<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: nonacosa</title>
    <description>The latest articles on DEV Community by nonacosa (@_cda735563962186822ba).</description>
    <link>https://dev.to/_cda735563962186822ba</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4255454%2Fb7f9aabe-9843-4c15-afa0-1d5038131471.png</url>
      <title>DEV Community: nonacosa</title>
      <link>https://dev.to/_cda735563962186822ba</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/_cda735563962186822ba"/>
    <language>en</language>
    <item>
      <title>How We Search 207 Million Spoken Lines Without Sponsor Reads (and Cut 6-Minute Clips in the Browser)</title>
      <dc:creator>nonacosa</dc:creator>
      <pubDate>Sat, 21 Nov 2026 17:19:09 +0000</pubDate>
      <link>https://dev.to/_cda735563962186822ba/how-we-search-207-million-spoken-lines-without-sponsor-reads-and-cut-6-minute-clips-in-the-browser-55o4</link>
      <guid>https://dev.to/_cda735563962186822ba/how-we-search-207-million-spoken-lines-without-sponsor-reads-and-cut-6-minute-clips-in-the-browser-55o4</guid>
      <description>&lt;p&gt;Long-form podcasts and university lectures hold some of the clearest explanations on the internet, buried inside 3-hour recordings.&lt;/p&gt;

&lt;p&gt;Over the past few months I built &lt;a href="https://pastnoise.com/en" rel="noopener noreferrer"&gt;PastNoise&lt;/a&gt;, a search and documentary-editing engine over 206.9 million spoken transcript lines from 483,987 videos across 2,494 channels. Instead of returning a wall of links or an AI text summary, it pulls primary-source spoken clips from multiple speakers, stitches them into a 6-minute bilingual film, and pairs each claim with Wikipedia or PubMed citations.&lt;/p&gt;

&lt;p&gt;Along the way, four concrete engineering problems bit us hard. Here is what broke and what actually fixed each one.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Substring matches and sponsor reads poison podcast search
&lt;/h2&gt;

&lt;p&gt;If you run a naive Elasticsearch &lt;code&gt;match&lt;/code&gt; query over 200M+ conversational lines, two things happen immediately:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Short words trigger noisy matches. Searching for &lt;code&gt;sleep&lt;/code&gt; or &lt;code&gt;AI&lt;/code&gt; pulls in casual mentions like "I didn't sleep on this deal" or host banter from the first 90 seconds of an episode.&lt;/li&gt;
&lt;li&gt;Sponsor reads score surprisingly high on health and tech queries because hosts repeat keywords like "magnesium", "focus", "security", and "sleep" inside mid-roll ad scripts.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;What worked for us:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Enforcing exact word-boundary checks (&lt;code&gt;\b&lt;/code&gt;) on query terms during candidate reranking, plus filtering out opening episode fluff (&lt;code&gt;first 60s&lt;/code&gt; greeting patterns and sponsor-code phrases like "use code", "percent off", "brought to you by").&lt;/li&gt;
&lt;li&gt;Scoring by speaker authority tier (&lt;code&gt;PRIMARY&lt;/code&gt; first-person researcher or founder vs. &lt;code&gt;SECONDARY&lt;/code&gt; commentary), so an interview with the actual scientist beats a reaction channel talking about them.&lt;/li&gt;
&lt;li&gt;Balancing chapters across distinct channels instead of pulling five clips from the same episode. You can inspect how a finished multi-voice cut looks on &lt;a href="https://pastnoise.com/en/film/phone-addiction-dopamine-habits" rel="noopener noreferrer"&gt;Your Phone Is a Dopamine Slot Machine&lt;/a&gt; or &lt;a href="https://pastnoise.com/en/film/nvidia-jensen-huang-gpu-ai" rel="noopener noreferrer"&gt;The Engine That Learned to Think&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  2. Syncing bilingual subtitles without DOM jitter
&lt;/h2&gt;

&lt;p&gt;YouTube's &lt;code&gt;IFrame Player API&lt;/code&gt; does not emit per-frame time events. If you poll &lt;code&gt;getCurrentTime()&lt;/code&gt; every 500ms, fast conversational speech visibly lags behind the transcript. If you let the subtitle box resize between 1-line and 3-line cues, the entire control bar below it jumps up and down.&lt;/p&gt;

&lt;p&gt;The fix was small and boring:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Poll &lt;code&gt;player.getCurrentTime()&lt;/code&gt; on an 80ms interval only while &lt;code&gt;isPlaying&lt;/code&gt; is true, matching the latest cue where &lt;code&gt;currentSec + 0.08 &amp;gt;= cue.startSec&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Lock the subtitle well to a fixed minimum height (&lt;code&gt;min-h-[96px]&lt;/code&gt; on desktop) so 1-line and 2-line bilingual cues never cause layout shift.&lt;/li&gt;
&lt;li&gt;When switching acts, mute and pre-buffer the next video behind a full-frame intertitle card for ~1.6s, then snap to &lt;code&gt;startSec&lt;/code&gt; and unmute. That eliminates the jarring YouTube red spinner between chapters.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  3. Running single-thread &lt;code&gt;ffmpeg.wasm&lt;/code&gt; in Next.js 15 without Webpack crashing
&lt;/h2&gt;

