvious main issue was for years you had to guess how someone else would phrase something if using longer phrases, since the longer the query the less likely you'll find a matching result. If you didn't get the desired results you then would either re-phrase it to yet another natural sentence or (increasingly since it shortcuts the process) pare back the query to simpler keywords to have a better chance of success finding those same core keywords in sites.

Eventually people got used to just entering keywords instead since it wasted time trying one's luck on more natural sentences for the reasons above. Obviously with better natural language tech (and with more user content generated and indexed) some queries became more successful but even today it can be a complete dice roll for many longer, naturally phrased queries whether anything useful is returned.

This might be true for a small subset or early or techie users, but from my experience watching some of my less tech-inclined friends and family using google search. I find it fascinating how much trouble they go through by typing a complete sentence with proper grammar and spelling. I honestly don't know which one is common, but I wouldn't be surprised if a majority of users type full sentences or long search queries.

In my experience the results are worse with full sentences, just as it was described.

Actually both are true. Google (or earlier search engines) did work on basis of keywords.

The ability to handle full sentences well came much later in one of google's major search updates - I think Hummingbird which came out in 2013.

Basically everyone in my office types full sentences to Google, and it drives me crazy.

Did you actually try the former query? It works pretty darn well now -- for me, it brings up recipes (and your HN comment, of course).

Disclaimer: My personal opinion, not that of my employer.

You are comparing a generic computer to the best human conversation. But people make phone calls when driving, make phone calls with background noise, make phone calls while speaking to other people, make phone calls even if they don't know exactly what they want and waste the businesses time, all the time.

They will hang up or be confused the first time. As soon as this becomes common, they will get used to it and actually waste way less time than with the usual human call and will look forward to the familiar efficient robotic voice instead of the third prank or drunk call of the night.

>It needs to sound human so that humans will interact with it correctly.

I disagree. It needs to sound understandable and easy to use so that humans will interact with it correctly. Making it more "human" is orthogonal to that. All the replies to my comments have something to do with interacting in a more human way. Does it ever provide value for simple use cases? I'm honestly asking.

(If I am calling to book a table and the restaurant suggests me another day as in your example, I would be as stumped as the computer and would need to check my calendar and call them back, so your human sounding assistant is providing no value anyway. Is that even a real use case?)

I don't know, maybe a understand-all-manages-everything computer would be used more happily by people, but I'm not convinced. For these simple use case, I still think a binary algorithm would be, on average, more efficient for both the caller and the business.

You say

> We have had this kind of technology for years.

and

> For these simple use case, I still think a binary algorithm would be, on average, more efficient for both the caller and the business.

The fact that we have had this technology for years, and nobody has been able to make it work as you describe, suggests that it's not as simple as you think.

My take: sometimes you can persuade users to learn a new UI, but consider the stereotypical grandparent that still can barely use email; teaching a new UI is hard, normal non-tech people don't like it, and it will take more than a generation to reach 100% penetration.

The existing UI of using a telephone and conversing like a human is probably not at 100% penetration either (e.g. considering deaf individuals), but I'd wager it's as close as any one UI is going to get.

> Because people will hang up on it like they do all other robocalls

A business that is essentially closing a sale with customers? No way, the call might not be from a real human but the money that the reservation will bring in is no different than any other.

Real humans don't behave like that. They behave in all sorts of weird ways when they know that they're interacting with a machine. They SHOUT, O-VER-E-NUN-CI-ATE and use unnatural sentence fragments, which craters the accuracy rate of a speech parsing algorithm that was trained on natural speech. They get instantly frustrated, because they assume that this machine will be just as stupid as the archaic IVR system that their bank uses. They assume it's a scam robocall or a prank and hang up without saying anything.

Acting like a human is by far the most efficient option for a system that is designed to do what Duplex does.

How are you supposed to tell whether an automated system is intelligent or not? Sometimes the voice prompt tells me to "speak in full sentences", but it doesn't understand what I'm saying anyway.

Google solved this by having it talk to you naturally so you never even consider responding differently. Seems like a winning idea.

Only non native speakers think duplex sounds human

I'm native and I think it sounds pretty damn close. I could see myself getting through a simple conversation with it without getting suspicious (that's assuming that I had absolutely no reason to believe it was a system from the start)

How are you supposed to tell whether a human is intelligent or not?

Why wouldn't the same process hold here?

That would be a fault of the system then.

That would be wasting time, because humans are used to talking to humans, not to robots. Do you enjoy it when you call a business and you have to sit through the annoying automated system? If you want to make the experience better for businesses, this is the way to go, although much harder to achieve. The "um"s aren't to trick the person, it's actually makes the conversation flow better.

If you try calling a restaurant and say "Hi. I'm trying to book for 4 at 20 tonight for my client. Is it possible? Please say Yes or No", I suspect you'll discover that it's in fact not more efficient than a typical human conversion.

Would you want Siri or the personal Google Assistant to talk this way?

A strict binary tree structure chirped at you in low-fi synth voice like this is really not that pleasant or efficient.

...because technology should adapt to humans, not the other way around.

These voice systems are annoying as hell, as shown by the fact that they have become a comedy trope similar to airline food.

It's bad enough to deal with them when you're initiating the call. Imagine being a business forced to interact with such a system because it brings in too much revenue to ignore.

I'd definitely spit in every soup served to a guest willing to annoy me just to save himself the hassle of a phone call.

Thats assuming simple restaurant reservations are the end game. I'd be willing to bet that they will be expanding the utility of this far beyond tasks like these. This is just a way to get the right kind of training data for future projects and products, while providing value to google users to keep them in the ecosystem.

Because talking to robocall like that is annoying as hell. Not to mention fail rate is high (at least for my voice).

what @Klathmon said. If I am business , I will hang up at "hi I am robot..."

Why? People call businesses all the time and get robots, why would it be different the other way around? As soon as the boss figures out you're hanging up on people trying to get reservations you can bet hangups will stop ASAP.

Most of us hang up on robots too though. Robocalls are a nightmare. The best argument against Duplex IMO is that it’s going to enable a new generation of hellish scams and robocalls, if it ever leaves Google. Presumably Google knows this and will keep the tech locked down. Still, it’s only going to be a matter of time before it’s open one way or another, and then we’ll have to have the “should I be told up from that it’s a robot?” conversation again.

Most of us aren't paid to answer the phone though... If my job was to answer the phone for people making reservations and I started to hang up on reservation bots I would expect to be fired immediately. I also hang up on real people who call me unsolicited for sales pitches, it's not the bot part that annoys me.

Agreed. I would hang up at "This call is being recorded". No. I don't want this. How dare you just record me out of the blue without consent?

Why don't you want your business to get money for your products and services?

Reply on news.ycombinator.com