Scaling Semantic Role Labeling and Semantic Parsing Across Languages

Video in TIB AV-Portal: Scaling Semantic Role Labeling and Semantic Parsing Across Languages

Formal Metadata

Title
Scaling Semantic Role Labeling and Semantic Parsing Across Languages
Title of Series
Author
License
CC Attribution - NonCommercial - NoDerivatives 3.0 Germany:
You are free to use, copy, distribute and transmit the work or content in unchanged form for any legal and non-commercial purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Identifiers
Publisher
Release Date
2021
Language
English
Production Year
2021

Content Metadata

Subject Area
Abstract
Sentence-level semantics is hampered by the lack of large-scale annotated data in non-English languages. In this talk I will focus on two key tasks aimed at enabling Natural Language Understanding, that is, Semantic Role Labeling (SRL) and semantic parsing, and put forward innovative approaches which we developed to scale across languages. I will show you how new, language-independent techniques – including new deep learning architectures and high-quality silver data creation, as well as a brand-new, wide-coverage, multilingual verb frame resource, namely VerbAtlas – will help significantly close the gap between English and low-resource languages, and achieve the state of the art across the board.
Keywords semantic parsing, semantic role labeling
Presentation of a group Group action Context awareness Parsing State of matter Natural language Multiplication sign Parameter (computer programming) Mereology Semantics (computer science) Computer programming Word Coefficient of determination Bit rate Pressure volume diagram Compiler Finitary relation Negative number Endliche Modelltheorie Logic gate Error message Position operator Identity management Physical system Area Parsing Knowledge base Parameter (computer programming) Price index Flow separation Connected space Arithmetic mean Process (computing) Repository (publishing) Order (biology) Natural number Mathematical singularity Right angle Task (computing) Resultant Reading (process) Point (geometry) Trail Server (computing) Sine Service (economics) Observational study Software industry Rule of inference Event horizon Element (mathematics) Twitter Wave packet Number Latent heat Goodness of fit Internetworking Software E-learning Metropolitan area network Task (computing) Context awareness Rule of inference Information Cellular automaton Projective plane Semantics (computer science) Predicate (grammar) Grass (card game) Frame problem Word Maize Event horizon Software Predicate (grammar) Personal digital assistant Speech synthesis Natural language Musical ensemble Object (grammar) Routing Address space
Complex (psychology) Presentation of a group Context awareness Group action Parsing PRINCE2 Archaeological field survey Bit error rate Parsing Bit rate Cohen's kappa Disk read-and-write head Software bug Data model Roundness (object) Different (Kate Ryan album) Single-precision floating-point format Finitary relation Videoconferencing System identification Abstraction Multiplication Physical system Theory of relativity Interior (topology) Sound effect Bending Parameter (computer programming) Sequence Wave Arithmetic mean Process (computing) Order (biology) System programming Representation (politics) Point (geometry) Wage labour Transformation (genetics) Connectivity (graph theory) Letterpress printing Control flow Mass Average Number Architecture Frequency Goodness of fit Performance appraisal Term (mathematics) Representation (politics) Associative property Metropolitan area network Computer architecture Form (programming) Standard deviation Graph (mathematics) Scaling (geometry) Information State of matter Coma Berenices System call Frame problem Compiler Word Error message Software Personal digital assistant Predicate (grammar) String (computer science) Game theory Resolvent formalism State of matter Natural language Multiplication sign 1 (number) Set (mathematics) Parameter (computer programming) Mereology Semantics (computer science) Word Programmer (hardware) Endliche Modelltheorie Algebra Logic gate Position operator Boss Corporation Parsing Price index Picture archiving and communication system Flow separation Connected space Prediction Bridging (networking) Self-organization output Right angle PRINCE2 Resultant Thomas Bayes Asynchronous Transfer Mode Server (computing) Wave packet 2 (number) Average Green's function Software testing output Proxy server Task (computing) Module (mathematics) Context awareness Hazard (2005 film) Military base Projective plane Predicate (grammar) Performance appraisal Calculation Formal grammar Surreal number Natural language Object (grammar) Sanitary sewer
Context awareness Group action Parsing Distribution (mathematics) State of matter PRINCE2 Natural language Correspondence (mathematics) Direction (geometry) Multiplication sign Programmable read-only memory 1 (number) Set (mathematics) Parsing Cohen's kappa Function (mathematics) Parallel port Mereology Independence (probability theory) Data model Strategy game Different (Kate Ryan album) Finitary relation Einbettung <Mathematik> Endliche Modelltheorie System identification Multiplication Parsing Sequel Theory of relativity Interior (topology) Sequence Demoscene Virtual machine Arithmetic mean Order (biology) System identification Summierbarkeit Representation (politics) Resultant Row (database) Point (geometry) Server (computing) Drop (liquid) Menu (computing) Raw image format Axonometric projection Graph coloring Twitter Number Wave packet Power (physics) Revision control Causality Robotics Term (mathematics) Maschinelle Übersetzung Graph (mathematics) Representation (politics) Software testing Configuration space Associative property Computer architecture Addition Distribution (mathematics) Graph (mathematics) Projective plane Semantics (computer science) Grass (card game) System call Compiler Performance appraisal Word Personal digital assistant Predicate (grammar) Thermal radiation Labour Party (Malta) Natural language Musical ensemble