&lt;p&gt;We built a pure-frontend social clip exporter so users can capture a chapter sequence at 1.5x or 2.0x speed using &lt;code&gt;getDisplayMedia&lt;/code&gt; + Region Capture (&lt;code&gt;CropTarget&lt;/code&gt;) and have the browser automatically retime it back to smooth 1.0x MP4 with &lt;code&gt;setpts&lt;/code&gt; and &lt;code&gt;atempo&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Importing &lt;code&gt;@ffmpeg/ffmpeg&lt;/code&gt; directly inside a Next.js client component threw &lt;code&gt;Cannot find module 'http://.../ffmpeg/ffmpeg-core.js'&lt;/code&gt; because Webpack tried to rewrite the dynamic worker import at bundle time. Also, multi-threaded &lt;code&gt;ffmpeg.wasm&lt;/code&gt; requires &lt;code&gt;Cross-Origin-Embedder-Policy: require-corp&lt;/code&gt;, which immediately breaks cross-origin YouTube iframes on the same page.&lt;/p&gt;

&lt;p&gt;The clean escape hatch:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Use the single-thread UMD build of &lt;code&gt;@ffmpeg/core&lt;/code&gt; (zero &lt;code&gt;SharedArrayBuffer&lt;/code&gt;, zero COEP header conflict with YouTube iframes).&lt;/li&gt;
&lt;li&gt;Copy &lt;code&gt;ffmpeg-core.js&lt;/code&gt; and &lt;code&gt;ffmpeg-core.wasm&lt;/code&gt; into &lt;code&gt;public/ffmpeg/&lt;/code&gt; during &lt;code&gt;pnpm build&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Load them from a static classic Web Worker at &lt;code&gt;public/ffmpeg-worker.js&lt;/code&gt; using plain &lt;code&gt;importScripts('/ffmpeg/ffmpeg-core.js')&lt;/code&gt;. Webpack never touches it, and retime runs locally in the user's browser in a few hundred milliseconds with zero server CPU cost.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  4. A subtle YouTube embed trap on &lt;code&gt;127.0.0.1&lt;/code&gt;
&lt;/h2&gt;

&lt;p&gt;During local QA, one Big Think clip (&lt;code&gt;sX2xsAzvm90&lt;/code&gt;) showed "This video is unavailable" inside our player even though its YouTube metadata clearly reported &lt;code&gt;playableInEmbed: true&lt;/code&gt; and &lt;code&gt;isFamilySafe: true&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;We tested the exact same embed across six &lt;code&gt;Referer&lt;/code&gt; values:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Referer&lt;/th&gt;
&lt;th&gt;YouTube Embed Response&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;http://127.0.0.1:3005/&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;UNPLAYABLE&lt;/code&gt; (Error 150, &lt;code&gt;auth&lt;/code&gt;)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;https://127.0.0.1:3005/&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;UNPLAYABLE&lt;/code&gt; (Error 150, &lt;code&gt;auth&lt;/code&gt;)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;http://localhost:3005/&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;OK&lt;/code&gt; (&lt;code&gt;isPlayable: true&lt;/code&gt;)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;https://pastnoise.com/&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;OK&lt;/code&gt; (&lt;code&gt;isPlayable: true&lt;/code&gt;)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Certain partner channels reject bare IP referrers (&lt;code&gt;127.0.0.1&lt;/code&gt;) at the embed edge while happily playing on &lt;code&gt;localhost&lt;/code&gt; and real HTTPS domains. If you ever hit mysterious Error 150s during local development, check whether you opened &lt;code&gt;127.0.0.1&lt;/code&gt; instead of &lt;code&gt;localhost&lt;/code&gt; before assuming the video is blocked.&lt;/p&gt;

&lt;h2&gt;
  
  
  Open index and live archive
&lt;/h2&gt;

&lt;p&gt;We also published an open JSON and Markdown index of our first 100 bilingual dossiers across 8 pillars on GitHub: &lt;a href="https://github.com/nonacosa/pastnoise-spoken-knowledge-index" rel="noopener noreferrer"&gt;nonacosa/pastnoise-spoken-knowledge-index&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;You can browse the full live archive by topic at &lt;a href="https://pastnoise.com/en/topic/tech" rel="noopener noreferrer"&gt;PastNoise Topic Hubs&lt;/a&gt; or test a custom question directly on &lt;a href="https://pastnoise.com/en" rel="noopener noreferrer"&gt;pastnoise.com&lt;/a&gt; (Chinese bilingual edition at &lt;a href="https://pastnoise.com/zh" rel="noopener noreferrer"&gt;pastnoise.com/zh&lt;/a&gt;). Happy to answer questions on the Elasticsearch index layout or the static &lt;code&gt;ffmpeg.wasm&lt;/code&gt; worker setup.&lt;/p&gt;

