0:00:15 | and the hybrid hamming layer i it from a liberal and i'm here to present |
---|
0:00:19 | a data set i collected and annotated with my colleagues at a little bit |
---|
0:00:25 | highness is actually here with me if you want to talk to him |
---|
0:00:30 | so |
---|
0:00:32 | there is a the motivation behind this dataset is that there is indeed |
---|
0:00:36 | for dialogue systems to be able to handle complex interactions |
---|
0:00:41 | one motivation comes from studies and e commerce and there is a paper by month |
---|
0:00:45 | later in twenty eleven |
---|
0:00:47 | where they show that users that come to an e commerce website problem sometimes with |
---|
0:00:51 | a very well defined cool |
---|
0:00:54 | in mind but sometimes they just come to shop around or they don't really know |
---|
0:00:58 | what they want just one to look for options |
---|
0:01:01 | there is also |
---|
0:01:03 | sorry some interest in the dialogue community and most notably there was a paper last |
---|
0:01:09 | or it's a dialogue by finding the and i mean distance |
---|
0:01:12 | i think it was based that's papers |
---|
0:01:14 | last year |
---|
0:01:15 | it's that has any idea that the state tracking for flexible interaction |
---|
0:01:19 | and is this in this paper they try to move a beyond the traditional |
---|
0:01:24 | linear slot filling paradigm and try to handle more complex |
---|
0:01:29 | conversations where you have different user goals and possibly across domains |
---|
0:01:33 | so we decided are so for this work actually didn't have a proper dataset to |
---|
0:01:38 | test their method because they there wasn't anything available |
---|
0:01:42 | so the |
---|
0:01:44 | modified an existing data set and so we decided to actually try to collect data |
---|
0:01:49 | and promote this kind of work for future dialogue systems |
---|
0:01:55 | so we collected one thousand two hundred and sixty nine human-human interactions and the travel |
---|
0:02:00 | domain |
---|
0:02:01 | we also propose a new time frame tracking and the dataset is fully annotated and |
---|
0:02:06 | publicly available at this url |
---|
0:02:11 | so when i talk about linear slot filling what i mean it's something like this |
---|
0:02:16 | is actually here dialogue from the dataset |
---|
0:02:19 | and here and so the user basically gives you some constraints you want to go |
---|
0:02:24 | somewhere from columbus it doesn't really know where |
---|
0:02:26 | then the wizard is the agent two plays the role of the dialogue system |
---|
0:02:31 | he proposes two options vancouver draw no then the user gives a bit more information |
---|
0:02:35 | about his constraints |
---|
0:02:36 | and then at the end of day and then the user asks |
---|
0:02:39 | for information about the offers from the wizard |
---|
0:02:42 | and that the and the user box the |
---|
0:02:45 | one of the proposed trips |
---|
0:02:46 | so here the user will never really changes during the dialogue it's very just drilling |
---|
0:02:51 | down some options |
---|
0:02:53 | and by nonlinear slot filling i mean something like this dialogue which is also from |
---|
0:02:58 | our data is that it was able to onto to support entirely on the slides |
---|
0:03:02 | are just cut the interesting part |
---|
0:03:04 | so here this is a representation of the different options that the user |
---|
0:03:09 | see the mouse you can okay |
---|
0:03:12 | so on the left |
---|
0:03:13 | the this is a representation of the different options and goals that the user might |
---|
0:03:18 | have during the dialogue |
---|
0:03:19 | so by nonlinear slot filling what i mean is that at the beginning the user |
---|
0:03:24 | is talking about or in some going to toronto |
---|
0:03:27 | and then and he explores a options and i think in green |
---|
0:03:32 | but at the end of the dialog the actually decides to go back to that |
---|
0:03:36 | you're on a trip and then |
---|
0:03:38 | so in this case |
---|
