[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article:line-llmops":3},{"meta":4,"markdown":137},{"type":5,"articleId":6,"slug":7,"title":8,"titleEn":9,"category":10,"summary":11,"publishedAt":12,"image":13,"vocabulary":14},"article","tech-line-llmops","line-llmops","LINEのLLMOps — 大規模言語モデル運用の取り組み","LINE's LLMOps — Operating Large Language Models at Scale","tech","An overview of LINE's LLMOps practice: HyperCLOVA Japanese-tuned models, GPU inference infrastructure with vLLM and batching, prompt management, evaluation pipelines, cost control, and RAG integration.\n","2026-04-27T00:00:00Z","https:\u002F\u002Fimages.yamiyomi.com\u002Ftech-line-llmops.png",[15,20,24,29,33,37,41,45,49,53,57,61,65,69,73,77,81,85,89,93,97,101,105,109,113,117,121,125,129,133],{"word":16,"reading":17,"meaning":18,"level":19},"大規模","だいきぼ","large-scale","N2",{"word":21,"reading":22,"meaning":23,"level":19},"運用","うんよう","operations",{"word":25,"reading":26,"meaning":27,"level":28},"推論","すいろん","inference","N1",{"word":30,"reading":31,"meaning":32,"level":19},"学習","がくしゅう","learning",{"word":34,"reading":35,"meaning":36,"level":28},"微調整","びちょうせい","fine-tuning",{"word":38,"reading":39,"meaning":40,"level":19},"評価","ひょうか","evaluation",{"word":42,"reading":43,"meaning":44,"level":28},"指標","しひょう","metric",{"word":46,"reading":47,"meaning":48,"level":19},"検索","けんさく","search",{"word":50,"reading":51,"meaning":52,"level":19},"拡張","かくちょう","extension",{"word":54,"reading":55,"meaning":56,"level":28},"埋め込み","うめこみ","embedding",{"word":58,"reading":59,"meaning":60,"level":28},"検索拡張生成","けんさくかくちょうせいせい","retrieval-augmented generation",{"word":62,"reading":63,"meaning":64,"level":19},"基盤","きばん","infrastructure",{"word":66,"reading":67,"meaning":68,"level":19},"構築","こうちく","construction",{"word":70,"reading":71,"meaning":72,"level":19},"計算","けいさん","computation",{"word":74,"reading":75,"meaning":76,"level":19},"資源","しげん","resources",{"word":78,"reading":79,"meaning":80,"level":19},"削減","さくげん","reduction",{"word":82,"reading":83,"meaning":84,"level":19},"効率","こうりつ","efficiency",{"word":86,"reading":87,"meaning":88,"level":19},"一括","いっかつ","batch",{"word":90,"reading":91,"meaning":92,"level":28},"並列","へいれつ","parallel",{"word":94,"reading":95,"meaning":96,"level":28},"量子化","りょうしか","quantization",{"word":98,"reading":99,"meaning":100,"level":19},"試行","しこう","trial",{"word":102,"reading":103,"meaning":104,"level":19},"配信","はいしん","distribution",{"word":106,"reading":107,"meaning":108,"level":19},"承認","しょうにん","approval",{"word":110,"reading":111,"meaning":112,"level":19},"履歴","りれき","history",{"word":114,"reading":115,"meaning":116,"level":19},"改善","かいぜん","improvement",{"word":118,"reading":119,"meaning":120,"level":19},"適用","てきよう","application",{"word":122,"reading":123,"meaning":124,"level":28},"制約","せいやく","constraint",{"word":126,"reading":127,"meaning":128