<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Christopher Phipps on Medium]]></title>
        <description><![CDATA[Stories by Christopher Phipps on Medium]]></description>
        <link>https://medium.com/@christopher.phipps?source=rss-8da58490a69b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*cgyFAqSJ5kcFA5-anRCkVQ.jpeg</url>
            <title>Stories by Christopher Phipps on Medium</title>
            <link>https://medium.com/@christopher.phipps?source=rss-8da58490a69b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 15:21:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@christopher.phipps/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Linguistics on TV — Criminal Minds s9 Recap]]></title>
            <link>https://medium.com/@christopher.phipps/linguistics-on-tv-criminal-minds-s9-recap-a8298f720350?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/a8298f720350</guid>
            <category><![CDATA[linguistics]]></category>
            <category><![CDATA[criminal-minds]]></category>
            <category><![CDATA[television]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Wed, 19 Aug 2020 17:54:03 GMT</pubDate>
            <atom:updated>2020-08-19T18:32:00.107Z</atom:updated>
            <content:encoded><![CDATA[<h3>Linguistics on TV — Criminal Minds s9 Recap</h3><p>Generally season 9 included more linguistics than season 8, but still just lingo nuggets. They never fully realized forensic linguistics or Dr. Alex Blake as a major story line.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*uMEIxt6h0SUEahBhkNzuOg.png" /></figure><p>RECAP</p><p><strong>Episode 4</strong> <em>To Bear Witness</em></p><p>Min 3:00 Dr. Blake uses etymology to finish crossword puzzle quickly. Then Dr. Reid mentions saccade. Not strictly linguistics, but eye tracking is a common technique in a lot of psycholinguistics experiments.</p><p>This was a frustrating hour of TV for a linguist. Full discussion of this episode <a href="https://medium.com/@christopher.phipps/linguistics-on-tv-criminal-minds-s9-to-bear-witness-907dccace79f">here</a>.</p><p><strong>Episode 5</strong> <em>Route 66</em><br>Min 15:50 Dr. Blake reads texts and interprets them. A little linguistic, and helps team determine elements of the crime (that daughter wanted her estranged dad to take her away from Kansas)</p><p><strong>Episode 6</strong> <em>In the Blood</em><br>Min 39 Dr. Blake honors her mother who got her into crossword puzzles which got her into linguistics. Character building. Every linguist I know has a story about what they got into linguistics, so I thought this was a really nice touch that I felt deeply.</p><p><strong>Episode 9</strong> <em>Strage Fruit</em><br>The best episode in the two season run I&#39;ve watched so far. Having nothing to do with linguistics, it&#39;s treatment of race, while still antiseptic and clean for a network audience, it is more interesting and compelling than all the other episodes I&#39;ve seen so far.</p><p><strong>Episode 10</strong> <em>The Caller</em><br>Min 23 using spectrogram to analyze spoke language<br>My Brazilian Portuguese is rusty, but</p><p><strong>Episode 11</strong> <em>Bully</em><br>Blake heavy episode where we meet her father and brother.</p><p>Min 14: “there is something odd about his syntax ... I can&#39;t explain it ...” That’s it. They do nothing else with this.<br>Min 27 Blake&#39;s brother underlines words, reminds her of family dispute resolution of writing letters</p><p>Min 38: brother says here you are this hot shot linguist and yet you never say anything</p><p><strong>Episode 12</strong> <em>The Black Queen</em><br>Min 3:15 while ignoring a sexual harassment training class they most certainly should be paying attention to (but are not), we learn the term &quot;Flarpy Blunderguff&quot;. This has nothing to do with linguistics, I just wanted to type Flarpy Blunderguff.</p><p>Min 10:40 Man says he&#39;d swear on stack of bibles, they find amethyst books, Blake says his comment is a &quot;linguistic tell&quot; (I&#39;m stretching here folks. At this point, if they say the word &quot;linguistic&quot; I&#39;m counting it)</p><p>Min 28 - Blake reads a blog post for a few seconds and immediately says &quot;based on the verbiage, syntax, and organizational structure I can tell you [who wrote it]&quot; hmm, as a matter of linguistic fact, no you can not. Another example of the magical expert trope</p><p><strong>Episode 14</strong> <em>200</em><br>Min 18 just a tiny bit of linguistics. Dr. Blake describes what an Iraqi man&#39;s name means.</p><p><strong>Episode 16</strong> <em>Gabby</em><br>Overall disturbing episode about child abduction</p><p>Min 3 neighbor uses sign language with child</p><p>Min 32 After suspect says she&#39;s pregnant, Dr. Blake says &quot;she mixed her past and present tenses so it shows unconscious deception.&quot; Deception detection</p><p>Min 34 Garcia to Dr. Blake &quot;Lady, I need your word wisdom&quot;<br>Dr. Blake says her use of &quot;light bug&quot; instead of &quot;lightbulb&quot;<br>Garcia uses an unknown software that provides an onscreen &quot;linguistic correlation results. Also onscreen is the term &quot;Computational Matrix HB-88/1028-206&quot;. I Googled that and only found images from other Criminal Minds episodes.<br>Garcia uses the correlation to find a chat log that leads to the suspect&#39;s motives</p><p>Min 37:20 having nothing to do with linguistics, the hillbilly stereotyping is so offensive it&#39;s almost comical</p><p><strong>Episode 17</strong> <em>Persuasion</em><br>Min 37 Dr. Reid says he used neurolinguistic programming to get info from a suspect. Not only does Dr. Blake NOT slap Dr. Reid upside his head, she goes along with it and helps define it. This is morally unacceptable.</p><p>I was tempted to not include this because of my and every sane linguist&#39;s hatred for the pseudoscience of neurolinguistic programming. But then I realized, all of Criminal Minds&#39;s linguistics is shit. Why should I treat this shit lingo babble any different than the others? So I begrudgingly include it.</p><p><strong>Episode 18</strong> <em>Rabid</em><br>Min 18 Dr. Blake tells police that the words &quot;rabies&quot; is derived from a Sanskrit word meaning to do violence. This information is completely useless and does nothing to further the case. But I guess it helps build her character a bit. Makes he look like kind of a pompous ass, in my opinion.</p><p>NOTE: parents who dies in a car accident is a recurring trope.</p><p><strong>Episode 20</strong> <em>Blood Relations</em><br>More offensive hillbilly stereotyping</p><p><strong>Episode 21</strong> <em>What Happens in Mecklinburg</em><br>A natural language processing lingo nugget: tech-ninja Garcia realized that a hyphenated name &quot;Lauren-Ann&quot; didn&#39;t hit on her initial search for &quot;Lauren&quot;. This realization helps connects dots to lead to the killer. While a known issue in the name matching world, and kinda obvious when you think about it, it&#39;s very possible some older systems still fail to account for this kind of thing. Especially some janky-ass government database. Which I may or may not have worked on. Okay, this plot point might actually be my fault.