You can subscribe to this list here.
2005 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
(1) |
Sep
|
Oct
|
Nov
(2) |
Dec
|
---|---|---|---|---|---|---|---|---|---|---|---|---|
2006 |
Jan
(4) |
Feb
|
Mar
|
Apr
(3) |
May
|
Jun
(5) |
Jul
(9) |
Aug
|
Sep
(3) |
Oct
|
Nov
(2) |
Dec
|
2007 |
Jan
(6) |
Feb
(5) |
Mar
(3) |
Apr
(7) |
May
(3) |
Jun
|
Jul
|
Aug
(7) |
Sep
(8) |
Oct
(6) |
Nov
(1) |
Dec
|
2008 |
Jan
|
Feb
(3) |
Mar
(9) |
Apr
(2) |
May
(2) |
Jun
(2) |
Jul
(10) |
Aug
(4) |
Sep
(12) |
Oct
(7) |
Nov
(29) |
Dec
(35) |
2009 |
Jan
(10) |
Feb
(16) |
Mar
(17) |
Apr
(20) |
May
(42) |
Jun
(19) |
Jul
(32) |
Aug
(8) |
Sep
(2) |
Oct
(2) |
Nov
|
Dec
(10) |
2010 |
Jan
(7) |
Feb
(8) |
Mar
(3) |
Apr
|
May
(5) |
Jun
(1) |
Jul
(2) |
Aug
(2) |
Sep
(1) |
Oct
|
Nov
(1) |
Dec
|
2011 |
Jan
(1) |
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
(2) |
Aug
|
Sep
(2) |
Oct
|
Nov
|
Dec
|
2012 |
Jan
|
Feb
(1) |
Mar
(10) |
Apr
(1) |
May
(3) |
Jun
|
Jul
(7) |
Aug
|
Sep
|
Oct
|
Nov
|
Dec
(1) |
2013 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
(1) |
2014 |
Jan
|
Feb
|
Mar
|
Apr
|
May
(1) |
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
2015 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
(2) |
Jul
(2) |
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
2016 |
Jan
|
Feb
|
Mar
|
Apr
|
May
(3) |
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
2018 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
(2) |
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
2019 |
Jan
|
Feb
|
Mar
(1) |
Apr
(1) |
May
(1) |
Jun
(1) |
Jul
|
Aug
|
Sep
(1) |
Oct
|
Nov
|
Dec
|
2020 |
Jan
|
Feb
|
Mar
(1) |
Apr
|
May
(1) |
Jun
|
Jul
(1) |
Aug
|
Sep
(1) |
Oct
|
Nov
|
Dec
|
2021 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
(1) |
Aug
|
Sep
(3) |
Oct
(4) |
Nov
|
Dec
|
2022 |
Jan
|
Feb
(1) |
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
2023 |
Jan
|
Feb
|
Mar
(1) |
Apr
|
May
|
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
(2) |
2024 |
Jan
(3) |
Feb
(1) |
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
|
Sep
|
Oct
|
Nov
|
Dec
|
From: Xu C. <Xu...@va...> - 2018-06-25 16:58:13
|
* https://www.google.com/url?q=https%3A%2F%2Fbit.ly%2F2MXyUY6&sa=D&sntz=1&usg=AFQjCNHroTSaw7bGyOpVv5hNNskqa8WQIQ * * |
From: Cathy J. <in...@tr...> - 2018-06-15 14:45:18
|
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "https://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"> <html xmlns="https://www.w3.org/1999/xhtml"> <head> <meta name="viewport" content="width=device-width" /> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8" /> <title>My Blog</title> <style type="text/css"> /* ------------------------------------- GLOBAL A very basic CSS reset ------------------------------------- */ * { margin: 0; font-family: "Helvetica Neue", Helvetica, Arial, sans-serif; box-sizing: border-box; /*font-size: 14px;*/ } img { max-width: 100%; height: auto; } body { -webkit-font-smoothing: antialiased; -webkit-text-size-adjust: none; width: 100% !important; height: 100%; line-height: 1.6; font-size: 14px; } /* Let's make sure all tables have defaults */ table td { vertical-align: top; } /* ------------------------------------- BODY & CONTAINER ------------------------------------- */ body { background-color: #f6f6f6; font-size: 14px; line-height: auto; } .body-wrap { background-color: #f6f6f6; width: 100%; } .container { display: block !important; max-width: 600px !important; margin: 0 auto !important; /* makes it centered */ clear: both !important; } .content { max-width: 600px; margin: 0 auto; display: block; padding: 20px; } /* ------------------------------------- HEADER, FOOTER, MAIN ------------------------------------- */ .main { background-color: #fff; border: 1px solid #e9e9e9; border-radius: 3px; } .content-wrap { padding: 20px; } .content-block { padding: 0 0 20px; } .header { width: 100%; margin-bottom: 20px; } .footer { width: 100%; clear: both; color: #999; padding: 20px; } .footer a { color: #999; } .footer p, .footer a, .footer unsubscribe, .footer td { font-size: 12px; } /* ------------------------------------- TYPOGRAPHY ------------------------------------- */ h1, h2, h3 { font-family: "Helvetica Neue", Helvetica, Arial, "Lucida Grande", sans-serif; color: #000; margin: 40px 0 20px 0; line-height: 1.2; font-weight: bold; } h1 { font-size: 32px; font-weight: 500; } h2 { font-size: 24px; } h3 { font-size: 18px; } h4 { font-size: 14px; font-weight: 600; } p, ul, ol { margin-top: 10px; margin-bottom: 15px; font-weight: normal; } p li, ul li, ol li { margin-left: 