,"level":28},"蓄積","ちくせき","accumulation",{"word":130,"reading":131,"meaning":132,"level":28},"文脈","ぶんみゃく","context",{"word":134,"reading":135,"meaning":136,"level":28},"抑制","よくせい","suppression","\n::para\nLINE（[現]{げん:current:N3}LINEヤフー）は、[日本]{にほん:Japan:N5}で[最も]{もっとも:most:N3}[広く]{ひろく:widely:N4}[使われて]{つかわれて:used:N4}いるメッセージングサービスを[運営]{うんえい:operating:N2}しながら、[独自]{どくじ:proprietary:N1}の[大規模]{だいきぼ:large-scale:N1}[言語]{げんご:language:N4}モデル「HyperCLOVA X」シリーズの[研究]{けんきゅう:research:N4}と[運用]{うんよう:operations:N4}を[進めて]{すすめて:advancing:N3}きました。[本]{ほん:this:N5}[記事]{きじ:article:N3}では、LLMを[本番]{ほんばん:production:N3}サービスで[安定]{あんてい:stably:N3}[稼働]{かどう:run:N1}させるための[取り組み]{とりくみ:efforts:N3}、いわゆる「LLMOps」の[全体]{ぜんたい:overall:N3}[像]{ぞう:picture:N2}を[概観]{がいかん:overview:N1}します。[詳細]{しょうさい:details:N1}は[公開]{こうかい:published:N4}[資料]{しりょう:materials:N3}や[時期]{じき:period:N3}により[異なる]{ことなる:differs:N1}ため、[一般的]{いっぱんてき:general:N2}な[傾向]{けいこう:trends:N2}として[読み]{よみ:read:N5}[取って]{とって:taking:N3}ください。\n\n#en\nLINE (now LINE Yahoo) has been advancing research and operation of its proprietary \"HyperCLOVA X\" series of large language models while running Japan's most widely used messaging service. This article gives an overview of the efforts — so-called \"LLMOps\" — for stably running LLMs in production. Details vary by published material and period, so please read this as general trends.\n::\n\n::heading\nHyperCLOVA Xという[基盤]{きばん:foundation:N1}モデル\n\n#en\nThe HyperCLOVA X Foundation Model\n::\n\n::para\nHyperCLOVA Xは、[元来]{がんらい:originally:N4}NAVER[側]{がわ:side:N3}で[韓国語]{かんこくご:Korean:N2}を[中心]{ちゅうしん:centered:N4}に[開発]{かいはつ:developed:N4}されたシリーズに、LINE[側]{がわ:side:N3}で[日本語]{にほんご:Japanese:N5}データを[大量]{たいりょう:massive:N2}に[追加]{ついか:added:N3}・[微調整]{びちょうせい:fine-tuned:N1}した[多言語]{たげんご:multilingual:N4}モデル[群]{ぐん:group:N2}と[位置]{いち:positioned:N3}[付けられて]{づけられて:positioned:N3}います。[日本語]{にほんご:Japanese:N5}における[敬語]{けいご:keigo:N2}・[業務]{ぎょうむ:business:N3}[文書]{ぶんしょ:documents:N4}・[固有]{こゆう:proper:N2}[名詞]{めいし:nouns:N2}など、グローバルモデルが[弱い]{よわい:weak:N2}[領域]{りょういき:areas:N2}を[強化]{きょうか:strengthening:N3}している[点]{てん:point:N3}が[特徴]{とくちょう:feature:N1}とされます。\n\n#en\nHyperCLOVA X is positioned as a multilingual model series originally developed at NAVER centered on Korean, then heavily augmented and fine-tuned by LINE with Japanese data. It is characterized as strengthening areas where global models tend to be weak, such as Japanese honorifics (keigo), business documents, and proper nouns.\n::\n\n::callout\n[業務]{ぎょうむ:business:N3}[ドメイン]{どめいん:domain}に[特化]{とっか:specialized:N3}したLLMは、[巨大]{きょだい:giant:N2}モデルの[生]{なま:raw:N5}[出力]{しゅつりょく:output:N4}より[小さい]{ちいさい::N5}ながら[安定]{あんてい:stable:N3}した[品質]{ひんしつ:quality:N4}を[出せる]{だせる:can produce:N5}ことが[期待]{きたい:expected:N3}されます。\n\n#en\nDomain-specialized LLMs are expected to produce stable quality — though smaller — than the raw output of giant general models.