[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article:smartnews-recommendation":3},{"meta":4,"markdown":67,"quiz":68},{"type":5,"articleId":6,"slug":7,"title":8,"titleEn":9,"category":10,"summary":11,"publishedAt":12,"image":13,"vocabulary":14,"quizId":66},"article","tech-smartnews-recommendation","smartnews-recommendation","SmartNewsの推薦システム — ニュースを最適化する機械学習","SmartNews Recommendation: ML for Personalized News","tech","SmartNews's recommendation system — click prediction vs diversity tradeoff, story clustering for deduplication, breaking news detection, hate\u002Fmisinformation filtering, A\u002FB testing infrastructure, user feedback learning, and partnership royalty model with publishers.\n","2026-04-28T00:00:00Z","https:\u002F\u002Fimages.yamiyomi.com\u002Ftech-smartnews-recommendation.png",[15,20,24,28,32,37,41,45,49,53,57,61],{"word":16,"reading":17,"meaning":18,"level":19},"推薦","すいせん","recommendation","N1",{"word":21,"reading":22,"meaning":23,"level":19},"速報","そくほう","breaking news",{"word":25,"reading":26,"meaning":27,"level":19},"重複","ちょうふく","duplication",{"word":29,"reading":30,"meaning":31,"level":19},"多様性","たようせい","diversity",{"word":33,"reading":34,"meaning":35,"level":36},"予測","よそく","prediction","N2",{"word":38,"reading":39,"meaning":40,"level":19},"集約","しゅうやく","aggregation",{"word":42,"reading":43,"meaning":44,"level":19},"検出","けんしゅつ","detection",{"word":46,"reading":47,"meaning":48,"level":19},"誤情報","ごじょうほう","misinformation",{"word":50,"reading":51,"meaning":52,"level":19},"提携","ていけい","partnership",{"word":54,"reading":55,"meaning":56,"level":36},"配信","はいしん","distribution",{"word":58,"reading":59,"meaning":60,"level":19},"偏り","かたより","bias",{"word":62,"reading":63,"meaning":64,"level":65},"新聞社","しんぶんしゃ","newspaper company","N3","tech-smartnews-recommendation-quiz","\n::para\n[スマートニュース]{すまーとにゅーす:SmartNews}は[二〇一二年]{にせんじゅうにねん:2012:N5}に[創業]{そうぎょう:founded:N1}され、[現在]{げんざい:currently:N3}は[日本]{にほん:Japan:N5}と[米国]{べいこく:United States:N3}の[両]{りょう:both:N3}[市場]{しじょう:markets:N3}で[利用]{りよう:used:N3}されている[ニュース]{にゅーす:news}[配信]{はいしん:distribution:N3}アプリです。[一日]{いちにち:one day:N5}に[数]{すう:several:N3}[万]{まん:tens of thousands:N5}[件]{けん:cases:N3}の[記事]{きじ:articles:N3}が[世界]{せかい:world:N4}[中]{じゅう:throughout:N5}から[流れ込む]{ながれこむ:flow in:N3}[中]{なか:among:N5}、[利用者]{りようしゃ:user:N3}[一人]{ひとり:one person:N5}[一人]{ひとり:one person:N5}に[最適]{さいてき:optimal:N3}な[記事]{きじ:articles:N3}[二十]{にじゅう:20:N5}[件]{けん:items:N3}[程度]{ていど:approximately:N3}を[選び]{えらび:select:N3}[出す]{だす:out:N5}のが[推薦]{すいせん:recommendation:N1}システムの[役割]{やくわり:role:N3}です。[本]{ほん:this:N5}[記事]{きじ:article:N3}では[クリック]{くりっく:click}[予測]{よそく:prediction:N2}と[多様性]{たようせい:diversity:N3}の[両立]{りょうりつ:balancing:N3}、[速報]{そくほう:breaking news:N3}[検出]{けんしゅつ:detection:N1}、[新聞社]{しんぶんしゃ:newspaper companies:N4}との[提携]{ていけい:partnership:N1}など[全体]{ぜんたい:overall:N3}[像]{ぞう:picture:N2}を[整理]{せいり:organize:N1}します。\n\n#en\nSmartNews was founded in 2012 and is currently a news distribution app used in both the Japanese and US markets. With tens of thousands of articles flowing in from around the world per day, the recommendation system's role is to select roughly twenty optimal articles for each individual user. This article organizes the overall picture, covering click prediction and diversity balancing, breaking news detection, partnerships with newspaper companies, and more.\n::\n\n::heading\n[クリック]{くりっく:click}[予測]{よそく:prediction:N2}モデル\n\n#en\nClick Prediction Model\n::\n\n::para\n[推薦]{すいせん:recommendation:N1}システムの[基本]{きほん:basic:N1}[層]{そう:layer:N2}は「この[利用者]{りようしゃ:user:N3}はこの[記事]{きじ:article:N3}を[読む]{よむ:to read:N5}か[否か]{いなか:or not:N3}」を[予測]{よそく:predict:N2}する[二]{に:two:N5}[値]{ち:value:N3}[分類]{ぶんるい:classification:N3}モデルです。[利用者]{りようしゃ:user:N3}の[過去]{かこ:past:N3}の[閲覧]{えつらん:browsing:N1}[履歴]{りれき:history:N1}、[滞在]{たいざい:dwell:N1}[時間]{じかん:time:N5}、[時間]{じかん:time:N5}[帯]{たい:band:N2}、[端末]{たんまつ:device:N1}、[記事]{きじ:article:N3}の[カテゴリ]{かてごり:category}や[掲載]{けいさい:posted:N1}[媒体]{ばいたい:media outlet:N1}など、[多数]{たすう:many:N3}の[特徴量]{とくちょうりょう:features:N1}を[入力]{にゅうりょく:input:N4}に[勾配]{こうばい:gradient:N1}[ブースティング]{ぶーすてぃんぐ:boosting}や[深層]{しんそう:deep:N2}[学習]{がくしゅう:learning:N4}モデルが[使われ]{つかわれ:used:N4}ます。