fengruixiang
/
hklii_samples
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

qwen3 前后5000 分析结果

main
fengruixiang 2026-05-15 09:15:23 +08:00
parent 22e5058de1
commit 3851474aa3
3 changed files with 50 additions and 52 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
AGENTS.md
View File

@ -1 +1,2 @@
人与助手之间的语言交互默认为简体中文。
python 要使用 uv run 运行。

49
en_cases_hkcfa/2025_HKCFA_20/summary_fast.yml
View File

@ -1,33 +1,33 @@
plaintiff:
- HKSAR
defendant:
- MAK KWONG YIU (麥光耀) (D1)
- CHAN LAI YEE (陳麗兒) (D2)
- WONG SHUK ON (黃淑安) (D3)
- LEE YICK MING (李易明) (D4)
- MAK KWONG YIU (麥光耀)
- CHAN LAI YEE (陳麗兒)
- WONG SHUK ON (黃淑安)
- LEE YICK MING (李易明)
jurisdiction_code: HKCFA
jurisdiction_name: Court of Final Appeal of the Hong Kong Special Administrative Region
case_location:
- 香港特別行政區
- Hong Kong Special Administrative Region
case_reason: >-
HKSAR appeals against the conviction of respondents for conspiracy to defraud, challenging the Court
of Appeal's decision (CACC No. 239 of 2021) regarding the validity of the prosecution's case on the
HKSAR appeals against the Court of Appeal's decision (CACC No. 239 of 2021) to overturn convictions
of defendants for conspiracy to conceal CISL's role as placing agent in a connected transaction.
case_object:
- criminal charges of conspiracy to defraud
- concealment of financial arrangements
- validity of placing agreements
- conviction
- right to have conviction upheld
- connected transaction determination
judgment_result:
- charge: Whether the placement/sub-placement agreements were continuing connected transactions (liability
issue)
result: No view expressed. Court did not hear arguments on this matter, therefore expressed no view
as to whether they were continuing connected transactions. (See [12] of Reasons for Verdict)
- charge: Conspiracy to conceal CISL's role as placing agent (liability issue)
result: Upheld. Appeal allowed, convictions restored. Court found concealment of CISL's agency constituted
unlawful connected transaction under Listing Rules, and directors' conflict of interest was central
to conspiracy charges.
judgment_summary: >-
HKSAR appealed the conviction of defendants for conspiracy to defraud, challenging the validity of prosecution's
case on concealment of financial arrangements. Core issues centered on whether agreements constituted
continuing connected transactions and if concealment justified criminal liability. The Court rejected
limiting concealment to sham transactions (Snook v London), citing Adams v The Queen, and emphasized
dishonesty as a required element. It did not address the agreements' status but up
HKSAR appeals against the Court of Appeal's overturning of convictions for conspiracy to conceal CISL's
role in a connected transaction. Core issues: upholding convictions and determining connected transaction
status. Court rejected narrow 'sham' test from Snook v London, citing Adams v The Queen, finding concealment
of CISL's agency constituted unlawful connected transaction under Listing Rules. Convictions upheld,
appeal allowed, with directors' conflicts central to charges. Hong Kong Stock Ex
involved_entities:
- entity_name: Mr Justice Ribeiro PJ
reason: Presiding judge in this case, responsible for fact-finding and adjudication.
@ -39,13 +39,8 @@ involved_entities:
reason: Presiding judge in this case, responsible for fact-finding and adjudication.
- entity_name: Sir William Young NPJ
reason: Presiding judge in this case, responsible for fact-finding and adjudication.
- entity_name: HKSAR
reason: Appellant in this criminal case, representing the Hong Kong Special Administrative Region.
- entity_name: Hong Kong Stock Exchange
reason: Institution involved in the regulatory framework governing connected transactions.
- entity_name: Diplock LJ
reason: Served as judge in Snook v London and West Riding Investments Ltd [4], articulated test for
determining sham transactions.
- entity_name: Lord Jauncey of Tullichettle
reason: Served as judge in Adams v The Queen [5], articulated legal principle regarding concealment
of director's breach of duties.
reason: Served as judge in Snook v London and West Riding Investments Ltd, articulated test for determining
sham transactions.
- entity_name: Hong Kong Stock Exchange
reason: Institution involved in regulatory framework for connected transactions.

