Text By the Bay 2015: David Murgatroyd, Identity Resolution in the Sharing Economy
Recording: Text By the Bay 2015: David Murgatroyd, Identity Resolution in the Sharing Economy
so uh yeah thanks a lot i am dave i'm the vp of engineering at basis technology and just a very brief thing to kind of get your head in the game of who basis is so we are at the basis of a lot of different folks nlp stacks we build a lot of sort of low-level nlp functionality entity extraction resolution we've got classification and sentiment in beta now and other technologies here that are going to be relevant to this kind of application oriented talk one of the cool things about being kind of at that this basis level is we get to see a lot of different verticals a lot of different uh how folks use it and so i'm going to be talking about the sharing economy here so i'm the second dave that stood in this place and talked about sharing economy dave holtz from airbnb uh talked about a little earlier on so sharing thrives on trust this is a point that's been made a few times today it thrives on reputation and nlp text analysis has a part to play in establishing that trust it may not be everything that we need there's other data you need too but when you have text you should take advantage of it i want to do a little bit of a history of the sharing economy so it actually goes back to ebay right so back in 2005 when ebay celebrated their 10th anniversary so now it's their 20th anniversary they had already found that 135 million people have learned to trust a complete stranger so the sharing economy has actually been going on for a while but technology has changed enough to make it accessible to a lot more use cases than just online auctions and as dave holt said earlier today the key dynamic or a key dynamic of the sharing economy is it's not so much that we have trust for institutions anymore we still do but we want to establish trust for our peers we want to establish reputation so that has helped companies like airbnb have huge growth and the number of guests that they've had stay the breadth of uh places where they have hosts so been a lot of cool growth there another uh sharing economy kind of success story this is blah blah car which isn't exactly like uber uber you can think of as more sort of a pseudo taxi blah blah car is in the uk primarily i think they're moving into the u.s and the idea there is you know i'm driving from edinburgh to london would you like to pay for a seat in my car and maybe then it's not so much that you just want to make sure that the person you're getting into the car with isn't going to kill you you might also want to make sure that they don't have like annoying taste in music it's like that uh i don't know if you've seen the rideshare sketch from portlandia on ifc but it's kind of like that so there's a lot of other sharing economy apps that are out there uh so flight car is one where you can leave your car at the airport at sfo and someone will instead of you having to pay for parking they will rent that car from you while you're on vacation uh neighborhood goods is where you're sharing kind of tools with people in your neighborhood you may be familiar with taskrabbit tasks that are going on um air pooler is you can hitch a ride with a pilot who's flying somewhere else in his or her small plane so maybe you can avoid having to play pay commercial fare by hitching right there and my favorite here is the one in the bottom right hand corner air pnp so air pnp is if you need to do number one or number two and you don't have a place to do that and you're willing to pay for it so actually the closest place right now to do your business in the cloud um this is actually one in boston there's one over uh uh south uh in mid market i think i'm like a few blocks away 250. so um so yeah it's a it's a pretty entertaining one actually the place that this really got big or as big as it got is for mardi gras in new orleans where you have a lot of you know folks walking down the street who put a lot in and not necessarily had the opportunity to get a lot out so so it's it's kind of fun to think about the variety of things and what kind of trust would you need to use someone's bathroom right um in fact for the sharing economy apps we sort of think of it as these different apps being at different points of the trust adoption curve and sort of their life cycle so like air pnp you know those of us that are feeling particularly adventurous maybe that's the kind of thing that we want to check out but ebay you know even grandmothers use ebay right so those safety mavens they're pretty comfortable and at least in my view maybe airbnb and home away are you know a little a little bit past the curve maybe uh blah blah cars a little bit earlier doggie vacay so you could think about you know whether you agree with this or whether there are other sharing apps and where you kind of put them on this curve and really think about which of these buckets you would consider yourself to be in you know your own risk tolerance for sharing a ride for five hours for someone or jumping in the back of their uber or staying on their couch versus in their you know separate apartment so the point of this talk is about how text is valuable to establishing trust in the sharing economy i want to start out talking about vetting talking about using sources outside of the sharing app to vet and to verify identity so dave this morning talked about using reviews within the sharing app and how to establish reputation with that so i'm talking about using data outside of that sharing app to establish trust even at that initial moment that you've signed up so there are different sources for this you can imagine someone saying well should we look at your fico score maybe not so great because fico's about credit history not so much about what kind of guest you're going to be offline so i don't know how many of you have had the experience of holding your driver's license or passport up to your webcam to verify your id for