</description>
      <category>performance</category>
      <category>search</category>
      <category>softwaredevelopment</category>
    </item>
    <item>
      <title>Building Cross-Lingual Semantic Video Search Over 200M+ Movie and YouTube Lines</title>
      <dc:creator>nonacosa</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:10:21 +0000</pubDate>
      <link>https://dev.to/_cda735563962186822ba/building-cross-lingual-semantic-video-search-over-200m-movie-and-youtube-lines-1hib</link>
      <guid>https://dev.to/_cda735563962186822ba/building-cross-lingual-semantic-video-search-over-200m-movie-and-youtube-lines-1hib</guid>
      <description>&lt;p&gt;Most video clip search engines only match exact English strings. That breaks down the moment you remember the gist of a scene, want a specific emotional beat for an edit, or think of an idea in Chinese, Japanese, or Spanish rather than English.&lt;/p&gt;

&lt;p&gt;That limitation is what pushed our team to build &lt;a href="https://winklingo.com/en?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;WinkLingo&lt;/a&gt;, an all-in-one platform for searching and exploring real spoken English across &lt;strong&gt;1,466 feature films (1.8M+ synced clips)&lt;/strong&gt; and &lt;strong&gt;485,000+ YouTube videos (over 200M spoken lines)&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why String Matching Falls Short for Real Video Discovery
&lt;/h2&gt;

&lt;p&gt;Exact substring matching works when you already know the exact wording of a quote. It falls apart across three everyday scenarios:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Video editors and film fans&lt;/strong&gt; looking for a reaction moment, tonal match, or YouTube channel covering a specific theme without scrubbing through hours of footage.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Global language learners&lt;/strong&gt; who know what they want to express in their native language, but need to hear how native speakers actually phrase and deliver it in real dialogue.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Writers and researchers&lt;/strong&gt; trying to recover a half-remembered film line or compare subtle differences in register across multiple contexts.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;To bridge that gap, we ran large-scale GPU pipelines to extract fine-grained cross-lingual semantic vectors and time-synced indices across hundreds of millions of spoken lines. Instead of forcing every query into exact English keywords, users can type a quote, a loose concept, or an everyday situation in virtually any native language and jump directly to the spoken timestamp with word-level synced subtitles.&lt;/p&gt;

&lt;h2&gt;
  
  
  Five Workflows Connected in One Browser Tab
&lt;/h2&gt;

&lt;p&gt;Rather than splitting discovery, dictionary lookup, and deep listening across five separate tabs or browser extensions, we connected the full loop directly in desktop and mobile browsers (with core search, chapters, and monologues open to try without an account):&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Cross-Lingual Semantic Search &amp;amp; Themed Scene Discovery
&lt;/h3&gt;

&lt;p&gt;Search by exact phrase or meaning across films and YouTube together, or browse curated real-conversation collections such as &lt;a href="https://winklingo.com/en/discover/small-talk?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;Movie Small Talk &amp;amp; Everyday Scenes&lt;/a&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Tap-to-Lookup Oxford Dictionary &amp;amp; Multi-Sense Distributions
&lt;/h3&gt;

&lt;p&gt;Click any word inside synced movie or YouTube subtitles for instant Oxford definitions and pronunciation, alongside how native speakers actually split meanings across real films (for example, see &lt;a href="https://winklingo.com/en/clip/whatever-you-say?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;real movie contexts for "whatever you say"&lt;/a&gt;).&lt;/p&gt;

&lt;h3&gt;
  
  
  3. 3,354 Four-Minute Movie Chapters &amp;amp; 20,730 Monologues
&lt;/h3&gt;

&lt;p&gt;Three-second clips help you catch a phrase, but real listening fluency requires multi-minute context. Every short hit links into guided four-minute scene breakdowns with idiom notes and connected-speech cues (such as &lt;a href="https://winklingo.com/en/insight/the-intern-2015/a-little-me-time?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;this scene chapter from &lt;em&gt;The Intern&lt;/em&gt; (2015)&lt;/a&gt;) plus &lt;a href="https://winklingo.com/en/insight/monologues?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;20,730 movie monologues&lt;/a&gt; built for line-by-line shadowing.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Video-Backed Spaced Review Flashcards
&lt;/h3&gt;

&lt;p&gt;Any spoken line or phrase can be saved directly into interactive video flashcards that keep the original actor audio and surrounding scene context, so review sessions reinforce real cadence rather than isolated text.&lt;/p&gt;

&lt;h3&gt;
  
  
  5. Native Multilingual Interfaces From Day One
&lt;/h3&gt;

&lt;p&gt;Because semantic retrieval works across native languages, we built dedicated multilingual portals and bilingual study views for &lt;a href="https://winklingo.com/zh?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;Chinese (/zh)&lt;/a&gt;, &lt;a href="https://winklingo.com/ja?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;Japanese (/ja)&lt;/a&gt;, and &lt;a href="https://winklingo.com/es?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;Spanish (/es)&lt;/a&gt; alongside English.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try It &amp;amp; Share Feedback
&lt;/h2&gt;