5px; list-style-position: inside; } /* ------------------------------------- LINKS & BUTTONS ------------------------------------- */ a { color: #348eda; text-decoration: underline; } .btn-primary, .newsletters_button { text-decoration: none; color: #FFF; background-color: #348eda; border: solid #348eda; border-width: 10px 20px; line-height: 2; font-weight: bold; text-align: center; cursor: pointer; display: inline-block; border-radius: 5px; text-transform: capitalize; } /* ------------------------------------- OTHER STYLES THAT MIGHT BE USEFUL ------------------------------------- */ .last { margin-bottom: 0; } .first { margin-top: 0; } .aligncenter { text-align: center; } .alignright { text-align: right; } .alignleft { text-align: left; } .clear { clear: both; } hr { border: 0; height: 0; border-top: 1px solid rgba(0, 0, 0, 0.1); border-bottom: 1px solid rgba(255, 255, 255, 0.3); } /* ------------------------------------- ALERTS Change the class depending on warning email, good email or bad email ------------------------------------- */ .alert { font-size: 16px; color: #fff; font-weight: 500; padding: 20px; text-align: center; border-radius: 3px 3px 0 0; } .alert a { color: #fff; text-decoration: none; font-weight: 500; font-size: 16px; } .alert.alert-warning { background-color: #FF9F00; } .alert.alert-bad { background-color: #D0021B; } .alert.alert-good { background-color: #68B90F; } /* ------------------------------------- INVOICE Styles for the billing table ------------------------------------- */ .invoice { margin: 40px auto; text-align: left; width: 80%; } .invoice td { padding: 5px 0; } .invoice .invoice-items { width: 100%; } .invoice .invoice-items td { border-top: #eee 1px solid; } .invoice .invoice-items .total td { border-top: 2px solid #333; border-bottom: 2px solid #333; font-weight: 700; } /* ------------------------------------- RESPONSIVE AND MOBILE FRIENDLY STYLES ------------------------------------- */ @media only screen and (max-width: 640px) { h1, h2, h3, h4 { font-weight: 600 !important; margin: 20px 0 5px !important; } h1 { font-size: 22px !important; } h2 { font-size: 18px !important; } h3 { font-size: 16px !important; } .container { width: 100% !important; } .content, .content-wrap { padding: 10px !important; } .invoice { width: 100% !important; } } </style> </head> <body itemscope itemtype="https://schema.org/EmailMessage"> <table class="body-wrap"> <tr> <td></td> <td class="container" width="600"> <div class="content"> <table class="main" width="100%" cellpadding="0" cellspacing="0"> <tr> <td class="alert alert-good"> Audio Transcription Service Provider </td> </tr> <tr> <td class="content-wrap"> <table width="100%" cellpadding="0" cellspacing="0"> <tr> <td class="content-block"> <!-- Content Goes Here --><p>Hello,</p> <p>Do you need someone reliable to transcribe both your short term and long term projects? Or do you need an accurate transcript for your audio or video?</p> <p>Allow us to transcribe your audio and provide you accurate transcripts and let us help you reach your business/project goals through the help of our transcription services.</p> <p>What are our goals with each transcript?</p> <ul> <li>Speed</li> <li>Accuracy</li> <li>Confidentiality</li> </ul> <p>Each transcript is properly formatted. Strict grammar and punctuation rules are adhered to and of course, file security is something we take very seriously.</p> <p>Have any transcription queries? Send me a message. Let's discuss what you need to get done. We will address any concerns you have.</p> <p>- Professional transcription</p> <p>- Accurate and thorough</p> <p>- Beautifully transcribed documents.</p> <p>- Grammar, spelling and jargon thoroughly checked</p> <p>We have transcribed within most industries:</p> <ul> <li>Medical transcription</li> <li>Technological</li> <li>Academic</li> <li>Lectures</li> <li>Business</li> <li>Groups</li> <li>Legal</li> <li>Research interviews</li> </ul> <p>more...</p> <p>Skilled with international accents and prompt response. Our pricing is better or comparable to individual service provider. In addition we also assist in APA Style formatting for research papers. Please note we don’t conduct research but assist only in formatting of the papers.</p> <p>Regards,<br /> Cathy Jones</p> <img alt="track" class="newsletters-tracking" src="http://transcriptionmasters.info?wpmlmethod=track&id=c1fb6cd411a237fa0cbb475c8982b67e" /> </td> </tr> </table> </td> </tr> </table> <div class="footer"> <table width="100%"> <tr> </tr> </table> </div> </div> </td> <td></td> </tr> </table> </body> </html> |