\n::\n\n::heading\n[推論]{すいろん:inference:N1}[基盤]{きばん:infrastructure:N1}：GPUプールとvLLM\n\n#en\nInference Infrastructure: GPU Pools and vLLM\n::\n\n::para\nLLMの[推論]{すいろん:inference:N1}は[非常]{ひじょう:exceedingly:N3}に[計算]{けいさん:computationally:N2}[コスト]{こすと:costly}が[高い]{たかい:high:N5}ため、GPUを[効率]{こうりつ:efficiently:N1}よく[共有]{きょうよう:share:N3}する[仕組み]{しくみ:mechanism:N3}が[不可欠]{ふかけつ:essential:N3}です。LINEは[社内]{しゃない:in-house:N3}に[共有]{きょうよう:shared:N3}GPUプールを[構築]{こうちく:built:N2}し、Kubernetes[上]{じょう:on:N5}で[各]{かく:each:N2}サービスからの[推論]{すいろん:inference:N1}リクエストを[受ける]{うける:receive:N3}[構成]{こうせい:configuration:N3}を[採用]{さいよう:adopting:N2}していると[語られて]{かたられて:said:N5}います。[推論]{すいろん:inference:N1}サーバーには、PagedAttentionによる[効率]{こうりつ:efficient:N1}[的な]{てきな:like:N4}KVキャッシュ[管理]{かんり:management:N2}と[継続]{けいぞく:continuous:N1}バッチングを[特徴]{とくちょう:characteristics:N1}とするvLLMなどのOSSが[広く]{ひろく:widely:N4}[利用]{りよう:used:N3}されている[傾向]{けいこう:trend:N2}があります。\n\n#en\nBecause LLM inference is computationally very costly, mechanisms for efficiently sharing GPUs are essential. LINE is reportedly building shared GPU pools in-house and adopting a configuration where inference requests from each service are received on Kubernetes. There is a trend of widely using OSS inference servers such as vLLM, which is characterized by efficient KV-cache management via PagedAttention and continuous batching.\n::\n\n::heading\nバッチングと[並列]{へいれつ:parallel:N2}[化]{か:-ization:N3}\n\n#en\nBatching and Parallelization\n::\n\n::para\nLLM[推論]{すいろん:inference:N1}の[一]{いち:one:N5}リクエストは[多く]{おおく:often:N4}の[計算]{けいさん:computation:N2}を[消費]{しょうひ:consumes:N3}しますが、GPUの[計算]{けいさん:compute:N2}[資源]{しげん:resources:N1}を[十分]{じゅうぶん:sufficiently:N5}に[使い切る]{つかいきる:fully utilize:N4}には[複数]{ふくすう:multiple:N2}リクエストの[一括]{いっかつ:batch:N1}[処理]{しょり:processing:N3}が[必要]{ひつよう:necessary:N3}です。Continuous Batchingは、[途中]{とちゅう:mid-flight:N3}リクエストの[隙間]{すきま:gaps:N1}に[新しい]{あたらしい:new:N4}リクエストを[挿入]{そうにゅう:insert:N1}することでスループットを[上げる]{あげる:raise:N5}[手法]{しゅほう:technique:N3}です。さらに、テンソル[並列]{へいれつ:parallel:N2}・パイプライン[並列]{へいれつ:parallel:N2}・[エキスパート]{えきすぱーと:expert}[並列]{へいれつ:parallel:N2}（MoE）といった[並列]{へいれつ:parallel:N2}[化]{か:-ization:N3}[手法]{しゅほう:techniques:N3}を[組み合わせて]{くみあわせて:combining:N3}、[巨大]{きょだい:giant:N2}モデルを[複数]{ふくすう:multiple:N2}GPUに[載せる]{のせる:place:N1}[構成]{こうせい:configuration:N3}も[行われます]{おこなわれます:performed:N5}。\n\n#en\nA single LLM inference request consumes substantial computation, yet fully utilizing GPU compute requires batching multiple requests. Continuous batching is a technique that raises throughput by inserting new requests into the gaps of in-flight ones. Furthermore, parallelization techniques — tensor, pipeline, and expert parallelism (MoE) — are combined to place giant models across multiple GPUs.