Cluster sampling Group action Parsing State of matter Natural language Multiplication sign Source code Set (mathematics) Insertion loss Parameter (computer programming) Mereology Different (Kate Ryan album) Ontology Information Endliche Modelltheorie Social class Intelligent Network Source code Boss Corporation Parsing Link (knot theory) Attribute grammar Parameter (computer programming) Statistics Element (mathematics) Demoscene Product (business) Type theory Arithmetic mean Googol Order (biology) Mathematical singularity Moving average Energy level Freeware Resultant Thomas Bayes Point (geometry) Frame problem Digital filter Link (knot theory) Sparse matrix Wave packet Number Latent heat Prototype Term (mathematics) Reduction of order Energy level Data structure Maß <Mathematik> Metropolitan area network Task (computing) Pairwise comparison Scaling (geometry) Information Expression Predicate (grammar) Grass (card game) System call Frame problem Causality Word Uniform resource locator Personal digital assistant Social class Natural language Family Tunis
into the invasion and i think you either you were checking my recent publications all are you were very lucky because through the wall with been doing modelling will and of the increasing while he is a low interest in more modern languages recently the past year. the years we've been working very hard on topics a very late to your or objectives in the group entrepreneurship about serve. and then to the action extraction and the in corning as much grass and as you'll see the two talks them a touching today which are some antique roles or. man to grow labeling and some of the bar sing and are very very related and the i guess i have a lot of interest for you and me of course and my team show this is a joint work with several seasons are in my group economy under the fabio. the gene and lost me and the also pulled a cricketer pod so an inn in in order to achieve natural language understanding than identify three main tasks that make it possible to connect service or realisations which are. the tracks the read every day to knowledge grass knowledge base is an explicit means that we find somewhere in structured repositories the first one i wasn't isn't the way shunned which aims to ask the shading meanings with the words occurring contacts to one. the social the explicit meanings within these words like not firing isn't be years and you want to choose the right meaning of far. this is a dog i'm not directly dutch in today. the second task is semantic role labeling which is also called shall semantic parsing and aims at earn identity fine and the annotating predicates with their arguments so for example in this sentence again employer five programs on the software company want to identify the credit of the predicates. fires year saner this is a turning it work frame or band and i didn't define the arguments. none are. part of this is the vende and mom and label them as well so for example here where the employer and employee in the position and this is a says them from some years ago with was identified in the frame and the arguments unfortunately the. no way and so was an understanding that this the event would be shooting project house which is obviously in some key issue not too if you want to fire employees from your comment and rules and so again you have some sense of some problems immigration problem because in choosing the right friendly get forgiven. verb or war in general you have to these and the good work and so you have to choose the right frame and finally the third task is semantic parsing which goes all the further one step further and transforms the center's into a full fledged a rough in which not only year. the air. a friend gate and arguments but we also have a new route that connects each item beach element of the sentence with other elements like for example software with company as the money fire of the company or we entrance he is like an order something that these the richard back in the sand then. his or internet and you can create a full fledged rough as will see later during his presentation so easy to do tasks and we are focusing on today and nine inches and the one is a recurring issue in many tasks in have been particularly in a new which is the show's it the of a resources in trend data. nine many languages. this results in the performance gaps you all know that doesn't have that local resources in english but they're not so many many other languages and so you will see that is a huge issue in the two tasks today i just did the fight and also does a lack of shared semantics across languages so it's hard to provide. as an asian that can be used to obstruct away from the specific language or interest and so will address these two issues in this talk of and the as a said was for one as early as monday passing while leaving out i wasn't is an invasion because of lack of time however there's a. a recent work again for my group on both of the child last year in trouble at these year