\n\n#en\nThe base layer of the recommendation system is a binary classification model that predicts \"will this user read this article or not.\" Many features are fed in — the user's past browsing history, dwell time, time of day, device, article category, the source media outlet — and gradient boosting or deep learning models are used.\n::\n\n::callout\n[クリック]{くりっく:click}されやすい[記事]{きじ:articles:N3}だけを[並べる]{ならべる:to line up:N2}と、ゴシップや[扇情]{せんじょう:sensational:N1}[的]{てき:-style:N4}な[見出し]{みだし:headlines:N5}に[偏り]{かたより:biased:N1}、[利用者]{りようしゃ:user:N3}の[長期]{ちょうき:long-term:N3}[満足]{まんぞく:satisfaction:N3}を[損なう]{そこなう:damage:N2}。\n\n#en\nLining up only articles likely to be clicked biases the feed toward gossip and sensational headlines, damaging long-term user satisfaction.\n::\n\n::heading\n[多様性]{たようせい:diversity:N3}との[両立]{りょうりつ:balancing:N3}\n\n#en\nBalancing With Diversity\n::\n\n::para\nクリック[予測]{よそく:prediction:N2}スコアだけで[並べ替える]{ならべかえる:to sort:N2}と、[同じ]{おなじ:same:N4}カテゴリの[記事]{きじ:articles:N3}（[例えば]{たとえば:for example:N3}スポーツや[芸能]{げいのう:entertainment:N2}）ばかりが[上位]{じょうい:top:N3}を[占める]{しめる:occupy:N2}[現象]{げんしょう:phenomenon:N2}が[起き]{おき:occur:N4}ます。スマートニュースは[政治]{せいじ:politics:N3}・[経済]{けいざい:economy:N3}・[国際]{こくさい:international:N3}・[科学]{かがく:science:N3}・[エンタメ]{えんため:entertainment}など[幅広い]{はばひろい:broad:N2}カテゴリから[均衡]{きんこう:balanced:N1}よく[配信]{はいしん:distribute:N3}するために、[再]{さい:re-:N2}[ランキング]{らんきんぐ:ranking}[段階]{だんかい:stage:N2}で[多様性]{たようせい:diversity:N3}[制約]{せいやく:constraint:N3}を[加え]{くわえ:add:N3}ます。「[同一]{どういつ:same:N4}カテゴリは[連続]{れんぞく:consecutively:N3}[三]{さん:three:N5}[件]{けん:items:N3}まで」「[同一]{どういつ:same:N4}[媒体]{ばいたい:media outlet:N1}は[一定]{いってい:certain:N3}[割合]{わりあい:ratio:N3}まで」といった[規則]{きそく:rules:N2}を[適用]{てきよう:apply:N3}し、[偏り]{かたより:bias:N1}を[抑え]{おさえ:suppress:N1}ます。\n\n#en\nSorting only by click prediction score causes a phenomenon where articles in the same category (for example sports or entertainment) dominate the top spots. To deliver a balanced mix from broad categories like politics, economy, international, science, and entertainment, SmartNews adds diversity constraints at the re-ranking stage. Rules like \"no more than three of the same category in a row\" or \"any one media outlet capped at a certain ratio\" are applied to suppress bias.\n::\n\n::heading\n[類似]{るいじ:similar:N3}[記事]{きじ:articles:N3}の[クラスタリング]{くらすたりんぐ:clustering}と[重複]{ちょうふく:duplication:N2}[排除]{はいじょ:elimination:N1}\n\n#en\nStory Clustering and Deduplication\n::\n\n::para\n[同じ]{おなじ:same:N4}[出来事]{できごと:event:N4}を[扱う]{あつかう:cover:N1}[記事]{きじ:articles:N3}が[各]{かく:each:N2}[新聞社]{しんぶんしゃ:newspaper company:N4}から[並行]{へいこう:in parallel:N2}に[投稿]{とうこう:posted:N1}されると、[利用者]{りようしゃ:user:N3}の[画面]{がめん:screen:N3}に[同じ]{おなじ:same:N4}ニュースが[何]{なん:how many:N5}[件]{けん:items:N3}も[並ぶ]{ならぶ:line up:N2}ことになります。スマートニュースは[本文]{ほんぶん:body text:N4}の[埋め込み]{うめこみ:embedding:N2}[ベクトル]{べくとる:vector}や[見出し]{みだし:headline:N5}の[類似]{るいじ:similarity:N3}[度]{ど:degree:N4}を[使って]{つかって:using:N4}[類似]{るいじ:similar:N3}[記事]{きじ:articles:N3}を[一]{ひと:one:N5}つの[クラスタ]{くらすた:cluster}に[集約]{しゅうやく:aggregate:N3}し、[代表]{だいひょう:representative:N3}[記事]{きじ:article:N3}[一]{いっ:one:N5}[件]{けん:item:N3}だけを[表示]{ひょうじ:display:N3}します。[残り]{のこり:remaining:N3}の[記事]{きじ:articles:N3}は「[他]{ほか:other:N3}の[報道]{ほうどう:reports:N3}」として[折りたたみ]{おりたたみ:fold:N3}[表示]{ひょうじ:display:N3}にします。\n\n#en\nWhen articles covering the same event are posted in parallel by each newspaper company, the same news ends up lined up multiple times on the user's screen. SmartNews uses body text embedding vectors and headline similarity to aggregate similar articles into a single cluster, displaying only one representative article. The remaining articles are shown collapsed as \"other reports.\"\n::\n\n::heading\n[速報]{そくほう:breaking news:N3}[検出]{けんしゅつ:detection:N1}\n\n#en\nBreaking News Detection\n::\n\n::para\n[地震]{じしん:earthquake:N2}・[政治]{せいじ:political:N3}[発表]{はっぴょう:announcement:N3}・[大]{おお:large:N5}[事件]{じけん:incident:N3}など[緊急]{きんきゅう:urgent:N1}[性]{せい:nature:N3}の[高い]{たかい:high:N5}[ニュース]{にゅーす:news}は、[通常]{つうじょう:normal:N3}の[推薦]{すいせん:recommendation:N1}フローを[飛ばして]{とばして:bypass:N3}[即座]{そくざ:immediately:N1}に[全]{ぜん:all:N3}[利用者]{りようしゃ:users:N3}に[配信]{はいしん:distribute:N3}する[必要]{ひつよう:need:N3}があります。