From: LinkedIn S. <sec...@li...> - 2016-05-23 10:02:42
|
Hi L., You've successfully changed your LinkedIn password. Thanks for using LinkedIn! The LinkedIn Team When and where this happened: Date:May 23, 2016, 6:02 PM Browser:Chrome Mobile Operating System:Android Approximate Location:Shenzhen, Guangdong, China Didn't do this? Be sure to change your password right away: https://www.linkedin.com/e/v2?e=1d5ygu-iojugm0s-y2&a=uas-request-password-reset&midToken=AQHKo-bTxEZFZg&ek=security_reset_password_notification This email was intended for L. Y. (Web Service Engineer at Xignite). Learn why we included this: https://www.linkedin.com/e/v2?e=1d5ygu-iojugm0s-y2&a=customerServiceUrl&midToken=AQHKo-bTxEZFZg&ek=security_reset_password_notification&articleId=4788 If you need assistance or have questions, please contact LinkedIn Customer Service: https://www.linkedin.com/e/v2?e=1d5ygu-iojugm0s-y2&a=customerServiceUrl&midToken=AQHKo-bTxEZFZg&ek=security_reset_password_notification © 2016 LinkedIn Corporation, 2029 Stierlin Court, Mountain View CA 94043. LinkedIn and the LinkedIn logo are registered trademarks of LinkedIn. |
From: LinkedIn S. <sec...@li...> - 2016-05-23 06:15:21
|
Hi L., You recently requested a password reset. To change your LinkedIn password, paste the following link into your browser: https://www.linkedin.com/e/rpp/82579134/crf-users%40lists%2Esourceforge%2Enet/3296877483998095975/?hs=true&tok=2MjrP12sBTjTg1 The link will expire in 24 hours, so be sure to use it right away. Thanks for using LinkedIn! The LinkedIn Team This email was intended for L. Y. (Web Service Engineer at Xignite). Learn why we included this: https://www.linkedin.com/e/v2?e=1d5ygu-iojmc875-a5&a=customerServiceUrl&midToken=AQHKo-bTxEZFZg&ek=security_password_reset&articleId=4788 If you need assistance or have questions, please contact LinkedIn Customer Service: https://www.linkedin.com/e/v2?e=1d5ygu-iojmc875-a5&a=customerServiceUrl&midToken=AQHKo-bTxEZFZg&ek=security_password_reset © 2016 LinkedIn Corporation, 2029 Stierlin Court, Mountain View CA 94043. LinkedIn and the LinkedIn logo are registered trademarks of LinkedIn. |
From: LinkedIn S. <sec...@li...> - 2016-05-19 10:34:50
|
Hi L., To make sure you continue having the best experience possible on LinkedIn, we're regularly monitoring our site and the Internet to keep your account information safe. We've recently noticed a potential risk to your LinkedIn account coming from outside LinkedIn. Just to be safe, you'll need to reset your password the next time you log in. Here's how: 1. Go to the LinkedIn website. 2. Next to the password field, click the "Forgot your password" link, and enter your email address. 3. You'll get an email from LinkedIn asking you to click a link that will help you reset your password. 4. Once you've reset your password, a confirmation email will be sent to the confirmed email addresses on your account. Thanks for helping us keep your account safe, The LinkedIn Team ..................................... This email was intended for L. Y. (Web Service Engineer at Xignite). Learn why we included this: https://www.linkedin.com/e/v2?e=1d5ygu-ioe5ujke-af&a=customerServiceUrl&midToken=AQHKo-bTxEZFZg&ek=email_password_invalidated_01&articleId=4788 © 2016 LinkedIn Corporation, 2029 Stierlin Court, Mountain View CA 94043. LinkedIn and the LinkedIn logo are registered trademarks of LinkedIn. |
From: Kartik K. P. <kar...@gm...> - 2015-07-31 01:17:53
|
Hi, I would like to know if its possible to train the CRF model in stochastic fashion( i.e. training the model with instances one-by-one). Since I am willing to train with huge training set, I was wondering if there is a support for on-line(or stochastic) training in the CRF package? Any input is appreciated. Thanks! -- Regards, Kartik Perisetla Carnegie Mellon University, Pittsburgh |
From: Kartik K. P. <kar...@gm...> - 2015-06-29 20:35:02
|
Hello, I am using this CRF package for my project. I have a question - I understand that FeatureTypesEachLabel is used when we want to fire that specific feature for each label type in the dataset. How can I configure it to only fire when a specific label is seen in the dataset. For example, lets say if label is 'category1' only then feature 'f1' has to fire or else not. I would appreciate if someone provide inputs. Thanks! -- Regards, Kartik Perisetla |
From: Rohit N. <roh...@ce...> - 2014-05-15 08:32:30
|