on how to scale a with and is a musician modeling willie in the absence of training day die on certain. certain amount. then so we're focusing on some of the crow labeling semantic parsing insurmountable early last month parsing and let me start on some article lately. some of the crew labeling you are addressing the boston went in to find automatically who did what to whom where when and how for example the week brown fox jobs all is a dog who wanted him to find a predicate chums want these and the years the predicate and a solution to write me with this. this word in context then we want to identify the arguments so here we have two arguments and finally want to classify these arguments by a say in what this text on her first to make even the pretty good news for example who is performing the action and over what this. his argument classification so not to perform these forecasts for the wider a lot of work that. stays there is a sin bass is quite helpful in some article italy so there are some advantages in using seen as a strong the gate or for many of these subclass so who's one is that if we were four pounds be shocking and we associate the most suitable. the speech with a given wooden context and willful him a rise the word then we can more easily. i didn't the fire brigade because for example most of the time pretty it is a work and also recently years the brigade cells while for argument to the asian pacific asian it can be very useful to have seen that information for example to know that the certain word is a direct object rob were ordered it's a subject of a verb so these can help improve. the in indignation and also has indication of the arguments and this is a this is all true however that also the the bandages on many area in order to achieve these the to inject to use that information in the ask certain that we just mentioned we need. to train systems on cars speech time being and owns mountains into the parsing and to the is we need training data in the language of interest however these data are expensive to create and is causing many languages as especially for resource languages this is a problem because we don't have. such resources and even if we have me we don't have enough data and avoid the on these outcomes is not a good and so you might even have an umbrella of the their use in order to performance around for example if we cause years the the yard about the study for english we have. why high performance lose some times even a well above ninety five percent even ninety eight percent is due a horse on while if we do the same for chinese will get lower performance is good but not dead re not not as great as compared to english exam even want to. parsing then there with her we were very low or and so english as for example of georgia where label attachments or which is around seventy seven percent which is so and so and the chinese as a more than ten points lower on this task. so it's true that for some languages we can take advantage of this information however is not so obvious that we can do this for all languages and then it's k. have easily indeed there's a a white be modeling will gallons a month call it being so you look at the results of steel the are. systems in some and all ugly we can see their the new state of the art error a show. when in on a in english man to call a glitch is around nine percent saying means the route one hundred items to be classified not for nine we make a mistake higher and is this error increases as we move to languages for which. the amount of training day eyes lower is increasingly lower to even twenty percent higher rate and so this error age. if we our age these error rates we see that we have a multilingual gap between english and the numbers marketing will get within english and i'm the worst language which in this case is a surprisingly german people positive data around fifteen percent and the.
the gap is than more than ten percent so. in this is a waltz. one person a multilingual smucker labeling from come true becoming a true. food. and so how can we close this gap in this is one of the key concerns in one of the readers the numbers and in today. and and the answer is yes we can close this gap and or we don't need seeing that's so we're going to propose a neurotic that sure that doesn't exploiting that information while at the same time scaling to as many languages as possible and this is a labor of it. first known economy by presented a calling last year in a december a trickle. the objective is a said he's too. remove the requirement whole on language disease features like glenn mars process the huge that the information and set seen as a stronger we must days land for future work that might use that information is available so the mantle is to have a new room or the. these language across the uk and the model takes his input ascendance thing goes through the three main modules neural neural modules one is about just would present in words are in context the other one is about to be sending words in context by having the college. business of on the predicate that we are working on where are focusing on their weekend before credit and education period he occasionally and the mission which we can then use to perform to create representations of the world who just the civic forgiven predicate and as a result we can use these forgive. and present it to classify those words as potential our demands and classify them with the right or war and so more needles. so we start from the bugs and those we do not use as