&lt;p&gt;You can test exact lookup, scene chapters, and monologues directly at &lt;a href="https://winklingo.com/en?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=semantic_video_search" rel="noopener noreferrer"&gt;winklingo.com/en&lt;/a&gt;. If you work on video search, multilingual embeddings, or editing workflows, we would love to hear what queries trip up semantic clip retrieval in your experience.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>search</category>
      <category>webdev</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Indexing 206M+ Podcast Subtitle Lines and Pairing Them with PubMed &amp; Wikipedia Citations</title>
      <dc:creator>nonacosa</dc:creator>
      <pubDate>Sat, 21 Nov 2026 07:37:32 +0000</pubDate>
      <link>https://dev.to/_cda735563962186822ba/indexing-206m-podcast-subtitle-lines-and-pairing-them-with-pubmed-wikipedia-citations-2ikk</link>
      <guid>https://dev.to/_cda735563962186822ba/indexing-206m-podcast-subtitle-lines-and-pairing-them-with-pubmed-wikipedia-citations-2ikk</guid>
      <description>&lt;p&gt;When I research a topic through long-form podcasts or university lectures (say, sleep architecture, creatine for cognition, or semiconductor supply chains), I keep hitting two annoying walls:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Searching YouTube drops you into three-hour episodes full of sponsor reads and detours.&lt;/li&gt;
&lt;li&gt;Dropping a link into an LLM summarizer strips away the speaker's actual tone, caveats, and source provenance.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Over the past few weeks I built &lt;a href="https://pastnoise.com" rel="noopener noreferrer"&gt;PastNoise&lt;/a&gt; to scratch that exact itch. Instead of generating another text summary, it searches across 206M+ timestamped spoken subtitle lines and stitches 5 to 6 non-duplicate primary speakers into a single continuous screening room, paired with side-by-side Wikipedia, PubMed, Crossref, and OpenAlex citations.&lt;/p&gt;

&lt;h3&gt;
  
  
  Three Small Engineering Choices That Made the Biggest Difference
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;1. Filtering sponsor reads before picking clips (not after)&lt;/strong&gt;&lt;br&gt;
Early builds looked fine until a flagship topic pulled two Hostinger and Tonal ad reads from top podcasts. Moving a strict sponsor/outro detector ahead of act selection and hydrating 16 candidate windows first (then backfilling clean distinct channels) dropped sponsor contamination to zero across all 58 indexed screenings.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Rendering full bilingual transcripts in SSR DOM with &lt;code&gt;#act-X-t-Y&lt;/code&gt; citation anchors&lt;/strong&gt;&lt;br&gt;
Single-page tabs that unmount transcript DOM hide 95% of spoken text from search engines and make it impossible to quote a speaker accurately. Keeping every chapter's timestamped transcript inside lightweight &lt;code&gt;&amp;lt;details&amp;gt;&lt;/code&gt; accordions with click-to-copy citation links lets readers jump straight to the exact second and verify the primary quote.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Preserving original subtitle sequence IDs (&lt;code&gt;origSeq&lt;/code&gt;) when splitting long cues&lt;/strong&gt;&lt;br&gt;
Auto-captions often pack 220+ characters into one cue. Splitting cues down to readable 96-character lines broke pre-translated subtitle lookups whenever the original sequence index passed 100. Keeping an explicit &lt;code&gt;origSeq&lt;/code&gt; field alongside sequential display IDs solved alignment cleanly for English, Simplified Chinese, and Traditional Chinese (via OpenCC at zero extra LLM cost).&lt;/p&gt;

&lt;h3&gt;
  
  
  Open Assets You Can Reuse (CC BY 4.0)
&lt;/h3&gt;

&lt;p&gt;If you prefer local notes or want clean benchmark data for citation-grounded RAG experiments, I exported the entire 58-topic curated archive under CC BY 4.0:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Obsidian Knowledge Vault (58 dossiers + 25 Thinker MOCs)&lt;/strong&gt;: &lt;a href="https://github.com/nonacosa/pastnoise-obsidian-vault" rel="noopener noreferrer"&gt;github.com/nonacosa/pastnoise-obsidian-vault&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hugging Face Dataset (491 ad-filtered clips with YouTube timestamps + DOIs)&lt;/strong&gt;: &lt;a href="https://huggingface.co/datasets/pkwenda/pastnoise-primary-citations" rel="noopener noreferrer"&gt;huggingface.co/datasets/pkwenda/pastnoise-primary-citations&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Free Hydration Utility&lt;/strong&gt;: &lt;a href="https://pastnoise.com/tools/galpin-equation-calculator" rel="noopener noreferrer"&gt;Galpin Equation Hydration Calculator&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Full Interactive Screening Room (free without login)&lt;/strong&gt;: &lt;a href="https://pastnoise.com" rel="noopener noreferrer"&gt;pastnoise.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Curious how other builders here handle long-form audio/video retrieval and hallucination checks without losing the original speaker's voice.&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>nextjs</category>
      <category>ai</category>
      <category>webdev</category>
    </item>
    <item>
      <title>Counting Films, Not Tokens: What Indexing 1,466 Movies Taught Us About Bilingual Phrase Dictionaries</title>
      <dc:creator>nonacosa</dc:creator>
      <pubDate>Fri, 20 Nov 2026 10:25:21 +0000</pubDate>
      <link>https://dev.to/_cda735563962186822ba/counting-films-not-tokens-what-indexing-1466-movies-taught-us-about-bilingual-phrase-dictionaries-3k16</link>
      <guid>https://dev.to/_cda735563962186822ba/counting-films-not-tokens-what-indexing-1466-movies-taught-us-about-bilingual-phrase-dictionaries-3k16</guid>
      <description>&lt;p&gt;Most vocabulary frequency lists lie to language learners for a simple mathematical reason: they count tokens instead of sources.&lt;/p&gt;

