Hi Jean!

It’s so nice you keep working and improving Harumi!
I’ve visited your site but I couldn’t find the link for downloading Harumi 2.
Can you help me to find out?

Hello Fatima, here are the links (you can see those links on my site just after “starcraft” picture) :

Harumi 2 alpha version in french :
http://jeffelson.e-monsite.com/medias/files/harumi-2-french-alpha-version.zip

Harumi 2 alpha version in english :
http://jeffelson.e-monsite.com/medias/files/harumi-2-english-alpha-version.zip

Regards,

Jean-François

  I’ve not updated Harumi for a while but hope to do so soon. Many upgrades have been made on my personnal version, espcially fixes and better interactions, even if it works only in french since I use Harumi every days for my convenance.
  The new interesting function will be the ability for Harumi to catch RSS flux and giving informations to the user out of Internet without any active request. For the fist time, Harumi will interact not only with user but also by reading the news (I’ve picked up a news site RSS flux for the moment)
Here is what you can expect from her :
Harumi : I don’t like storms because power outages occur…
Harumi : Mid-Atlantic storms leave power outages, tossed boats 

How does it work? 1) Harumi controls RSS flux from google news, searching for the last news. As soon as she grabs a new title, she explores her database to match it with the closest stimulus available. In this case, storm+power outages are detected in the database so she “knows” the answer(s) linked to this stimulus is/are suited to the RSS flux. This way, the user will have the feeling not only to be informed, but also to have a comment from Harumi

Jean-François, what is “RSS flux”?  It seems to be something French.  I could only find this definition: “RSS flux is a text file whose content is produced automatically based on information that is updated and published on a website.”  It seems you are doing what I would like to do.  I work with web feeds a lot.  Have you seen my project at http://meta-guide.com ?

Marcus Endicott : an RSS flux is a simple internet site with only text informations. It is very usefull to automatically grab datas. For example, to check the news, you can open say http://www.reuters.com but you’ll have a profusion of pictures, videos, adds… which are very complicated to analyse in real time by a computer.
On the other hand, the RSS Reuters site countains only text informations, meaning titles, description of articles, aka datas easy to handle for a machine.

> http://en.wikipedia.org/wiki/Web_feed

I suspect that “flux” means “feed” in French.  Can you confirm that you are working with standard RSS XML (above) ?  If so, I’m very interested in how you are parsing it to NLP.  What AI engine are you using, pattern matching ?

Marcus Endicott : You are right, it’s a feed, not a “flux” (in english).

  Regarding AI engine, I’m using actually my own since Harumi was entirely programmed by me from scratch in delphi (an evolution from turbo pascal with graphic interface)

  To implement RSS reading in Harumi, I first added a function which records as a long string the internet site with RSS feed, then I defined ineresting “edges” out of the string marking beginning and end of title and description. Those edges are editable by user since they appear in the memory file of Harumi, an excell like sheet which countains her “brain”.
This way, anyone can edit the RSS site and proper edges to grab intersting datas/informations.

  Harumi then cut title coming from RSS feed in words (string list) then create a list of possible combinaisons. This list is then matched by fuzzy logic with stimulus and answers database from her memory. This is how you Harumi can say
“I don’t like storms because power outages occur…” because she saw “Mid-Atlantic storms leave power outages, tossed boats” in RSS feed.

Marcus, I just added a new tool to our project tool-stack, RapidMiner. You can use RapidMiner to scrape RSS-feeds and then run all sorts of machine learning algos on the retrieved information, ultimately generating any data structure you want. I suspect you might even be able to generate AIML directly from it, as it supports exporting to XML-structured files.

For a quick introduction into RapidMiner you might look here: http://vancouverdata.blogspot.nl/2010/11/text-analytics-with-rapidminer-loading.html

> http://www.meta-guide.com/home/knowledgebase/best-rapidminer-videos

Best RapidMiner Videos

= = =

HP, that’s a hot tip, thanks!  I’ve installed RapidMiner on my new machine, and will let you know how it goes.  I’ve made a webpage above on “Best RapidMiner Videos”, which also includes tutorials.

Hi!
Last version of Harumi released today (07.08.2012) !

RSS feeds available and editable (sheet 6 lines 462 to 466)
Click on newspaper icon and Harumi will extract informations from yahoo news rss feed.
Note that french version works better for the moment.

I also upgraded the way Harumi add informations to a stimulus.
Example :
User : Paul is my best friend (then click on student icon in order to add this information as a new stimulus/entry)
Harumi : Ok
User : his phone is 01 02 03 (then press enter)
Harumi : Paul is your best friend? (A way for Harumi to ask confirmation that the phone number is Paul’s one)
User : yes (note that you don’t have to write “yes” in order to record the phone number. Nevertheless, in case you click on “No” icon, Harumi will delete “his phone is ...”

