4

The concept of AI collaboration remains somewhat controversial in the translator community. Much has already been said about the impact of AI on more language-centered tasks such as translation, but what about AI in a broader sense? An increasing number of linguists now collaborate in AI training tasks such as image annotation or speech recognition, categorizing or annotating the data which ‘teaches’ AI systems. But is this an interesting new sideline for linguists, or something of a step down in the world?

 

On the one hand, AI training work does not necessarily require the same degree of qualification or experience that translation does. Yet in many ways the skill set is quite similar. AI data training relies on many of the qualities that are innate to linguists, such as strong analytical skills, attention to detail, and critical thinking. The fact that not every collaborator will complete a given task to the same level of quality also suggests that there is a relatively high degree of finesse involved even in fairly routine AI tasks. In this sense, we could argue that AI collaboration provides a wealth of interesting opportunities for skilled and experienced translators, even if only as a sideline to language work.

 

On the other hand, there are those who argue that the often fairly repetitive nature of AI training work can make it frustrating for translators, who might easily miss the more creative or ‘human’ elements of language-focused work. The fact that AI work is generally open to a much larger pool of people than translation work can also mean that translators, who may have extensive experience or education, feel that it is less valued than more ‘traditional’ language work, and therefore less worthy of their consideration. This could lead to linguists feeling that AI training work is not an appropriate outlet for their talents.

 

Having presented some of the arguments for and against machine learning work, it’s now over to our translator community. Does AI training work inspire you or fill you with dread? Do you feel it’s worthy of your time as a linguist? Feel free to add your voice to the discussion below!

20件のコメント

  • 22
    Avatar
    kvstegemann

    First of all, I still think that "AI" is a complete misnomer. The appropriate term would be "ML" (machine learning). "AI" suggests that these systems are intelligent, but they are not. They cannot make intelligent decisions, they can only create results from enormous amounts of data they are fed, and patterns in these.

    Anyway, for the individual freelance translator, the most interesting question is whether you can earn enough money with this kind of work. I have no qualms to do any kind of work I'm qualified for as long as I get a serious hourly rate out of it. I'm a tech guy and I have an overall optimistic view about new tech developments. I also have the advantage that in my case no "passion for language" stands in the way, since I have none.

    However, the rates I have seen so far for this kind of jobs seem to be so low that I never found any incentive to start. The work involved seems to be rather repetitive, too. In one case, I tried to calculate the speed I needed to do the task and I found that I would have to complete one task in about six seconds, in order to get a reasonable hourly rate out of it. Six seconds to read a list of several sentences and assess it. Six hundred of these tasks in one hour. Imagine that for a full working day of 8 hours, doing 4800 tasks in one day. I believe that I would need pain pills after the first day, and after the first week I'd probably look for a bullet instead.

    Therefore, the answer is rather simple. At the current rates, these tasks are out of the question. Here's a hint: I can translate about 500 words per hour with no sweat. You can easily figure my average hourly rate at Gengo pro rates or even higher rates I get elsewhere. Now try to translate this hourly rate to a reasonable rate for the "AI" training tasks in a way that allows enough time to do the tasks responsibly and carefully, and figure what kind of rate should be paid for these tasks in order to make them attractive for a translator. It's as easy as that.

  • 5
    Avatar
    squeck94

    As an engineer having supervised quite a few data annotators in a big company (funnily enough, they were and still are employed by Lionbridge), I can say many of them had studied translation. You are right in saying the skill set is similar -- translators can do the job and they can do it quite well, provided that there is (and they can feel) at least some sense of purpose and, I can't stress this enough, the labeling guidelines are clear and well documented.

    The most difficult thing was defining clear guidelines and communicating quickly and effectively with everyone involved when they were not. If this is not done, frustration quickly sets in. Communication and documentation are key, and I think Gengo should focus on enabling quick and effective communication with clients before anything else.

    Other things to keep in mind are giving the same work to different people (you need labeling to be consistent no matter who is doing it) and technical training (training vs test sets, when can you throw away bad data etc.)