</p><p><strong>Episode 22</strong> <em>Fatal</em><br>Dr. Blake analyzes a note left by a killer which reads &quot;YOU HAVE LESS THAN A DAY TO LIVE. THERE IS NO WAY TO PREVENT THE INEVITABLE.&quot;: Dr. Blake says The syntax of the note is interesting. Instead of a simple &quot;stop it&quot; you get &quot;prevent the inevitable&quot;.</p><p>So a whole episode centered around Greek philosophy but they can&#39;t squeak in single etymology quip. Especially since the killer suffers from (spoilers) <em>mesothelioma</em>, from the ... wait for it, Greek!</p><p><a href="https://www.etymonline.com/word/meso-?ref=etymonline_crossreference">meso</a>-<br><em>before vowels mes-, word-forming element meaning &quot;middle, intermediate, halfway,&quot; from Greek mesos &quot;middle, in the middle; middling, moderate; between&quot;</em></p><p><a href="https://www.etymonline.com/word/epithelium?ref=etymonline_crossreference">epithelium</a> (n.)<br><em>1748, Modern Latin (Frederick Ruysch), from Greek epi &quot;upon&quot; (see epi-) + thēlē &quot;teat, nipple,&quot; from suffixed form of PIE root *dhe(i)- &quot;to suck.&quot; Related: Epithelial.</em></p><p>This is a slam dunk opportunity to give your linguist a line. How do you miss this? C&#39;mon people. You&#39;re just not trying.</p><p>And sadly that’s it. Dr. Alex Blake leaves the show to go teach at Harvard. No more forensic linguistics, that I know of.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8298f720350" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Linguistics on TV — Criminal Minds s9 ‘To Bear Witness’]]></title>
            <link>https://medium.com/@christopher.phipps/linguistics-on-tv-criminal-minds-s9-to-bear-witness-907dccace79f?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/907dccace79f</guid>
            <category><![CDATA[aphasia]]></category>
            <category><![CDATA[television]]></category>
            <category><![CDATA[linguistics]]></category>
            <category><![CDATA[screenwriting]]></category>
            <category><![CDATA[lobotomy]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Mon, 03 Aug 2020 20:09:23 GMT</pubDate>
            <atom:updated>2020-08-04T15:47:10.057Z</atom:updated>
            <content:encoded><![CDATA[<h3>Linguistics on TV — Criminal Minds ‘To Bear Witness’</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ig4nVn87M51Wv1VIVaecFw.png" /></figure><p>Reader, I tell you there may be no other hour of TV more frustrating to a linguist than the fourth episode of the ninth season of Criminal Minds ‘To Bear Witness’.</p><blockquote><strong>Logline</strong>: Mateo Cruz is named section chief over the BAU as the team hunts for a sadistic perp in Baltimore who renders his victims unable to communicate.</blockquote><p>Gasp! Unable to communicate? How?</p><p>Here’s how. The bad guy performs a rudimentary lobotomy. The victims are able to think clearly and believe they are speaking to various people, but we soon find out it is all in their heads. They can think words and sentences, but they cannot articulate them in speech or write them.</p><p>This is AWESOME! This is linguistics! This is going to be the best hour of TV for linguists ever.</p><p>And then disappointment sets in. Dr. Alex Blake, the BAU’s very own specialist with a PhD in linguistics from Berkeley, is never asked to consult on the victim’s communication issues. Not even hinted at. The episode spends a considerable amount of time with the primary victim, even explaining various communicative disorders associated with lobotomies and teaching the victim to use a spelling board to help get information. But the linguist is NEVER involved!!!</p><p>ARRGGHHHHH!!!!</p><p>Why 1) have a linguist character and 2) devote a whole plot to a linguistics issue, and never put the two together? I was screaming at the screen so many times during this episode. Mind you. Dr. Blake is in the episode. She’s running around Baltimore pointing guns at people while her colleague, a non-linguist who’s primary character trait is running around pointing guns at people, sits quietly in a hospital room playing amateur linguist.</p><p>My best guess is that they wanted a non-linguist to interact with the victim in order to be a surrogate for the audience. This is a common screenwriters trick. When you have a technical issue involved in a plot, you have one character play dumb so it can be explained to that character, and hence explained to the audience without disrupting the flow of the plot. But it strikes me as trivial to have the linguist leading the interrogation of the victim and explaining as she goes to another agent standing by. This would have made so much more sense.</p><p>It’s not even worth fisking the many interesting linguistics nuggets they could have put into this episode. They fumbled the ball.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=907dccace79f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Linguistics on TV — Criminal Minds Season 8 Recap]]></title>
            <link>https://medium.com/@christopher.phipps/linguistics-on-tv-criminal-minds-season-8-recap-4df8c70e7f26?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/4df8c70e7f26</guid>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Tue, 21 Jul 2020 20:06:36 GMT</pubDate>
            <atom:updated>2020-07-21T20:11:07.832Z</atom:updated>
            <content:encoded><![CDATA[<h3>Linguistics on TV — Criminal Minds Season 8 Recap</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lil-KSA_Fe7Kya1DDQnQmA.png" /></figure><p>I watched every episode of Criminal Minds season 8, the one where they introduce the character of Dr. Alex Blake, forensic linguist.</p><p>Unfortunately, Dr. Blake is not in the season as much as I’d like. She only plays a pivotal role in a small number of episodes, though she appears in many of them.</p><p>The writers don’t seem to grok how to utilize the technical skills of a linguist to analyze evidence. They did a pretty good job in <a href="https://medium.com/@christopher.phipps/linguistics-on-tv-criminal-minds-58e0f85f93fc">the very first episode</a> of the season with dialect analysis, but since then, not much.</p><p>After that, linguistics only gets used to either build character or solve a crime in five episodes. For the most part, linguistics is offered in lingo-nugget form.</p><p><strong>Episode 15</strong> <em>Broken</em></p><blockquote>The episode begins with Dr. Blake giving a linguistics lecture about the theory of variation. Unfortunately the lecture only lasts about 30 seconds when she’s interrupted and leaves. In my opinion, the lecture content confuses several different linguistics topics from typology, sociolinguistics, and creoles, but I’m not gonna nitpick the one linguistics lecture they actually show.</blockquote><blockquote>The episode ends with a creep watching video of Dr. Blake giving the earlier linguistics lecture and we see him holding her book “The Route of Linguistics: Contemporary Analytics of Language and Grammar” but nothing more about the book is shown.