Imminent release of Harumi update. Features below :

A) when a sentence is given to Harumi, she has to “understand” the meaning of it.
If I say something like “I went to the beach with my wife and childrens and while I was swimming, I saw a shark”,
any human beeing is able to understand that what is important is “I saw a shark” and not “I went to the beach” or “wife”.
On the other hand, if I say “My wife loves me”, then wife becomes an important word in the sentence.

I wanted to give the ability to Harumi to hierarchise informations.
Here is how I suceeded :
Harumi cuts any user request in words and 2 words combinaisons (since my computer is too slow to treat more than 2 words combinaisons).
Then she calculates the occurence for each of those words/combinaisons.

The best word/combinaison is the one with the rarest occurence in the database as long as rare doesn’t mean 0.
Results with this simple idea were immediatly far better for treating datas in a large database.

B) Shortest stimulus is better
Here is another kind of problem I solved :
I created as stimulus :
my photos
my photos during 2012 holidays
my brother’s photos

When I asked Harumi “show me my photos”, she opened the 3 repertory linked to the 3 stimulus.
I decided that the best stimulus when many are available, is the SHORTEST one.

C) Answers backed as stimulus
When many answers are linked to a stimulus, it appears that some datas inside answers are not reachable directly from stimulus. Example :
Stimulus column       answer 1                   answer 2                                 answer 3
terminator           it’s a james cameron movie   Sarah Connor is John’s Connor mother     Arnold is the terminator

In case you ask “do you know james cameron movie with arnold?”, you won’t have any answer since answers are out of reach.
So I decided to put answers back as stimulus, which modify stimulus column as :
terminator *** james cameron movie sarah connor john arnold

Note that “***” mark is used to hierarchise before and after main stimulus for some cases like :

terminator *** james cameron movie sarah connor john arnold
hal 9000 *** 2001 space odyssey terminator

if you ask “who is terminator?”, Harumi will use first stimulus because terminator is the main subject while it appears as a secondary subject in second line.

D) Context and “ghosts stimulus”
I created a list of “ghost stimulus” which are used in combinaison of words only. Ghost stimulus are what Harumi just said, what she sees on webcam, geolocalisation….
It helps Harumi to figure out the context of the discussion

Hello people, Harumi’s last version released today (09.12.2012) !

When I told Harumi, “I have a Dr. appointment on Friday, September 14 at 8:50 am., the program started with endless Pop-ups stating, ” ‘85000’ is not a valid integer value.

I think I caught a brief glimpse of Harumi’s text indicating something to the effect, “I’ll try to remember that for you.” or something similar. I’ll wait to see.

I was able to hold the ALT key while repeatedly pressing the F4 function key, until all the active windows of the Pop-ups were closed (along with Harumi’s window too…:( )

Up until the error, we were getting along pretty well and I was teaching “her” lots of new info. I do notice a difference from the old version. Please keep going forward in developing the program. It has a tremendous entertaining and educational value.

Thanks!

Art : Alarms and appointment function works only in french since it was an old function that I didn’t edit for at least a whole year, if not two.
On the other hand, you noticed that the way she learns is very different now, because Harumi put answers datas back on the stimulus in order to treat more efficiently user requests. Moreover, “visible thinking” (icon with the brain) helps the user to get how Harumi “understands” requests. Note that if you are not satisfied with Harumi answer when you know a better one exists in her brain, you can click on “no icon”, then she’ll try another words combinaison suited to your precedent request. Note that at the moment you click on “no icon”, the precedent words combinaison she picked up is added in sheet 2 in the list of combinaisons to avoid.
Thanks for your support!

Jeff

New version of Harumi released today (10.07.2012)

I modified heavily the way Harumi analyses sentences. You will notice a better precision to pick up the best answer available in the database. Harumi is also able to understand despite minor syntax errors in user request.
Moreover, the way she records data with “learning icon” is far more efficient. No more “+” necessary at the beginning of a sentence.
Just write the sentence and press the icon, she’ll ask you informations to put the sentence at the right place.
For example :
User : Paul is Jack’s brother (then press learning icon)
Harumi : Do speak about Linda? (because the precedent subject was about Linda)
User : (Click on “No icon”)
Harumi : Does this has something to do with Jack Nicholson (I suppose she knows something about the actor)
User : (Click on “No icon” again)
Harumi : All right! Paul is Jack’s brother (A new entry appears as a new stimulus)

Note that you can click yes to put the sentence as a new answer available for an older stimulus.
Thanks to test it and speak about Harumi to friends if you enjoy using her!