&lt;p&gt;If a courtroom thriller repeats a legal phrase thirty-five times across ninety minutes, a raw token counter ranks that expression alongside everyday speech. A learner memorizes it, walks into a coffee shop or watches a different movie, and never hears it again.&lt;/p&gt;

&lt;p&gt;When we built the spoken-phrase index for &lt;a href="https://winklingo.com/en/data/most-common-movie-phrases" rel="noopener noreferrer"&gt;WinkLingo&lt;/a&gt;, we decided to measure phrase prevalence by counting distinct films rather than raw occurrences. Today the public library indexes &lt;code&gt;1,466&lt;/code&gt; feature films, &lt;code&gt;1,819,649&lt;/code&gt; time-synced dialogue clips, &lt;code&gt;3,354&lt;/code&gt; interactive scene chapters, &lt;code&gt;20,730&lt;/code&gt; monologues, and &lt;code&gt;485,620&lt;/code&gt; YouTube videos containing &lt;code&gt;202,967,728&lt;/code&gt; spoken lines. Across that corpus, &lt;code&gt;7,129&lt;/code&gt; multi-word expressions meet our indexability threshold (with &lt;code&gt;6,890&lt;/code&gt; live in the current public ranking snapshot), and single words or half-formed fragments are left out completely.&lt;/p&gt;

&lt;p&gt;Switching the sorting unit from "lines spoken" to "distinct films" fixed genre skew almost overnight. It also exposed two harder engineering problems when we rolled out Spanish (&lt;code&gt;EN -&amp;gt; ES&lt;/code&gt;) meanings across all &lt;code&gt;7,129&lt;/code&gt; expressions: single-gloss dictionary collapse and Latin-script validation blind spots.&lt;/p&gt;

&lt;p&gt;Here is what the numbers actually look like, where flat translations fail learners, and what we had to change in the validation pipeline.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why One Film Equals One Vote
&lt;/h2&gt;

&lt;p&gt;Our ranking query uses a plain rule: a phrase gets one point per distinct film whose dialogue contains it. Repeating &lt;code&gt;"you know"&lt;/code&gt; twenty times inside one character's monologue still counts as one film. Ties break by how many of those lines have been verified by a human or model check, then alphabetically by phrase name.&lt;/p&gt;

&lt;p&gt;You can see the spread across grammatical categories in the full index of &lt;code&gt;7,129&lt;/code&gt; expressions:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Category&lt;/th&gt;
&lt;th&gt;Indexable Phrases&lt;/th&gt;
&lt;th&gt;Calibrated Against Senses&lt;/th&gt;
&lt;th&gt;Calibrated Share&lt;/th&gt;
&lt;th&gt;Top Film Count&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Idioms (&lt;code&gt;idiom&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;3,872&lt;/td&gt;
&lt;td&gt;2,315&lt;/td&gt;
&lt;td&gt;59.8%&lt;/td&gt;
&lt;td&gt;1,418&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Everyday expressions (&lt;code&gt;expression&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;1,158&lt;/td&gt;
&lt;td&gt;693&lt;/td&gt;
&lt;td&gt;59.8%&lt;/td&gt;
&lt;td&gt;837&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Phrasal verbs (&lt;code&gt;phrasal_verb&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;1,063&lt;/td&gt;
&lt;td&gt;711&lt;/td&gt;
&lt;td&gt;66.9%&lt;/td&gt;
&lt;td&gt;1,423&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Polite formulas (&lt;code&gt;formula&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;561&lt;/td&gt;
&lt;td&gt;460&lt;/td&gt;
&lt;td&gt;82.0%&lt;/td&gt;
&lt;td&gt;1,234&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Slang (&lt;code&gt;slang&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;212&lt;/td&gt;
&lt;td&gt;148&lt;/td&gt;
&lt;td&gt;69.8%&lt;/td&gt;
&lt;td&gt;708&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reactions (&lt;code&gt;attitude&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;168&lt;/td&gt;
&lt;td&gt;121&lt;/td&gt;
&lt;td&gt;72.0%&lt;/td&gt;
&lt;td&gt;866&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Proverbs (&lt;code&gt;proverb&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;95&lt;/td&gt;
&lt;td&gt;49&lt;/td&gt;
&lt;td&gt;51.6%&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;7,129&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;4,497&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;63.1%&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;1,423&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Notice that proverbs top out at just &lt;code&gt;21&lt;/code&gt; films (&lt;em&gt;enough is enough&lt;/em&gt;), whereas phrasal verbs, idioms, and polite formulas regularly clear &lt;code&gt;1,200&lt;/code&gt; films out of &lt;code&gt;1,466&lt;/code&gt;. In the &lt;a href="https://winklingo.com/es/data/most-common-movie-phrases" rel="noopener noreferrer"&gt;live Spanish ranking table&lt;/a&gt;, the top five expressions look like spoken glue rather than textbook grammar:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;come on&lt;/code&gt; appears in &lt;code&gt;1,423&lt;/code&gt; films (&lt;code&gt;21,186&lt;/code&gt; raw lines; &lt;code&gt;14,645&lt;/code&gt; verified lines).&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;you know&lt;/code&gt; appears in &lt;code&gt;1,418&lt;/code&gt; films (&lt;code&gt;15,623&lt;/code&gt; raw lines).&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;see you&lt;/code&gt; appears in &lt;code&gt;1,342&lt;/code&gt; films (&lt;code&gt;6,812&lt;/code&gt; raw lines).&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;go on&lt;/code&gt; appears in &lt;code&gt;1,341&lt;/code&gt; films (&lt;code&gt;6,199&lt;/code&gt; raw lines).&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;get it&lt;/code&gt; appears in &lt;code&gt;1,297&lt;/code&gt; films (&lt;code&gt;6,172&lt;/code&gt; raw lines).&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Ranking by distinct films solves the frequency distortion, yet looking at those top five rows immediately surfaced an embarrassing product bug.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Single-Gloss Trap in Bilingual Dictionaries
&lt;/h2&gt;