    I would do some data annotation work, provided the rates are good and considering the work can be done from anywhere at any time. I "value" it less than translation work, but datasets are playing an increasingly large and important role in the world and the quality of data is absolutely important to get good results.

    Good luck!

    EDIT: I agree with kvstegemann, the rates have to be good for it to be worth it (and I think clients are willing to pay good money to avoid hiring full-time in-house annotators). However, I haven't seen any AI projects in my language pair so I don't know anything about the rates on Gengo.

    squeck94により編集されました
  • 2
    Avatar
    Sohail Moghal

    At the moment, I totally agree with 'kvstegemann'. Rates are absolutely key here. But, as squeck94 said, I too haven't yet received any AI project in my language, so I'll be able to provide further feedback after I do. Take care.

  • 10
    Avatar
    Trustlator

    I agree with the comments above. I was 'lucky' enough to be invited for a couple of fairly bulky AI-projects containing around 6000 tasks. After completing 12 tasks in two different projects, I quickly realized I was making $3 per hour at my speed. It ended right there as I knew I wasn't going to be fast enough to earn a decent amount of money no matter how much I would 'practice'. Second thing is indeed the quite boring and stress-inducing nature of the work. Repetitiveness and the need to go faster while maintaining quality make for a dreadful experience.

    So as things stand right now, this type of work may be more suited to people in certain regions with extremely low labor costs.

    In light of the above, the question of 'prestige' seems premature at this point. We can discuss whether AI is a "step down" for a translator once the rates are a step up.

     

     

  • 2
    Avatar
    kevin.thoma88

    I've tried my hand at some AI projects on Gengo and actually mostly enjoyed them (if the guidelines were clear). I like some repetitive work now and then to take a break from translating and other more creative things I do.

    That said, I do have to agree with the others. With the current rates I can only conclude that it isn't worth my time, even if I don't mind the work itself. I understand that it's basically "unskilled labour", but I have to be realistic about the projects I choose.

    kevin.thoma88により編集されました
  • 4
    Avatar
    phil

    Don't these AI projects have the ultimate goal of replacing human translators? Why would human translators want to help machines replace them?

  • 1
    Avatar
    Chris

    @phil

    If you are doing translation work on any bigger platform, especially one utilizing a workbench, you are very likely to be feeding some database that will be used to train AI/MT. Even translating anything were both source and translation will be publicly available might be used that way.
    So it's more about to which degree you want to help or can avoid it.

    Chrisにより編集されました
  • -1
    Avatar
    Katrina Paterson

    Thanks to everyone for all of your engaging comments so far! It's fascinating to read what you all have to say about this now very topical issue.

    @phil and @chris, I guess that when I wrote the original post I was thinking more about non-language-related machine learning work such as bounding box annotation and that sort of thing (the type of tasks that anyone can have a go at, but that linguists in particular might possibly be good at, or drawn to). However, I'm sure we'd all really enjoy to hear more about machine learning tasks that directly relate to translation too (thanks to @kvstegemann for your note on terminology) ;)

    Equally well, I'm more than interested to see what anybody else has to say on any of the issues that we've touched on above, or other questions that relate to AI/machine learning in a broader sense.

    Don't hesitate to keep your comments coming!

  • 11
    Avatar
    mdj_cm

    $3 per hour and we are kindly asking you why you aren't interested... But don't forget that with good practice, you may earn 10 USD per hour, at unsustainable high speed!

    Beside the pay rate, which is ultimately the most decisive factor, what sort of AI projects are you referring to?

    AI projects that are exclusively designed to reduce costs and replace us (mostly underpaid workers) in the long run (for the sole benefit of investors and corporate clients)?

    Or meaningful AI projects that are designed to improve our lives and/or that of future generations? In which case we may all participate pro bono!

    In the ideal world; we would first ask ourselves if a particular project is meaningful or/and relevant to us, the "laborers". I understand we can omit that part though in today's world and stick to what unites us here in the first place: money, and fair trade if I may add.

    The AI subject can be a very sensitive one among laborers, mostly for the 2 reasons cited above.