I would like to thank Prof. Sunita Sarawagi for the magical code. I am referring to the 2001 paper Automatic segmentation of text into structured records by Prof. Sunita Sarawagi and others. A nested(k parallel) crf is used in it and the inner crfs are pruned by merging two parallel chains into one. A self-loop is made on any of the middle nodes in the process, as opposed to having a self loop on the end that is available in the package. Any help is greatly appreciated. Regards, Rohit Nandwani nan...@gm... |
From: Tao (陶. <tao...@my...> - 2013-12-31 02:44:04
|
Hi, My training set contains few Boolean features but many number features. But the FeatureType class doesn’t seem to support this. What should I do? Thanks Tao |
From: Martin K. <mkr...@cn...> - 2012-12-20 15:04:54
|
CALL FOR PARTICIPATION: CHEMDNER task: Chemical compound and drug name recognition task ( http://www.biocreative.org/tasks/biocreative-iv/chemdner ) The CHEMDNER task (part of The BioCreative IV competition) is a community challenge on named entity recognition of chemical compounds. CRFs were used successfully as a method for named entity recognition (NER) by teams that participated in previous BioCreative challenges. We expect that the CRF package will be a useful resource also for the chemical compound name recognition task. We thus encourage CRF Packaged users to participate at the chemical compound named entity recognition task of BioCreative IV. (1) TASK GOAL AND MOTIVATION The goal of this task is to promote the implementation of systems that are able to detect mentions in text of chemical compounds and drugs. The recognition of chemical entities is also crucial for other subsequent text processing strategies, such as detection of drug-protein interactions, adverse effects of chemical compounds or the extraction of pathway and metabolic reaction relations. A range of different methods have been explored for the recognition of chemical compound mentions including machine learning based approaches, rule-based systems and different types of dictionary-lookup strategies. The Weka framework has been successfully explored by several participating teams for previous biomedical text mining task posed in the context of the BioCreative challenge. We foresee a considerable interest in the result of this task by the NLP/text mining community on one side, as well as by the bioinformatics, drug discovery/biomedicine and chemoinformatics communities on the other side. As has been the case in previous BioCreative efforts (resulting in high impact papers in the field), we expect that successful participants will have the opportunity to publish their system descriptions in a journal article. (2) CHEMDNER TRACK DESCRIPTION The CHEMDNER is one of the tracks posed at the BioCreative IV community challenge (http://www.biocreative.org). We invite participants to submit results for the CHEMDNER task providing predictions for one or both of the following subtasks: a) Given a set of documents, return for each of them a ranked list of chemical entities described within each of these documents [Chemical document indexing sub-task] b) Provide for a given document the start and end indices corresponding to all the chemical entities mentioned in this document [Chemical entity mention recognition sub-task]. For these two tasks the organizers will release training and test data collections. The task organizers will provide details on the used annotation guidelines; define a list of criteria for relevant chemical compound entity types as well as selection of documents for annotation. (3) REGISTRATION Teams can participate in the CHEMDNER task by registering for track 2 of BioCreative IV. You can register additionally for other tracks too. To register your team, go to the following page that provides more detailed instructions: http://www.biocreative.org/news/biocreative-iv/team/ Mailing list and contact information: You can post questions related to the CHEMDNER task to the BioCreative mailing list. To register for the BioCreative mailing list, please visit the following page: http://biocreative.sourceforge.net/mailing.html (4) WORKSHOP CHEMDNER is part of the BioCreative evaluation effort. The BioCreative Organizing Committee will host the BioCreative IV Challenge evaluation workshop (http://www.biocreative.org/events/biocreative-iv/CFP/) at NCBI, National Institutes of Health, Bethesda, Maryland, on October 7-9, 2013 (5) CHEMDNER TASK ORGANIZERS Martin Krallinger, Spanish National Cancer Research Center (CNIO) Obdulia Rabal, University of Navarra, Spain Julen Oyarzabal, University of Navarra, Spain Alfonso Valencia, Spanish National Cancer Research Center (CNIO) (6) REFERENCES - Vazquez, M., Krallinger, M., Leitner, F., & Valencia, A. (2011). Text Mining for Drugs and Chemical Compounds: Methods, Tools and Applications. Molecular Informatics, 30(6-7), 506-519. - Krallinger M, et al. The Protein-Protein Interaction tasks of BioCreative III: classification/ranking of articles and linking bio-ontology concepts to full text. BMC Bioinformatics. 