0:03:39 | and the user goal changes during the dialogue but the user also goes from one |
---|
0:03:43 | able to the other and if we want to be able to actually broke the |
---|
0:03:47 | drawing a package for this trees are we need to remember it |
---|
0:03:51 | so let's that of into the details of the datasets freeze the domain so it's |
---|
0:03:55 | a travel domain we had trouble packages with a round trip flight and a hotel |
---|
0:04:00 | this is an example of a package so you had you hold our |
---|
0:04:03 | the flights with their time and the dates |
---|
0:04:07 | and for the hotel we had are the category which is the number of stars |
---|
0:04:11 | we also have guessed readings on a scale of and |
---|
0:04:14 | of one to ten and amenities and vicinity so |
---|
0:04:18 | on the rows |
---|
0:04:20 | those are the first one is |
---|
0:04:22 | a bit too small to read |
---|
0:04:24 | but it vicinity so vicinity of the hotel you have something like shopping malls museums |
---|
0:04:31 | but is universities airports et cetera so that |
---|
0:04:36 | the distribution |
---|
0:04:37 | and on the o |
---|
0:04:39 | a button graph we had the number of amenities burr hotels so the amenities could |
---|
0:04:45 | be breakfast wifi |
---|
0:04:47 | whether the what has a spot those kind of things |
---|
0:04:50 | and so that for most hotels we have more than one and automatically so that |
---|
0:04:55 | the users |
---|
0:04:56 | had something some ground |
---|
0:04:58 | some matter to compare to what else one against each other |
---|
0:05:02 | and we had two hundred and sixty eight hotels and one o nine cities in |
---|
0:05:06 | total |
---|
0:05:10 | so for this dataset we hired |
---|
0:05:12 | twelve participants to collect the entire data |
---|
0:05:17 | are over twenty days don't our data collection last |
---|
0:05:20 | the twenty day i'll for of the participants |
---|
0:05:24 | it entire data collection and the other ones where hired for just one week |
---|
0:05:30 | and each dialogue was performed ugly a chat on slack |
---|
0:05:34 | so we had about that was a pairing up to user is |
---|
0:05:39 | and then they can they were able to chat so when the user what spare |
---|
0:05:43 | to a wizard you would get a task |
---|
0:05:45 | and we generated those that is based on templates like this one |
---|
0:05:49 | so are basically we tell the user his goal |
---|
0:05:52 | and to generate those are tasks from the templates we just replace the placeholders for |
---|
0:05:57 | the different entities with values that we randomly true from the database |
---|
0:06:02 | and |
---|
0:06:03 | two very the task |
---|
0:06:05 | we actually |
---|
0:06:07 | word error probability for each template |
---|
0:06:10 | so for this template would say |
---|
0:06:12 | and has a probability of additive |
---|
0:06:15 | point five to succeed |
---|
0:06:16 | so that means that when we actually wary the database with the entities |
---|
0:06:22 | well fifty |
---|
0:06:23 | present of the time it will every turn results and fifty percent of the time |
---|
0:06:27 | it want to return results |
---|
0:06:29 | and when it won't return results we would give to the user we would either |
---|
0:06:32 | tell the user to close the dialogue |
---|
0:06:34 | or we would give him some alternative like if nothing much easier constraint then tried |
---|
0:06:39 | increasing your budget by twelve hundred |
---|
0:06:41 | dollars |
---|
0:06:42 | so as i said we only had twelve participants and we collected a bit more |
---|
0:06:46 | than a thousand dialogues |
---|
0:06:47 | so to keep it interesting for them |
---|
0:06:50 | we tried to tell them to play roles and try to very the way they |
---|
0:06:55 | speak to the to the wizard and to anchorage just a bit more we also |
---|
0:07:00 | growed sound fine |
---|
0:07:02 | templates like this one so that was at the time when pocket mango was very |
---|
0:07:08 | popular so we told them to pretend that there are pokemon hunter and they're really |