</blockquote><blockquote><strong><em>Episode 17</em></strong> The Gathering<br>min 32:00 Comparing writing styles, Dr. Blake briefly reads a page then says “the semantics are different. They were all composed by different authors.” While authorship analysis is legit linguistics, it’s also largely computational these days. Generally it’s is not a good idea to make a judgement like this from a cursory reading of a few pages. But teevee is gonna teevee. Every police procedural does this kind of thing with many forensic sciences where an expert looks at a piece of evidence, forms an immediate judgment (because they’re “smart”), and they’re always right in the end.</blockquote><p><strong>Episode 18</strong> <em>Restoration</em></p><blockquote>Crime scene killer’s writing: “look up to the sky” and Dr. Blake says “linguistically could mean God judges the vicim for his sins”</blockquote><blockquote>That’s the whole line. As is, there’s nothing specific to the field of linguistics about interpreting a quote. I think they could have punched it up with maybe metaphor analysis, especially since we later learn she wrote her dissertation on metaphor, or maybe a quip about connotation versus denotation.</blockquote><blockquote>Unfortunately, even Dr. Blake’s minimal analysis is immediately discarded as irrelevant because the quote is only used as a connection to a previous case.</blockquote><p><strong>Episode 19 </strong><em>Pay It Forward</em></p><blockquote>Vague textual analysis is used to identify a false confession. I’ll leave aside whether or not this is good linguistic analysis, my point only is that it counts as an example of using linguistics to solve a crime.</blockquote><p><strong>Episode 22 </strong><em>#6</em></p><blockquote>Dr. Blake says, during a call with her husband, “the worst thing you can do to a linguist is not communicate”. Her husband got offer at Harvard and he says the only way he would agreed is if Harvard gave Dr. Blake a position in linguistics department. They said okay. This gives us some character building. It’s not much, but I’ll take it. Spoiler Alert: she does not take the Harvard position, in order to remain an FBI agent.</blockquote><blockquote>Later Dr. Reid says Dr. Blake’s dissertation was on the fluidity of metaphor. But there’s no discussion about it. I’ll go so far as to count this as an actual linguistic Easter egg because metaphor is a very Berkeley topic and we know she did her PhD at Berkeley from the first episode which introduced her. That had to be intentional, right?</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4df8c70e7f26" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Linguistics on TV — Criminal Minds]]></title>
            <link>https://medium.com/@christopher.phipps/linguistics-on-tv-criminal-minds-58e0f85f93fc?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/58e0f85f93fc</guid>
            <category><![CDATA[linguistics]]></category>
            <category><![CDATA[criminal-justice]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[screenwriting]]></category>
            <category><![CDATA[television]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Mon, 13 Jul 2020 23:04:11 GMT</pubDate>
            <atom:updated>2020-07-19T04:01:45.598Z</atom:updated>
            <content:encoded><![CDATA[<h3>Linguistics on TV — Criminal Minds</h3><p>The TV show Criminal Minds had a forensic linguist character for two seasons (2012–2014) named Dr. Alex Blake. I watched the first episode with her in it (s8, e1 — “The Silencer”) to see how they handled the character with respect to 1) character introduction and 2) furthering the story. After <a href="https://twitter.com/lousylinguist/status/1282449764798218240">tweeting a few thoughts</a>, I put together this longer reaction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/729/1*SMMbROq8GPRNoUB2klDgiQ.png" /></figure><blockquote><strong>Logline: </strong>When a Texas prison inmate escapes during a routine transfer and a guard is subsequently found with his mouth sewn shut, the BAU works with the U.S. Marshals Service to determine whether or not a serial killer known as “The Silencer” has resurfaced. Meanwhile, the team welcomes its newest member, linguistics expert Alex Blake (Jeanne Tripplehorn), into the fold.</blockquote><p><strong>CHARACTER INTRODUCTION</strong></p><p>Here’s her introduction:</p><blockquote><strong>Penelope:</strong> Hey, how’s the new? Is she nice?</blockquote><blockquote><strong>Agent Jareau:</strong> Yeah, she is.</blockquote><blockquote><strong>Penelope: </strong>I’m just asking. Reid, do you know when she was recruited she was 24?</blockquote><blockquote><strong>Reid: </strong>I’ve guest lectured in her forensic linguistics class before.</blockquote><blockquote><strong>Penelope: </strong>Of course you have, ’cause she’s a professor and an agent, which is so impressive. Mm-hmm.</blockquote><blockquote><strong>Agent Jareau: </strong>No, really, she’s great.</blockquote><blockquote><strong>Penelope: </strong>And driven, clearly. Berkeley grad, double major, Bureau star of the Unabomber case. And she teaches at Georgetown. When she’s not doing that, she’s SSAing at the Washington Field Office, and she’s a Ph.D. Does that mean I have to call her doctor? Does that mean I have to call her doctor? What?! I did some research on her ’cause I couldn’t sleep on the plane. And my point is, clearly, she is smart and capable, but is she nice? [Clears throat] Oh, dear God, this is happening, where I talk and the person is behind… Me.</blockquote><blockquote><strong>Dr. Blake: </strong>The origin of “nice” is 12th century middle English, meaning foolish or stupid. I hope you’re referring to the modern use of the adjective. You must be Penelope. I’m Alex Blake.</blockquote><blockquote><strong>Penelope: </strong>Nice to meet you. Really, actually nice in the nicest of ways nice.</blockquote><p>In her very first line, Dr. Blake provides the etymology of the word “nice”.</p><p><strong>Snarky First Impression: </strong>that’s what we linguists do, recite the OED to strangers. It’s how we make friends.</p><p><strong>Nuanced Impression: </strong>actually, plenty of linguists love etymology and do in fact recite word histories (as <a href="https://twitter.com/lynneguist/status/1282452969510903809">Dr. Lynne Murphy</a> and others were quick to point out to me), but the way this is handled to introduce Dr. Blake makes her come across as condescending and belligerent. There’s no chance Dr. Blake actually thinks Penelope is using a 900 year old sense of “nice”. I get that they needed to establish her linguistics credibility right off the bat, but I think they could have handled this more generously. Especially frustrating because it reinforces the <em>language police</em> reputation that linguists hate so much.</p><p>The preceding line from Penelope comes across as a recitation of a resume. It’s clunky, fast exposition. It got the job done, but I feel like we didn’t need all this info shoveled at us all at once. Much of these character details could have been dropped occasionally throughout the script. But in one short interaction, we got her linguistics and FBI background plus a lingo-nugget and now we’re off to solve a crime.