2011;12 Suppl 8:S3 - Corbett, P., Batchelor, C., & Teufel, S. (2007). Annotation of chemical named entities. BioNLP 2007: Biological, translational, and clinical language processing, 57-64. - Klinger, R., Kolářik, C., Fluck, J., Hofmann-Apitius, M., & Friedrich, C. M. (2008). Detection of IUPAC and IUPAC-like chemical names. Bioinformatics, 24(13), i268-i276. - Hettne, K. M., Stierum, R. H., Schuemie, M. J., Hendriksen, P. J., Schijvenaars, B. J., Mulligen, E. M. V., ... & Kors, J. A. (2009). A dictionary to identify small molecules and drugs in free text. Bioinformatics, 25(22), 2983-2991. - Yeh, A., Morgan, A., Colosimo, M., & Hirschman, L. (2005). BioCreAtIvE task 1A: gene mention finding evaluation. BMC bioinformatics, 6(Suppl 1), S2. - Smith, L., Tanabe, L. K., Ando, R. J., Kuo, C. J., Chung, I. F., Hsu, C. N., ... & Wilbur, W. J. (2008). Overview of BioCreative II gene mention recognition. Genome Biology, 9(Suppl 2), S2. |
From: Nikita M. <nik...@gm...> - 2012-07-27 09:44:00
|
Hi, I am a student from IIT Kharagpur. I am working on a project related to web-query segmentation and I am using the CRF code provided at http://sourceforge.net/projects/crf/ . I was hoping if a sample file(working example) for semi-Markov CRFAppl(main), is available. It would be really helpful if it could be provided. Thanks and Regards! Nikita Mishra |
From: Hamid R. H. <ha....@ie...> - 2012-07-26 17:59:39
|
Hello, OK, it seems that in Segment trainer and Nested Trainer, in order to get rid of overflows all metrics are calculated in a log space. Which means that I can't define features that return negative values. I can shift the features but then they wont represent the underlying characteristics correctly. Is there any workaround ? Thanks |
From: Hamid R. H. <ha....@gm...> - 2012-07-16 12:11:53
|
Hello, I would like to know why during the computation of alpha and beta, you take the log? Is that because of the overflow/underflow problem? If yes, then when you calculate Mi*alpha and Mi*beta you first take the exponent of the matrix/vector elements and then multiply the exponents (and not adding the logs) which brings about the precision errors. Regards |
From: Hamid R. H. <ha....@ie...> - 2012-07-15 16:57:28
|
Hello, I would like to know why during the computation of alpha and beta, you take the log? Is that because of the overflow/underflow problem? If yes, then when you calculate Mi*alpha and Mi*beta you first take the exponent of the matrix/vector elements and then multiply the exponents (and not adding the logs) which brings about the precision errors. Regards |
From: Hamid R. H. <ha....@ie...> - 2012-07-03 20:03:33
|
and what is the purpose of variable "base" ? On Tue, Jul 3, 2012 at 1:58 PM, Hamid Reza Hassanzadeh < ha....@gm...> wrote: > Thanks a lot. > > > On Tue, Jul 3, 2012 at 1:24 PM, Sunita Sarawagi <su...@ii...> wrote: > >> Please follow the definition of Mi_YY.zMult. This function has been >> over-ridden so that it actually does the right thing for matrix entries >> containing log of the actual values. >> >> >> On 07/03/2012 09:31 PM, Hamid Reza Hassanzadeh wrote: >> >> Hello, >> I have hard times understanding the following lines of codes seen in both >> SegmentTrainer and NestedTrainer which compute the Betas in Log space, I >> would appreciate it if you can help me on that, >> >> initMDone = >> computeLogMi(dataSeq,i,i+ell,featureGenNested,lambda,Mi_YY,Ri_Y,reuseM,initMDone); >> tmp_Y.assign(Ri_Y); >> tmp_Y.assign(beta_Y[i+ell], sumFunc); >> Mi_YY.zMult(tmp_Y, beta_Y[i],1,1,false); >> >> OK, in general to compute beta_Y[i] we should do this, >> computeMi(featureGenerator,lambda,dataSeq,i,Mi_YY,Ri_Y); >> tmp_Y.assign(beta_Y[i]); >> tmp_Y.assign(Ri_Y,multFunc); >> Mi_YY.zMult(tmp_Y, beta_Y[i-1]); >> >> Which makes sense to me, but in Log space how can you multiply Mi_YY to >> tmp_Y and add the result to beta_Y? They are in log space. >> >> Regards >> >> >> >> >> ------------------------------------------------------------------------------ >> Live Security Virtual Conference >> Exclusive live event will cover all the ways today's security and >> threat landscape has changed and how IT managers can respond. Discussions >> will include endpoint security, mobile security and the latest in malware >> threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ >> >> >> >> _______________________________________________ >> Crf-users mailing lis...@li...https://lists.sourceforge.net/lists/listinfo/crf-users >> >> > |
