語義作成語言(Semantic Composition Language, SCL)是一種類似Lispエ³語言。雖然LLM広尹內底無SCLエ³詳細資料，ㆬ³コㅗㇷ¹尹ロㆁ有法度處理チッ¹種類Lisp語言。用EBNF來表示表達式文法(expression grammar)嘛是無問題。有一クァ模型エ³広語義作成チッ¹エ⁷作成性(Compositionality)エ³概念是Frege代先提出エ³，Wikipedia內底コㅗㇷ¹有コㆁ是コㅗㇷ¹卡早エ³Boole所提出エ³。
ラㇴエ³タㆁ²カ³一个散文(prose)prompt翻譯做一个SCL prompt。嘛エ³タㆁ²カ³一个SCL prompt翻倒뚜ㆁ²ラィㇷ做一个散文prompt，也着是自然語言エ⁷prompt。散文適合快速AI輸入，SCL適合可再使用エ⁷工課流程。
Shorthand是速記アㇷ³是簡略表記エ³意思。譬喻広，漢字着是一種速記符號，像南ジㇷ，極ジㇷ。注音符號用ㄇ來表示規个音節ㄇㄚ，用ㄎ來表示音節ㄎㄜ，着是一種簡略表記。使用者若無想ベㇷ¹用速記ネㇷ？使用者猶原是エ³用エ作成一个類Lispエ³樹仔，不免用着速記表記，只要有カ³動詞エ³項構造保留키ㇷラィㇷ着好。
下腳是無用着速記エ³一个例:
recommend user
filter (recommend user)
select (filter (recommend user))
(let ((x (filter (recommend user)))
(select x))
動詞エ³項構造對類Lisp作成來広是核心イェㇴ⁵ジㇴ²。類Lisp作成治基礎上倚靠動詞エ³項數，伊接受有型別的項，カㇷ¹巢狀logic(例， filter(recommend(user)))。
速記表記カㇴ⁷ナ⁷是一種構文上快速輸入カㇷ¹使用者操作利便エ⁷一个方便エ⁷層次-syntactic sugar，構文糖仔也。不是一个核心依存。
一个速記表達式
Movies → Sci-fi → Thought-provoking → Netflix
一个速記成員、部件
Sci-fi
所以速記表記有增加啥乜好處ネㇷ？
let + recommend + user + filter + select，取代巢形碼。治SCL語境內底エ³速記表記是代表一个卡長エ⁷表達式アㇷ³是表達鏈エ³實腹エ³別名アㇷ³是省略語。伊カ³結構複雜性藏키ㇷラィㇷ，但是保留語義上エ⁷clarity。
伊是一个語義上抽象化層次，一个方便エ⁷提示記述構文。
カ³下腳エ³列表下ルェㇷ¹AI對話內底，エ³得着詳細解說。アㇴ¹ツァ˜カ³表達map到速記。
ChatGPT産生
How Expressions Map to Shorthand.
1. Direct Verb Mapping
2. Chained Pipeline Mapping
3. Parameterized Mapping
4. Let-bound Structure
カ³下腳エ³列表下ルェㇷ¹AI對話內底，エ³得着詳細解說。SCL內底有啥乜款エ³速記。
ChatGPT産生
Types of Shorthand in SCL
1. Verb-Noun Shorthand
2. Attribute Shorthand
3. Chain Shorthand
4. Let-bound Shorthand
5. Custom Semantic Tokens
一个速記表達式的象徵形(symbolic form)實際上着是一个操作式(operation formula)。治轉換做最終解析化エ³表達式進前-治解析エ³時，伊提供語義logic。
速記表達式治解析エ³時ツㇴ⁷，エ³去ホㆁ⁵翻譯做象徵形，治轉換做Lisp樹仔進前去띠ッ¹띠ㅗㇷ³ヒッ¹个結構化formatエ³操作logic。
速記表達式卡倚使用者，伊エ³象徵形是卡倚Lisp語言。
ラㇴベㇷ¹用下腳エ³用例來看解析エ³step。伊是一个表示AND logicエ³序列式filtering，後壁接一个表示OR logicエ³分岐式filtering。
ChatGPT産生
4. Interactive Exploration (Branch + Merge)
Breadcrumb Flow:
Home → Movies → Sci-fi → Thought-provoking → Netflix
↘ Runtime: Short (<90m)
↘ Tone: Dark
↘ Compare with: Apple TV+
最後カ³feedback loop接키ㇷリㇷ，着變做下腳アㇴ¹ネ，有解析step，コㅗㇷ¹有Lisp code:
Copilot産生
1. Sequential Filtering
Start by narrowing the movie pool step by step:
- 1. recommend + user → candidate_movies
- 2. filter + genre("sci-fi") + platform("netflix") → netflixscifi
- 3. filter + tone("thought-provoking") + netflixscifi → thoughtful_pool
---
2. Branching & Merge
Explore two refinement paths in parallel and recombine:
- 4. filter + runtime<90 + thoughtfulpool → shortbranch
- 5. filter + tone("dark") + thoughtfulpool → darkbranch
- 6. merge(shortbranch, darkbranch) → mergedfocuspool
---
3. Feedback Loop: Actor Preference
Incorporate user affinity for Jennifer Lawrence:
- 7. rate(mergedfocuspool) → detect user preference
- 8. filter + actor("Jennifer Lawrence") + mergedfocuspool → jlawpreferredpool
---
Final Lisp-Style Composition
`lisp
(let* (
(candidates (recommend user))
(netflix-sci-fi (filter genre("sci-fi") platform("netflix") candidates))
(thoughtful (filter tone("thought-provoking") netflix-sci-fi))
(short-branch (filter runtime<90 thoughtful))