    Considering you cite pros that should actually justify a reasonable pay rate: "strong analytical skills, attention to detail, and critical thinking." and "there is a relatively high degree of finesse involved".

    Something isn't right.

    Kind regards

     

     

     

    mdj_cmにより編集されました
  • 1
    Avatar
    Margherita Ruggiero

    I agree with what you all say about very law earnings - but this is happening in general for all of the work that can be done through crowdsourcing as it is possible to count on a wide number of people willing to earn just a few dollars for an extra job, and in the case of translations they are not necessarily professional translators.

    Our job as translators is more complex, we not only translate from one human language into another, we translate from one culture to another  - which, by the way, is something machines are not ready to do yet.

    What if in the long run our future lay in the ability to translate from human to machine languages and viceversa? Science fiction? I am giving it a thought...

  • 0
    Avatar
    Katrina Paterson

    Hi @Margherita Ruggiero, thanks for your very interesting last comment. What do you think the machine languages would look or sound like? And if we were to translate from machine languages to human languages, where would the machine language or the machine speech have originated from? Would it mean that machines would effectively generate their own independent speech, and we would translate this into a language that humans understand? I'm also very intrigued by your question and while maybe it does sound like science fiction for now, who knows what the future holds?

  • 0
    Avatar
    Katrina Paterson

    To everyone else that's commented, thanks again. I see that the questions of rates and of machines replacing humans are both (understandably) quite popular topics so I just wanted to clarify a couple of points. The first is that when I refer to machine learning I'm referring to the entire broad category of tasks that this encompasses, such as image recognition, speech recognition, text categorisation, sentiment analysis, and not necessarily just language-related tasks. The second point is that my intention was to talk about industry trends as a whole, not about Gengo or one specific company or companies. Given that AI/machine learning is a massive, wide-ranging topic, I'd be very interested to read any kind of thoughts that people have, in general, about the concept, including ideas that we haven't discussed yet, if you have new things to add. 

    To address the question of the replacement of humans in particular, do any of you have any thoughts about possible new opportunities that machine learning might open up for humans?

    Katrina Patersonにより編集されました
  • 5
    Avatar
    kvstegemann

    @Katrina, I think that opportunities might arise if/when we understand the different strengths and weaknesses of humans and machines. Machines are capable of processing incredible amounts of data in incredible speed, but they cannot produce original thoughts. Humans are able to make intelligent decisions, but they need time and can easily overlook important facts among the mass of input. That's why I made a point of getting the definitions right in my first posting: Humans still have a monopoly on intelligence, and I don't expect this to change. Combining human intelligence with machine learning might be a viable option to create new business models.

    That's not a new idea. "Amazon Mechanical Turk" tries to market a similar concept for 15 years now. However, AMT has quickly deteriorated to a human exploitation scheme, making people work for pennies, so that we are back to posting 1 again. It seems that this kind of mechanization of humans combines easily with the selection of the lowest bidder, of the people who are easiest to exploit. Finding automation applications that actually need highly educated and more specialized humans seems to be difficult, or maybe not so attractive for business.

    Here's one idea. One thing that machines will never be able to do is expressing the wishes, desires, and needs of humans. Whenever a machine needs to decide between several alternatives on something, and the main deciding factor is how much the given alternative would appeal to a human (reader, consumer, voter, ...), this decision could be transferred to a human. Often, decisions like this might depend on culture and location, so that linguists might be suited well for this.

  • 0
    Avatar
    Katrina Paterson

    @kvstegemann, I'm really enjoying reading all of your comments.

    Do you think that machines will ever be capable of producing original thoughts? I've heard people theorise that at some point, if given enough data, machines will able to 'train themselves' in ways that we don't understand and we will ultimately lose control of them. Do you think there's any grounding in this concept or is it venturing into the realms of science fiction?