52
hk_case_extractor.py
View File

@ -365,35 +365,37 @@ def gather_all(text: str) -> dict[str, str]:
"""為每個 group 召回對應的上下文片段
優化策略：
1. 基礎信息（當事人）：直接取開頭2000字符，不使用關鍵詞召回
2. 判決結果：重點關注尾部4000字符 + 關鍵詞召回
3. 其他字段：保持關鍵詞召回策略
1. 基礎信息（當事人）：直接取開頭5000字符，不使用關鍵詞召回
2. 事由與標的（reason_object）：直接取開頭5000字符，不使用關鍵詞召回
3. 判決結果：取開頭5000字符 + 尾部5000字符
4. 其他字段：保持關鍵詞召回策略
"""
out: dict[str, str] = {}
# 1. 當事人信息：直接從開頭2000字符提取
out["parties"] = text[:2000]
# 1. 當事人信息：直接從開頭5000字符提取
out["parties"] = text[:5000]
out["_parties_hits"] = "0" # 不使用關鍵詞，標記為0
# 2. 判決結果：優先使用尾部4000字符，再補充關鍵詞召回
tail_text = text[-4000:] if len(text) > 4000 else text
# 如果尾部不足4000字，用關鍵詞召回補充
if len(tail_text) < 4000:
kws = KEYWORD_GROUPS["judgment_result"]
ctx, hits = gather_chunks(text, kws, half_window=500, max_total=6500)
out["judgment_result"] = ctx
out["_judgment_result_hits"] = str(hits)
else:
out["judgment_result"] = tail_text
out["_judgment_result_hits"] = "0" # 直接使用尾部，不計算關鍵詞命中
# 2. 事由與標的：直接從開頭5000字符提取
out["reason_object"] = text[:5000]
out["_reason_object_hits"] = "0" # 不使用關鍵詞，標記為0
# 3. 其他字段：使用關鍵詞召回
# 3. 判決結果：取開頭5000字符 + 尾部5000字符
head_text = text[:5000] if len(text) > 5000 else text
tail_text = text[-5000:] if len(text) > 5000 else ""
# 如果文本足夠長，拼接頭尾；否則只用全文
if tail_text and head_text != tail_text:
out["judgment_result"] = head_text + "\n\n[…]\n\n" + tail_text
else:
out["judgment_result"] = head_text
out["_judgment_result_hits"] = "0" # 直接截取，不計算關鍵詞命中
# 4. 其他字段：使用關鍵詞召回
params: dict[str, tuple[int, int]] = {
"reason_object": (500, 6000),
"entities": (400, 6500),
"analysis": (500, 6500),
}
for group in ["reason_object", "entities", "analysis"]:
for group in ["entities", "analysis"]:
kws = KEYWORD_GROUPS[group]
hw, mt = params[group]
ctx, hits = gather_chunks(text, kws, half_window=hw, max_total=mt)
@ -554,9 +556,9 @@ def extract_parties(client: OllamaClient, context: str, lang: str = 'zh') -> dic
fewshot = PARTIES_FEWSHOT_ZH if lang == 'zh' else PARTIES_FEWSHOT_EN
if lang == 'zh':
user = f"{fewshot}\n\n請從以下判決書開頭部分抽取：\n```\n{context[:2000]}\n```"
user = f"{fewshot}\n\n請從以下判決書開頭部分抽取：\n```\n{context[:5000]}\n```"
else:
user = f"{fewshot}\n\nPlease extract from the following judgment header:\n```\n{context[:2000]}\n```"
user = f"{fewshot}\n\nPlease extract from the following judgment header:\n```\n{context[:5000]}\n```"
return client.chat_json_with_retry(system, user, PARTIES_SCHEMA)
@ -658,10 +660,10 @@ def extract_reason_object(client: OllamaClient, context: str, lang: str = 'zh')
if lang == 'zh':
user = (f"{fewshot}\n\n"
f"請從以下片段（多處關鍵詞召回拼接）抽取：\n```\n{context[:6000]}\n```")
f"請從以下判決書開頭部分抽取：\n```\n{context[:5000]}\n```")
else:
user = (f"{fewshot}\n\n"
f"Please extract from the following segments (keyword-based retrieval):\n```\n{context[:6000]}\n```")
f"Please extract from the following judgment header:\n```\n{context[:5000]}\n```")
out = client.chat_json_with_retry(system, user,
schema,
@ -774,10 +776,10 @@ def extract_judgment_result(client: OllamaClient, context: str, lang: str = 'zh'
if lang == 'zh':
user = (f"{fewshot}\n\n"
f"請從以下判決書尾部（命令/裁定部分）抽取：\n```\n{context[-4000:]}\n```")
f"請從以下判決書片段（開頭5000字符 + 尾部5000字符）抽取：\n```\n{context}\n```")
else:
user = (f"{fewshot}\n\n"
f"Please extract from the following judgment ending (order/disposition section):\n```\n{context[-4000:]}\n```")
f"Please extract from the following judgment segments (first 5000 + last 5000 characters):\n```\n{context}\n```")
return client.chat_json_with_retry(system, user,
JUDGMENT_RESULT_SCHEMA,