---|
0:07:12 | wanna go to the city because there is a very rare pokemon there and that |
---|
0:07:16 | they should find a good package to do that |
---|
0:07:20 | so |
---|
0:07:21 | to keep it interesting we are created such templates and we then kind of |
---|
0:07:26 | throughout the day data collection so that they would have different tasks and they did |
---|
0:07:31 | they would they would stay engaged in the data collection |
---|
0:07:39 | we also gave some instructions to the user to make sure that we collected dialogues |
---|
0:07:43 | that we could use so we told them to not use too much and comments |
---|
0:07:47 | buying but also to use some so that you know what it's data bit realistic |
---|
0:07:53 | so we told them to make personally the lectures and |
---|
0:07:57 | and |
---|
0:07:59 | we also told them to feel free to and the conversation at any time because |
---|
0:08:03 | we wanted them to feel like they're real users |
---|
0:08:05 | and for that we also created some templates that would |
---|
0:08:10 | and courage to select one of the templates words |
---|
0:08:13 | you're a pop star you're an absolute geneva and you want accept anything under five |
---|
0:08:17 | stars |
---|
0:08:18 | so sometimes you know there would be we act like a different just close the |
---|
0:08:22 | dialogue and leave so that was interesting for us to have different cases the |
---|
0:08:26 | successful dialogues in there are lots where the user would just three |
---|
0:08:29 | we also told them to try to spell things directly to keep not too complicated |
---|
0:08:35 | and we told them to |
---|
0:08:38 | try to determine what they can get for their money so that they would really |
---|
0:08:41 | exploring the options compare the hotels and |
---|
0:08:45 | try to figure out what's in the database |
---|
0:08:49 | so on the wizard side so the agent |
---|
0:08:53 | playing the role of the dialogue system at the beginning of each dialogue they get |
---|
0:08:57 | a link to search interface that look like that |
---|
0:09:01 | so on the left |
---|
0:09:04 | you have although searchable fields and on the right you have the results |
---|
0:09:08 | and for each search the wizard will always get up to ten results so from |
---|
0:09:12 | zero to ten |
---|
0:09:15 | and you can also see |
---|
0:09:17 | the little tab on top |
---|
0:09:18 | so basically what we did is that |
---|
0:09:21 | every time the user would change i've been strange so it might so here it's |
---|
0:09:25 | for which cd baltimore |
---|
0:09:28 | if the user would say then okay what about to run all then we create |
---|
0:09:32 | this search and you have so that if the user wants to go back to |
---|
0:09:35 | baltimore |
---|
0:09:36 | the wizard can do it easily and wouldn't have to repeat the search over again |
---|
0:09:42 | and we also gave instructions to see whether it |
---|
0:09:45 | those where whites |
---|
0:09:46 | critical for us to be able to have a dataset where we can actually try |
---|
0:09:50 | to imitate the wizard behaviour |
---|
0:09:52 | so we told them to be polite and not jump |
---|
0:09:56 | and on the role played by the user |
---|
0:09:58 | claim that a mistake |
---|
0:10:00 | and this the start point also relates to that we told them your knowledge of |
---|
0:10:05 | the world is only a limited by the database because we don't want the wizard |
---|
0:10:08 | to start talking about pokemon |
---|
0:10:10 | or things that we can't we don't wanna dialogue system to do so we just |
---|
0:10:14 | pull them to |
---|
0:10:15 | you know that the user is gonna play a role in be kind of funny |
---|
0:10:18 | but try to just |
---|
0:10:19 | talk like a dialogue system basically |
---|
0:10:23 | i we also tell them to told them to try to spell things correctly for |
---|
0:10:27 | nlg |
---|
0:10:28 | and now the second point we told them to very the way a cancer |
---|
0:10:33 | the user and we told them that sometimes |
---|