(dark-branch (filter tone("dark") thoughtful))
(merged (merge short-branch dark-branch))
(jlaw-filtered (filter actor("Jennifer Lawrence") merged))
(apple-tv (filter platform("apple tv+") (filter tone("dark") (filter runtime<90 (recommend user)))))
(result (compare apple-tv jlaw-filtered))
)
result)
`
序列式過濾是一種逐カィ²追加條件エ⁷過濾。逐カィ²エ³使用者提示ロㆁ以漸進方式カ³結果集阨化(ウェㇷ³ファ³)。譬喻広:
쩨着是絞り込みサーチ検索(Refined Search)エ³核心，逐漸施加有層次エ³條件。不知台灣話是不是エ³用エカ⁷翻譯做絞入式搜尋檢索。
利用터ㆁ³煞尾エ³frameworkカ³任何一个速記表達式エ³解析step印추ァィㇷ。
Show the parsing steps for this shorthand expression
各層次エ³運作。
ChatGPT産生
[[ 1. Natural Language Expression ]
"Show me some thought-provoking sci-fi movies on Netflix—
either short or dark— and compare them to what's on Apple TV+."
↓ (interpreted by NL-to-shorthand module)
[ 2. Shorthand Expression ]
Home → Movies → Sci-fi → Thought-provoking → Netflix →
Branch 1: Runtime: Short (<90m) →
Branch 2: Tone: Dark →
Merge →
Compare with: Apple TV+
↓ (parsed into structured symbolic operations)
[ 3. Symbolic Form ] ← Operation formulas are defined and applied here
let + recommend + user
+ filter + genre("sci-fi")
+ filter + tone("thought-provoking")
+ filter + platform("netflix")
+ branch(filter + runtime<90, filter + tone("dark"))
+ merge
+ compare(platform("apple tv+"))
↓ (converted into executable computation structure)
[ 4. Parsed Expression (Lisp-style Tree) ]
(let ((base (platform "netflix"
(tone "thought-provoking"
(genre "sci-fi"
(recommend user)))))
(short (runtime<90 base))
(dark (tone "dark" base))
(merged (merge short dark)))
(compare (platform "apple tv+") merged))
譬喻広有一个無完整速記:
Home → Movies → Sci-fi
カㇷ¹伊エ³象徵形(操作式):
let + genre("sci-fi") +
象徵形터ㆁ³煞尾是一个加號。用伊做輸入。ヒッ¹个尾溜エ⁷加號是指出チッ¹條鏈仔アㇷ¹未完整。伊カ³建議器提示広アィ²提出有效エ⁷延續。
解析器建立一个Lisp風格エ³樹仔:
(genre "sci-fi")
쩨代表一組科幻電影，對續ロㅗㇷラィㇷエ³操作嘛準備好ア³。
建議器エ³産生쩨四款エ³建議，附加エ⁷filter，分岐，合併/比較，選擇。以目前エ³速記表達式來看，有下腳7種選項타ㆁ⁷選:
Grok産生
Suggestions for 'let + genre("sci-fi") +':
- platform("netflix")
- tone("thought-provoking")
- runtime<90
- actor("Keanu Reeves")
- award("oscar")
- branch(runtime<90, tone("dark"))
- select
若準広使用者カ³鏈仔訓ホ⁷長，
速記:
Home → Movies → Sci-fi → Netflix
象徵形:
let + genre("sci-fi") + platform("netflix") +
解析器カ³樹仔update做アㇴ¹ネ:
(platform "netflix" (genre "sci-fi"))
建議器コㅗㇷ¹根據updateエ³上下文(sci-fi movies on Netflix)獻出精煉エ⁷しあげ:
- tone("dark")
- runtime<90
- actor("Keanu Reeves")
- award("oscar")
- branch(runtime<90, tone("dark"))
- select
チッ¹个過程反覆運作，建議器エ³根據鏈仔狀態カㇷ¹項數規則カ³建議創ホ⁷ハㇷ゚³ス。
建議器是運作治操作式チッ¹緣。
ChatGPT加上Grok産生
---
# Movie-Selecting Framework
## **Folder Structure**
```
/PromptFramework/
/Verbs/: recommend.md, select.md, filter.md, rate.md, merge.md, compare.md
/Nouns/: user.md, movie.md, genre.md, actor.md, rating.md, platform.md, tone.md, runtime.md, award.md
/Rules/: operation_formula.md
/Templates/: prompt_templates.md
/SemanticParsingLayer/: suggestor.md, parser.md
```
---
## A. Verbs Folder
### File: recommend.md
* **Purpose:** Retrieve candidate movies for a user.