スマートニュースは[短時間]{たんじかん:short time:N2}に[同じ]{おなじ:same:N4}[話題]{わだい:topic:N4}の[記事]{きじ:articles:N3}が[急増]{きゅうぞう:surge:N3}する[現象]{げんしょう:phenomenon:N2}や、[特定]{とくてい:specific:N3}[キーワード]{きーわーど:keywords}の[出現]{しゅつげん:appearance:N3}[頻度]{ひんど:frequency:N1}を[監視]{かんし:monitor:N1}して[速報]{そくほう:breaking news:N3}[性]{せい:nature:N3}を[検出]{けんしゅつ:detect:N1}します。[検出]{けんしゅつ:detected:N1}されるとプッシュ[通知]{つうち:notification:N4}や[専用]{せんよう:dedicated:N2}[枠]{わく:slot:N1}で[強制]{きょうせい:forcibly:N3}[的]{てき:-ly:N4}に[露出]{ろしゅつ:expose:N1}します。\n\n#en\nNews with high urgency — earthquakes, political announcements, major incidents — must bypass the normal recommendation flow and be distributed to all users immediately. SmartNews monitors phenomena where articles on the same topic surge in a short time, and the appearance frequency of specific keywords, to detect breaking-news characteristics. When detected, the news is forcibly exposed via push notifications or dedicated slots.\n::\n\n::heading\n[ヘイト]{へいと:hate}・[誤情報]{ごじょうほう:misinformation:N3}フィルタリング\n\n#en\nHate and Misinformation Filtering\n::\n\n::para\n[推薦]{すいせん:recommended:N1}される[記事]{きじ:articles:N3}が[差別]{さべつ:discriminatory:N3}[的]{てき:-style:N4}な[表現]{ひょうげん:expression:N3}を[含んだり]{ふくんだり:contain:N2}、[事実]{じじつ:facts:N3}に[反する]{はんする:contrary:N3}[誤]{ご:false:N3}[情報]{じょうほう:information:N3}を[広める]{ひろめる:to spread:N4}ことは[配信]{はいしん:distribution:N3}[事業]{じぎょう:business:N4}[者]{しゃ:operator:N4}の[信頼]{しんらい:trust:N3}を[根本]{こんぽん:fundamentally:N2}から[揺るがし]{ゆるがし:shake:N1}ます。スマートニュースは[機械]{きかい:machine:N2}[学習]{がくしゅう:learning:N4}による[自動]{じどう:automatic:N4}[判定]{はんてい:judgment:N3}と[人手]{ひとで:human:N4}の[審査]{しんさ:review:N1}を[組み合わせて]{くみあわせて:combine:N3}、[ヘイト]{へいと:hate}スピーチや[誤情報]{ごじょうほう:misinformation:N3}の[疑い]{うたがい:suspicion:N3}がある[記事]{きじ:articles:N3}を[配信]{はいしん:distribution:N3}[対象]{たいしょう:target:N2}から[除外]{じょがい:exclude:N3}します。[第三]{だいさん:third:N1}[者]{しゃ:party:N4}[ファクト]{ふぁくと:fact}[チェック]{ちぇっく:check}[機関]{きかん:organizations:N3}との[連携]{れんけい:collaboration:N1}も[行われて]{おこなわれて:carried out:N5}います。\n\n#en\nIf recommended articles contain discriminatory expressions or spread misinformation contrary to fact, the trust of the distribution operator is fundamentally shaken. SmartNews combines automatic judgment by machine learning with human review to exclude articles suspected of hate speech or misinformation from distribution. Collaboration with third-party fact-checking organizations is also carried out.\n::\n\n::heading\n[利用者]{りようしゃ:user:N3}フィードバックによる[学習]{がくしゅう:learning:N4}\n\n#en\nLearning From User Feedback\n::\n\n::para\n[各]{かく:each:N2}[記事]{きじ:article:N3}には「[興味]{きょうみ:interest:N1}なし」を[示す]{しめす:indicate:N3}[親指]{おやゆび:thumb:N3}[下]{した:down:N5}ボタンが[配置]{はいち:placed:N3}されており、[押された]{おされた:pressed:N3}[記事]{きじ:articles:N3}と[似た]{にた:similar:N3}[特徴]{とくちょう:characteristics:N1}を[持つ]{もつ:have:N4}[記事]{きじ:articles:N3}は[今後]{こんご:going forward:N5}その[利用者]{りようしゃ:user:N3}には[出にくく]{でにくく:less likely to appear:N5}なります。クリックしなかった[記事]{きじ:articles:N3}を[暗黙]{あんもく:implicit:N1}の[負]{ふ:negative:N3}[例]{れい:example:N3}とする[手法]{しゅほう:method:N3}と[組み合わせ]{くみあわせ:combined:N3}、[利用者]{りようしゃ:user:N3}の[嗜好]{しこう:preferences:N1}を[継続]{けいぞく:continuously:N1}[的]{てき:-ly:N4}に[更新]{こうしん:update:N3}します。\n\n#en\nEach article has a thumbs-down button placed to indicate \"not interested,\" and articles with characteristics similar to ones that were pressed become less likely to appear for that user going forward. Combined with methods that treat unclicked articles as implicit negative examples, the user's preferences are continuously updated.\n::\n\n::heading\nA\u002FBテスト[基盤]{きばん:infrastructure:N1}\n\n#en\nA\u002FB Testing Infrastructure\n::\n\n::para\n[新しい]{あたらしい:new:N4}モデルや[多様性]{たようせい:diversity:N3}[制約]{せいやく:constraint:N3}の[調整]{ちょうせい:adjustment:N1}を[本番]{ほんばん:production:N3}に[投入]{とうにゅう:deploy:N3}する[前]{まえ:before:N5}には、[必ず]{かならず:always:N3}A\u002FBテストで[既存]{きそん:existing:N1}モデルとの[差]{さ:difference:N3}を[計測]{けいそく:measure:N2}します。