is normally done the leaders are any particular fly or freddie gets busy can bend been drawn up and the predicate so we just are we present the you could send the ends on a word by word bases and and the and provide a lead and ferguson did. insular a position of these worlds then we the create a regular were would representations by on. a utilizing whole the various pretty good positions and senses in the scientists are finally and once we perform an education is immigration we are the. basically be exploited the relations between the predicate and its arguments in order to perform at thirty patients received so let's look into the various models indeed else away using a b b organisations have that each work care which were piece is a part of. but even san francisco and so long. these go through one of the many trouble a base models. and which are pre-training and are we just take the last four lays of these and catherine age the you the representations a resulting from the last eight years then this goes through a fully connected in a year's wish foundation average bullying and there's a result we can combine. the new car her representations of even the word into a single latent representation and so these doesn't consider anything about some article it is just the way to present a word make its form. how do we have our own text to all who are representations for each word in ascendance these girls through a biased g.m.. which. the group used a contextual representation of the word the together with is called at the nation with the rear the computed representation on as as a set earlier and these than ngos are through a print in the education model. that which just determines whether this work to even the fact that this could see the the complex from the ascendance what these worries a break it will not so each word is classified as predator not to and those words of a classified as predicates then are busy. and gated my eyes second model will be connected to porn and that determines whether they spread the cage which which sends which frame this predicate is the noting in this context for example if these were the fire. the work can be a sentence that i mentioned earlier and this would choose between the various meanings of fire and so now that we no way we each of these are words is a bread the good and the effects of credit which sense it's the all teams. and now we can for each of these bread the gates we can compute a predator were representation that takes again each quarterback stall visualize were representation for each word to and contaminates the predicate for presentation with each of the moon. words in context a while for the brain get itself would just double this the very same representation so now we have pretty good aware of positions that have a renegade representation together with the work of actual as would a person patients and again these goes through a ballast g.m. and new and. the out with a mallet game is again cabin eighty with this representation and then we perform argument an indication classification. and be a as a result of these forces. so this is the each and every brigade that was identified earlier. and so for example where the single printed it would be his bosses michael company in these are predicated before the mercy. the very same by a sequence of calculations so as not how we trade isn't a new role architecture so let's look in let's look at the evaluation to perform which was dependency based as you might know there are two ways of performing some on to grow labeling one of the bends e bay's and one is on. based independently based s. around we care when we are to classify an argument which is really which is the news jointly with a given predicates is this is the predicate in this is the certain argument we don't want to classify each and every. word in the spa but we just under eighty five the synthetic head of the spawn for example if we say and all the. the software programmer programmer is the head and so we just want to time programmer as the right. with the rights the role as usual the the brigade's fiery for example can and so an independent the bases are a lot we normally using gold standard there was a survey from the coral two thousand and nine the shared a task a year what we did was too. try several different transformer based architectures birth frozen bird find to and where we would defy the waves of the birds model axel m. roberto not but also a ballast game based the approach the hell hole and so on. in all these cases we call p.r. modal were where we were blaze the language model with the. the corresponding model of a mention year when they are best approach on which is based on and you have these language models and we see that given the simplicity are also the end to end the architecture and effectiveness of the approach we can. the debt once we train on the calls and nine train data and the interest only respond that's it. comparing is the state of the art we outperform and no case he is the new committee doors and it's interesting to see their. the best a resolves are tainted by when of course the trouble of basic church or architecture is used to and on english in the best results are tainted by a berk fight went fine tuned to an excellent were bafta. um. now let now this is however the results on english and you can see the results are quite high because the english is a really easy due to the high quantity over the top for training and so on so now evaluation on marketing when crossing was around however is more.