airbnb so i did that and then also online so you can you know link your airbnb account to your facebook account and other social media accounts i think maybe your linkedin one and this in turn becomes a signal of trust in the airbnb and other analogous uh apps you know view of a host so this is the actual woman uh whose apartment i stayed in back in february in puerto rico with my family and she had her verified id she had linked her passport there she also linked facebook so she had some good reviews so this all helped me feel comfortable in having my wife and two little daughters and i stay at her place now you might want to know so where does text fit in here and this is actually something we know firsthand because airbnb uses one of our text products as part of the pipeline i'm about to tell you about so when you held that driver's license up to that webcam what happened so a picture was taken of that and the name on that was ocrd off and other things about it like the shape of it and other images on it i'm sure we're sort of vetted to look like you know is this a valid id but the part i want to talk about is the text part so the text part uh maybe in this case instead of jane citizen it might have pulled like you know miss i recognize the t as a y here for the ocr uh filter in then sitting there on your airbnb profile dear jane citizen maybe you actually had janie and you spelled your citizen name correctly so this is a case where having something as simple as name matching actually just like jeff was talking about for the ancestry.com case this maybe has some other complexities to it this is an important capability it's one that airbnb uses right now it was kind of cool when i held my uh driver's license up to know that my code was running behind the scenes um but you might need to go beyond that so if janie's actually a host and she lives in seoul and the property record for her apartment is in korean you may need to go one more step and actually use a cross language name match so even something as simple as name matching can be valuable in connecting these different sources of online and offline identities so uh i want to talk about what else you could do with text so not only can you use names to match ids versus profiles you might also have heard about this story from last summer about these airbnb squatters so a host her name is corey had a guest named maxime no offense to other maxines in the audience who actually ended up squatting at her airbnb apartments for much longer than he had paid for so here's the timeline so maxime showed up on may 25th last year and he had made a reservation to stay till july 8th and he paid till june 24th you pre-paid that then on june 24th airbnb tried to collect a balanced due from maxime and that attempt to collect did not succeed so the host texted maxime and like two weeks after that and said your reservation is over i'm gonna cut off your power and maxim replied i guess he was aware this was in palm springs so he was aware of the laws in terms of like occupancy i think once you've been somewhere for 30 days if you can show that being evicted affects your livelihood then there has to be this whole court order that goes through it i don't understand all the legal details apparently he did so he threatens to press charges corey takes this to the press and there's this legal eviction press coverage finally he gets out of there so this obviously is a case where you know not a happy thing right this is if cory could go back and say no you cannot come maxime that would be great now could that trouble have been predicted well it turns out the same guy had swindled people on kickstarter out of forty thousand dollars like three months earlier here's the timeline for kickstarter so back in 2013 so whatever that is um you know seven eight months before he came yeah his reservation started with with corey he hit a kickstarter goal for a video game that he was supposed to be working on so he had two progress updates in november then one in december then just one over the course of two months and then there had been three months at the on the day that he checked into this airbnb with corey since his last kickstarter update so what could have happened if his airbnb account was linked either purposely by his opt-in or by some monitoring of publicly available kickstarter data could corey have seen well this guy has a release goal coming up in like you know i guess arguably a few days at the beginning of june or at least sometime in june and he hasn't made an update since february 28th she might have decided to request the full payment in advance or maybe she would have decided to cancel his reservation or at least kind of ask him about that or possibly airbnb and there's there's this whole question in the sharing economy of what information do you push to whom when but possibly airbnb could have said oh well this doesn't look so great and maybe they would have uh flagged something to their own internal employees to kind of check up on maxine now there's a lot of other names that maxime actually use through kickstarter so here's a kickstarter backer named chris chen and if you go on you go today onto the kickstarter page for their video game you can see this they're like stay away from this guy he's using all these other name variants so if you could have used some kind of text analysis to connect these names across these different accounts across these different places that's manifesting you might have been able to to flag that a bit earlier so that's that's another kind of application of text analytics that can help the sharing economy not just verifying identity but also connecting accounts and connecting information to try and say well if on this site the person's reputation score is going down maybe we want to know about it over on this other side so here's another one and this one is uh the story is real but the connection to airbnb is false um so just keep that in mind i'm not saying that this guy described in this article actually