[評価]{ひょうか:evaluation:N1}[指標]{しひょう:metrics:N1}は[クリック]{くりっく:click}[率]{りつ:rate:N1}だけでなく、[滞在]{たいざい:dwell:N1}[時間]{じかん:time:N5}、[次の]{つぎの:next:N3}[セッション]{せっしょん:session}までの[復帰]{ふっき:return:N2}[率]{りつ:rate:N1}、[多様性]{たようせい:diversity:N3}[指標]{しひょう:metric:N1}、[親指]{おやゆび:thumb:N3}[下]{した:down:N5}[率]{りつ:rate:N1}など[多面]{ためん:multi-faceted:N3}[的]{てき:-ly:N4}に[評価]{ひょうか:evaluate:N1}します。クリックは[増えた]{ふえた:increased:N3}が[復帰]{ふっき:return:N2}[率]{りつ:rate:N1}が[落ちた]{おちた:dropped:N3}場合、[本当の]{ほんとうの:true:N3}[満足]{まんぞく:satisfaction:N3}は[下がっている]{さがっている:declining:N5}と[判断]{はんだん:judge:N3}します。\n\n#en\nBefore deploying a new model or diversity-constraint adjustment to production, the difference from the existing model is always measured via A\u002FB test. Evaluation metrics emphasize multiple facets — not only click rate, but dwell time, return rate to the next session, diversity metrics, and thumbs-down rate. If clicks increased but the return rate dropped, true satisfaction is judged to be declining.\n::\n\n::heading\n[新聞社]{しんぶんしゃ:newspaper companies:N4}との[提携]{ていけい:partnerships:N1}と[ロイヤリティ]{ろいやりてぃ:royalty}モデル\n\n#en\nNewspaper Partnerships and Royalty Model\n::\n\n::para\nスマートニュースは[各]{かく:each:N2}[新聞社]{しんぶんしゃ:newspaper companies:N4}・[出版]{しゅっぱん:publishing:N2}[社]{しゃ:companies:N4}と[正式]{せいしき:formal:N3}な[配信]{はいしん:distribution:N3}[契約]{けいやく:contracts:N1}を[結び]{むすび:conclude:N1}、[読まれた]{よまれた:read:N5}[記事]{きじ:articles:N3}[数]{すう:count:N3}や[広告]{こうこく:ad:N3}[収益]{しゅうえき:revenue:N1}に[応じて]{おうじて:according to:N1}[ロイヤリティ]{ろいやりてぃ:royalties}を[支払う]{しはらう:pay:N3}[仕組み]{しくみ:mechanism:N3}を[構築]{こうちく:built:N2}しています。これは[出版]{しゅっぱん:publisher:N2}[側]{がわ:side:N3}の[収益]{しゅうえき:revenue:N1}を[支える]{ささえる:support:N3}と[同時に]{どうじに:while also:N4}、[質]{しつ:quality:N4}の[高い]{たかい:high:N5}[記事]{きじ:articles:N3}を[継続]{けいぞく:continuously:N1}[的]{てき:-ly:N4}に[供給]{きょうきゅう:supply:N3}してもらうための[基盤]{きばん:foundation:N1}でもあります。[広告]{こうこく:ad:N3}や[サブスク]{さぶすく:subscription}に[頼らない]{たよらない:not relying:N3}[出版]{しゅっぱん:publishing:N2}[社]{しゃ:companies:N4}の[新しい]{あたらしい:new:N4}[収益]{しゅうえき:revenue:N1}[源]{げん:source:N1}を[作り出す]{つくりだす:create:N4}[試み]{こころみ:attempt:N4}と[位置付けられて]{いちづけられて:positioned:N3}います。\n\n#en\nSmartNews concludes formal distribution contracts with each newspaper and publishing company, and has built a mechanism that pays royalties according to the number of articles read and ad revenue. This both supports publisher-side revenue and serves as a foundation for continuously sourcing high-quality articles. It is positioned as an attempt to create a new revenue source for publishers that does not rely on ads or subscriptions.\n::\n\n::heading\n[低]{てい:low:N2}レイテンシでの[応答]{おうとう:response:N1}\n\n#en\nLow-Latency Response\n::\n\n::para\n[数]{すう:tens of:N3}[万]{まん:tens of thousands:N5}の[記事]{きじ:articles:N3}から[利用者]{りようしゃ:user:N3}[個別]{こべつ:individual:N2}に[二十]{にじゅう:20:N5}[件]{けん:items:N3}[程度]{ていど:approximately:N3}を[即座]{そくざ:immediately:N1}に[返す]{かえす:to return:N3}ためには、[二]{ふた:two:N5}[段階]{だんかい:stage:N2}[推薦]{すいせん:recommendation:N1}が[基本]{きほん:basic:N1}です。[第]{だい:nth:N1}[一]{いち:first:N5}[段階]{だんかい:stage:N2}で[新着]{しんちゃく:newly arrived:N4}や[人気]{にんき:popular:N5}や[利用者]{りようしゃ:user:N3}[履歴]{りれき:history:N1}に[基づいて]{もとづいて:based on:N1}[候補]{こうほ:candidates:N2}[数百]{すうひゃく:hundreds:N3}に[絞り]{しぼり:narrow down:N1}、[第]{だい:nth:N1}[二]{に:second:N5}[段階]{だんかい:stage:N2}で[重い]{おもい:heavy:N4}クリック[予測]{よそく:prediction:N2}モデルと[多様性]{たようせい:diversity:N3}[再]{さい:re-:N2}[ランキング]{らんきんぐ:ranking}を[適用]{てきよう:apply:N3}します。[全件]{ぜんけん:all items:N3}にモデルを[掛ける]{かける:to apply:N3}と[応答]{おうとう:response:N1}[時間]{じかん:time:N5}が[爆発]{ばくはつ:explode:N2}するため、[現実]{げんじつ:realistic:N3}[的]{てき:-ly:N4}な[計算]{けいさん:computation:N2}コストで[精度]{せいど:accuracy:N3}を[出す]{だす:to produce:N5}[工夫]{くふう:device:N3}が[施されて]{ほどこされて:applied:N1}います。\n\n#en\nTo instantly return roughly twenty articles per individual user from tens of thousands, two-stage recommendation is the basic approach. The first stage narrows candidates down to hundreds based on freshness, popularity, and user history; the second stage applies a heavy click-prediction model and diversity re-ranking. Applying the model to all items would explode response time, so devices for producing accuracy at realistic computational cost are applied.