&lt;p&gt;If you store one summary gloss per phrase (&lt;code&gt;definition_es&lt;/code&gt; on the parent row) and display it beside a movie rank, polysemous expressions turn into traps.&lt;/p&gt;

&lt;p&gt;Out of &lt;code&gt;5,105&lt;/code&gt; indexable phrases mapped to active dictionary senses (&lt;code&gt;7,210&lt;/code&gt; active senses in &lt;code&gt;dict_senses&lt;/code&gt;, average &lt;code&gt;1.4&lt;/code&gt; senses per mapped phrase), &lt;code&gt;3,731&lt;/code&gt; map to one sense. The remaining &lt;code&gt;1,374&lt;/code&gt; phrases (&lt;code&gt;26.9%&lt;/code&gt;) split across &lt;code&gt;2&lt;/code&gt; to &lt;code&gt;12&lt;/code&gt; distinct meanings in real film dialogue. Specifically, &lt;code&gt;150&lt;/code&gt; phrases carry &lt;code&gt;4&lt;/code&gt; or more senses, and &lt;code&gt;45&lt;/code&gt; phrases carry &lt;code&gt;5&lt;/code&gt; or more.&lt;/p&gt;

&lt;p&gt;Because early dictionary pipelines often write a literal or textbook first meaning onto the parent phrase record, the flat Spanish gloss regularly disagrees with how screenwriters and actors actually use the phrase:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Expression&lt;/th&gt;
&lt;th&gt;Distinct Films&lt;/th&gt;
&lt;th&gt;Mapped Senses&lt;/th&gt;
&lt;th&gt;Flat Parent Gloss (&lt;code&gt;definition_es&lt;/code&gt;)&lt;/th&gt;
&lt;th&gt;Dominant Cinema Sense (Verified Film Share)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;come on&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;1,423&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;td&gt;salir a escena&lt;/td&gt;
&lt;td&gt;"¡venga!; ¡vamos!; ¡date prisa!; ¡esfuérzate más!" (75.0%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;get it&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;1,297&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;ser castigado o regañado&lt;/td&gt;
&lt;td&gt;comprender un chiste, un punto, una situación (65.0%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;take care&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;909&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;cuídate; adiós&lt;/td&gt;
&lt;td&gt;encargarse de un problema, tarea o persona para resolverlo (55.0%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;get out of&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;229&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;librarse de algo&lt;/td&gt;
&lt;td&gt;salir de; irse de un lugar (85.0%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;i beg your pardon&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;133&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;Le pido disculpas&lt;/td&gt;
&lt;td&gt;¿cómo dice? (muestra ofensa por lo dicho o el modo) (45.0%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;break down&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;125&lt;/td&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;averiarse&lt;/td&gt;
&lt;td&gt;derrumbarse y echarse a llorar (41.7%) · echar abajo (16.7%)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Look at &lt;code&gt;come on&lt;/code&gt; at rank #1. Literally, "come on" can mean an actor walking onto a stage (&lt;em&gt;salir a escena&lt;/em&gt;). In &lt;code&gt;75%&lt;/code&gt; of sense-mapped film appearances, though, characters are urging someone on (&lt;em&gt;¡vamos!&lt;/em&gt; / &lt;em&gt;¡date prisa!&lt;/em&gt;). Meanwhile &lt;code&gt;get it&lt;/code&gt; at rank #5 had &lt;em&gt;ser castigado o regañado&lt;/em&gt; ("get scolded") sitting on the summary field even though &lt;code&gt;65%&lt;/code&gt; of mapped film clips use it to mean "understand a joke or point."&lt;/p&gt;

&lt;p&gt;On &lt;a href="https://winklingo.com/en/clip/break-down" rel="noopener noreferrer"&gt;&lt;code&gt;break down&lt;/code&gt;&lt;/a&gt; (&lt;code&gt;125&lt;/code&gt; films, &lt;code&gt;148&lt;/code&gt; total lines, &lt;code&gt;109&lt;/code&gt; verified lines), the distortion cuts two ways. The flat parent gloss &lt;em&gt;averiarse&lt;/em&gt; ("for a machine or car to break down") accounts for &lt;code&gt;1&lt;/code&gt; out of &lt;code&gt;12&lt;/code&gt; sense-mapped films (&lt;code&gt;8.3%&lt;/code&gt;). Emotional collapse (&lt;em&gt;derrumbarse; echarse a llorar&lt;/em&gt;) leads at &lt;code&gt;41.7%&lt;/code&gt; (&lt;code&gt;5&lt;/code&gt; films), followed by physically smashing something open (&lt;em&gt;echar abajo&lt;/em&gt;, &lt;code&gt;16.7%&lt;/code&gt;, &lt;code&gt;2&lt;/code&gt; films, like the police breaking down the door in &lt;em&gt;Trainspotting&lt;/em&gt;).&lt;/p&gt;

