Update README.md

README.md

@@ -16,9 +16,17 @@ size_categories:
 
 总计 92,421 句剧情对白（带有角色标签）+旁白，从崩坏3的“主线1黄昏、少女、战舰”到“主线第二部03间章：一个梦游者的苦痛”
 
-本数据集经由 [honkai_impact_3rd_game_playthrough](https://huggingface.co/datasets/mrzjy/honkai_impact_3rd_game_playthrough) 视频数据集-> OCR 文字识别 -> VLM 结构化解析 -> 后处理 pipeline 得到。
+本数据集从 [honkai_impact_3rd_game_playthrough](https://huggingface.co/datasets/mrzjy/honkai_impact_3rd_game_playthrough) 视频数据集出发，经过 AI pipeline 最终获取结构化的文本剧情语料。
 
-所有没有识别出的 character role 记为 "\<unknown\>"
+AI pipeline 概述如下：
+
+1. 分P下载视频（使用 [BBDown](https://github.com/nilaoda/BBDown) 下载 [BiliBili崩三剧情视频](https://www.bilibili.com/video/BV12W411h76f/)）
+2. 视频帧分割（每1秒取一帧画面）
+3. 逐帧 OCR 检测文本（使用 [Paddle-OCR](https://github.com/PaddlePaddle/PaddleOCR)）
+4. 逐帧 VLM 结构化解析（使用 [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6)，输入为帧图像 + OCR结果）
+5. 基于规则的后处理
+    - 规范化 VLM 输出（e.g., 去噪、排除格式有问题的输出）
+    - 基于编辑距离等方法的连续帧的信息去重与归并（e.g., 由于播放对白需要时间，话说一半的帧信息会被 merge 到后续话说完整的帧信息中，合并成一帧））
 
 由于是通过 AI pipeline，所以难免存在误差（识别错误等），但此数据质量依旧还算尚可
 
@@ -121,4 +129,70 @@ You can see that sometimes the role names are not consistent (e.g., 德丽莎女
 萝莎莉娅 331
 长光 302
 羽兔 293
+```
+
+# VLM Prompt
+
+```
+PROMPT = """This is an image of RPG game. Given associated OCR result, please help us identify the existence of story narrations and dialogues and extract them in structured format.
+This is the associated OCR results:
+```ocr
+{ocr}
+```
+
+There are two types of story content you should extract:
+
+- Narration: single line or paragraph of narration, telling the story background and plots
+- Dialogue: dialogue contents spoken by a character. The speaker character name and spoken content must co-appear in the image.
+
+Note:
+
+- Be strict with OCR texts, you are NOT allowed to fabricate contents that are not captured by OCR results.
+- The OCR often separate multiline texts, and it's your task to concatenate consecutive lines if necessary.
+- There might be noisy textual contents (e.g., advertisement, UI elements, combos, etc.), which are not our interest.
+- There might be texts indicating state/environment information (e.g., location, time, source, state, etc), you can extract them as well in environment field.
+
+Please output your response in JSON structure in one of the 3 following ways:
+
+1. In case of no desired content (neither dialogue nor narration), output a JSON dict whose type is null.
+
+```json
+{{"type": null}}
+```
+
+2. In case of dialogue
+
+```json
+{{
+    "type": "dialogue",
+    "role": "<speaker name>",
+    "content": "<spoken content>",
+    "state": "<state/environment info, null if there isn't any>"
+}}
+```
+
+3. In case of narration
+
+```json
+{{
+    "type": "narration",
+    "content": "<narrative content>"
+}}
+```"""
+```
+
+# VLM code snippet
+
+```
+# generate
+for batch in tqdm(batches):
+    msgs = [
+        [{"role": "user", "content": [Image.open(b["frame_path"]), format_template(b["ocr"])]}]
+        for b in batch
+    ]
+    outputs = model.chat(
+        image=None,
+        msgs=msgs,
+        tokenizer=tokenizer
+    )
 ```