0:10:36 | they can try to say something that is a bit impromptu so imagine if you're |
---|
0:10:40 | having a dialogue and then the middle of it the wizard with say hello |
---|
0:10:44 | doesn't make sense |
---|
0:10:45 | and we did that because we wanted to have so |
---|
0:10:48 | we have a lot of experience in training dialogue systems with reinforcement learning and the |
---|
0:10:52 | problem with that is that if you only have |
---|
0:10:55 | positive examples and you don't know |
---|
0:10:57 | what a mistake looks like so something that you shouldn't do at some point of |
---|
0:11:01 | the dialogue it's it makes it a bit hard |
---|
0:11:03 | and as a way to |
---|
0:11:06 | measure how |
---|
0:11:08 | how that |
---|
0:11:09 | was there are in the in the dataset we ask |
---|
0:11:12 | the user to read the dialogue at the end of each dialogue |
---|
0:11:16 | and we told them to base the rating only on the wizard behaviour so if |
---|
0:11:20 | they didn't get any results because there wasn't any result in the database |
---|
0:11:24 | but the wizard was helpful and we told them to give a maximum score |
---|
0:11:28 | so we had suppose on the scale of one to five and those are available |
---|
0:11:31 | as the dataset |
---|
0:11:32 | and as we can say as we can see there are a few most of |
---|
0:11:36 | them have |
---|
0:11:37 | the maximal score of five but somehow |
---|
0:11:40 | lower scores because the wizard was not completely operators and the actions that were not |
---|
0:11:45 | very helpful |
---|
0:11:49 | then other statistics of the corpus this is the proportion of dialogue |
---|
0:11:53 | through dialogue length so number of turns in a dialogue as you can see |
---|
0:11:58 | the |
---|
0:12:00 | for of the dataset is around |
---|
0:12:03 | fifteen turns bird the averages that fifty turns per dialogue so even though we have |
---|
0:12:07 | only one thousand three hundred sixty nine dialogues we have about twenty thousand turns in |
---|
0:12:12 | total |
---|
0:12:15 | a then this is the number of dialogue act |
---|
0:12:19 | this is the distribution of dialogue act types in the dataset so we had about |
---|
0:12:22 | twenty dialogue act types |
---|
0:12:26 | and the number of dialogue acts per turn so during one turn because it's human |
---|
0:12:32 | dialogues and |
---|
0:12:34 | there was more than one dialogue act per turn very often as you can see |
---|
0:12:38 | about three percent of the time |
---|
0:12:40 | there is more than one dialogue act type opportunity |
---|
0:12:44 | so |
---|
0:12:46 | that is that isn't in frames so once a frame but we |
---|
0:12:49 | so and i said what we really want to do is |
---|
0:12:52 | remember everything that the user has |
---|
0:12:54 | tool this during the dialogue so that we can |
---|
0:12:59 | get back to one option if the user decides to put that option in the |
---|
0:13:01 | n |
---|
0:13:02 | so we took inspiration from state tracking and the definition of a state and a |
---|
0:13:08 | dialog state tracking challenge in this challenge they define the state by the user constraints |
---|
0:13:14 | and at the user requests so everything that the user's task if he asks for |
---|
0:13:19 | the price or for the |
---|
0:13:22 | the name of the what out that that's a request |
---|
0:13:25 | and we also added things that we |
---|
0:13:28 | saw in the dataset and that we needed |
---|
0:13:30 | one is user binary questions so those are questions where you have |
---|
0:13:36 | so the user is |
---|
0:13:37 | a request is like the user is asking for price |
---|
0:13:40 | a binary question is when the user asks is the price |
---|
0:13:43 | two thousand dollars for instance so that's the yes no answer |
---|
0:13:47 | and we also had comparison request |
---|
0:13:50 | where the user as |
---|
0:13:51 | to compare something between two or tells you can ask if there is what do |
---|
0:13:55 | a cheaper than hotel be for instance |
---|
0:13:58 | and so those are examples of frames and the how their related so those two |
