back to article 39 episodes of 'CSI' used to build AI's natural language model

A group of University of Edinburgh boffins have turned CSI:Crime Scene Investigation scripts into a natural language training dataset. Their aim is to improve how bots understand what's said to them – natural language understanding. Drawing on 39 episodes from the first five seasons of the series, Lea Frermann, Shay Cohen and …

  1. Anonymous Coward
    Anonymous Coward

    They might pick...

    The show _CSI Miami_. This show is also predictable, and (surprise) overacted by its principle actor.

    Unfortunately, it still shows up on a TV Channel I watch, and I need to switch channels. I still get the promos which are just as ....

    Oh, well.

    1. The Man Who Fell To Earth Silver badge

      Re: They might pick...

      39 episodes of 'CSI' ? Sounds more like this is training AS (Artificial Stupidity).

  2. frank ly

    Unknown knowns?

    "... in the one episode of the 39 involving a suicide, ... the model kept guessing right to the end."

    Was the model 'aware' that it's possible for victim and perpetrator to be the same person? This might be an oversight on the part of the designers of the model.

  3. Pete 2 Silver badge


    It has always struck me as amusing that the vast majority of the data broadcast / streamed for a TV programme is video. However, that contributes almost nothing to the informational content of a programme. As this is mostly in the tiny little proportion of audio sent along with it. A situation that is even more apparent with HD and 4K: large increases in the (mostly content-free) video, very little if any change in the audio - and none at all in what is arguably the most important part of any TV drama: the script.

    I would expect that at least some of the clues, picked up by the audience in trying to guess "whodunnit" would be visual. But that these would be inaccessible to the AI-thingy. So it comes as no surprise that people guess better than computers. It might also be that the human audience knows the "rules of the game" that the perp. won't be revealed before the first ad-break.

    Surely a better source of training material would be to use radio programmes, or to only have the audience listening to the sound (and not viewing the screen). That would give a more equal basis for comparison.

    1. Anonymous Coward
      Anonymous Coward

      Re: visuals?

      I beg to differ....

      A TV show (no matter how lame the story line and acting might be) is a complex interaction of visual, speech and musical clues... remove any "channel" and you're likely going to miss a lot of the action.

      Just for fun try the following with an episode you have not seen before and see if you really "get" the story:

      - Watch the show without any sound

      - Listen the TV show without the image

      Even weirder is watching a show without the soundtrack... you suddenly realise how much the music "cues you in" on the action to come... and that without it the show is really "flat".

      1. Simon Harris

        Re: visuals?

        Or you could do without the visuals and the sound, and scan a bunch of Agatha Cristies, Morses, Rebuses, etc.

        Come to think of it, University of Edinburgh - they should have used Rebus!

      2. elgarak1

        Re: visuals?

        Ahhh... No.

        Modern TV drama have not a lot of time to be creative with visual storytelling. They are extremely dialog-heavy for exposition. In fact, I have done the second of your options, and there are lots of shows that I can follow the story just fine, just by dialog.

        CSI (the original) is a prime example of how it tumbles down. The first few seasons were fairly intelligent story-telling, if formulaic. But a different formula from other procedural crime dramas, which made it fresh. Then the formula settled in, it got boring, and viewers left, so in later seasons they needed to dumb down the show to keep their viewers. It removed every single bit of visual story-telling, and moved every bit of exposition to the dialog. And. to. make. sure. everyone. gets. it. they. started. to. talk. slowly. and. repeated. everything. Repeat. everything. and. slowly. so. everyone. understands. what's. going. on.

      3. Stuart Castle Silver badge

        Re: visuals?

        Re: A TV show (no matter how lame the story line and acting might be) is a complex interaction of visual, speech and musical clues... remove any "channel" and you're likely going to miss a lot of the action.

        Depends on the show (and in particular, the director). If you watch any soap, the drama tends to be in what the characters say rather than do, and the visuals don't actually change much. Certain other TV shows are like that as well. For instance, you can generally get the gist of what is happening in Doctor Who by listening to the dialogue, if not the finer detail. Then there are other shows where the soundtrack is almost secondary to the visuals, and you would have little or no clue regarding what is happening if you weren't watching (e.g. Legends of Tomorrow).

    2. Fruit and Nutcase Silver badge

      Re: visuals?

      @Pete 2

      I would expect that at least some of the clues, picked up by the audience in trying to guess "whodunnit" would be visual.

      Like the big name guest star in an episode of Columbo

      1. elgarak1

        Re: visuals?

        Columbo told the story from the perp's view.

        CSI and later procedurals are really easy to predict by the 'bigness' of an actor, though. A well-known name with just a two-sentence witness statement in the second act? Guilty. Will come back in the last act.

        Unknown actor gets grilled for three pages of script and has all evidence against them? Nope, they're not it.

    3. Fuzz

      Re: visuals?

      If the AI had access to the script then the clues would be present. A script isn't just the dialogue.

      Also I'm guessing that CSI contains a lot of exposition with characters walking around explaining the plot.

      Also I think that the human testers would also just be reading the script rather than being forced to watch episodes of CSI.

    4. myhandler

      Re: visuals?

      It's essential it knows that humans take their sunglasses off to think.

      1. Simon Harris

        Re: visuals?

        ... and that it knows the exact duration of the pause before the punchline.

    5. Mark 85

      Re: visuals?