\n::\n\n::heading\n[量子化]{りょうしか:quantization:N2}と[蒸留]{じょうりゅう:distillation:N2}\n\n#en\nQuantization and Distillation\n::\n\n::para\n[本番]{ほんばん:production:N3}[環境]{かんきょう:environment:N1}では[精度]{せいど:accuracy:N3}を[極端]{きょくたん:extremely:N1}に[下げない]{さげない:without lowering:N5}まま[コスト]{こすと:cost}を[抑える]{おさえる:control:N1}ため、[重み]{おもみ:weights:N4}のINT8／FP8[量子化]{りょうしか:quantization:N2}や、[大]{だい:large:N5}モデルから[小]{しょう:small:N5}モデルへの[知識]{ちしき:knowledge:N3}[蒸留]{じょうりゅう:distillation:N2}も[並行]{へいこう:in parallel:N2}して[行われます]{おこなわれます:performed:N5}。これにより、[応答]{おうとう:response:N1}[速度]{そくど:speed:N3}とGPUメモリ[使用]{しよう:usage:N4}[量]{りょう:amount:N2}を[改善]{かいぜん:improve:N1}し、[最終]{さいしゅう:eventually:N3}[的に]{てきに:like:N4}は[一]{いち:one:N5}[トークン]{とーくん:token}あたりの[コスト]{こすと:cost}を[下げる]{さげる:lower:N5}ことが[目的]{もくてき:goal:N4}です。\n\n#en\nTo control cost in production without dramatically lowering accuracy, INT8\u002FFP8 weight quantization and knowledge distillation from large to small models are performed in parallel. This improves response speed and GPU memory usage, with the ultimate goal of lowering cost per token.\n::\n\n::heading\nプロンプト[管理]{かんり:management:N2}とバージョニング\n\n#en\nPrompt Management and Versioning\n::\n\n::para\nLLMアプリケーションでは「プロンプト」がコードと[同等]{どうとう:equivalent:N3}に[重要]{じゅうよう:important:N3}な[資産]{しさん:asset:N3}になります。LINEを[含む]{ふくむ:including:N2}[多く]{おおく:many:N4}の[企業]{きぎょう:companies:N1}は、プロンプトをGitで[管理]{かんり:managed:N2}するか、[専用]{せんよう:dedicated:N2}のプロンプト[管理]{かんり:management:N2}サービスを[内製]{ないせい:in-house:N1}するなどして、バージョン・[承認]{しょうにん:approval:N2}フロー・[実験]{じっけん:experiment:N3}[履歴]{りれき:history:N1}を[保持]{ほじ:retain:N1}する[仕組み]{しくみ:mechanism:N3}を[整える]{ととのえる:setting up:N1}[傾向]{けいこう:trend:N2}があります。これにより、「[特定]{とくてい:certain:N3}のプロンプトを[更新]{こうしん:updating:N3}したら[品質]{ひんしつ:quality:N4}が[落ちた]{おちた:dropped:N3}」といった[退行]{たいこう:regression:N3}を[追跡]{ついせき:trace:N2}できるようになります。\n\n#en\nIn LLM applications, prompts become assets equivalent in importance to code. Many companies, including LINE, tend to set up mechanisms that retain versions, approval flows, and experiment history by managing prompts in Git or building in-house prompt management services. This makes it possible to trace regressions like \"quality dropped after updating a certain prompt.\"\n::\n\n::heading\n[評価]{ひょうか:evaluation:N1}パイプライン\n\n#en\nEvaluation Pipelines\n::\n\n::para\nLLMの[品質]{ひんしつ:quality:N4}は[従来]{じゅうらい:traditional:N1}の[精度]{せいど:accuracy:N3}・F[値]{ち:value:N3}だけでは[測れず]{はかれず:not measurable:N2}、[人手]{ひとで:human:N4}[評価]{ひょうか:evaluation:N1}・LLM-as-a-Judge・[ベンチマーク]{べんちまーく:benchmarks}（JGLUEやJapanese MT-Benchなど）・[安全性]{あんぜんせい:safety:N3}[評価]{ひょうか:evaluation:N1}を[組み合わせる]{くみあわせる:combining:N3}[必要]{ひつよう:needed:N3}があります。LINEはCIに[評価]{ひょうか:evaluation:N1}[ジョブ]{じょぶ:jobs}を[組み込み]{くみこみ:embedding:N3}、プロンプトやモデルを[更新]{こうしん:updated:N3}するたびに[一連]{いちれん:series:N3}の[評価]{ひょうか:evaluation:N1}が[自動]{じどう:automatically:N4}[実行]{じっこう:run:N3}される[体制]{たいせい:setup:N3}を[整えて]{ととのえて:building:N1}いると[考えられて]{かんがえられて:thought:N4}います。