From: Hamid R. H. <ha....@gm...> - 2012-07-03 17:58:10
|
Thanks a lot. On Tue, Jul 3, 2012 at 1:24 PM, Sunita Sarawagi <su...@ii...> wrote: > Please follow the definition of Mi_YY.zMult. This function has been > over-ridden so that it actually does the right thing for matrix entries > containing log of the actual values. > > > On 07/03/2012 09:31 PM, Hamid Reza Hassanzadeh wrote: > > Hello, > I have hard times understanding the following lines of codes seen in both > SegmentTrainer and NestedTrainer which compute the Betas in Log space, I > would appreciate it if you can help me on that, > > initMDone = > computeLogMi(dataSeq,i,i+ell,featureGenNested,lambda,Mi_YY,Ri_Y,reuseM,initMDone); > tmp_Y.assign(Ri_Y); > tmp_Y.assign(beta_Y[i+ell], sumFunc); > Mi_YY.zMult(tmp_Y, beta_Y[i],1,1,false); > > OK, in general to compute beta_Y[i] we should do this, > computeMi(featureGenerator,lambda,dataSeq,i,Mi_YY,Ri_Y); > tmp_Y.assign(beta_Y[i]); > tmp_Y.assign(Ri_Y,multFunc); > Mi_YY.zMult(tmp_Y, beta_Y[i-1]); > > Which makes sense to me, but in Log space how can you multiply Mi_YY to > tmp_Y and add the result to beta_Y? They are in log space. > > Regards > > > > > ------------------------------------------------------------------------------ > Live Security Virtual Conference > Exclusive live event will cover all the ways today's security and > threat landscape has changed and how IT managers can respond. Discussions > will include endpoint security, mobile security and the latest in malware > threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ > > > > _______________________________________________ > Crf-users mailing lis...@li...https://lists.sourceforge.net/lists/listinfo/crf-users > > |
From: Hamid R. H. <ha....@gm...> - 2012-07-03 16:01:14
|
Hello, I have hard times understanding the following lines of codes seen in both SegmentTrainer and NestedTrainer which compute the Betas in Log space, I would appreciate it if you can help me on that, initMDone = computeLogMi(dataSeq,i,i+ell,featureGenNested,lambda,Mi_YY,Ri_Y,reuseM,initMDone); tmp_Y.assign(Ri_Y); tmp_Y.assign(beta_Y[i+ell], sumFunc); Mi_YY.zMult(tmp_Y, beta_Y[i],1,1,false); OK, in general to compute beta_Y[i] we should do this, computeMi(featureGenerator,lambda,dataSeq,i,Mi_YY,Ri_Y); tmp_Y.assign(beta_Y[i]); tmp_Y.assign(Ri_Y,multFunc); Mi_YY.zMult(tmp_Y, beta_Y[i-1]); Which makes sense to me, but in Log space how can you multiply Mi_YY to tmp_Y and add the result to beta_Y? They are in log space. Regards |
From: Hamid R. H. <has...@ga...> - 2012-05-26 19:44:11
|
Hello, I would like to know if there is any feature caching has been implemented in CRF package. Regards, -- Hamid Reza Hassanzadeh, PhD Student Bioinformatics Lab Center for Bioinformatics and Computational Genomics Joint Georgia Tech and Emory Wallace H Coulter Department of Biomedical Engineering, School of Computational Science and Engineering at Georgia Institute of Technology email: has...@ga... phone: +1 404 719 0810 ; office: KACB 1343 group: http://opal.gatech.edu/genemark/ |
From: Hamid R. H. <ham...@ya...> - 2012-05-07 20:29:31
|
Hi, Here is some updates about my code which is still very slow, I have only 3 labels, 130 features (68 feature types which increases to 130 due to different combination of y's). As training I'm using 5 training sample each having around 1500 residues. I'm using semi-markov model which uses nestedTrainer for training. The maxMermory is 3500. Under these conditions each single iteration of training takes around 2 hours. Also my features are all simple (just additions and subtractions) since I precalculate the possible outputs and store them in a large array. Any ideas? Regards ________________________________ From: Sunita Sarawagi <su...@cs...> To: Hamid Reza Hassanzadeh <has...@ga...> Cc: crf...