</p><p>FURTHERING THE STORY</p><p>While there is some vaguely language-related stuff involving sewing mouths shut and writings left at the scenes of crimes, I didn’t feel these were specifically instances of linguistics per se. More like psychological profiling. By my analysis, there are nine times the script used linguistics specifically to further the story. They generally fell into three categories:</p><ol><li>Narrow down the suspect pool = 5 instances</li><li>Character building = 3 instances</li><li>Narrow down a location = 1 instance</li></ol><p>I’m listing these in chronological order to give a sense of the way they sprinkled these throughout the script.</p><blockquote>1. Dr. Blake quotes etymology of “nice” — character building, discussed above.</blockquote><blockquote>2. Dr. Blake knows ‘primary languages’ of Texas are Spanish, English, and German — helps narrow down suspect pool. While suspect was in jail in Texas, he mostly reads in French (not Spanish, English, or German), suggesting he’s not from Texas but rather he grew up in New Orleans. These dots are not connected right away.</blockquote><blockquote>3. Team uses handwriting analysis — narrows down suspect pool by suggesting personality traits. Probably not linguistics per se, more psychology, but I’ll allow it.</blockquote><blockquote>4. Suspect had high reading comprehension but low writing style (“inferior” … cringe) — narrows down suspect pool by suggesting his writing style mirrors ‘what he grew up hearing’. This logic was muddled for me and they didn’t explain it. Seems like some ugly assumptions about speaking styles are built into this point.</blockquote><blockquote>5. Perpetrator teased about a speech impediment as a teen — narrows down suspect pool as it allows the team to eliminate several other suspects.</blockquote><blockquote>6. IPA gets defined — character building. Dr. Blake connects dots of different language use modalities to “highlight his desire to communicate’. More on this below.</blockquote><blockquote>7. Experimental cochlear implant — narrows down suspect pool by identifying the one suspect who had one. Also character building because the killer’s cochlear implant was faulty and caused him pain.</blockquote><blockquote>8. Perpetrator uses sign language — character building. Through flashbacks we see the killer trying to use sign language to communicate with his abusive mom. Dr. Blake will later use sign language to communicate directly with the killer.</blockquote><blockquote>9. Texas South English dialect — narrows down location of killer. The phrase “waiting on the taste of honey” is used to identify a prison neighbor of the killer whose house is searched. Of that phrase, Dr. Blake says “It’s got Southern US roots. But more specifically, it’s Texas South, a sub-dialect of English found in the central part of this state.” I like the idea of using dialect to identify speakers. Unfortunately Dr. Blake doesn’t explain which linguistic features indicate that dialect (it wasn’t obvious to me).</blockquote><p>Generally it’s gratifying to see linguistics used to further the plot and build characters. This episode’s use of linguistics was not great, but I give them credit for trying.</p><p>CALL OUT</p><p>So, I have to call out #6. Here’s the fuller context:</p><blockquote>“I’ve found that those who understand something as detailed as [International Phonetic Alphabet] are also proficient in sign language.”</blockquote><p>What. A. Load. Of. Crap. This is a doosie. I’m happy to give screenwriters leeway to tell their stories. They’ve only got 40 minutes or so to get across a lot of information. But we can ask that their claims, especially technical claims, at least pass the smell test. This literally makes no sense, and no linguist would ever say something like this.</p><p>Also, a dirty secret within linguistics (one of many) is that linguists who study sign language <a href="https://twitter.com/Linguischick/status/1282530704153284609">struggle to find their place</a> in mainstream linguistics. It is not given the respect as a field of study that it deserves.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=58e0f85f93fc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[TV Linguistics — Pronouncify.com and the fictional Princeton Linguistics department]]></title>
            <link>https://medium.com/@christopher.phipps/tv-linguistics-pronouncify-com-and-the-fictional-princeton-linguistics-department-5534a7111904?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/5534a7111904</guid>
            <category><![CDATA[tv-series]]></category>
            <category><![CDATA[screenwriting]]></category>
            <category><![CDATA[writing]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Mon, 17 Feb 2020 21:12:50 GMT</pubDate>
            <atom:updated>2020-02-17T21:12:50.397Z</atom:updated>
            <content:encoded><![CDATA[<h3>TV Linguistics — Pronouncify.com and the fictional Princeton Linguistics department</h3><p>[reposted from <a href="https://thelousylinguist.blogspot.com/2020/02/tv-linguistics-pronouncifycom-and.html">11/20/10</a>]</p><p>I spent Thursday night on a plane so I missed 30 Rock and the most linguistics oriented sit-com episode since … since … um … okay, the most linguistics oriented sitcom episode EVER! But thanks to the innerwebz, I have caught up on my TV addiction.</p><p>The set-up has Jack Donaghy being the voice of Pronouncify.com, (I BEG you to sign up, PLEASE!!!!) a website that demonstrates the correct pronunciation of all English words. Apparently, when Jack was a poor undergrad at Princeton, he was hired by the “Linguistics Department” to pronounce every word in an English dictionary to preserve the correct pronunciation for generations to come. But they sold his readings, and hence his voice is now the voice of Pronouncify.com (as well as the first perfect microwave…).</p><p>Here is as faithful a transcript of the critical dialogue as I can muster:</p><p>Jack: Those bastards!<br>Liz: Who bastards?<br>Jack: Part of my Princeton scholarship included work for the Linguistics department. They wanted me to record every word in the dictionary to preserve the perfect American accent in case of nuclear war. Well, the cold war ended, and Princeton began selling the recordings.<br>Liz: So people can just buy your voice?<br>Jack: Ohhhh, the things it’s been dragged into. Thomas the Tank Engine; Wu-Tang songs…</p><p>This must have been the glory days before the hippies took over and started “protecting” undergrads from “exploitation.” Whatever…</p><p>In any case, it’s understandable that this trivial tid-bit of academic minutia blew right by most people, but it is a fact of the world we live in that Princeton University does not have a linguistics department per se. They do offer an Undergraduate Program in Linguistics in which students can “pursue a Certificate in Linguistics,” but this is not an official department as far as I understand it. Jack, if he is the same age as the actor Alec Baldwin, would have been at Princeton in late 1970s. Maybe they had a full fledged department back then, I honestly don’t know.</p><p>You can watch the episode College at NBC, or wherever else you prefer. BTW, there’s an awesome ode to color perception conundrums at the end as well. It’s all kinda linguisticee/cog sciencee (I never know how to add the -ee morpheme?).</p><p>Random after-point: Near the end of Thursday’s episode of Community, Dean Pelton actually utilized the Shakespearean subjunctive construction Would that X were Y… He says “Would that this hoodie were a time hoodie” around the 19:20 mark (see Hamlet, would it were not so, you are my mother). Just thought that was kinda awesome.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5534a7111904" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to create a custom NLU annotation without writing a line of code]]></title>