---|
0:14:03 | hotels are children of the |
---|
0:14:06 | the bowl |
---|
0:14:07 | frame |
---|
0:14:08 | as you can see |
---|
0:14:09 | and something you in our dataset is that |
---|
0:14:14 | frames can be created by users but also by whether it's so every time the |
---|
0:14:18 | wizard makes a proposition for hotel we create a frame because we want to remember |
---|
0:14:22 | it in case the user wants to book this hotel |
---|
0:14:27 | so we had a we |
---|
0:14:29 | made up a few rules for frame creation after analysing the dataset and seeing what |
---|
0:14:34 | makes sense |
---|
0:14:35 | and for frame creation |
---|
0:14:37 | we create a new frame every time the user changes a value so here at |
---|
0:14:42 | the beginning the user is to go to atlantis so that's one frame |
---|
0:14:46 | and then on these are utterance the user asked to go to never land and |
---|
0:14:51 | sold or destination cities change the we create an you separate frame with this value |
---|
0:14:56 | for the destination city |
---|
0:14:57 | actually changes a more entities here but we need to just have one tend to |
---|
0:15:02 | change to creating you frame |
---|
0:15:04 | and so that's one type of frame creation but we also create a new frame |
---|
0:15:10 | one the wizard makes a proposition for hotel and we put in this frame all |
---|
0:15:14 | the properties of the hotel |
---|
0:15:16 | so that gives you are frequencies of those behaviours |
---|
0:15:20 | in the dataset |
---|
0:15:21 | as for changing frames |
---|
0:15:24 | as you can see it's all user controls |
---|
0:15:27 | because we want |
---|
0:15:29 | the wizard to really be an assistant and |
---|
0:15:31 | just a dialogue system to really be an assistant and propose things but then the |
---|
0:15:35 | user controls what we're talking about the user controls the topic and the |
---|
0:15:40 | in the dialogue so the user or only has the power to change the frame |
---|
0:15:45 | that were talking about |
---|
0:15:46 | and so that happens |
---|
0:15:48 | which in you frame when the user proposes a new values a leafy changes the |
---|
0:15:52 | destination city then we automatically switch to that new frame |
---|
0:15:57 | if the user decides to consider an option a hotel and ask more information about |
---|
0:16:01 | those this option then we also switch to that option is a frame corresponding to |
---|
0:16:06 | that option |
---|
0:16:07 | and we can also switch to an earlier frame if the user says for instance |
---|
0:16:12 | and the dialogue that actually earlier okay let's go back to toronto package then we |
---|
0:16:16 | switch to the frame corresponding to the toronto package |
---|
0:16:21 | we also have annotations for dialogue acts and slots |
---|
0:16:26 | so the dialogue acts |
---|
0:16:28 | we have general purpose function still kind of typical dialogue act inform offer compare |
---|
0:16:34 | we also have dialogue act specific for frame tracking with the which is which frame |
---|
0:16:38 | that in the case when the user switches to are a frame |
---|
0:16:42 | then a for the slots we have all the fields in the database we also |
---|
0:16:47 | have specific ask the slots describing specific aspects of the dialogue |
---|
0:16:53 | while one is intense so the intent of the user is to book for instance |
---|
0:16:57 | action is their counterparts on the on the wizard side so the wizard book a |
---|
0:17:01 | hotel we annotated as action equal book |
---|
0:17:05 | and count is when the user gives the number of hotels in the database corresponding |
---|
0:17:10 | to the user constraints are sometimes the wizard will they i have stream or tell |
---|
0:17:14 | them about a more since the we would |
---|
0:17:16 | we would annotated with count peoples three |
---|
0:17:20 | and then we have specific |
---|
0:17:23 | slot-types |
---|
0:17:23 | to report |
---|
0:17:25 | the creation and a modification of that of a frames |
---|
0:17:28 | so we actually |
---|