    I agree with what you say about humans being able to make intelligent decisions in a way that machines can't, now, and I instinctively see what you're saying about human wishes, desires and needs being beyond the comprehension of machines. At the same time, I'm reading Yuval Noah Harari's '21 Lessons for the 21st Century' at the moment and he argues that what we call 'intuition' is actually more like pattern recognition. At one point, he says that 'If you think AI needs to compete against the human soul in terms of mystical hunches - that sounds impossible. But if AI really needs to compete against neural networks in calculating probabilities and recognising patterns - that sounds far less daunting.' (p. 31 of the aforementioned book)

    Can we reduce our deepest hopes, fears, desires and so on to 'pattern recognition' or is there something more to our character as humans? I believe that our behavior is based on something more than rational cost-benefit analyses, otherwise how could we explain many of the big questions of the human condition, such as why do so many people fall in love with people who are totally incompatible with them?

    But then, people that defend the widespread adoption of machine learning often argue that machines can overcome many of the implicit biases that humans carry and/or the human tendency to make mistakes (such as overlooking important facts, as you mentioned in your comment). I'm not sure how this factors into the argument of the pros versus cons of machine learning, but it's an interesting idea, I think.

    I'd be really interested to learn more about what you have to say.

  • 6
    Avatar
    Nelson Bras

    Mmmm…. Let’s consider it in practical terms…

     

    English verbs vs Portuguese verbs:

    ENGLISH

    “Do”

    - do, does, doing, did, done (5 possibilities)

    PORTUGUESE

    “Fazer”

    Faço, fazes, faz, fazemos, fazeis, fazem, fazia, fazias, fazíamos, fazíeis, faziam, farei, farás, fará, faremos, fareis, farão, faria, farias, faríamos, faríeis, fariam, fiz, fizeste, fez, fizemos, fizestes, fizeram, fizera, fizeras, fizéramos, fizéreis, faça, faças, façamos, façais, façam, fizesse, fizesses, fizéssemos, fizésseis, fizessem, fizer, fizeres, fizermos, fizerdes, fizerem, faz, faze, faça, façamos, fazei, fazer, fazeres, fazermos, fazerdes, fazerem, fazendo, feito. (59 possibilities!!!)

    This is just one of many reasons why using machine translation to translate from English into romantic complex languages will never be as easy as the other way around.

    Another reason would be the fact that 99% of the times MT will give you a Brazilian translation, which may be entirely different from European Portuguese. E.g.:

    en: pt-br / pt-pt

    • Bus: ônibus / autocarro
    • Cell Phone: celular / telemóvel
    • Train: trem / comboio
    • Juice: suco / sumo
    • Goal Keeper: goleiro / guarda-redes
    • Bathroom: banheiro / casa de banho
    • T-shirt: camiseta / camisola
    • Ice Cream: sorvete / gelado

    Some terms may even lead you to severe problems, due to considerable differences of meaning between pt-pt and pt-br. E.g.:

    • “Puto”- a “kid” in Portugal / a male prostitute in Brazil;
    • “Bicha” – a "cue" in portugal / an offensive way to refer to homosexuals in Brazil…
    • etc.

    99% of the times, the AI doesn’t get it right, and you have to delete everything and redo it from scratch, meaning that you will be paid half-wage to do something much harder and more time-consuming than a “normal” translation. If you are a Portuguese native and want to have a good laugh… try Google Translator with en-pt – almost every result is absolutely hilarious!

     

    I rest my case.

  • 0
    Avatar
    Margherita Ruggiero

    I guess it is very difficult to make predictions too far in the future - we do not know for example how our languages will develop, if they will keep their rich vocabularies. In Italy, younger people tend to use a limited number of words. I have just recently moved from Rome to Milan, where I found people use more frequently English words in their everyday conversations, so if this is a general trend in the next years, we will end up with less complex languages in favour of more practical languages and this might influence the capacities of machines to translate and generate languages. People will have to get used to work every day with AI colleagues, and interact with them, this will have an effect on language I think. It is happening already, when we use Alexa or Google, we tend to ask questions with simple combinations of words. Economic competitions as well as wars have always been propellent for scientific and technological improvements. So in the long run, it is possible that we will get used to weird machine language combinations and adapt, who knows. In the short run instead, humans are not being replaced by machines but by other humans potentiated with machines, so that one person can do more work (but probably and unfortunately paid less than before). 