first thing because here at what your interest in the inning is to perform as are rare in multiple languages and there were five different languages catalan check german spanish and chinese and say you can see that the current state of the reuse date of the artois seen acts aware performs much. slower than english you can see even even more than ten points lower than english were german for example and several points lower for spanish and chinese and so on but now if he used the excellent real america which is the. just from a basic language model which tends to perform best across languages you can see a considerable improvement performances and this makes it clear their you by choosing the right the new architecture and with the modeling wall on with marketing will complex well as would a person. patients you can achieve high and considerably higher performance isn't close the marketing in his earshot costing wall as around setting in which you you for your when you are training is certain language a new test in another language our this can be done on the for those. those pairs of languages were the as are ailing venturi is the same man used to achieve promising results. like for example you can achieve a g. almost eighty two percent as surreal performance in both languages without having a tie in nine the language of interest at least smelly so. this is a the current and state of the art way of scaling semantic really been marketing but now i would like to focus on the other tasks that i mentioned that the bag is crucial for national and which understanding so moving from the connections between. the renegade and its arguments to the connections of the between the various components in the entire sentence so this is in a sense a super past all as around because that it works with all the components of a good sense visibly for the center's the little prince. clap his hands not only do we want to connect the frame the verbal frame clout to its arguments but we also want to specify that for example he though is more the fire brains and the other hand the launch the prince and so we want to end it if i all the relationships are expressed be innocent ones. however we need a gain of formalism to express his need is a gruff bay's new representation and one of the most popular form is of this kind is called the obstruct me representation. the question is can we do this marketing will again so how does this scale marketing willing can we use aymar absurd mean representation across languages and so the first thing is there are a market is not an entirely well so it was not born as an interesting well and it's actually heavily biased was invalid. wish you number of respects for example i like with english as a rail in makes use of probably friends sets to these and the good the credits so this club zero one is actually one of the meanings encoded in the problem was the river for a verbal frames. need to be used by the man to our server to associate the right meaning with this were in context so this is one of the ways you can clap one the meanings of her lap and this is language the city so aymar is tightly bound to the problem. friends say how at the same time a marlins to obstruct the way firms and not to get using crecy is a chance to be agnostic about the way we create meaning starting from pacs and so there is actually work that stage the a mars northern injuring were. is close to it and also that is true there in other languages we have other. trains that border is for example expand have or land in chinese the chinese problem and and neither are now none of these is connected to the broad bank the english project resource so this creates a problem however we cannot just the aim are guidelines to cover costing were aspects. the our own be addressed the numbers languish pacific resources so the question is what can we do to create a representation of the is as similar as possible to the aymar english a mark as a nation but there can be produced for our language is like for example here we have the time span. issue and we see that lu redoubling taliban and also spanish we don't express the fact that the hands belong to the greens. we don't say he is hands or sues or sewer money or says mass just say the hands his liking issues and the mill brains collapse the hands round and his hands so we don't make these relationships place it and even worse in a sense your mom. more different in spanish we don't say that the we don't have an extra words to express the the prince's is lido and we just use a single word to say busy paedo to say that the greens is little and so this is again blizzard in the representation and so. we want to find a way to scale up winning wall-e. and again you can guess that year when we have very little they'd high non english language is so what to do them well. if you look at the english for presentation one possible way or performing surmounted parsing martin in one way is to stick to the english representation so they use these as the reference for presentation while what pros the same sentence is in our language is an. and so in terms of the new being expressed by these three sentences the meaning is more or less to say so what we want to do is to say they're these meanings year that we like to obtain a deal the book which we don't have trained data we like to just make an assumption that we could represent these meaning six. pressing business long wished languages via the english aymara representations and this would be a good proxy to the semantics of the non english languages. one doesn't mean should challenge their third makes it hard to learn. in the new name are parsing to perform the task of which is this earth were too old aymar lime and so the point is there. in the northern to come up with a gruff representations the hazards with his ends expresses the semantics a sentence i need to find a connection between the various words in the sand dams and the various knows that i'm going to produce only to perform some and. word to log alignment and this is very helpful especially for want to propagate is information from one language to another how it's hard to protect this information across languages and say i have a data set in english that associates in a graph with the sentence. then the probe and even this information to maybe and my parlours sentence in another language is not so easy because i would need to take the non english language translation and then realign these knows the right words in a sense it's not an easy task. and also fortunately there's no available or it is not so much available costing whole data for a mark parsing so i don't have many aig zandt also learn english sentence is aligned wave a corresponding english demography. so not even in the towel spanish him around the northern english in our process does not know how many examples of these kinds of is not easy to train system on this drive past and so this is work that who visited the amount the last november when with the june of last year. the body in and this is a way to enable crossing wall a mark parsi and so here are what we're going to lose to these pools of these were these were two loans alignment the just mentioned and creates your data that enables high quality crossly more m.r.. which makes it possible to work. on other languages languages of english. wilson the video also provide for the insides of mobs the new way that a non english sentence is transformed into a new english a mark rafa and the shoes the these polls is because of course.
