{
  "id": 2266,
  "title": "Enhanced Disambiguated Product Mentions Data Coming",
  "url": "/competitions/cprod1/discussion/2266",
  "author_name": "",
  "post_date": "2012-07-27T17:52:35.747Z",
  "votes": null,
  "comment_count": 4,
  "views": 2966,
  "content": "<p>Dear contestants,&nbsp;here is an early notice of some upcoming changes.</p>\r\n<p>Our text annotators have taken another pass through the data and they have created substantially cleaner disambiguated product mention datasets. Given that there have been few leaderboard submissions to date and significant time to train models we plan to\r\n make the change to this data early next week - likely Tuesday morning. Only one of your files on hand changes: training-disambiguated-product-mentions.csv. The leaderboard and final evaluation will also change. The main change that you will notice is the inclusion\r\n of several new mentions. Other outcomes included updates to the list of products for some terms and some small boundary modifications for existing mentions.</p>\r\n<p>Also, along with the new data, we will release the code to a second baseline system that trains a CRF-based sequential tagging model. Recall that the existing baseline system simply extracts product terms from the training disambiguated product mentions\r\n and naively applies these directly to the test set. The new baseline will train a statistical model based on some simple features. The feature generator will be Perl-based and will use\r\n<a title=\"MALLET home page\" href=\"http://mallet.cs.umass.edu/\" target=\"_blank\">MALLET</a>&nbsp;to train and test a CRF.</p>\r\n<p>We hope that these changes will make the resulting solutions more relevant and to make the challenge more accessible.</p>\r\n<p>Best,</p>\r\n<p>The CPROD1 Team</p>",
  "messages": [
    {
      "id": "12746",
      "postDate": "07/27/2012 17:52:35",
      "content": "<p>Dear contestants,&nbsp;here is an early notice of some upcoming changes.</p>\r\n<p>Our text annotators have taken another pass through the data and they have created substantially cleaner disambiguated product mention datasets. Given that there have been few leaderboard submissions to date and significant time to train models we plan to\r\n make the change to this data early next week - likely Tuesday morning. Only one of your files on hand changes: training-disambiguated-product-mentions.csv. The leaderboard and final evaluation will also change. The main change that you will notice is the inclusion\r\n of several new mentions. Other outcomes included updates to the list of products for some terms and some small boundary modifications for existing mentions.</p>\r\n<p>Also, along with the new data, we will release the code to a second baseline system that trains a CRF-based sequential tagging model. Recall that the existing baseline system simply extracts product terms from the training disambiguated product mentions\r\n and naively applies these directly to the test set. The new baseline will train a statistical model based on some simple features. The feature generator will be Perl-based and will use\r\n<a title=\"MALLET home page\" href=\"http://mallet.cs.umass.edu/\" target=\"_blank\">MALLET</a>&nbsp;to train and test a CRF.</p>\r\n<p>We hope that these changes will make the resulting solutions more relevant and to make the challenge more accessible.</p>\r\n<p>Best,</p>\r\n<p>The CPROD1 Team</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "12751",
      "postDate": "07/27/2012 18:37:51",
      "content": "<p>While this is certainly good news, I would have appreciated better quality in the crawled data more.</p>\r\n<p>Many entries contain completely unrelated website text like:</p>\r\n<blockquote style=\"padding-left:30px\">Respond to this message &lt;&lt; Previous Topic | Next Topic &gt;&gt;</blockquote>\r\n<p>There are many such examples. You also seem to completely ignore any non-ascii character; a russian sample in the data shows up as:</p>\r\n<blockquote style=\"padding-left:30px\">- &lt;P&gt; VIDEOMAX &lt;P&gt; &gt; &lt;P&gt; &gt; &lt;P&gt; FAQ &lt;P&gt; VIDEOMAX &lt;P&gt; . &lt;s&gt; &lt;P&gt; : 07--2009 : &lt;P&gt; : 134 &lt;P&gt; ( 0 ) &lt;P&gt; : &lt;P&gt; : 11--2010 12:27 &lt;P&gt; - &lt;P&gt; ! . &lt;s&gt; &lt;P&gt; .2 &lt;P&gt; , , . &lt;s&gt; &lt;P&gt; , &lt;P&gt; - &lt;P&gt; . &lt;s&gt; &lt;P&gt; , &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt;\r\n &lt;P&gt; - &lt;P&gt; . &lt;s&gt; &lt;P&gt; , . &lt;s&gt; &lt;P&gt; , &lt;P&gt; - &lt;P&gt; . &lt;s&gt; &lt;P&gt; - 11--2010 12:27 &lt;P&gt; Spinoz &lt;P&gt; ( ) : 15--2008 : &lt;P&gt; : &lt;P&gt; : 3884 &lt;P&gt; Spinoz &lt;P&gt; ( 0 ) &lt;P&gt; : 11--2010 13:18 &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt; &lt;P&gt; - &lt;P&gt; 3/4 &lt;P&gt; 4 - &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt; &lt;P&gt; ( , &lt;P&gt; ) . &lt;s&gt; &lt;P&gt; - . &lt;s&gt;\r\n &lt;P&gt;</blockquote>\r\n<p>Will the evaluation data be as &quot;dirty&quot; as that too?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "12762",
      "postDate": "07/28/2012 04:34:43",
      "content": "<p>What features does the sequence-tagging model use?</p>\r\n<p>regards</p>\r\n<p>shankar.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "12768",