    Margherita Ruggieroにより編集されました
  • 0
    Avatar
    Lara Fernandez

    Hey all - thank you very much for this discussion! The concept of AI training data, or ML, as Kay-Viktor brilliantly put it, is definitely a controversial one, and many people have different views on what it means for AI to become an increasingly present part of our everyday lives :) 

    I see comments that refer to working on PEMT (post-edited machine translation), which is a very valid topic in and of its own (guys, please feel free to start your own discussions and debates, as well!), but I wanted to make sure that we're all on the same page here when we say AI tasks. Especially, because some members of the community may not have been invited to these types of projects yet, and instinctively associate "AI" with the provided MT for certain types of jobs. 

    What we at Gengo call AI tasks, are tasks that go well beyond translation, such as sentiment analysis, text/data categorization, entity annotation, image annotation (bounding boxes, etc), keypoint annotation, transcription and video captioning, and more that are being added as we expand our offerings :) You can read in detail about them here. Some of you above have brought up great points based on your experiences with these types of tasks and your stance as a translator. For those of you who have participated yet into a project involving these types of tasks, we would love to hear more about how you'd feel about being invited to these types of projects more in terms of "as a translator, is this a type of work that appeals to you, why or why not?" As Katrina already stated, there are lots of controversial opinions around this, ranging from "as a freelancer, I like to work on different fields" to "as a translator, I want to work on translation only". 

    We'd love to broaden the discussion in terms of freelancing in a wide variety of tasks versus specialization, and your thoughts about it! Feel free to think industry wide -- this is not necessarily a Gengo specific topic!

  • 0
    Avatar
    Katrina Paterson

    Indeed - I very much agree with Lara's point that the theme of 'AI' or 'machine learning' (thanks again, @kvstegemann) is hugely broad and it encompasses lots of different concepts, some of which are directly relevant or of interest to us as translators and others of which are quite different indeed. I guess that for those of you that have tried some of the tasks that Lara mentioned (text/data categorisation, image annotation and so on), it would be very interesting to hear what your thoughts are. For those that haven't and are curious, Lara's link gives a really interesting overview and provides some good excellent for this discussion.

    I think we'd also be curious to know people's general thoughts about AI/machine learning as a topic, since it's so frequently talked-about at the moment and will no doubt become even more so. If there's anything that you've heard or read and you'd like to add your voice to the discussion, please don't hesitate to do so!

  • 0
    Avatar
    Katrina Paterson

    @Nelson Bras, I do very much see your point about the relative complexity of Portuguese compared to English and my perception is also that machine translations into English are generally better than machine translations out of English. However, the technology is evolving all of the time (for better or for worse depending on where your interests lie). Do you think some language combinations are more 'safe' than others? I ask this out of genuine curiosity as I work into English and my personal feeling is that into-English machine translation has changed a lot in recent years. I would be very interested to learn what the opposite situation is like.

  • 0
    Avatar
    Katrina Paterson

    @Margherita Ruggiero - it's very interesting to read your comments about the language of younger people in Italy changing, especially about the adoption of English expressions. Do you think the English influence comes from travel, or from contact with non-Italian people, or from the internet? Your comment reminded me that in some ways we could perhaps also argue that the internet and technology are changing people's speaking and writing habits, not only in the sense of the different sorts of expressions used on the internet (slang, people writing in their non-native languages and so on) but also in the sense that people often seem to engage less, or at least to engage in different ways, with things they read online. People's online attention spans are a lot shorter and so I think a lot of online content has been created to be snappier and more attention-grabbing. Do you think consuming this type of content will have an impact on people's language habits?

    Another thing I'm curious about is the way that travel and communication affect language, and particularly the way that communication between native and non-native speakers can affect language. English is (for now at least) one of the main 'global' languages of communication but the way in which it is used varies hugely between native speakers in different English-speaking countries and also between native and non-native speakers. Do you think this also happens in Italian, when people from different regions speak to one another? Do you think that one day both human-human and human-machine communication will be more 'global' and generalised, and less specific to particular places?

サインインしてコメントを残してください。