and there are some things that you can easily express in english and the arts that expressed in another now which you'll find is in the paper because of my time here to talk about them. alone so we are mobile overview of these more though called the your movies model called the x.l. aymar is that are we have sentences and their translations into a number of languages so first we apply sick the sec in bordeaux that produces. his identifies and produces the least of concepts and so for example for these were given sentence in indian language we could use a sequence of concepts these companies include. no bank english problem predicates and also other words which we go concepts to as a was his only wish an researcher i wouldn't hold and concerts because the zombie his words anyway but anyway this is a what is than in any of our cars and so we just produced east's of words or predicates from. probably in than ever with the by finest attention will perform relation identification so we determine that is collapsing no one knows your. haven't even a certain relationship with some other nodes that were produced are in the cause of the review should face so this is a is of course excludes other knowles with which it's not new relationship for example claps here one is not new tissue that no direct mission leno and so on and is a. it makes impossible to produce a graph that would resent the meaning of the eagles and now how his and cars are in tuition performs again and i were were making news over minus the am some so we take the sentence the now the non english language and want to. use a sequence of concepts in english for english aymar and then the second step that were not mentions year that connecting them with relationships so how will this well the first thing is that their we are in cold this trend. and then smiled by stemming border that produces a representation for each word in conflicts. gavin and so then the thanks to any quarter pain show we can produce in a sense the context of. of these words are in the sand dams and pointed dead and at the for initially we start with the asimo and in these are other who bore in which is the color we want to print used the sequence of concepts so we start with the a the starts the mall. and then want to produce one after the other numbers concepts not have to choose what to produce. now we are going to use the context the and the current he then stayed so the need to be easy because they have as some to learn to the complex than the. currently the state of the beholder and tunis hide an urban we each new world who want to produce which new concept want to produce and we use our vocabulary concepts and to decide what is the next what is the wall what will the next to the concert be this least year. and always lead his book early will include both prabang many gays and words in the vocabulary. but not only eared do we do that but we also have to decide whether we want to generate a word from this distribution or call the one from a the the in previews are our calls from the recorder and so on. with some probe he needs he were going to cope the previews concert from this least on some other probably were generating. any of these words and so as a result of painting final distribution and this will tell us then what we're going to produce as the next hour and we continue. nine. so this is for the cause of radiation step no debts and the reason why we call been is that an urban because of the relation of indignation stare per week might need to have multiple copies of the same concept because a nerd the costs of my d.. and reported multiple times in the sand pants and we might want to have the entrance he is too was previews concert previously produced concert in the sand dunes. oh and not focusing on the relation and evasion part because as low time and it's basically by a fire engine like is used in africa apples and topic parsing so there's not much year to understand. learn that is a singleton contribution which is having begun to train is architecture on so in order to trade architecture in need pairs of senators easy non english language is and the corresponding concepts and also we asians for the by fina pension and to connect. these words and so in order to graze they got because these are not available we are a work of two different an additional projection approach is so obviously would need to work to start from the malaysian project innovations from english on to other languages and within days through different approaches one is based on piles of. and as he is in which we used a parallel manually produced are many created parle centres is that his goal gold translations are created by humans while associate being english aymar grass or in which are. however seen the past so basically even wish yeah yeah because we have policies and as a will have a sentence in english and the corresponding translations in other languages but for this sentence year we don't have a gold a mark raf so we had to use the name are parcel. to produces shield her margaret because another is no parallel new corporals when the menu he created. in march last year a third of these to the west mercia to have told him are grousing and so use the main they does it mean to indeed as a username are crossing and associates i am are english a madrassa the english were a sentence he is and then because of four. fortunately these sentences do not belong to foreign corpus we need to translate this sentence is automatically by a machine translation system so here is that we have both a mark raf so and silver translated sentences who want to see which of these two works best in the first case there said we day. maybe even wish and thence goes through an english to him are our server with produce a scene were. one aim are graphic and then as social receiver english him on rough with the translated the the gold translations and from other languages in the goal a marvelous either translations instead what we do is we have a gold aymar graph associated with the english sentence and then these. goes through of the mother the translation the provides translations of the mission and dancing to the other languages in than their we also are back trouble a decent the easy into english she and i went to the translation i was all the more not in those cases in which it. it's not your motivation we had just remove that item from that corresponding language know the other cases we just associate the an automatic translation with the old english a mark raw and sumo ages we have created a data set that contains a non english sentences and the. for spawning mission a mark rough now the experimental set up was to use for one am are they the sense that were created for forty four languages cheney is german town and spanish music stars mash and ones or to perform evaluation projects these on the. these innovations for the first the projection strategy using new robot which is as a corpse of arlen sentence is a the several cases using a mark to as the gold rush a marked drop an updated the set. and we tried a number of bargains in which we must train an english and best on any other language or train on the language and wrestlers in language training in knowing which is in test in the language of interest and train in two languages english trust language of interest interest in the language we trust and.