            <link>https://medium.com/@christopher.phipps/how-to-create-a-custom-nlu-annotation-without-writing-a-line-of-code-e5f43d1fd6b9?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e5f43d1fd6b9</guid>
            <category><![CDATA[watson-knowledge-studio]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[tutorial]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ibm-watson]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Tue, 05 Dec 2017 19:32:24 GMT</pubDate>
            <atom:updated>2018-01-04T21:03:22.768Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*nVQO0JvNp3CDuV0bdlcUtA.png" /></figure><p><a href="http://ibm.biz/NaturalLanguageUnderstanding">IBM Watson Natural Language Understanding</a> (NLU) is a cloud-based Watson service that enables developers to easily and quickly extract and analyze metadata from unstructured text.</p><p>In this post, we’ll look at how we can customize data from IBM Watson Natural Language Understanding by adding human annotations using Watson Knowledge Studio, and how that impacts our results.</p><p><strong><em>If you haven’t already, </em></strong><a href="https://www.ibm.com/account/us-en/signup/register.html?a=IBMWatsonKnowledgeStudio"><strong><em>sign up for free access</em></strong></a><strong><em> to Watson Knowledge Studio.</em></strong></p><p>I’ve taken the text from this post: <a href="https://www.ibm.com/blogs/think/author/aleksandramoj/">https://www.ibm.com/blogs/think/author/aleksandramoj/</a></p><p>And saved it as a txt file. I’ll use this as the baseline of data to annotate for a new entity which I’ll name Social Issues.</p><p>After uploading my sample data (the <a href="https://www.ibm.com/blogs/think/2017/06/science-for-social-good/">blog post</a>), I’m going to teach Watson how to identify entities called “social issues.” But, I have to initially tell it what I mean by “social issues.”</p><p>I’ll use the Human Annotation tool to begin this process. That means I’ll act as the trainer that tells the system what words need to be tagged as social issues in the document.</p><p>Once we annotate this information with what terms in the document are relevant to social issues, it will allow us to use NLU to recognize the exact terms, and should be able to identify associated “social issues” terms within other documents.</p><p>We’ll test that theory in a moment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yUCyKuDBm-pwEbak8Beb7w.png" /></figure><p>Now I am ready to annotate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WqwR-7LISK7nmpC4C30TYw.png" /></figure><p>Select the entity you want to train, which for us is Social_issues.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*BOfMVs9JTi7VXaxFru9emw.png" /></figure><p>Now I am ready to tag items that I think refer to social issues. After I’ve tagged all of the content that I think is relevant to social issues, the text looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ggJtt2KMC_rfSQmb9V5_uw.png" /></figure><p>Since I’ve finished tagging my terms, I’m going to accept the changes by saving the file, closing it and adding ‘accept’ to its status. This means I’m satisfied with the way the system has been trained and it is now considered the ‘ground truth’.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*LPKj526NNlSSb555e2E7RA.png" /></figure><p>I’ll then move to the annotator component tab and select machine learning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/421/1*9kd16FIAjCVqNCvvZ3JYpA.png" /></figure><p>Since this is a small data set of one text file and we have no test set or blind set, I’ll select the file and set it to 100%</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fq5S858pVznX4H9c6Y4kxw.png" /></figure><p>I then selected Next -&gt; Train and let Watson NLU do its thing:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/383/1*fwG3NdCn7wRmBpnQbmtI0A.png" /></figure><p>Now, since I trained Watson on the terms I thought were social issues, let’s see how it learned. I’m using a new blog topic, <a href="https://www.ibm.com/blogs/think/2016/12/watson-cancer-care/">https://www.ibm.com/blogs/think/2016/12/watson-cancer-care/</a>. I applied my model ID that deployed my custom entities.</p><p>Even though there was only a limited sample as shown above, when I tested the new URL in Bluemix using entities and my custom entity model, I retrieved the following:</p><p><a href="https://gist.github.com/thelousylinguist/10c4aab00a70da39c00c6eb54b3dcf94">https://gist.github.com/thelousylinguist/10c4aab00a70da39c00c6eb54b3dcf94</a></p><p>Watson NLU identified “<strong>cancer care</strong>”, “<strong>treatment possible</strong>”, “<strong>medical practice</strong>” and “<strong>treatment for</strong>” as social issues even though none of those terms were tagged in the original training!</p><p>So, even with a small data set (one blog post), you can see the power of teaching Watson the language of your domain for advanced text analytics.</p><p><em>Originally published at </em><a href="https://developer.ibm.com/dwblog/2017/create-custom-nlu-annotation-without-writing-line-code/"><em>developer.ibm.com</em></a><em> on October 20, 2017.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e5f43d1fd6b9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A linguist asks some questions about word vectors]]></title>
            <link>https://medium.com/@christopher.phipps/a-linguist-asks-some-questions-about-word-vectors-5ba8afce8edd?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/5ba8afce8edd</guid>
            <category><![CDATA[word-embeddings]]></category>
            <category><![CDATA[linguistics]]></category>
            <category><![CDATA[word2vec]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[vector]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Wed, 18 Oct 2017 21:29:35 GMT</pubDate>
            <atom:updated>2017-10-19T00:16:09.830Z</atom:updated>