@li... Sent: Saturday, March 24, 2012 1:53 AM Subject: Re: [Crf-users] Running Time Are you using straight CRF with only level-1 Markov dependency, or a higher-order Markov dependency, or a semi-crf? Is there no dependence on y in the feature function, or do you have a wrapper around this feature that takes a cross-product of this feature with all possible y-s? Hamid Reza Hassanzadeh wrote: > Dear CRF users, > I'm testing the crf package with a sequence of some 30000 residues length with a simple base feature over each residue. The feature is simply like this: > > public boolean startScanFeaturesAt(...) > { > ... > if (x(pos-5:pos)=="xxxxxx" && x(pos:pos+2)=="yyy") > return true; > ... > } > > And the execution time is as long as 1hour while this normally takes some 5 minutes with Mallet. The situation becomes even much worse when I use a longer training sequence. > > Am I making a mistake? When I feed the training file as test file I get 100% accuracy for both cases (mallet and crf). Thus there should not be any mistake. In that case what can be the cause of this huge difference? > > ------------------------------------------------------------------------------ > This SF email is sponsosred by: > Try Windows Azure free for 90 days Click Here http://p.sf.net/sfu/sfd2d-msazure > _______________________________________________ > Crf-users mailing list > Crf...@li... > https://lists.sourceforge.net/lists/listinfo/crf-users > |
From: Girish M. <gi...@cs...> - 2012-05-07 05:53:02
|
Hi, I have a query similar to one that I saw unanswered in the archives ( http://sourceforge.net/mailarchive/message.php?msg_id=18752494<http://sourceforge.net/mailarchive/forum.php?forum_name=crf-users&max_rows=25&style=nested&viewmonth=200803>). I am trying to use your CRF implementation for POS tagging. I am able to get the initial code up & running on the Penn Treebank ATIS corpus. I am aiming at using orthographic features from the data for training as well. For example, apart from supplying the word I would also include features indicating capitalization (caps) as well as common English suffixes (e.g. -ing and -s) as well as features for words that start with a number. Currently I am able to get the program running on data in the below format, where each line consists of a word token separated from its part of speech by a | delimiter with sentences separated by blank lines. <Word1>|<number_indicating_POS_tag> I would like to modify the above data & use it in the below format where I have 2 features for each word: <Word1> <feature1> <feature2>|<number_indicating_POS_tag> It would be really helpful if you could tell me how these features should be passed to the CRF module while training & testing. I have gone through the code & previous posts on the mailing list but such orthographic features which are a part of the training data itself do not seem to have been considered ever. Thanks, Girish http://xkcd.com/979 |
From: Hamid R. H. <has...@ga...> - 2012-04-02 03:31:13
|
Dear all, I've previously added some features to CRF package, and I'm doing some tests now, but the program fails to satisfy some tests, here is the most strange one, input: a sequence of consecutive tokens with their associated labels labels: only 2 labels (lab1, lab2) features: only 2 simple features: feature 1: public boolean startScanFeaturesAt(DataSequence data, int prevPos, int pos) { //In other words I take no account of the input observation, I know this is strange { edgeNum = 0; if (edgeIter == null) { setEdgeIter(); } if (edgeIter != null) edgeIter.start(); return hasNext(); } } public void next(FeatureImpl f) { edgeIsOuter = edgeIter.nextIsOuter(); Edge e = edgeIter.next(); Object name=""; if (featureCollectMode()) { if (labelNames == null) { name = "E."+model.label(e.start); } else { name = labelNames[model.label(e.start)]; } } if (edgeIsOuter) { setFeatureIdentifier(model.label(e.start)*model.numberOfLabels() +model.label(e.end) + model.numEdges(), model.label(e.end),name,f); } else { setFeatureIdentifier(edgeNum,e.end,name,f); } f.ystart = e.start; f.yend = e.end; if (e.end==1) //Label 1 f.val = 1; else f.val=0; edgeNum++; } if (y_end==lab1) f.val=1; else f.val=0; feature 2: exactly the same as feature 1 except these lines if (e.end==0) //Label 2 f.val = 1; else f.val=0; These are the some lines of codes from my two feature classes. What I intend to do is that if label at position i is 0 feature 1 should fire otherwise feature 2 must fire. According to the training algorithm of CRF the expectation value of features in the training set should be equal to the expectation value feature with respect to the model. That means that if in my training file 50% of the times I observe label 1 and 50% label 2, then the test set should be labelized ! 50% of times with label 1 and 50% with label 2. Unfortunately after I run the program I get 100% label 1. What is wrong then? Am I making a mistake in the implementation ? Regards |