\n::\n\n::heading\nおわりに\n\n#en\nConclusion\n::\n\n::para\nスマートニュースの[推薦]{すいせん:recommendation:N1}は「[クリック]{くりっく:click}されやすい[記事]{きじ:articles:N3}を[出す]{だす:show:N5}」という[単純]{たんじゅん:simple:N2}な[最適化]{さいてきか:optimization:N3}にとどまらず、[多様性]{たようせい:diversity:N3}・[速報]{そくほう:breaking news:N3}[性]{せい:nature:N3}・[誤情報]{ごじょうほう:misinformation:N3}の[排除]{はいじょ:elimination:N1}・[出版]{しゅっぱん:publisher:N2}[社]{しゃ:companies:N4}との[提携]{ていけい:partnership:N1}を[同時]{どうじ:simultaneously:N4}に[考慮]{こうりょ:consider:N1}した[複合]{ふくごう:composite:N2}[システム]{しすてむ:system}です。[機械]{きかい:machine:N2}[学習]{がくしゅう:learning:N4}と[ジャーナリズム]{じゃーなりずむ:journalism}の[両方]{りょうほう:both:N3}の[視点]{してん:perspectives:N1}が[揃って]{そろって:together:N1}[初めて]{はじめて:for the first time:N3}、[利用者]{りようしゃ:user:N3}の[長期]{ちょうき:long-term:N3}[満足]{まんぞく:satisfaction:N3}を[支える]{ささえる:support:N3}[ニュース]{にゅーす:news}[体験]{たいけん:experience:N4}が[実現]{じつげん:realized:N3}します。\n\n#en\nSmartNews's recommendation is not limited to the simple optimization of \"showing articles that are likely to be clicked.\" It is a composite system that simultaneously considers diversity, breaking-news characteristics, misinformation exclusion, and publisher partnerships. Only when both machine learning and journalism perspectives come together is a news experience that supports long-term user satisfaction realized.\n::\n",{"id":66,"title":69,"titleEn":70,"topicPath":10,"questions":71},"テック確認テスト — SmartNews 推薦システム","Tech Quiz — SmartNews Recommendation System",[72,99,122,146,169,193],{"id":73,"articleId":6,"question":74,"options":77,"correctLabel":83,"explanation":94,"tags":97},"tech-smartnews-recommendation-quiz-q01",{"en":75,"jp":76},"What best explains why SmartNews's recommendation system does not sort solely by click-prediction score?","SmartNewsの[推薦]{すいせん:recommendation}システムが[クリック]{くりっく:click}[予測]{よそく:prediction}スコアだけで[並べ替えない]{ならべかえない:does not sort}[主な]{おもな:main}[理由]{りゆう:reason}として[最も]{もっとも:most}[適切]{てきせつ:appropriate}なものはどれか？",[78,82,86,90],{"label":79,"jp":80,"en":81},"ア","[計算]{けいさん:computation}コストが[高すぎて]{たかすぎて:too high}[応答]{おうとう:response}できないため。","Because computational cost is too high to respond.",{"label":83,"jp":84,"en":85},"イ","ゴシップや[扇情]{せんじょう:sensational}[的な]{てきな:-style}[見出し]{みだし:headlines}に[偏り]{かたより:biased}、[利用者]{りようしゃ:user}の[長期]{ちょうき:long-term}[満足]{まんぞく:satisfaction}を[損なう]{そこなう:damages}ため。","Because the feed becomes biased toward gossip and sensational headlines, damaging long-term user satisfaction.",{"label":87,"jp":88,"en":89},"ウ","[新聞社]{しんぶんしゃ:newspaper companies}との[契約]{けいやく:contract}で[禁止]{きんし:prohibited}されているため。","Because it is prohibited by contracts with newspaper companies.",{"label":91,"jp":92,"en":93},"エ","[米国]{べいこく:US}[市場]{しじょう:market}では[クリック]{くりっく:clicks}が[計測]{けいそく:measurable}できないため。","Because clicks cannot be measured in the US market.",{"en":95,"jp":96},"Sorting only by click prediction lets gossip and sensational articles dominate the top, lowering long-term satisfaction, so diversity constraints are added. (A) is a different topic handled by two-stage recommendation; (C) and (D) are factually wrong.","クリック[予測]{よそく:prediction}だけで[並べる]{ならべる:lining up}とゴシップや[扇情]{せんじょう:sensational}[的な]{てきな:-style}[記事]{きじ:articles}が[上位]{じょうい:top}を[占め]{しめ:occupy}、[長期]{ちょうき:long-term}[満足]{まんぞく:satisfaction}を[下げる]{さげる:lowers}ため、[多様性]{たようせい:diversity}[制約]{せいやく:constraint}が[加えられ]{くわえられ:added}ます。