            <content:encoded><![CDATA[<p>I have at best a passing familiarity with word vectors, strictly from a 30,000 foot view. I’ve never directly used them outside a handful of toy tutorials (though they have been embedded in a variety of tools I use professionally). There is something very linguisticee about word vectors, so I’m generally happy with their use. The basic idea goes back aways, but was most famously articulated by J. R. Firth’s pithy phrase “<em>You shall know a word by the company it keeps</em>” (c. 1957).</p><p>This is a set of questions I had when reading Chris McCormick’s tutorial <a href="http://mccormickml.com/2016/04/19/word2vec-tutorial-the-skip-gram-model/">Word2Vec Tutorial — The Skip-Gram Model</a>. My ideal audience is NLPers who have a good understanding of the SOTA of word vectors.</p><p>Quick review of the tutorial:</p><ul><li><strong>Key Take Away</strong>: Word Vectors = the output probabilities that relate how likely it is to find a vocabulary word nearby a given input word.</li><li><strong>Methods</strong>: ngrams for training, order doesn’t matter</li><li><strong>Insight</strong>: If two words have similar contexts, then the network is motivated to learn similar word vectors for the two words.</li></ul><p><strong>Why sentences?</strong></p><ul><li>When choosing word window size, word vectors respect sentence boundaries. Why?</li><li>By doing this, vectors are modeling edited, written language.</li><li>Vectors are assuming the semantics of sentences are coherent in some way.</li><li>Vectors are, in essence, relying on sentences to be semantically coherent. As long as they are, the method works, but when they aren’t, how does this method break down?</li><li>This is an assumption often broken in spoken language with dissfluencies, topic shifts, interruptions, etc</li><li>How else might spoken language models differ?</li><li>How might methods change to account for this</li><li>Do words from previous sentences have semantic relationships to the words in following sentences? (Yes. Because discourse.)</li><li>Why miss out on that information?</li></ul><p><strong>Why the middle?</strong></p><ul><li>Vectors use words “in the middle of a sentence”? Why?</li><li>Do words that tend to occur in the middle of sentences differ in meaningful ways than those that do not? (Possibly. E.g, verbs and prepositions rarely begin or end sentences. Information structures affects internal ordering of constituents in interesting ways).</li><li>Did they look at the rate of occurrences of each unique lexical entry in-the-middle and near the peripheries?</li></ul><p><strong>Synthetic Languages</strong>:</p><ul><li>How well would this word vectors work for a synthetic language like Mohawk?</li><li>What pre-processing steps might need to be added/modified?</li><li>Would those modifications be enough to get similar/useful results?</li></ul><p><strong>Antonyms</strong>:</p><ul><li>Quote: “<em>the network will likely learn similar word vectors for the words “ant” and “ants”</em></li><li>Can the same method learn the difference between “ant” and “ants”?</li><li>I haven’t read it yet, but these folks seem to use word vectors to learn antonyms: <a href="https://www.aclweb.org/anthology/N15-1100">Word Embedding-based Antonym Detection using Thesauri and Distributional Information</a>.</li></ul><p>This was a fun read.</p><p>The original tutorial:<br> McCormick, C. (2016, April 19). Word2Vec Tutorial — The Skip-Gram Model. Retrieved from <a href="http://www.mccormickml.com/">http://www.mccormickml.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5ba8afce8edd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nuts and Bolts of Applying Deep Learning (Andrew Ng)]]></title>
            <link>https://medium.com/@christopher.phipps/nuts-and-bolts-of-applying-deep-learning-andrew-ng-67883aba9238?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/67883aba9238</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[andrew-ng]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Fri, 15 Sep 2017 22:15:43 GMT</pubDate>
            <atom:updated>2017-09-15T22:16:36.500Z</atom:updated>
            <content:encoded><![CDATA[<p>I recently watched Andrew Ng’s excellent lecture from 2016 <em>Nuts and Bolts of Applying Deep Learning</em> and took notes. I post them as a helpful resource for anyone who wants to watch the video.</p><p><a href="https://www.youtube.com/watch?v=F1ka6a13S9I">https://www.youtube.com/watch?v=F1ka6a13S9I</a></p><p>I broke it into the following sections</p><ol><li>End-to-End DL for rich output</li><li>Buckets of DL</li><li>Bias and variance</li><li>Applied machine learning work flow</li><li>New Era of ML</li><li>Build a Unified Data Warehouse</li><li>The 70/30 Split Revisited</li><li>Comparing to Human Level Performance</li><li>How do you define human level performance?</li><li>How do you build a career in machine learning?</li><li>AI is the new electricity</li></ol><p><strong>Intro</strong></p><ol><li>End to end DL — work flow</li><li>Bias and variance has changed in era of deep learning</li><li>DL been around for decades, why do they work well now?</li></ol><ul><li>Scale of data and computation</li><li>Two teams</li><li>AI Teams</li><li>Systems team</li><li>Sit together</li><li>Difficult for any one human to be sufficiently expert in multiple fields</li></ul><p><strong>End-to-End DL for rich output</strong></p><ul><li>From first three buckets below</li><li>Traditional ML models output real numbers</li><li>End-to-end DL can out put more complex things than numbers</li><li>Sentence captions for images</li><li>Speech-to-text</li><li>Machine translation</li><li>Synthesize new images (13:00)</li><li>End-to-End DL not the solution to everything.</li><li>End-to-end = having just a DL between input and output</li><li>Rules for when to use (13:35)</li><li>Old way: audio — — → phonemes → transcript</li><li>New DL way: audio — — — — — — — — -&gt; transcript</li><li>Makes for great PR, but only works some times (15:31)</li><li>Achilles heel — need lots of labeled data</li><li>Maybe phonemes are just a fantasy of linguists (15:48)</li><li>Advantage of old non-end-to-end architecture is it allows you to manually add more information to the processing (18:16)</li><li>Also, for self-driving cars, no one has enough data (right now) to make end-to-end work) (20:42)</li><li>Common problem — after first round of dev, ML not working that well, what do you do next?</li><li>Collect more data</li><li>Train longer</li><li>Different architecture (e.g., switch to NNs)</li><li>Regularization</li><li>Bigger model</li><li>More GPUs</li><li>Skill in ML engineer is knowing how to make these decisions (22:33)</li></ul><p><strong>Buckets of DL</strong></p><ol><li>General models</li></ol><ul><li>Densely connected layers — FC</li><li>Sequence models — 1D (RNN, LSTM, GRU, attention)</li><li>Image models — 2D, 3D (Convo nets)</li><li>Other — unsupervised, reinforcement</li></ul><ol><li>First three buckets driving market advances</li><li>But “Other” bucket is future of AI</li></ol><p><strong>Bias and variance — evolving</strong></p><ol><li>Scenario: build human level speech rec system</li></ol><ul><li>Measure human level error — 1</li><li>Training set error — 5%</li><li>Dev set — 6%</li></ul><ol><li>Bias = difference between human error level and your system’s</li><li>TIP: For bias problems try training a bigger model (25:21)</li><li>Variance (overfitting): if Human 1%, Training 2%, Dev 6%</li><li>TIP: for variance, try adding regularization, early stopping, best bet = more data</li><li>Both high bias and high variance: if Human 1%, Training 5%, Dev 10%</li><li>“<em>sucks for you</em>” (direct quote 26:30)</li></ol><p><strong>Applied machine learning work flow</strong></p><ol><li>Is your training error high</li></ol><ul><li>Yes</li><li>Bigger model</li><li>Train longer</li><li>New architecture</li><li>Repeat until doing well on training set</li></ul><ol><li>Is dev error high?