\n\n#en\nLLM quality cannot be measured by traditional accuracy and F-score alone — human evaluation, LLM-as-a-Judge, benchmarks (JGLUE, Japanese MT-Bench, etc.), and safety evaluation must be combined. LINE is thought to be building a setup where evaluation jobs are embedded in CI, with a battery of evaluations running automatically each time prompts or models are updated.\n::\n\n::callout\n[評価]{ひょうか:evaluation:N1}は「[一]{いち:one:N5}[度]{ど:time:N4}[作って]{つくって:built:N4}[終わり]{おわり:finished:N4}」ではなく、[本番]{ほんばん:production:N3}データの[ドリフト]{どりふと:drift}に[合わせて]{あわせて:matching:N3}[継続]{けいぞく:continuously:N1}[更新]{こうしん:updated:N3}される「[生]{なま:living:N5}データセット」として[扱う]{あつかう:treat:N1}ことが[重要]{じゅうよう:important:N3}です。\n\n#en\nEvaluation should be treated not as \"build once and done,\" but as a \"living dataset\" continuously updated to match production data drift.\n::\n\n::heading\nRAGとメッセージング[文脈]{ぶんみゃく:context:N1}の[統合]{とうごう:integration:N1}\n\n#en\nRAG and Integration with Messaging Context\n::\n\n::para\nRAG（Retrieval-Augmented Generation、[検索]{けんさく:retrieval:N1}[拡張]{かくちょう:augmented:N1}[生成]{せいせい:generation:N3}）は、ユーザーの[質問]{しつもん:question:N4}に[関連]{かんれん:relevant:N3}する[文書]{ぶんしょ:documents:N4}を[ベクトル]{べくとる:vector}[検索]{けんさく:search:N1}で[取り出し]{とりだし:retrieve:N3}、プロンプトに[埋め込んで]{うめこんで:embedding:N2}LLMに[渡す]{わたす:passing:N3}[手法]{しゅほう:technique:N3}です。LINEのようにユーザー[履歴]{りれき:history:N1}・FAQ・[公式]{こうしき:official:N3}[アカウント]{あかうんと:account}[情報]{じょうほう:information:N3}など[膨大]{ぼうだい:vast:N1}な[内部]{ないぶ:internal:N3}コンテンツを[持つ]{もつ:has:N4}[企業]{きぎょう:company:N1}にとって、[適切]{てきせつ:appropriate:N3}な[ベクトル]{べくとる:vector}[埋め込み]{うめこみ:embedding:N2}モデルと[ベクトル]{べくとる:vector}データベース（OpenSearchやMilvus、pgvectorなど）の[選定]{せんてい:selection:N3}が[重要]{じゅうよう:critical:N3}になります。\n\n#en\nRAG (Retrieval-Augmented Generation) is a technique that retrieves documents relevant to a user's question via vector search, embeds them in the prompt, and passes them to the LLM. For a company like LINE with vast internal content — user history, FAQs, official account information — the choice of an appropriate vector embedding model and vector database (OpenSearch, Milvus, pgvector, etc.) becomes critical.\n::\n\n::heading\n[安全性]{あんぜんせい:safety:N3}と[抑制]{よくせい:control:N1}\n\n#en\nSafety and Mitigation\n::\n\n::para\n[個人]{こじん:personal:N2}[情報]{じょうほう:information:N3}や[差別]{さべつ:discriminatory:N3}[的]{てき:like:N4}[表現]{ひょうげん:expressions:N3}、[誤った]{あやまった:false:N3}[医療]{いりょう:medical:N2}[助言]{じょげん:advice:N3}など、LLMが[出力]{しゅつりょく:output:N4}してはいけない[領域]{りょういき:areas:N2}があります。LINEは[入力]{にゅうりょく:input:N4}と[出力]{しゅつりょく:output:N4}の[両方]{りょうほう:both:N3}に[安全]{あんぜん:safety:N3}フィルタを[挟み]{はさみ:inserting:N2}、ガードレールモデルや[ルール]{るーる:rule}ベースの[抑制]{よくせい:control:N1}を[併用]{へいよう:concurrently using:N1}していると[見られて]{みられて:seen:N5}います。さらに、[個人]{こじん:personal:N2}[情報]{じょうほう:information:N3}[保護]{ほご:protection:N1}[法]{ほう:law:N3}の[要]{よう:requiring:N3}[配慮]{はいりょ:care:N1}[個人]{こじん:personal:N2}[情報]{じょうほう:information:N3}にあたるデータをプロンプトに[投入]{とうにゅう:feed:N3}しないよう、[匿名]{とくめい:anonymization:N1}[化]{か:-ization:N3}や[仮名]{かめい:pseudonymization:N1}[加工]{かこう:processing:N3}を[適用]{てきよう:apply:N3}する[層]{そう:layer:N2}も[必要]{ひつよう:needed:N3}とされます。