0:17:31 | automatically annotated the frames and the content in the under frames based on those slots |
---|
0:17:36 | so those slots are it for each new frame we give a to a new |
---|
0:17:39 | idea |
---|
0:17:41 | reference so every time the user preferences the past frame |
---|
0:17:45 | and read and write |
---|
0:17:47 | so i'm gonna go faster here |
---|
0:17:49 | so that's an example of how we used read and write |
---|
0:17:53 | for read it's |
---|
0:17:54 | basically it so we sorry wherein frame five here the "'cause" the active frame is |
---|
0:17:59 | frame five |
---|
0:18:01 | but the wizard five talks about |
---|
0:18:04 | values that were provided in frame for so reread those values from frame for and |
---|
0:18:09 | we would put them in figure five |
---|
0:18:10 | and for right it's on the last utterance |
---|
0:18:14 | duh wizard provides new information |
---|
0:18:17 | about a frame that we already talked about before so we write this information and |
---|
0:18:21 | the preview in frame for |
---|
0:18:24 | even though we're the currently active frame is |
---|
0:18:27 | the frame number six a basis it's a bit |
---|
0:18:30 | complicated like that but |
---|
0:18:31 | it's basically a way to track of all the values and then |
---|
0:18:35 | dynamically |
---|
0:18:38 | populate the content of the frames |
---|
0:18:42 | so i statistics are some statistics of frame changes in the dataset |
---|
0:18:46 | the average number of frame changes |
---|
0:18:49 | created per dialogue is six point seven |
---|
0:18:52 | and the average number of frame switches is a three point |
---|
0:18:56 | fifty eight and we get a we have a lot of variability between the daleks |
---|
0:19:00 | as you can see here |
---|
0:19:01 | so we observe do the behaviour that we wanted to observe |
---|
0:19:06 | we also trying to see so we had five experts annotating the dataset and we |
---|
0:19:12 | evaluating how well they agreed on the annotation |
---|
0:19:17 | and we got a reasonable agreements |
---|
0:19:21 | so we propose baselines with for this dataset one is an nlu baseline that was |
---|
0:19:27 | choose to you kind of how hard piano your task was |
---|
0:19:30 | we adapted model from arnold and colleagues published in twenty sixteen |
---|
0:19:37 | and we predict dialogue act type and slot |
---|
0:19:39 | and slot values and we get about eighty percent accuracy so |
---|
0:19:45 | it's all already pretty good but there is room for improvement |
---|
0:19:48 | so for frame tracking ripple for the task |
---|
0:19:51 | so if you want to create a dialogue system that's gonna be able to |
---|
0:19:55 | g |
---|
0:19:56 | in memory all the frames talked about during the dialogue you'll have to do it |
---|
0:19:59 | to create the frames dynamically as throughout the dialogue but we decided to take the |
---|
0:20:04 | first step |
---|
0:20:05 | of having a simple task |
---|
0:20:06 | so if you know all the frames created so far you have the new user |
---|
0:20:11 | utterance |
---|
0:20:11 | and the nlu annotation for this user utterance so you know the dialogue acts in |
---|
0:20:16 | the slot types |
---|
0:20:17 | and the task consists of for each |
---|
0:20:19 | slot |
---|
0:20:23 | find the frame that it references so here for instance |
---|
0:20:27 | that's efficiency nipples mine reference to frame number one |
---|
0:20:31 | budget a post you cheaper actually makes was created new frame |
---|
0:20:35 | and flexibly view of the steeple true refers to the current frame |
---|
0:20:42 | are we proposed a rule based baseline that was very simple and that we just |
---|
0:20:47 | we just observed some behaviour in the and the dataset and so we propose a |
---|
0:20:52 | very simple baseline so basically if the user can forms a new value we create |
---|
0:20:57 | a new frame |
---|
0:20:59 | we switch to a previous frame if we find the mouse is that the user |
---|
0:21:03 | is talking about in one of the previous frame |
---|
0:21:05 | and basically |
