Word AI Feature Request

Discussion in 'Apex Lounge' started by msharron, Jul 9, 2017.

Share This Page

  1. msharron

    msharron Member

    Joined:
    May 7, 2016
    Messages:
    25
    Likes Received:
    13
    Right now I'm balls deep in the semantic web.
    @cardine - I've maintained a subscription to WordAI since my Wickedfire days even through I've not used it a great deal.
    These times they are a changing, so I've got some feature requests for you or at the very least I'd be interested in getting a dialogue started.

    Here's what I'd like to see:

    Natural Language Processing (sorry I cant see that in the present system from the output evening with the Turing spinner).
    Entity Extraction + Entity Protection
    Topic Analysis / specify topic / restriction of vocabulary based on text topic classification/selection.
     
  2. msharron

    msharron Member

    Joined:
    May 7, 2016
    Messages:
    25
    Likes Received:
    13
    The ability to maybe connect WordAI to TextRazor or Wordlift.io's API.
     
  3. cardine

    cardine Administrator Staff Member

    Joined:
    Dec 9, 2015
    Messages:
    1,064
    Likes Received:
    1,026
    Hey! I think I saw your support ticket where you mentioned many of the same things. I talked to Dana (who I think responded to you) about it but I'll add my own two cents:
    For this I'm not sure exactly what you mean as Natural Language Processing simply means "computer programs that can process text" - which WordAi definitely does. Reading between the lines I think you more specifically are saying that you wish WordAi did a better job processing text. I certainly understand and emphasize with that sentiment - I think WordAi Version 4 was a huge step in that direction and I think the infrastructure for Version 4 has set the stage for Version 5 to make even more dramatic advances forward.

    We actually have this already. If you go to the "protected words" part of the account you'll see that we have the option to protect titles and proper nouns, which includes extracted entities. Entity extraction is used throughout WordAi but we don't display any of that externally to the user.

    We do this as well, although it certainly is not perfect. We do not have a specific topic classifier but rather we train that into our entire synonym selection algorithm. We do it that way because in a sentence like "I saw a saw" or "I suspect the suspect is guilty because he seemed suspect" or "After I cut my friend in line I tripped and cut my knee" the same word can have multiple definitions even if the sentence is one topic. If there are any specific examples you have where you think WordAi is consistently off for a topic then definitely send us a support ticket and we can make sure to try to give the WordAi "brain" more examples with those subjects so that it improves for those synonyms.

    WordAi Version 3 was built around our own internal version of something sort of similar to TextRazor. With that being said TextRazor looks to be fairly low quality. For instance, if you enter "The suspect looked suspect" it will give you definitions that lead to a sentence like "The defendant looked defendant" because it gets the wrong definition for the "suspect" - it also gets the definition for "looked" wrong that likely would have led to it getting replaced by something like "conveyed".

    The WordAi Version 3 engine was doing a lot of the same work and would have unequivocally got those definitions right - in fact, we had a well-placed source at Google who let us benchmark our engine against Google's own engine and our engine gave higher quality results than theirs (the caveat was our engine ran over 1000 times slower which would obviously not scale up to the amount of text Google needs to process).

    However for Version 4 we moved away from that technology entirely and have shifted to a much less brittle way of approaching things. In general the TextRazor approach requires huge comprehensive synonym databases for each definition that are impossible to keep up with how language or niches naturally evolve. Plus those synonym databases have their own issues... for instance imagine "help" had "assist" as a synonym... so a sentence like "I decided to help him out" would end up with "I decided to assist him out"... even though the two words mean the same thing they can't be replaced in that instance.

    The current Version 4 technology doesn't use any synonym databases and it simply learns new ways of writing things as it reads real articles. This much more organic process allows it to naturally keep up and over time it will teach itself what mistakes it is making or not making. So the current "brain" simply knows that a sentence like "I decided to assist him out" doesn't make sense without the need to add any special logic or database checks or anything like that (which is important because there are millions of other examples of things similar to the "help" vs "assist" issue).

    Now with all of that being said WordAi is definitely not perfect right now - however I think it is going in the right direction and the current architecture (which is very similar to the research DeepMind is doing) is going to allow us to make huge advances in the near future.
     
  4. msharron

    msharron Member

    Joined:
    May 7, 2016
    Messages:
    25
    Likes Received:
    13
    Great response.

    I had pegged wordai as part of a work flow for vocab content and lost confidence last week as the noun protection / word protection was non functional.

    I'll have a play again tonight and see what it can do.

    I sometimes get better results when using smaller blocks of text which is curious.

    Text razor for me is primarily an LSI KW tool.

    I've had the pleasure of training a few AI chat bots, usually with a :( when the bot screws up. Do you have anything similar?
     
  5. cardine

    cardine Administrator Staff Member

    Joined:
    Dec 9, 2015
    Messages:
    1,064
    Likes Received:
    1,026
    Definitely try again and if you are having problems with this even after turning that setting on let me know and I can look further into it.

    Even though I don't think Text Razor makes sense for spinning I definitely can see where it would have application for LSI or other text analytics stuff. A little over 3 years ago I was very close to releasing a joint venture product that involved releasing a product very similar to Text Razor but at the time I didn't see a solid commercial application for it (and I am generally reluctant to do joint ventures). The site and product were all created and ready to go and at the last second I decided not to pull the trigger. Maybe Text Razor will prove me wrong!

    I haven't done any AI chat bots recently but I have made quite a few of them in the past - some were done for fun but there were also quite a few specialized chat bots that made me a lot of money. I think if we spent as much time on a chatbot as we have on WordAi we'd likely have a nearly perfect chatbot by now. Even now if I was given 10 months and sufficient motivation (a truckload of cash) I think we could build an unbelievably high quality chatbot - the technology WordAi Version 4 uses right now would make for an absolutely perfect fit for a chatbot, and I think a perfect chatbot is a far easier problem than a perfect spinner.

    Unfortunately I don't see any strong commercial application for a generic chatbot right now. I think that's a big reason why they aren't nearly as prevalent as people thought they would be. But the technology is definitely out there right now to create an extremely high quality chat bot.
     
  6. msharron

    msharron Member

    Joined:
    May 7, 2016
    Messages:
    25
    Likes Received:
    13
    What I was asking is if there's a way to train WordAI if it gets something wrong?
     
  7. cardine

    cardine Administrator Staff Member

    Joined:
    Dec 9, 2015
    Messages:
    1,064
    Likes Received:
    1,026
    Oh... I think I misunderstood the direction you were going in with that post!

    The answer is yes, although there is no way for users to directly access it. If users select/unselect synonyms that info gets sent back to WordAi's brain. At the same time we are continually feeding WordAi's brain millions of articles and it is continually learning from that. Realistically the millions of articles its reading everyday results in a far bigger impact than the selected/unselected synonyms that users do.
     
  8. msharron

    msharron Member

    Joined:
    May 7, 2016
    Messages:
    25
    Likes Received:
    13
    Right the problem with that approach is a lot of people using the tool are probably not paying strict attention to the quality of its output.