\n\n#en\nThere are areas LLMs must not output, such as personal information, discriminatory expressions, and false medical advice. LINE is seen to insert safety filters on both input and output, concurrently using guardrail models and rule-based mitigation. Furthermore, a layer that applies anonymization and pseudonymization is needed to avoid feeding data corresponding to \"special-care-required personal information\" under the Act on the Protection of Personal Information into prompts.\n::\n\n::heading\nコスト[管理]{かんり:management:N2}\n\n#en\nCost Management\n::\n\n::para\nLLMの[運用]{うんよう:operations:N4}コストは、[利用]{りよう:usage:N3}[量]{りょう:volume:N2}に[応じて]{おうじて:according to:N1}[線形]{せんけい:linearly:N2}に[増える]{ふえる:increases:N3}わけではなく、GPUの[占有]{せんゆう:occupation:N2}[時間]{じかん:time:N5}・[モデルサイズ]{もでるさいず:model size}・[コンテキスト]{こんてきすと:context}[長]{ちょう:length:N5}に[応じて]{おうじて:depending on:N1}[非]{ひ:non-:N3}[線形]{せんけい:linearly:N2}に[増加]{ぞうか:increases:N3}します。LINEは[複数]{ふくすう:multiple:N2}モデル（[巨大]{きょだい:giant:N2}・[中型]{ちゅうがた:medium:N2}・[軽量]{けいりょう:lightweight:N2}）を[用途]{ようと:use case:N3}に[応じて]{おうじて:depending on:N1}[使い分け]{つかいわけ:routing:N4}、ルーティング[層]{そう:layer:N2}で「[簡単]{かんたん:simple:N2}な[要約]{ようやく:summary:N3}は[小さな]{ちいさな:small:N5}モデル」「[難しい]{むずかしい:hard:N3}[推論]{すいろん:reasoning:N1}は[大きな]{おおきな:large:N5}モデル」と[振り分ける]{ふりわける:routes:N1}[構成]{こうせい:configuration:N3}が[一般的]{いっぱんてき:common:N2}と[考えられます]{かんがえられます:thought:N4}。\n\n#en\nLLM operating cost does not grow linearly with usage — it grows non-linearly with GPU occupancy time, model size, and context length. LINE is thought to commonly route requests through a layer that uses different-size models (giant, medium, lightweight) per use case — \"simple summaries to a small model, hard reasoning to a large model.\"\n::\n\n::heading\nおわりに\n\n#en\nConclusion\n::\n\n::para\nLLMOpsは「モデルを[作って]{つくって:built:N4}デプロイすれば[完了]{かんりょう:done:N2}」という[世界]{せかい:world:N4}ではなく、[評価]{ひょうか:evaluation:N1}・[安全]{あんぜん:safety:N3}[抑制]{よくせい:mitigation:N1}・コスト[管理]{かんり:management:N2}・RAGとデータパイプラインの[整備]{せいび:building:N1}まで[含む]{ふくむ:including:N2}[継続]{けいぞく:continuous:N1}[的な]{てきな:like:N4}[運用]{うんよう:operations:N4}[行為]{こうい:practice:N1}です。LINEのような[巨大]{きょだい:giant:N2}メッセージング[基盤]{きばん:platform:N1}を[持つ]{もつ:having:N4}[企業]{きぎょう:company:N1}は、[蓄積]{ちくせき:accumulated:N1}データと[文脈]{ぶんみゃく:context:N1}を[活かす]{いかす:leveraging:N3}ことで[独自]{どくじ:unique:N1}の[価値]{かち:value:N1}を[作り]{つくり:creating:N4}[出せる]{だせる:can produce:N5}[立場]{たちば:position:N4}にあると[言える]{いえる:can say:N4}でしょう。\n\n#en\nLLMOps is not a world where \"building a model and deploying it is done\" — it is a continuous operational practice including evaluation, safety mitigation, cost management, and the buildout of RAG and data pipelines. A company like LINE, with a giant messaging platform, can be said to be in a position to create unique value by leveraging accumulated data and context.\n::\n"]