      In this case, you're right. It seems to follow the old Perry Mason novels, radio and then TV. Very formulaic and transitioned well from novels to TV. So the novels and even radio versions would work well.

      1. Fruit and Nutcase Silver badge

        Re: visuals?

        Star Trek - Original Series...

        Landing Party members wearing red will get vaporised

  4. Mark York 3 Silver badge


    Just as well they didn't start showing it Person Of Interest - Especially with regards to Samaritan.

  5. Anonymous Coward
    Anonymous Coward

    CSI is Natural language????

    If CSI is "natural language" then I am the Pope of England, why not "Downton Abbey" or maybe closer to the mark "The Wire" ...

    1. TRT Silver badge

      Re: CSI is Natural language????

      Obviously from a good family, well off, signs of recent malnourishment though, rope marks on the wrist, gun shot wounds... it looks like this project was part of the Google family - but someone took it... *removes glasses*... behind the woodshed.

      1. Aladdin Sane

        Re: CSI is Natural language????


  6. Slx

    Feed it a couple of seasons of classic Coronation Street and it will probably refuse to interact other than using a Bet Lynch or Vera Duckworth avatar and will have a strong desire to get into a bar brawl with anyone who tries anything!

  7. Anonymous Coward
    Anonymous Coward

    natural language model

    NATURAL? There's NOTHING natural in how they speak in that (...)! :)

    1. Forget It
      Thumb Up

      Re: natural language model

      I agree it is highly unnatural to speak so fast in turn without anyone ever um'ing or erring.

  8. Terry 6 Silver badge

    Or Casualty

    Now that's predictable. I don't even watch it, but I know that if I see some happy-go-lucky individual(s) going about their normal business appear on the screen just as I exit the room they will have come to some sticky end by the time I return. (Wife has now banned me from saying "He's a gonner" as I make my exit).

    1. TRT Silver badge

      Re: Or Casualty

      We usually have a bet on that - will the "ordinary Joe" (usually someone with a chainsaw / table saw / other power tool / 1cwt box of fireworks / getting into a motor vehicle) be (1) a victim or (2) a cause. Occasionally someone goes for (3) a new member of staff, but that's cheating because they often announce new cast members in TV Tripe or some other rag.

  9. Anonymous Coward
    Anonymous Coward

    "I'll create a GUI interface...

    ...using visual basic to track the killers IP address". I half expected them to include this priceless quote from CSI NY.

    1. Simon Harris

      Re: "I'll create a GUI interface...

      Presumably the AI now thinks that you can zoom in onto a single pixel and read the licence plate from it.

      1. Doctor_Wibble

        Re: "I'll create a GUI interface...

        Absolute nonsense, they clearly go through a really really complicated depixelatificational 'enhance' step first. This step takes mere moments if someone is just leaving the building, or for a really suspenseful episode, 37 minutes and 12 seconds, just in time for an arrest and the much-loved "Epilog".

    2. Aitor 1

      Re: "I'll create a GUI interface...

      I was just checking that this was in the comments, well done.

      1. Aladdin Sane

        Re: "I'll create a GUI interface...

        At least it's not the 2 people, 1 keyboard from NCIS.

        1. Simon Harris

          Re: "I'll create a GUI interface...

          "2 people, 1 keyboard"

          Is that the sequel to "2 girls, 1 cup" ?

          (don't google if it you don't know what that is - specially from work!)

          1. TRT Silver badge

            Re: "I'll create a GUI interface...

            If they trained it using NCIS it would refuse to work until it had at least 2 Catpows.

  10. Anonymous Coward
    Anonymous Coward

    I'd go with


    One of the few you often know who d'unnit at the beginning. Then once the "One more thing...." line is spoken, the case is cracked.

  11. Anonymous Coward
    Anonymous Coward

    Why not a 70's Brit cop show?

    Knick im.

    Shat it you slaaaaag

    you're goin daan sunshine

    Oi coppa...

    1. Aladdin Sane

      Re: Why not a 70's Brit cop show?

      Or Life on Mars?

      "Don't move, you're surrounded by armed bastards!"

    2. Charles 9

      Re: Why not a 70's Brit cop show?

      Guess it's just me and the timing, but when I think crime drama, I keep thinking Murder, She Wrote.

  12. Alan Bourke

    So they're using a television programme displaying no intelligence

    for something that isn't artifical intelligence.


  13. dokterdave

    Its the familiar actor.


  14. Ironclad


    39 Episodes of CSI, crikey.

    This must be the traumatic germ that ultimately causes Skynet to declare war on humans after it's natural language sub-mind suffers a series of horrifying flashbacks.

    1. TRT Silver badge

      Re: Skynet

      That's probably SVU or Spooks. Two cop/spy shows that highlight the utter, utter depravity and viciousness of the human race.

  15. Baldrickk

    My sister...

    ...despite having little to no knowledge of the minor 'celebrities' brought in each week became good enough at the guessing game to be able to name the perpetrator about before the first ad-break, every time - except for when they had not yet appeared on screen.

    She still enjoyed watching it however...

  16. Matthew Taylor

    An end to drudgery

    A great example of AI relieving humanity of mindless, repetitive and boring tasks.

  17. This post has been deleted by its author

  18. Stevie


    Only problem is that now the AI has false expectation of DNA evidence lab turnaround time and will shout at real life SOCOs for slacking off.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like