</li></ol><ul><li>Yes</li><li>Add data</li><li>Regularization</li><li>New architecture</li><li>Repeat until doing well on training set</li></ul><ol><li>Done</li></ol><p><strong>New Era of ML</strong></p><ol><li>We now know whatever problem you are facing (high bias or high variance) you have at least one action you can take to correct</li><li>No longer a bias/variance trade-off (29:47)</li><li>“Dumb” formula of bigger model/more data is easy to implement even for non-experts and is enough to do very well on a lot of problems (31:09)</li><li>More data has led to interesting investments</li></ol><ul><li>Data synthesis — Growing area</li><li>Examples-</li><li>OCR at Baidu</li><li>Take random image</li><li>Random word</li><li>Type random word in Microsoft Word</li><li>Use random font</li><li>You just created training data for OC</li><li>Still takes some human intervention, but lots of progress</li><li>Speech recognition</li><li>Take clean audio</li><li>Add random noise to background for more data</li><li>E.g., add car noise</li><li>Works remarkably well</li><li>NLP</li><li>Take ungrammatical sentences and auto-correct</li><li>Easy to create ungrammatical sentences programmatically</li><li>Video games in RL</li></ul><ol><li>Data synthesis has a lot of limits (36:24)</li></ol><ul><li>Why not take cars from Grand Theft Auto and use that as training data for self-driving cars</li><li>20 cars in video game enough to give “realistic” impression to player</li><li>But 20 cars is very impoverished data set for self-driving cars</li></ul><p><strong>Build a Unified Data Warehouse</strong></p><ol><li>Employees can be possessive of “their” data</li><li>Baidu- it’s not your data, it’s company data</li><li>Access rights can be a different issue</li><li>But warehouse everything together</li><li>Kaggle</li></ol><p><strong>The 70/30 Split Revisited</strong></p><ol><li>In academia, common for test/train to come from same distribution</li><li>But more ommon in industry for test and train to come from different distributions</li></ol><ul><li>E.g., speech rec at Baid</li><li>Speech enabled rear view mirror (in China)</li><li>50,000 hours of regular speech data</li><li>Data not from rear-view mirror interactions though</li><li>Collect another 10 hours of rear-view mirror scenario</li><li>What do you do with the original 50,000 hours of not-quite right data?</li><li>Old method would be to build a different model for each scenario</li><li>New era, one model for all data</li><li>Bad idea, split 50,000 into training/dev, use 10,000 as test. DON’T DO THIS.</li><li>TIP: Make sure dev and test are from same distro (boosts effectiveness)</li><li>Good Idea: make 50,000 train, split 10,000 into dev/test</li><li>Dev set = problem specification</li><li>Me: “dev set = problem you are trying to solve”</li><li>Also, split off just 20 hours from 50,000 to create tiny “dev-train” set</li><li>this has same distro as train</li></ul><ol><li>Mismatched train and dev set is problem that academia doesn’t work on much</li></ol><ul><li>some work on domain adaptation, but not much (44:53)</li></ul><ol><li>New architecture fix = “hail mary” (48:58)</li><li>Takes a long time to really grok bias/variance</li></ol><ul><li>People who really understand bias/variance deeply are able to drive rapid progress in machine learning (50:33)</li></ul><p><strong>Common Theme — Comparing to Human Level Performance</strong></p><ol><li>Common to achieve human level performance, then level off</li><li>Why?</li></ol><ul><li>Audience: <em>Labels come from humans</em></li><li>Audience: <em>Researchers get satisfied with results</em> (the <em>laziness </em>hypothesis)</li><li>Andrew: theoretical limits (aka <strong>optimal error</strong> rate, Bayes rate)</li><li>Some audio so bad, impossible to transcribe (phone call from a rock concert)</li><li>Some images so blurry, impossible to interpret</li><li>Humans are really good at some things, so once you surpass human level accuracy, there’s not much room left to improve (54:38)</li></ul><ol><li>While worse than humans, still ways to improve</li></ol><ul><li>Get labels from humans</li><li>Error analysis</li><li>Estimate bias/variance effects</li></ul><ol><li>For tasks that humans are bad at (say 30% error rate), really hard to find guidance on how to improve</li></ol><p><strong>How do you define human level performance?</strong></p><ol><li>Quiz: Which is the most useful definition? (101:000</li></ol><ul><li>Example: Medical image reading</li></ul><ol><li>Typical non-doctor error — 3%</li><li>Typical doctor — 1%</li><li>Expert doctor — 0.7%</li><li>Team of expert doctors — 0.5%</li></ol><ul><li>Answer: Team of expert doctors is best because ideally you are using human performance to proxy <strong>optimal error</strong> rate.</li></ul><p><strong>What can AI do?</strong> (106:30)</p><ol><li>Anything that a typical person can do in less than one second.</li></ol><ul><li>E.g., Perception tasks</li><li>Audience: if a human can do it in less than a second, you can get a lot of data</li></ul><p><strong>How do you build a career in machine learning</strong> (111:00)</p><ol><li>Andrew says he does not have a great answer (me: but he does have a good one)</li></ol><ul><li>Taking a ML course</li><li>Attend DL school</li><li>Work on project yourself (Kaggle)</li><li>Mimic PhD student process</li><li>Read a lot of papers (20+)</li><li>Replicate results</li><li>Dirty work</li><li>Downloading/cleaning data</li><li>Re-running someone’s code</li><li>Don’t only do dirty work</li><li>PhD process + Dirty work = reliable</li><li>Keep it up for a year</li><li>Competency</li></ul><p><strong>AI is the new electricity</strong> (118:00)</p><ol><li>Transforms industry after industry</li><li>Get in on the ground floor</li><li>NOTE: this is the title of his follow-up talk, which has a video link at the end of the one above.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67883aba9238" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[NLPers: How would you characterize your linguistics background?]]></title>
            <link>https://medium.com/@christopher.phipps/nlpers-how-would-you-characterize-your-linguistics-background-cb48c703bb31?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb48c703bb31</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Sun, 09 Apr 2017 19:13:01 GMT</pubDate>
            <atom:updated>2017-04-09T19:13:01.694Z</atom:updated>