アは[二]{ふた:two}[段階]{だんかい:stage}[推薦]{すいせん:recommendation}で[対応]{たいおう:handled}される[別]{べつ:different}の[論点]{ろんてん:topic}、ウとエは[事実]{じじつ:factually}[誤り]{あやまり:wrong}です。",[98,31],"click-prediction",{"id":100,"articleId":6,"question":101,"options":104,"correctLabel":79,"explanation":117,"tags":120},"tech-smartnews-recommendation-quiz-q02",{"en":102,"jp":103},"What best describes the benefit of adding diversity constraints at the re-ranking stage?","[多様性]{たようせい:diversity}[制約]{せいやく:constraint}を[再]{さい:re-}[ランキング]{らんきんぐ:ranking}[段階]{だんかい:stage}で[加える]{くわえる:adding}ことの[利点]{りてん:benefit}として[最も]{もっとも:most}[適切]{てきせつ:appropriate}なものはどれか？",[105,108,111,114],{"label":79,"jp":106,"en":107},"[同一]{どういつ:same}カテゴリや[同一]{どういつ:same}[媒体]{ばいたい:media outlet}の[記事]{きじ:articles}が[上位]{じょうい:top}を[占めない]{しめない:do not occupy}よう[偏り]{かたより:bias}を[抑え]{おさえ:suppress}、[幅広い]{はばひろい:broad}カテゴリから[均衡]{きんこう:balanced}よく[配信]{はいしん:distribute}できる。","It suppresses bias so that articles of the same category or media outlet do not dominate the top, and enables balanced delivery from broad categories.",{"label":83,"jp":109,"en":110},"[転置]{てんち:inverted}インデックスを[使わず]{つかわず:without using}に[全]{ぜん:all}[記事]{きじ:articles}を[即座]{そくざ:immediately}に[並べ替えられる]{ならべかえられる:can be sorted}。","All articles can be instantly sorted without using an inverted index.",{"label":87,"jp":112,"en":113},"[新聞社]{しんぶんしゃ:newspaper companies}への[ロイヤリティ]{ろいやりてぃ:royalty}[支払い]{しはらい:payment}を[ゼロ]{ぜろ:zero}にできる。","Royalty payments to newspaper companies can be reduced to zero.",{"label":91,"jp":115,"en":116},"[速報]{そくほう:breaking news}[検出]{けんしゅつ:detection}を[完全]{かんぜん:fully}に[置き換えられる]{おきかえられる:can replace}。","It can fully replace breaking-news detection.",{"en":118,"jp":119},"Diversity constraints suppress bias via rules like 'no more than three of the same category in a row' or 'any one media outlet capped at a certain ratio.' (B) contradicts two-stage recommendation; (C) and (D) are separate topics from diversity constraints.","[多様性]{たようせい:diversity}[制約]{せいやく:constraint}は「[同一]{どういつ:same}カテゴリは[連続]{れんぞく:consecutively}[三]{さん:three}[件]{けん:items}まで」「[同一]{どういつ:same}[媒体]{ばいたい:media outlet}は[一定]{いってい:certain}[割合]{わりあい:ratio}まで」といった[規則]{きそく:rules}で[偏り]{かたより:bias}を[抑え]{おさえ:suppress}ます。イは[二]{ふた:two}[段階]{だんかい:stage}[推薦]{すいせん:recommendation}と[矛盾]{むじゅん:contradicts}、ウとエは[多様性]{たようせい:diversity}[制約]{せいやく:constraint}とは[別]{べつ:separate}の[論点]{ろんてん:topic}です。",[31,121],"re-ranking",{"id":123,"articleId":6,"question":124,"options":127,"correctLabel":79,"explanation":140,"tags":143},"tech-smartnews-recommendation-quiz-q03",{"en":125,"jp":126},"Which most accurately describes what SmartNews does to organize multiple articles covering the same event?","[同じ]{おなじ:same}[出来事]{できごと:event}を[扱う]{あつかう:cover}[複数]{ふくすう:multiple}の[記事]{きじ:articles}を[整理]{せいり:organize}するために、SmartNewsが[行う]{おこなう:performs}[処理]{しょり:processing}として[最も]{もっとも:most}[正確]{せいかく:accurate}な[説明]{せつめい:description}は？",[128,131,134,137],{"label":79,"jp":129,"en":130},"[本文]{ほんぶん:body text}[埋め込み]{うめこみ:embedding}[ベクトル]{べくとる:vector}や[見出し]{みだし:headline}[類似]{るいじ:similarity}[度]{ど:degree}で[類似]{るいじ:similar}[記事]{きじ:articles}を[一]{ひと:one}つの[クラスタ]{くらすた:cluster}に[集約]{しゅうやく:aggregate}し、[代表]{だいひょう:representative}[記事]{きじ:article}[一]{いっ:one}[件]{けん:item}だけ[表示]{ひょうじ:display}する。","Aggregates similar articles into one cluster using body-text embedding vectors and headline similarity, displaying only one representative article.",{"label":83,"jp":132,"en":133},"[最初]{さいしょ:first}に[投稿]{とうこう:posted}された[記事]{きじ:article}[以外]{いがい:other than}を[全て]{すべて:all}[削除]{さくじょ:delete}する。","Deletes all articles other than the one posted first.",{"label":87,"jp":135,"en":136},"[新聞社]{しんぶんしゃ:newspaper companies}に[手動]{しゅどう:manual}で[統合]{とうごう:integration}を[依頼]{いらい:request}する。","Manually requests newspaper companies to integrate the articles.",{"label":91,"jp":138,"en":139},"[全]{ぜん:all}[記事]{きじ:articles}をそのまま[並べる]{ならべる:lines up}ことで[多様性]{たようせい:diversity}を[最大]{さいだい:max}[化]{か:-ize}する。","Lines up all articles as-is to maximize diversity.",{"en":141,"jp":142},"SmartNews clusters similar articles using body-text embeddings and headline similarity, displays only one representative, and folds the rest as 'other reports.' (B) wrongly causes loss of other companies' articles; (C) is unrealistic; (D) is the opposite of deduplication.","SmartNewsは[本文]{ほんぶん:body text}の[埋め込み]{うめこみ:embedding}[ベクトル]{べくとる:vector}や[見出し]{みだし:headline}[類似]{るいじ:similarity}[度]{ど:degree}で[類似]{るいじ:similar}[記事]{きじ:articles}を[クラスタリング]{くらすたりんぐ:cluster}し、[代表]{だいひょう:representative}[一]{いっ:one}[件]{けん:item}だけを[表示]{ひょうじ:display}、[残り]{のこり:remaining}は「[他]{ほか:other}の[報道]{ほうどう:reports}」として[折りたたみ]{おりたたみ:fold}ます。