came to stay but this is a story about a guy named terry who was arrested for being a peeping tom now you could imagine some sharing economy sites monitoring news looking for negative news about their account holders so here and this is stuff that text and text analytics can do now you know we can find names we can connect names to database entries we can learn that omaha okay this guy's address is in omaha nebraska we can know that a 46 year old man was born at least at the time of this article in 1962 so if you had terry's profile on whatever sharing economy website and you had a news monitoring service you might be able to connect those together and look for this kind of negative news and be able to say hey you person who has this reservation here's some negative news you know use some sentiment analysis to realize that this story is not necessarily a positive one and especially the way it treats terry is not positive so um so that's a a challenge uh to the sharing economy apps is how do you get this data to connect all together across these these different sites and i think nlp um is a big answer to that so there are some some challenges with this obviously some pieces of this puzzle are you know really hard text analytics tasks to do at scale um like entity linking entity resolution relationship extraction that stuff that we're working on and some of which we have products for but getting it getting it really good so that it's good enough to maybe cancel someone's airbnb reservation is a challenge there's also this question of what to show to users versus the trust and security team at one of these companies and how to allow users even to indicate their own trust and safety like when do they want to be notified what's their own risk tolerance like and how do you you know present this information that says well the kickstarter campaign isn't looking so good especially in the case that information can sometimes enable discrimination so we got to be careful that we don't just give people an excuse to you know be you know be discriminatory for reasons that really aren't justified i mean you can look at you know this is just one report on terry there are other studies that say that subconsciously we have these different biases as much as we hate it so um so these are some some questions here you know if people are opting in uh what expectations do they have for the use of the data that they're connecting if i connect my facebook account to my airbnb profile what what uh expectations do i have and if i do go monitor the news for the people who have signed up with my sharing economy app or if i do connect the kickstarter campaign um should i feel differently about that than if the host just went and googled whoever was coming to stay with them i'm not an airbnb host i'm just a guest but i don't know maybe hosts do regularly kind of google the people who are coming to stay with them to kind of add that layer well maybe they would like the sharing economy app to do that for them um so so these are some questions to deal with and they're kind of policy ethics questions and i realize that's not the main thing we're here today to talk about but we want to think about that that bigger picture as well as as the good deep learning stuff that we're talking about and the other uh you know nice text analytics algorithmic business value kinds of questions so um so those are two places kind of where text analytics can come in both in doing initial vetting when you've signed up with your id and also providing vigilance when you're looking at news sources when you're looking at sources on other sites one other that i think is an interesting place to think about is for these sharing economy apps that's and i guess maybe tinder isn't quite a sharing economy out but a place where people more have shades of preference than they do like no i do not want that person to stay with me and to think about it's almost like you know approaching it's like going from airbnb to eharmony.com or something right what what are the shades in there for different sharing economy apps that that you might want to identify for the blah blah cars of the world where you're going to spend five hours with the person maybe you want to know a little bit about their personality um so there's this vision in describing character you know like the thing on the left which is oh there's you know just sort of a straightforward enumeration where you can say what kind of person this is or what the connection is between them but really it can be pretty complicated like the enneagram on the on the right and um so it's something that's pretty hard to model so i'm just sort of putting this out there as kind of a challenge for us what's the right way to to handle it you know how do we elicit these fine grain preferences how do we make those kind of preferences visible to allow people to make informed decisions um and how do we respect those preferences but also not be discriminatory in the app itself so that's that's definitely a challenge and not enable discrimination now you could imagine using text to try and characterize things like sentiment analysis or attribute analysis of the characteristics of the author of the text or expressing it explicitly and you might combine those using thresholds on different traits collaborative filtering there's a lot of things that could could play in there so in conclusion what i wanted to talk about was how text analytics can help the sharing economy so there's things like name matching for doing vetting of ids there's things like entity resolution entity linking for pulling in negative news for pulling in other websites and maybe even there's things like sentiment analysis or relationship extraction to get those finer grained attributes of character to kind of align with preferences that might be important for summing some of those sharing economy apps so with that i'll say thank you and open it up for questions great anybody want to talk about this stuff all right thanks a lot everybody you