            <content:encoded><![CDATA[<p>That was the poll question my hero Professor Emily Bender posed on Twitter March 30th. 573 tweets later, a truly epic thread had been created pitting some of the most influential NLPers in the world head to head in a contest of wit and debate that is truly what the field of NLP needed. Unfortunately, Twitter proves hapless at being able to view the thread as a single conversation using their app. But, student Sebastian Meilke used a Chrome extension called Treeverse to put together the whole thread into a single, readable format, complete with sections!</p><p>If you are at all interested in NLP or linguistics, this is a must read: <a href="http://sjmielke.com/nlp-cl-megathread.htm">NLP/CL Twitter Megathrea</a>.</p><p>I would be remiss if I didn’t note my own small role in this. Emily’s poll was sparked by my own Tweet where I said I was glad the Linguistics Society of America is starting their own Society for Computational Linguistics because “the existing CL and NLP communities have gotten farther and farther off the linguistics path”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/548/1*SGcEvKIZ07Hsm2h7bcOn7A.png" /></figure><p>Happy reading.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb48c703bb31" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using IBM Watson Knowledge Studio to Train Machine Learning Models]]></title>
            <link>https://blog.chatbotslife.com/using-ibm-watson-knowledge-studio-to-train-machine-learning-models-33da7b2b319b?source=rss-8da58490a69b------2</link>
            <guid isPermaLink="false">https://medium.com/p/33da7b2b319b</guid>
            <category><![CDATA[how-to]]></category>
            <category><![CDATA[ibm-watson]]></category>
            <category><![CDATA[cognitive-computing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Christopher Phipps]]></dc:creator>
            <pubDate>Wed, 22 Mar 2017 19:08:28 GMT</pubDate>
            <atom:updated>2017-03-24T18:32:23.402Z</atom:updated>
            <content:encoded><![CDATA[<p>Using the Free Trial version of IBM’s <a href="https://www.ibm.com/us-en/marketplace/supervised-machine-learning">Watson Knowledge Studio</a>, I just annotated a text and created a machine learning model in about 3 hours without writing a single line of code. The mantra of WKS is that you don’t program Watson, you teach Watson.</p><p>For demo purposes I chose to identify personal relationships in Shirley Jackson’s 1948 short story <a href="https://en.wikipedia.org/wiki/The_Lottery">The Lottery</a>. This is a haunting story about a small village and its mindless adherence to an old, and tragic tradition. I chose it because 1) it’s short and 2) it has clear person relationships like brothers, sisters, mothers, and fathers. I added a few other relations like AGENT_OF (which amounts to subjects of verbs) and X_INSIDE_Y for things like pieces of paper inside a box.</p><blockquote><strong><em>Caveat</em></strong><em>: This short story is </em><strong>really</strong><em> short: 3300 words. So I had no high hopes of getting a </em><strong>good</strong><em> model out of this. I just wanted to go through an entire machine learning work flow from gathering text data to outputting a complete model </em><strong><em>without</em></strong><em> writing a single line of code. And that’s just what I did.</em></blockquote><p><strong>WORK FLOW</strong></p><p>I spent about 30 minutes prepping the data. E.g., I broke it into 20 small snippets (to facilitate test/train split later), I also edited some quotation issues, spelling, etc).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/132/0*gV8ZKoLTrkLy1m6a.PNG" /></figure><p>It uploaded into WKS in seconds (by simply dragging and dropping the files into the browser tool). I then created a Type System to include entity types such as these:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*zEcyKgheHy7Poytv.PNG" /></figure><p>And relation types such as these:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/364/0*hF0Oz_xHx287Y7pF.PNG" /></figure><p>I then annotated the 20 short documents in less than two hours (as is so often the case, I re-designed my type system several times along the way; luckily WKS allows me to do this fluidly without having to re-annotate).</p><p>Here’s a shot of my entity annotations:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*p0CIAOA1IEh8zvc0.PNG" /></figure><p>Here’s a shot of my relation annotations:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*iIYX3_Gnlbqmd7MK.PNG" /></figure><p>I then used these manually annotated documents as ground truth to teach a machine learning model to recognize the relationships automatically using a set of linguistic features (character and word ngrams, parts-of-speech, syntactic parses, etc). I accepted the WKS suggested split of documents as 70/23/2:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*HchvJoYFKGoUXEly.PNG" /></figure><p>I clicked “Run” and waited:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/285/0*aAFTiZzUTHE3hx-e.PNG" /></figure><p>The model was trained and evaluated in about ten minutes. Here’s how it performed on entity types:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*AQdF_9qf3YwAwfNw.PNG" /></figure><p>And here’s how it performed on relation types:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*ZIrqRwMkyy-tP723.PNG" /></figure><p>This is actually not bad given how sparse the data is. I mean, an F1 of .33 on X_INSIDE_Y from only 9 examples on a first pass. I’ll take it, especially since that one is not necessarily obvious from the text. Here’s one example of the X_INSIDE_Y relation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*QszaPStYpS80kH31.PNG" /></figure><p>So I was able to train a model with 11 entity types and 81 relation types on a small corpus in less than three hours start to finish without writing a single line of code. I did not program Watson. I taught it</p><iframe src="https://cdn.embedly.com/widgets/media.html?url=https%3A%2F%2Fpowered.by.rabbut.com%2Fp%2FqnHY%3Fc%3D0&amp;src=https%3A%2F%2Fpowered.by.rabbut.com%2Fp%2FqnHY%3Fc%3D0&amp;type=text%2Fhtml&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;schema=rabbut" width="500" height="500" frameborder="0" scrolling="no"><a href="https://medium.com/media/47a0a8491046a6d8bdbb86880dd8fc1e/href">https://medium.com/media/47a0a8491046a6d8bdbb86880dd8fc1e/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/725/1*bQlRSzFHJEmF4Q7PyrLgng.gif" /></figure><figure><a href="https://chatbotslife.com/bot-communities-mastermind-group-d2dae9876709#.53x0py6ou"><img alt="" src="https://cdn-images-1.medium.com/max/255/1*6XUspT9JOSq0w0Fi35HIaA.png" /></a></figure><figure><a href="https://m.me/ChatbotsLife"><img alt="" src="https://cdn-images-1.medium.com/max/255/1*c1LDMH5vbnIz9rmAka8Hwg.png" /></a></figure><figure><a href="https://chatbotslife.com/how-to-get-a-free-chatbot-b1fb9dfe109#.z9dtp2sy0"><img alt="" src="https://cdn-images-1.medium.com/max/255/1*D0Jf3dI6ZThtqcfwDYY7mg.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=33da7b2b319b" width="1" height="1" alt=""><hr><p><a href="https://blog.chatbotslife.com/using-ibm-watson-knowledge-studio-to-train-machine-learning-models-33da7b2b319b">Using IBM Watson Knowledge Studio to Train Machine Learning Models</a> was originally published in <a href="https://blog.chatbotslife.com/">Chatbots Life</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>