[INDOLOGY] {भारतीयविद्वत्परिषत्} Artificial Neural Network for Sanskrit programming

Amba Kulkarni ambapradeep at gmail.com
Thu Jul 16 02:48:09 EDT 2015


A correction:

Please read "The performance figures for 56 tags were as follows" as "The
performance was as follows".

-- Amba Kulkarni


On 16 July 2015 at 11:27, Amba Kulkarni <ambapradeep at gmail.com> wrote:

> Dear Dhaval Patel,
>
> Thanks for sharing this.
>
> I have a question: Is the test data different from the training?
>
> I assume you have referred to the following two works:
>
>    1. Statistical Constituency parser for Sanskrit compounds
>    <http://sanskrit.uohyd.ernet.in/faculty/amba/PUBLICATIONS/samaasa_const_parser_icon2011.pdf>
>    Amba Kulkarni and Anil Kumar, ICON 2011, Chennai Dec 18-19
>    2. Clues from Astadhyayi for compound identification
>    <http://sanskrit.uohyd.ernet.in/faculty/amba/PUBLICATIONS/scti-5scls.pdf>
>    Amba Kulkarni and Anil Kumar, 5th international SCLS 2013, Mumbai
>
> The first one reports the performance of the system, using only simple
> probabilities.
>
> The performance figures for 56 tags were as follows:
>
> Rank      With 55 tags       With 8 tags
> 1                   63.0%               72.7%
> 2                   10.9%               13.2%
> 3                     7.2%                 9.5%
>
> where Rank indicates the position of the correct solution among all
> possible solutions.
>
> The second one uses clues from the A.s.taadhyaayii to detect the compound
> type of rare compounds, and the performance was 61%  times we got the first
> guess correct, and if we allow 3 guesses, 81% one among the 3 guesses were
> correct.
>
> With regards,
> Amba Kulkarni
>
>
> On 16 July 2015 at 00:03, dhaval patel <drdhaval2785 at gmail.com> wrote:
>
>> Respected scholars,
>> Recently I have modified one Artificial neural net code for
>> identification of samAsas in Sanskrit language.
>> https://github.com/drdhaval2785/SamaasaClassification is the code
>> location.
>>
>> The results are very encouraging.
>> Without feeding any rule to the computer, the following is the
>> classification result
>>
>> 1. Major 5 samAsa types classification - 70%
>> 2. Minor 55 samAsa subtypes classification - 55 %.
>> 3. Major 5 samAsa types classification taking the first two entries - 85
>> %.
>>
>> Probability of a fluke would be 1 - 20 %, 2 - 0.2 %, 40 %.
>>
>> So, the machine learning is statistically significant, if not good enough.
>>
>>
>> The database which was used was scraped from
>> http://sanskrit.uohyd.ernet.in/Corpus/SHMT/Samaas-Tagging/ and randomly
>> shuffled to homogenize the dataset.
>>
>> The tool was developed for samAsa classification initially, but now
>> generalized for any string classification problem.
>>
>> Hope the scholars would like the tool.
>>
>> For those interested in Artificial neural networks, the link is
>> http://neuralnetworksanddeeplearning.com/
>>
>>
>>
>> --
>> Dr. Dhaval Patel, I.A.S
>> Collector and District Magistrate, Anand
>> www.sanskritworld.in
>>
>> --
>> निराशीर्निर्ममो भूत्वा युध्यस्व विगतज्वरः।। (भ.गी.)
>> ---
>> You received this message because you are subscribed to the Google Groups
>> "भारतीयविद्वत्परिषत्" group.
>> To unsubscribe from this group and stop receiving emails from it, send an
>> email to bvparishat+unsubscribe at googlegroups.com.
>> To post to this group, send email to bvparishat at googlegroups.com.
>> Visit this group at http://groups.google.com/group/bvparishat.
>> For more options, visit https://groups.google.com/d/optout.
>>
>
>
>
> --
> आ नो भद्रा: क्रतवो यन्तु विश्वत: ll
> Let noble thoughts come to us from every side.
> - Rig Veda, I-89-i.
> Assoc Prof.
> Department of Sanskrit Studies
> University of Hyderabad
> Prof. C.R. Rao Road
> Hyderabad-500 046
>
> (91) 040 23133802(off)
>
> http://sanskrit.uohyd.ernet.in/scl
> http://sanskrit.uohyd.ernet.in/faculty/amba
>
>


-- 
आ नो भद्रा: क्रतवो यन्तु विश्वत: ll
Let noble thoughts come to us from every side.
- Rig Veda, I-89-i.
Assoc Prof.
Department of Sanskrit Studies
University of Hyderabad
Prof. C.R. Rao Road
Hyderabad-500 046

(91) 040 23133802(off)

http://sanskrit.uohyd.ernet.in/scl
http://sanskrit.uohyd.ernet.in/faculty/amba
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.indology.info/pipermail/indology_list.indology.info/attachments/20150716/d527e19e/attachment.html>


More information about the INDOLOGY mailing list