---|
0:21:08 | very simple rules are those of some for |
---|
0:21:12 | switching to frames |
---|
0:21:13 | and of so the performance was bad because rules are not enough to do this |
---|
0:21:17 | task |
---|
0:21:19 | we kind of breaking down based on |
---|
0:21:22 | different cases and the dataset so it |
---|
0:21:25 | for frame switching |
---|
0:21:26 | if the user provides a slot so it's as they are let's go back to |
---|
0:21:29 | toronto package |
---|
0:21:31 | then we get about forty five percent performance |
---|
0:21:34 | if the user replies to a previous frame but without specifying a specific slot |
---|
0:21:39 | then it's harder because we don't it's harder to understand what the users talking about |
---|
0:21:45 | after a wizard after the wizard proposes a hotel so that after an offer |
---|
0:21:50 | most of time the user will ask for more information about this hotel so |
---|
0:21:55 | very often we would switch to that frame so what that's easier also to predict |
---|
0:21:59 | and it's easier than one there is no offers so we get a lower performance |
---|
0:22:04 | there |
---|
0:22:05 | and for frame creation we can predict that no frame is greeted but it's harder |
---|
0:22:10 | to predict when the frame is created |
---|
0:22:13 | and as followup work we |
---|
0:22:16 | okay so we had a paper was the better model that |
---|
0:22:21 | outperform the baseline by a lot |
---|
0:22:24 | we presented it workshop at a c l very recently |
---|
0:22:28 | and so to conclude this is the new human dataset to study complex state tracking |
---|
0:22:34 | we have turn level annotation of dialogue act slots and phrase we also propose a |
---|
0:22:38 | new task which is frame tracking and some baseline |
---|
0:22:41 | thanks for your attention |
---|
0:22:49 | the first minutes for questions |
---|
0:23:02 | fixed would talk could utilize the language variability |
---|
0:23:06 | but it's a few but anyway |
---|
0:23:10 | over one thousand dialogues actually the user actually filled or increasing the |
---|
0:23:18 | so by just eyeballing we didn't really |
---|
0:23:21 | compute anything but by just looking at the dialogue they really playing the really get |
---|
0:23:25 | into a they play the roles and they just change their language sometimes it goes |
---|
0:23:30 | from very polite to more |
---|
0:23:33 | like young speaking it there's a lot of variability thanks to |
---|
0:23:51 | possible to combinations so it is to monitor from you to see would be |
---|
0:23:58 | to generate will fall |
---|
0:24:00 | so it's of combinations able to do something over it sorry if |
---|
0:24:05 | but only from you |
---|
0:24:08 | to work well |
---|
0:24:10 | so that's |
---|
0:24:11 | that's something we decided not to deal with the we actually asked to always talk |
---|
0:24:16 | about one thing at a time |
---|
0:24:20 | but with the true for example the system |
---|
0:24:22 | is it should have seen from small words |
---|
0:24:24 | huh |
---|
0:24:27 | we would have would have all right |
---|
0:24:42 | to thank you for interesting to before i just quickly you and the u |
---|
0:24:51 | three point but the among the you can use you pixels detailed results tools to |
---|
0:25:00 | promote collagen dreams |
---|
0:25:04 | so we record all those urges and that the end result of the such as |
---|
0:25:11 | we |
---|
0:25:12 | that's an idea that we had we have we haven't really try to see if |
---|
0:25:16 | it's really reliable but |
---|
0:25:18 | because |
---|
0:25:18 | everything was not searchable database as well so that's probably had and we're actually |
---|
0:25:25 | that something when it's a we're collecting more dialogue right now to make it bigger |
---|
0:25:29 | and now we're gonna make all the field in the database searchable |
---|
0:25:33 | so that we can record of those searches and then do something like that |
---|
0:25:39 | just one more question |
---|
0:25:44 | all clusters let's take the speaker again |
---|