&lt;p&gt;Even after mapping all &lt;code&gt;7&lt;/code&gt; senses, our API originally shipped with a hardcoded &lt;code&gt;SENSE_LIMIT = 4&lt;/code&gt;. Any calibrated phrase with &lt;code&gt;5&lt;/code&gt; to &lt;code&gt;12&lt;/code&gt; meanings silently dropped its fifth through seventh senses, cutting off &lt;code&gt;25%&lt;/code&gt; of the corpus evidence on &lt;code&gt;break down&lt;/code&gt; alone. Raising &lt;code&gt;CALIBRATED_SENSE_LIMIT&lt;/code&gt; to &lt;code&gt;8&lt;/code&gt; for calibrated rows (while keeping &lt;code&gt;4&lt;/code&gt; for uncalibrated drafts) stopped that truncation, and updating our ranking UI to pull from calibrated senses rather than raw legacy glosses prevented misleading Spanish defaults like &lt;em&gt;salir a escena&lt;/em&gt; from appearing next to #1.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Spanish Learners Actually Hear in Day-0 Live Clips
&lt;/h2&gt;

&lt;p&gt;Data cleanup only matters if learners can inspect real speech rather than isolated glosses. In our current live Spanish sitemap (&lt;code&gt;576&lt;/code&gt; indexable URLs, including &lt;code&gt;299&lt;/code&gt; phrase clip pages and &lt;code&gt;247&lt;/code&gt; interactive scene breakdowns), we pulled &lt;code&gt;50&lt;/code&gt; high-utility expressions across five conversational domains to check how sense shares line up with real dialogue cues.&lt;/p&gt;

&lt;p&gt;Take boundary-setting and social friction. Textbooks teach polite apologies, whereas film dialogue leans heavily on direct pushback: &lt;a href="https://winklingo.com/es/clip/none-of-your-business" rel="noopener noreferrer"&gt;&lt;code&gt;none of your business&lt;/code&gt;&lt;/a&gt; shows up across &lt;code&gt;162&lt;/code&gt; films (&lt;code&gt;192&lt;/code&gt; clips, including &lt;em&gt;Dune: Part Two&lt;/em&gt; saying &lt;code&gt;"This is none of your business."&lt;/code&gt;), &lt;code&gt;leave me alone&lt;/code&gt; across &lt;code&gt;229&lt;/code&gt; films, and &lt;code&gt;i beg your pardon&lt;/code&gt; across &lt;code&gt;133&lt;/code&gt; films where &lt;code&gt;45%&lt;/code&gt; of clips express offended disbelief rather than a polite pardon.&lt;/p&gt;

&lt;p&gt;When a learner jumps from a single 2-second quote into the surrounding context via the &lt;a href="https://winklingo.com/en/movie-quote-finder" rel="noopener noreferrer"&gt;Movie Quote Finder&lt;/a&gt;, hearing three different actors stress &lt;code&gt;that's it&lt;/code&gt; (&lt;code&gt;866&lt;/code&gt; films, &lt;code&gt;5&lt;/code&gt; active senses: &lt;code&gt;40%&lt;/code&gt; calling a stop to something, &lt;code&gt;20%&lt;/code&gt; expressing disappointment, &lt;code&gt;20%&lt;/code&gt; confirming someone is right) does more work than memorizing a bilingual flashcard table.&lt;/p&gt;

&lt;h2&gt;
  
  
  Validating Spanish Strings Without Breaking Short Glosses
&lt;/h2&gt;

&lt;p&gt;Shipping Japanese localization earlier had given us an easy script-detection heuristic: valid Japanese prose almost always contains hiragana or katakana (&lt;code&gt;[\u3040-\u30ff]&lt;/code&gt;), and leaked simplified Chinese characters fail &lt;code&gt;windows-31j&lt;/code&gt; encoding checks. If an API payload slipped back into English or simplified Chinese, a single character-class check caught it cleanly.&lt;/p&gt;

&lt;p&gt;Spanish broke that assumption because English and Spanish share the Latin alphabet. When we wired up &lt;code&gt;5,443&lt;/code&gt; Spanish AI guides (&lt;code&gt;ai_phrase_guide&lt;/code&gt;), &lt;code&gt;5,556&lt;/code&gt; phrase explainer entries (&lt;code&gt;phrase-explainers&lt;/code&gt;), and &lt;code&gt;7,210&lt;/code&gt; Spanish sense definitions (&lt;code&gt;dict_senses&lt;/code&gt;), two opposite failure modes appeared within a single week.&lt;/p&gt;