From: Sunita S. <su...@cs...> - 2012-03-24 05:55:49
|
Are you using straight CRF with only level-1 Markov dependency, or a higher-order Markov dependency, or a semi-crf? Is there no dependence on y in the feature function, or do you have a wrapper around this feature that takes a cross-product of this feature with all possible y-s? Hamid Reza Hassanzadeh wrote: > Dear CRF users, > I'm testing the crf package with a sequence of some 30000 residues length with a simple base feature over each residue. The feature is simply like this: > > public boolean startScanFeaturesAt(...) > { > ... > if (x(pos-5:pos)=="xxxxxx" && x(pos:pos+2)=="yyy") > return true; > ... > } > > And the execution time is as long as 1hour while this normally takes some 5 minutes with Mallet. The situation becomes even much worse when I use a longer training sequence. > > Am I making a mistake? When I feed the training file as test file I get 100% accuracy for both cases (mallet and crf). Thus there should not be any mistake. In that case what can be the cause of this huge difference? > > ------------------------------------------------------------------------------ > This SF email is sponsosred by: > Try Windows Azure free for 90 days Click Here > http://p.sf.net/sfu/sfd2d-msazure > _______________________________________________ > Crf-users mailing list > Crf...@li... > https://lists.sourceforge.net/lists/listinfo/crf-users > |
From: Hamid R. H. <has...@ga...> - 2012-03-24 03:09:43
|
Dear CRF users, I'm testing the crf package with a sequence of some 30000 residues length with a simple base feature over each residue. The feature is simply like this: public boolean startScanFeaturesAt(...) { ... if (x(pos-5:pos)=="xxxxxx" && x(pos:pos+2)=="yyy") return true; ... } And the execution time is as long as 1hour while this normally takes some 5 minutes with Mallet. The situation becomes even much worse when I use a longer training sequence. Am I making a mistake? When I feed the training file as test file I get 100% accuracy for both cases (mallet and crf). Thus there should not be any mistake. In that case what can be the cause of this huge difference? |
From: Sunita S. <su...@cs...> - 2012-03-22 20:30:23
|
The model class only ensures that the ZZZ-->YYY feature will not be present and therefore the score of this transition is zero, not -infinity. If you want a hard constraint, you will have to do something more involved. Use SegmentCRF instead of CRF Each instance should implement CandSegDataSequence instead of DataSequence. The method constraints() in the above class should return any non-null value for Iterator. For example, it can be an empty iterator. Hamid Reza Hassanzadeh wrote: > Guys, > I have difficulty with defining my own Model. I derived my own class from Model.java and overridden the hasnext() function and some others in order to block some of transitions. That is instead of using a Complete Graph, I want data tagging obeys some rules. For example if token x_i is labeled ZZZ then the next label can not be YYY. But unfortunately after the program is run, I see that the tagged file includes such transitions. My question now is that is that due to an implementation bug or this can happen in reality? > > Regards > > ------------------------------------------------------------------------------ > This SF email is sponsosred by: > Try Windows Azure free for 90 days Click Here > http://p.sf.net/sfu/sfd2d-msazure > _______________________________________________ > Crf-users mailing list > Crf...@li... > https://lists.sourceforge.net/lists/listinfo/crf-users > |
From: Sunita S. <su...@cs...> - 2012-03-22 09:17:18
|
Well, except for the BSegment* package that implement my ICML 2006 paper (Efficient inference on sequence segmentation models), the other packages are quite exploratory and can be ignored. Xu Chu wrote: > Hi all > > I just checked out CRF. > I found there are some newly added features to it. > > Specifically there are four new packages, iitb.AStar, iitb.BSegment, iitb.BSegmentCRF, iitb.KernelCRF. But there aren't enough documentation. Can anyone tell me what are those for? Any reference paper about those packages? > > Thanks a lot > > > Regards, > Xu > > > > ------------------------------------------------------------------------------ > This SF email is sponsosred by: > Try Windows Azure free for 90 days Click Here > http://p.sf.net/sfu/sfd2d-msazure > _______________________________________________ > Crf-users mailing list > Crf...@li... > https://lists.sourceforge.net/lists/listinfo/crf-users > |