* **Valency:** 1–2
* **Examples:**
* `recommend + user`
* `recommend + user + filter`
### File: select.md
* **Purpose:** Choose final movie from a filtered list.
* **Valency:** 1
* **Examples:**
* `select from recommendation`
* `recommend + user + filter + select`
### File: filter.md
* **Purpose:** Narrow movie sets by attributes.
* **Valency:** 1
* **Filterable Attributes:** genre, actor, tone, runtime, platform, award
* **Examples:**
* `filter + genre`
* `recommend + user + filter + runtime<90`
* `filter + tone:dark`
### File: rate.md
* **Purpose:** Assign a score or evaluation to a movie.
* **Valency:** 1
* **Examples:**
* `rate movie`
* `rate movie with user`
### File: merge.md
* **Purpose:** Combine results from parallel filter branches.
* **Valency:** 2+
* **Example:** `merge(short, dark)`
### File: compare.md
* **Purpose:** Compare result sets, movies, or attributes (symmetric).
* **Valency:** 2
* **Example:** `compare(platform("netflix"), platform("apple tv+"))`
---
## B. Nouns Folder
| Noun | Examples |
| ------------ | ------------------------------------------- |
| **user** | `user:John`, `user:anonymous` |
| **movie** | `"Inception"`, `"The Matrix"` |
| **genre** | sci-fi, drama, comedy |
| **actor** | Leonardo DiCaprio, Emma Stone, Keanu Reeves |
| **rating** | 8.5, PG-13, R |
| **platform** | netflix, hulu, apple tv+, prime video |
| **tone** | thought-provoking, dark, uplifting |
| **runtime** | `runtime<90`, `runtime<120` |
| **award** | oscar, bafta, cannes |
---
## C. Rules Folder
### File: operation_formula.md
* **Purpose:** Store symbolic operation formulas for movie selection.
* **Contents:**
* **Sequential Recommendation Chain:**
`genre("sci-fi") + platform("netflix") → filtered_movies`
`recommend + user + filter → filtered_movies`
`recommend + user + filter + select → final_choice`
* **Filtering Logic:**
`filter + genre` → subset by genre
`filter + tone` → subset by mood
`filter + runtime<90` → short movies
* **Branch + Merge:**
`let + A + branch(B, C) + merge → intersected result`
`let + base + compare(other_set)`
* **Symbolic Chain Composition:**
`let + A + B + C + branch(...) + merge + compare(...)` → Lisp-style nested evaluation
---
## D. Templates Folder
### File: prompt_templates.md
* **Purpose:** Provide user-facing and symbolic shorthand templates.
* **Contents:**
* **User-Facing Shorthand:**
```
Home → Movies → Sci-fi → Thought-provoking → Netflix →
Branch 1: Runtime: Short (<90m) →
Branch 2: Tone: Dark →
Merge →
Compare with: Apple TV+
```
* **Framework Shorthand:**
* `let + recommend + user`
* `recommend + user + filter + select`
* **Symbolic Chain Composition:**
* `genre(...) + platform(...) + tone(...)`
* `let + moodToGenre("uplifting") + platform(...) + runtime<90`
* `let + analyze("Palm Springs") + themeMatch + toneShift(...) + platform(...) + compare(...)`
* **Branch + Merge Templates:**
* `let + A + branch(B, C) + merge`
* `let + A + platform(...) + compare(...)`
---
## E. SemanticParsingLayer Folder
### File: suggestor.md
* **Purpose:** Parse symbolic chain fragments, suggest completions using valency + rules.
* **Example:**
Input: `genre("sci-fi") +`
Suggestions: platform(...), tone(...), filter(...), compare(...)
### File: parser.md
* **Purpose:** Convert symbolic `+` composition into Lisp-style tree; support binding, branching, merging.
* **Contents:**
**Parsing Steps (Symbolic → Lisp):**
1. **Tokenize** the input expression by `+` symbols.
2. **Determine valency** of each verb from `/Verbs/`.
3. **Apply functions right-to-left**: each verb consumes arguments according to valency.
4. **Name intermediate results** using `let` for reusable subresults.
5. **Handle branching and merging** using `branch(...)` and `merge(...)`.
6. **Compose final Lisp tree** reflecting sequential and nested operations.
**Example:**
```
Input: recommend + user + filter + select
Step 1: recommend + user → candidate_movies
Step 2: filter(candidate_movies) → filtered_movies
Step 3: select(filtered_movies) → final_choice
Lisp:
(let ((x (filter (recommend user))))
(select x))
```
---