&lt;p&gt;First, naive English stopword detectors rejected legitimate Spanish explanations. Any bilingual guide naturally quotes the target English phrase inside Spanish prose, such as explaining why &lt;code&gt;"none of your business"&lt;/code&gt; sounds blunt or contrasting &lt;code&gt;"break down the door"&lt;/code&gt; with &lt;code&gt;"break down in tears"&lt;/code&gt;. If your validator scans raw text for English words like &lt;code&gt;the&lt;/code&gt;, &lt;code&gt;of&lt;/code&gt;, &lt;code&gt;your&lt;/code&gt;, or &lt;code&gt;break&lt;/code&gt;, clean Spanish paragraphs fail QA.&lt;/p&gt;

&lt;p&gt;Our Java guide filter fixed this by masking short quoted snippets up to &lt;code&gt;8&lt;/code&gt; words inside guillemets (&lt;code&gt;«...»&lt;/code&gt;) or standard quotation marks (&lt;code&gt;"..."&lt;/code&gt; / &lt;code&gt;“...”&lt;/code&gt;) before counting language markers. After masking quotes, the filter rejects any run of more than &lt;code&gt;5&lt;/code&gt; consecutive English words, blocks stray CJK Han characters (&lt;code&gt;\u4e00-\u9fff&lt;/code&gt;) and kana, and accepts strings that either carry Spanish diacritics (&lt;code&gt;áéíóúüñ¿¡&lt;/code&gt;) or score higher on non-overlapping Spanish markers than English markers.&lt;/p&gt;

&lt;p&gt;Second, applying prose rules to short 1-to-3-word dictionary glosses caused false rejections in the frontend sitemap gate. Consider real Spanish definitions from our corpus:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;"fracasar; fallar"&lt;/code&gt; (no accents, no articles, zero Spanish punctuation)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;"amante"&lt;/code&gt; or &lt;code&gt;"aprovechar"&lt;/code&gt; or &lt;code&gt;"casi"&lt;/code&gt; (single unaccented words)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;"(in)suficiente"&lt;/code&gt; (dictionary prefix shorthand inside parentheses)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of those pass a diacritic check, and none contain Spanish articles like &lt;code&gt;el&lt;/code&gt; or &lt;code&gt;la&lt;/code&gt;. We had to strip parenthetical prefixes &lt;code&gt;(in)...&lt;/code&gt; before scanning, ignore short quoted English phrases, and match Spanish verb infinitives and inflectional suffixes (&lt;code&gt;-ando&lt;/code&gt;, &lt;code&gt;-iendo&lt;/code&gt;, &lt;code&gt;-ción&lt;/code&gt;, &lt;code&gt;-ar/-er/-ir&lt;/code&gt;, plus short &lt;code&gt;-o/-a/-e&lt;/code&gt; endings) whenever zero English-echo stopwords were present. We also caught stray fullwidth punctuation from upstream bilingual runs—for instance, a single fullwidth semicolon &lt;code&gt;；&lt;/code&gt; sitting inside an otherwise clean Spanish gloss (&lt;code&gt;"entendido；lo tengo"&lt;/code&gt;)—by banning the entire &lt;code&gt;\uff00-\uffef&lt;/code&gt; Unicode block alongside CJK ranges.&lt;/p&gt;

&lt;h2&gt;
  
  
  Practical Takeaways for Bilingual Corpus Builders
&lt;/h2&gt;

&lt;p&gt;If you are indexing subtitle corpora or building bilingual learner tools from movie dialogue, four rules saved us the most rework:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Count distinct source works rather than raw subtitle lines. One repetitive movie or sitcom episode will distort a token-ranked top 100 list.&lt;/li&gt;
&lt;li&gt;Store and serve sense distributions (&lt;code&gt;share&lt;/code&gt; per film) instead of a single summary translation on polysemous verbs and idioms. Once a phrase crosses two senses (&lt;code&gt;26.9%&lt;/code&gt; of our mapped index), a flat gloss misleads users on at least one major use case.&lt;/li&gt;
&lt;li&gt;Raise display limits above &lt;code&gt;4&lt;/code&gt; senses for calibrated entries. Cutting off at four hides real spoken usage on high-frequency phrasal verbs like &lt;code&gt;break down&lt;/code&gt; (&lt;code&gt;7&lt;/code&gt; senses) or &lt;code&gt;that's it&lt;/code&gt; (&lt;code&gt;5&lt;/code&gt; senses).&lt;/li&gt;
&lt;li&gt;Separate prose validators from short-gloss validators when two languages share an alphabet. Mask quoted target-language examples first, normalize parenthetical shorthands like &lt;code&gt;(in)suficiente&lt;/code&gt;, and ban fullwidth punctuation alongside CJK blocks.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We also packaged &lt;code&gt;50&lt;/code&gt; live Spanish-indexed expressions from this dataset—grouped across reactions, multi-sense phrasal verbs, social boundaries, turning points, and everyday urgency—into a 14-slide clickable PDF field guide with direct links to the synced movie scenes on &lt;a href="https://winklingo.com/es" rel="noopener noreferrer"&gt;winklingo.com/es&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>dataengineering</category>
      <category>nlp</category>
      <category>linguistics</category>
      <category>spanish</category>
    </item>
  </channel>
</rss>