concede that compared to the only cost wing war in our cars are available in the leader church is a martin bigger. we want a considerable improvement in terms of performance is specially when he used the transition approach and that's why. the it's better to err noisy eating foods and gold our troops than vice versa and so this is why in translation approach works best because here in stand in the power cords approach we are going to use as a out to do is use her while she was off which is some automobile cause you would like to. you have or more output to be to fulfil the model to be trained on so you can see that we get. mall most of the time ten to fifteen and in grew up point improvement about the state of the art. so now i would counter last part of the talk and i we have led the quarterly does not so much time so i will need is a huge sums lives but just the e.u. an idiot the point is that what i don't know what i like of these approaches that we produce is we are able to parse sentences in all languages window.
why is that these are some of the bosses are at language to see if he can be used language pacific resources in laying am wars is the concepts so important part is to provide some group as a nation that the is as in the band as possible from the language.
any said we're not doing this because we use english dizzy credit sense sale way using a english aymar grass and so on and so are actually two years ago right before the lockdown we have proposed a new resource for on credit means. and this was based on the fact that the existing resources prabang that is heavily used its magical evidence mounted possible so for a minute and bourbon and have a number of issues liking big call rage read the case as if he calls a man takes me sing work last month weeks and these are all tied to the city. the language is mostly english so with good for the new resource older model us with under john and someone economy and a female be two thousand and ninety nine and this was entirely missing bowl way even though why. expensive in terms of work because this was done annually so what we need was to cause year old the young were the world's inserts into class of them based on the fact that they'd turn to appear in similar scenes and but it would seem look what does he wants to engage in action so for example my classroom all the. the word of her bills in services have to do with the gene we could create a frame coal used by. we need so for many other friends like for example googling has all the concerts in one and that have to do with. there's ways of cooking and in terms of roles instead of using roasted our own law meaningful in not not generous enough just selected twenty five rose from of urban and which is another use which source so that we could provide meaningful also designed the ban and of the. the frame of interest lack of our patients age and resolve this nation and so on. and so for example this is the prototypical argument structure we might well as an agent the patient the location and for cooking as an agent a patient an instrumental resolved and so on. hundred there's no time for the comparison to serve either said the point is to hand readable. roles will bear to roast adonis the the specific to the frame and not crossed the main like are in our case some on knox was under the bay's the light in the world what happens in bourbon and and the informative and preserving prabang which. users said very popular have roles cold eyed zero one and so on in these also not very readable wine as you say this is an agent or this is the experience or when action this is much more clearly the ball. these things tell us their own and thanks to her mother's would uncover the whole said the world meanings compared to all the other resources which are more leslie needed in this respect we only have less than five hundred frames and so the various types of actions and not in the order of thousands or sing done. luckily based and the number of arguments is also limited but these are crossed the main and so they're very informative. also very importantly were brothers is linked to worm and and as a result linked to model and which makes it more opening wall so you can parse with them opening way you know language is basically a socially back to problem so that you can we use the data sets of are used for some relabeling. and when the same time using this more expressive first trains for example we have these centres in different languages the verizon did the frame that is evoked by these birds as male the ons not lost on tea and the country is of course a link is in the racing bubble it seems that there's a. these are the very same her mother's fray the smell free and the there are arguments are in all languages the experience or and the sting was a need and the of the scene tax and the way these are express so this is very very cool and bring hortons sudoku new. i didn't have three are different works n one is the crossing will marketing walls multiple labeling to or would use the marketing will ever a gap between english and other languages and where it is these from fifteen point five percent to around ten percent so are about so. one third reduction. when. we are good for an approach to create seabird they die and train is his them to perform crossing will a mark are seeing in many languages and is that she is this state of the art on the basket and finally i briefly and overviews and you resource called about last which makes it. possible to parse. reforms multi-core labeling needs mountain parsing by organizing a verb meaning seeing to language in abandoned frain some twelve rolls the generalised across worsen frame son and which provides the english to work in and marvel at and problem and finally and his lawyer mission could have mentioned he will find in the paper which is the scenes of level. all information and their preference an election of reference he is and so on so this is a more or less all mother met the five has been recently we need is the which goes marketing will and. makes it possible to a scale mothering when all these tasks and. doing the future boy in more and more semantic and so bored in a smart really been surmounted are so we need to replace the world's next languages with some antics and much more this is open to for the attention these are some weeks think so.
Feedback