イは[他社]{たしゃ:other companies}[記事]{きじ:articles}の[消失]{しょうしつ:loss}を[招き]{まねき:causes}[誤り]{あやまり:wrong}、ウは[非]{ひ:non-}[現実]{げんじつ:realistic}[的]{てき:-ly}、エは[重複]{ちょうふく:duplication}[排除]{はいじょ:elimination}と[逆]{ぎゃく:opposite}です。",[144,145],"clustering","deduplication",{"id":147,"articleId":6,"question":148,"options":151,"correctLabel":83,"explanation":164,"tags":167},"tech-smartnews-recommendation-quiz-q04",{"en":149,"jp":150},"Which best describes SmartNews's breaking-news detection mechanism?","SmartNewsの[速報]{そくほう:breaking news}[検出]{けんしゅつ:detection}の[仕組み]{しくみ:mechanism}に[関する]{かんする:about}[記述]{きじゅつ:description}として[最も]{もっとも:most}[適切]{てきせつ:appropriate}なものはどれか？",[152,155,158,161],{"label":79,"jp":153,"en":154},"[編集]{へんしゅう:editorial}[部]{ぶ:department}が[全件]{ぜんけん:all items}を[手動]{しゅどう:manually}でタグ[付け]{づけ:tagging}している。","An editorial department manually tags all items.",{"label":83,"jp":156,"en":157},"[短時間]{たんじかん:short time}に[同じ]{おなじ:same}[話題]{わだい:topic}の[記事]{きじ:articles}が[急増]{きゅうぞう:surge}する[現象]{げんしょう:phenomenon}や[特定]{とくてい:specific}[キーワード]{きーわーど:keywords}の[出現]{しゅつげん:appearance}[頻度]{ひんど:frequency}を[監視]{かんし:monitor}し、[検出]{けんしゅつ:detected}されると[通常]{つうじょう:normal}フローを[飛ばして]{とばして:bypassing}プッシュ[通知]{つうち:notifications}や[専用]{せんよう:dedicated}[枠]{わく:slot}で[全]{ぜん:all}[利用者]{りようしゃ:users}に[配信]{はいしん:distribute}する。","Monitors phenomena where same-topic articles surge in a short time and the appearance frequency of specific keywords; when detected, bypasses the normal flow and distributes to all users via push notifications or a dedicated slot.",{"label":87,"jp":159,"en":160},"[気象]{きしょう:weather}[庁]{ちょう:agency}APIだけを[用い]{もちい:using}、[他]{た:other}の[速報]{そくほう:breaking news}は[扱わない]{あつかわない:does not handle}。","Uses only the Meteorological Agency API and does not handle other breaking news.",{"label":91,"jp":162,"en":163},"[利用者]{りようしゃ:user}の[親指]{おやゆび:thumb}[下]{した:down}ボタンの[押下]{おうか:press}[数]{すう:count}が[増加]{ぞうか:increased}したかどうかで[判定]{はんてい:judges}する。","Judges based on whether the press count of the user's thumbs-down button has increased.",{"en":165,"jp":166},"Breaking-news characteristics are detected by monitoring phenomena where same-topic articles surge in a short time and the appearance frequency of specific keywords, then forcibly exposed to all users. (A) is unrealistic; (C) is too narrow; (D) is about negative-feedback learning and unrelated.","[速報]{そくほう:breaking news}[性]{せい:nature}は[同じ]{おなじ:same}[話題]{わだい:topic}の[記事]{きじ:articles}が[短時間]{たんじかん:short time}に[急増]{きゅうぞう:surge}する[現象]{げんしょう:phenomenon}や[特定]{とくてい:specific}[キーワード]{きーわーど:keywords}の[出現]{しゅつげん:appearance}[頻度]{ひんど:frequency}を[監視]{かんし:monitor}して[検出]{けんしゅつ:detect}し、[全]{ぜん:all}[利用者]{りようしゃ:users}に[強制]{きょうせい:forcibly}[的]{てき:-ly}[露出]{ろしゅつ:expose}します。アは[非]{ひ:non-}[現実]{げんじつ:realistic}[的]{てき:-ly}、ウは[範囲]{はんい:scope}が[狭すぎ]{せますぎ:too narrow}、エは[負]{ふ:negative}フィードバック[学習]{がくしゅう:learning}の[話]{はなし:topic}で[無関係]{むかんけい:unrelated}です。",[168,44],"breaking-news",{"id":170,"articleId":6,"question":171,"options":174,"correctLabel":79,"explanation":187,"tags":190},"tech-smartnews-recommendation-quiz-q05",{"en":172,"jp":173},"Which best describes how SmartNews handles user feedback (the thumbs-down button)?","SmartNewsの[利用者]{りようしゃ:user}フィードバック（[親指]{おやゆび:thumb}[下]{した:down}ボタン）の[扱い]{あつかい:treatment}として[最も]{もっとも:most}[適切]{てきせつ:appropriate}なものは？",[175,178,181,184],{"label":79,"jp":176,"en":177},"[押された]{おされた:pressed}[記事]{きじ:articles}と[似た]{にた:similar}[特徴]{とくちょう:characteristics}を[持つ]{もつ:have}[記事]{きじ:articles}を[今後]{こんご:going forward}[出にくく]{でにくく:less likely to appear}し、クリックしなかった[記事]{きじ:articles}を[暗黙]{あんもく:implicit}の[負]{ふ:negative}[例]{れい:example}として[継続]{けいぞく:continuously}[的に]{てきに:-ly}[嗜好]{しこう:preferences}を[更新]{こうしん:updates}する。","Articles with characteristics similar to pressed ones become less likely to appear, and unclicked articles are treated as implicit negative examples to continuously update preferences.",{"label":83,"jp":179,"en":180},"[親指]{おやゆび:thumb}[下]{した:down}ボタンは[統計]{とうけい:statistics}[専用]{せんよう:dedicated}で、[配信]{はいしん:distribution}には[一切]{いっさい:not