      "postDate": "07/28/2012 08:07:38",
      "content": "<p>In response to question about the sample &quot;noisy&quot; text item: the evaluation-text.json file that we will release on Sept 17th was drawn from the same population as the three other json text files already provided, so, yes, some text items will contain its\r\n fair share of painfully noisy text items like the one you pointed out. Hmmm, heuristically excluding these types of items may help a learner to not have to do double duty...</p>\r\n<p>The CPROD1 Team</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "12787",
      "postDate": "07/30/2012 13:42:59",
      "content": "<p>Despite the fact that it will probably nudge my trivial entry out of the green, I'm looking forward to more consistent annotations:</p>\r\n<p style=\"padding-left:30px\">141361c53f2549232d5993a511a31da7<br>\r\n<br>\r\nI have TVersity working with my <strong>HR20-700</strong> but will be...</p>\r\n<p style=\"padding-left:30px\">product mention -&gt; 6-6,0</p>\r\n<p>but then</p>\r\n<p style=\"padding-left:30px\">7b465aa355535d76323c0f89dfbec8b5<br>\r\n<br>\r\nQuote : Originally Posted by MichaelLAX... I find the situation of recording from the newer\r\n<strong>HR20-700</strong> even better than the older model...</p>\r\n<p style=\"padding-left:30px\">product mention -&gt; 28-28,xyT-4D9J4GE</p>\r\n<p>What's a poor algorithm to do, find the products or model the fallible human?</p>\r\n<p>John</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 12751,
      "author_name": "lucasbeyer",
      "author_url": "",
      "post_date": "07/27/2012 18:37:51",
      "content": "<p>While this is certainly good news, I would have appreciated better quality in the crawled data more.</p>\r\n<p>Many entries contain completely unrelated website text like:</p>\r\n<blockquote style=\"padding-left:30px\">Respond to this message &lt;&lt; Previous Topic | Next Topic &gt;&gt;</blockquote>\r\n<p>There are many such examples. You also seem to completely ignore any non-ascii character; a russian sample in the data shows up as:</p>\r\n<blockquote style=\"padding-left:30px\">- &lt;P&gt; VIDEOMAX &lt;P&gt; &gt; &lt;P&gt; &gt; &lt;P&gt; FAQ &lt;P&gt; VIDEOMAX &lt;P&gt; . &lt;s&gt; &lt;P&gt; : 07--2009 : &lt;P&gt; : 134 &lt;P&gt; ( 0 ) &lt;P&gt; : &lt;P&gt; : 11--2010 12:27 &lt;P&gt; - &lt;P&gt; ! . &lt;s&gt; &lt;P&gt; .2 &lt;P&gt; , , . &lt;s&gt; &lt;P&gt; , &lt;P&gt; - &lt;P&gt; . &lt;s&gt; &lt;P&gt; , &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt;\r\n &lt;P&gt; - &lt;P&gt; . &lt;s&gt; &lt;P&gt; , . &lt;s&gt; &lt;P&gt; , &lt;P&gt; - &lt;P&gt; . &lt;s&gt; &lt;P&gt; - 11--2010 12:27 &lt;P&gt; Spinoz &lt;P&gt; ( ) : 15--2008 : &lt;P&gt; : &lt;P&gt; : 3884 &lt;P&gt; Spinoz &lt;P&gt; ( 0 ) &lt;P&gt; : 11--2010 13:18 &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt; &lt;P&gt; - &lt;P&gt; 3/4 &lt;P&gt; 4 - &lt;P&gt; . &lt;s&gt; &lt;P&gt; . &lt;s&gt; &lt;P&gt; ( , &lt;P&gt; ) . &lt;s&gt; &lt;P&gt; - . &lt;s&gt;\r\n &lt;P&gt;</blockquote>\r\n<p>Will the evaluation data be as &quot;dirty&quot; as that too?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 12762,
      "author_name": "mailshanx",
      "author_url": "",
      "post_date": "07/28/2012 04:34:43",
      "content": "<p>What features does the sequence-tagging model use?</p>\r\n<p>regards</p>\r\n<p>shankar.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 12768,
      "author_name": "cprod1",
      "author_url": "",
      "post_date": "07/28/2012 08:07:38",
      "content": "<p>In response to question about the sample &quot;noisy&quot; text item: the evaluation-text.json file that we will release on Sept 17th was drawn from the same population as the three other json text files already provided, so, yes, some text items will contain its\r\n fair share of painfully noisy text items like the one you pointed out. Hmmm, heuristically excluding these types of items may help a learner to not have to do double duty...</p>\r\n<p>The CPROD1 Team</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 12787,
      "author_name": "j1o1h1n",
      "author_url": "",
      "post_date": "07/30/2012 13:42:59",
      "content": "<p>Despite the fact that it will probably nudge my trivial entry out of the green, I'm looking forward to more consistent annotations:</p>\r\n<p style=\"padding-left:30px\">141361c53f2549232d5993a511a31da7<br>\r\n<br>\r\nI have TVersity working with my <strong>HR20-700</strong> but will be...</p>\r\n<p style=\"padding-left:30px\">product mention -&gt; 6-6,0</p>\r\n<p>but then</p>\r\n<p style=\"padding-left:30px\">7b465aa355535d76323c0f89dfbec8b5<br>\r\n<br>\r\nQuote : Originally Posted by MichaelLAX... I find the situation of recording from the newer\r\n<strong>HR20-700</strong> even better than the older model...</p>\r\n<p style=\"padding-left:30px\">product mention -&gt; 28-28,xyT-4D9J4GE</p>\r\n<p>What's a poor algorithm to do, find the products or model the fallible human?</p>\r\n<p>John</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "12746": "",
    "12751": "",
    "12762": "",
    "12768": "",
    "12787": ""
  },
  "source": "meta"
}