at all}[反映]{はんえい:reflected}されない。","The thumbs-down button is for statistics only and is not reflected in distribution at all.",{"label":87,"jp":182,"en":183},"[押された]{おされた:pressed}[記事]{きじ:articles}を[全]{ぜん:all}[利用者]{りようしゃ:users}から[削除]{さくじょ:remove}する。","Removes pressed articles from all users.",{"label":91,"jp":185,"en":186},"[新聞社]{しんぶんしゃ:newspaper companies}への[ロイヤリティ]{ろいやりてぃ:royalty}[計算]{けいさん:calculation}にのみ[使用]{しよう:used}される。","Used only for calculating royalties to newspaper companies.",{"en":188,"jp":189},"Combining thumbs-down presses with implicit negative examples from unclicked articles continuously updates each user's preferences. (B) wrongly denies reflection in distribution; (C) confuses individual preferences with global removal; (D) is unrelated to royalties.","[親指]{おやゆび:thumb}[下]{した:down}ボタンの[押下]{おうか:press}と[クリック]{くりっく:click}しなかった[暗黙]{あんもく:implicit}[負]{ふ:negative}[例]{れい:example}を[組み合わせて]{くみあわせて:combining}、[利用者]{りようしゃ:user}[個別]{こべつ:individual}の[嗜好]{しこう:preferences}を[継続]{けいぞく:continuously}[的に]{てきに:-ly}[更新]{こうしん:updates}します。イは[配信]{はいしん:distribution}への[反映]{はんえい:reflection}を[否定]{ひてい:denies}し[誤り]{あやまり:wrong}、ウは[個別]{こべつ:individual}[嗜好]{しこう:preferences}と[全体]{ぜんたい:overall}[削除]{さくじょ:removal}を[混同]{こんどう:confuses}、エは[ロイヤリティ]{ろいやりてぃ:royalty}とは[無関係]{むかんけい:unrelated}です。",[191,192],"feedback","personalization",{"id":194,"articleId":6,"question":195,"options":198,"correctLabel":83,"explanation":211,"tags":214},"tech-smartnews-recommendation-quiz-q06",{"en":196,"jp":197},"Which best describes SmartNews's relationship with newspapers and publishing companies?","SmartNewsの[新聞社]{しんぶんしゃ:newspaper companies}・[出版]{しゅっぱん:publishing}[社]{しゃ:companies}との[関係]{かんけい:relationship}に[関する]{かんする:about}[記述]{きじゅつ:description}として[最も]{もっとも:most}[適切]{てきせつ:appropriate}なものは？",[199,202,205,208],{"label":79,"jp":200,"en":201},"[許可]{きょか:permission}を[得ず]{えず:without obtaining}に[記事]{きじ:articles}を[転載]{てんさい:reposting}し、[読まれた]{よまれた:read}[件数]{けんすう:count}を[非]{ひ:non-}[公開]{こうかい:disclosed}にしている。","Reposts articles without obtaining permission and keeps the read counts non-disclosed.",{"label":83,"jp":203,"en":204},"[正式]{せいしき:formal}な[配信]{はいしん:distribution}[契約]{けいやく:contracts}を[結び]{むすび:concludes}、[読まれた]{よまれた:read}[記事]{きじ:articles}[数]{すう:count}や[広告]{こうこく:ad}[収益]{しゅうえき:revenue}に[応じて]{おうじて:according to}[ロイヤリティ]{ろいやりてぃ:royalties}を[支払う]{しはらう:pays}[仕組み]{しくみ:mechanism}を[構築]{こうちく:built}している。","Concludes formal distribution contracts and has built a mechanism that pays royalties according to articles read and ad revenue.",{"label":87,"jp":206,"en":207},"[出版]{しゅっぱん:publishing}[社]{しゃ:companies}が[全]{ぜん:all}[記事]{きじ:articles}を[無償]{むしょう:for free}で[提供]{ていきょう:provide}することを[条件]{じょうけん:condition}としている。","Requires publishing companies to provide all articles for free.",{"label":91,"jp":209,"en":210},"[広告]{こうこく:ad}・[サブスク]{さぶすく:subscription}[収益]{しゅうえき:revenue}に[完全]{かんぜん:fully}に[依存]{いぞん:dependent}し、[出版]{しゅっぱん:publishing}[社]{しゃ:companies}[側]{がわ:side}には[一切]{いっさい:nothing at all}[還元]{かんげん:returned}しない。","Fully depends on ad and subscription revenue and returns nothing at all to publishing companies.",{"en":212,"jp":213},"SmartNews concludes formal distribution contracts with each newspaper, paying royalties according to read counts and ad revenue, and is positioned as an attempt to create a new revenue source for publishers. (A) contradicts the reality of the contracts; (C) and (D) are the opposite of the royalty model.","SmartNewsは[各]{かく:each}[新聞社]{しんぶんしゃ:newspaper companies}と[正式]{せいしき:formal}[配信]{はいしん:distribution}[契約]{けいやく:contract}を[結び]{むすび:concluding}、[読まれた]{よまれた:read}[件数]{けんすう:count}や[広告]{こうこく:ad}[収益]{しゅうえき:revenue}に[応じた]{おうじた:according to}[ロイヤリティ]{ろいやりてぃ:royalty}を[支払い]{しはらい:paying}、[出版]{しゅっぱん:publishing}[側]{がわ:side}の[新しい]{あたらしい:new}[収益]{しゅうえき:revenue}[源]{げん:source}を[作り出す]{つくりだす:creating}[試み]{こころみ:attempt}と[位置付けられて]{いちづけられて:positioned}います。アは[契約]{けいやく:contract}の[実態]{じったい:reality}と[反し]{はんし:contradicts}、ウとエは[ロイヤリティ]{ろいやりてぃ:royalty}モデルと[逆]{ぎゃく:opposite}です。",[215,216],"partnerships","royalty"]