239 lines
10 KiB
C#
239 lines
10 KiB
C#
using CloudBuilder.AI.Entity;
|
||
using CloudBuilder.Topshelf.Python;
|
||
using System;
|
||
using System.Collections.Generic;
|
||
using System.Diagnostics.Eventing.Reader;
|
||
using System.Linq;
|
||
using System.Reflection.Metadata;
|
||
using System.Text;
|
||
using System.Text.RegularExpressions;
|
||
using System.Threading.Tasks;
|
||
using static System.Net.Mime.MediaTypeNames;
|
||
|
||
namespace CloudBuilder.Topshelf.Utility
|
||
{
|
||
public class SpeakerAnalysisHelper
|
||
{
|
||
public string GetPerson(Hanlp2 hanlp, string text)
|
||
{
|
||
string depRoot = hanlp.GetDepRoot(text);
|
||
|
||
List<NamedEntity> ners = hanlp.GetHanlpNer(text);
|
||
|
||
List<List<MeaningRepresentationParsingEntity>> srls = hanlp.GetHanlpSrl(text);
|
||
|
||
List<HanlpConstituencyNode> cons = hanlp.GetHanlpCon(text);
|
||
|
||
SpeakerAnalysisHelper speakerFeatureLearner = new SpeakerAnalysisHelper();
|
||
|
||
List<string> ps = new List<string>();
|
||
|
||
string arg1 = string.Empty;
|
||
|
||
string person = GetArg0(hanlp, depRoot, ners, srls, cons, ref arg1);
|
||
|
||
//同边同时有的，放在最前面
|
||
if (!string.IsNullOrEmpty(person) && ners != null && ners.Where(x => x.Type.Contains("PERSON")).Count() > 0)
|
||
{
|
||
if (ners.Where(x => x.Text.Contains(person)).Any()) ps.Add(person);
|
||
}
|
||
//如果前面的不一样，那么加入人名
|
||
if (ners != null && ners.Where(x => x.Type.Contains("PERSON")).Count() > 0)
|
||
{
|
||
foreach (string namedEntity in ners.Where(x => x.Type.Contains("PERSON")).Select(x => x.Text).Distinct().ToArray())
|
||
{
|
||
if (namedEntity.Length > 1)
|
||
{
|
||
if (!ps.Contains(namedEntity))
|
||
ps.Add(namedEntity);
|
||
}
|
||
else
|
||
{
|
||
HanlpConstituencyNode[] nodes;
|
||
HanlpConstituencyNode nodef;
|
||
string feature = null;
|
||
|
||
nodes = HanlpResultHelper.GetNodeByLabel(cons, namedEntity);
|
||
if (nodes == null || nodes.Length == 0) continue;
|
||
|
||
string[] nps = new string[] { "NP", "DNP" };
|
||
//如果存在，找到父节点上的NP
|
||
foreach (HanlpConstituencyNode node in nodes)
|
||
{
|
||
nodef = HanlpResultHelper.GetNodeTopLabelById(cons, node.FatherId, nps);
|
||
if (nodef != null)
|
||
{
|
||
feature = HanlpResultHelper.GetNodeBottomLabelById(cons, nodef.ItemId, nps);
|
||
|
||
if (string.IsNullOrEmpty(feature)) continue;
|
||
}
|
||
else if (string.IsNullOrEmpty(feature))
|
||
{
|
||
feature = node.Label;
|
||
}
|
||
//如果下一个是“NP、DNP”，可以串起来，跟上面的GetNodeBottomLabelById类似
|
||
if (feature.Length == 1)
|
||
{
|
||
HanlpConstituencyNode[] nodesSon;
|
||
nodef = HanlpResultHelper.GetNodeById(cons, node.ItemId + 1);
|
||
if (nps.Contains(nodef.Label))
|
||
{
|
||
nodesSon = HanlpResultHelper.GetNodeSonById(cons, nodef.ItemId);
|
||
if (nodesSon != null && nodesSon.Where(x => x.Children == 0).Any())
|
||
{
|
||
feature += string.Concat(nodesSon.Select(x => x.Label));
|
||
}
|
||
}
|
||
}
|
||
|
||
if (feature.Length == 1)
|
||
{
|
||
var bs = HanlpResultHelper.GetNodeBrothersById(cons, node.ItemId);
|
||
|
||
if (bs != null && bs.Where(x => x.Children == 0).Any())
|
||
{
|
||
feature += string.Concat(bs.Select(x => x.Label));
|
||
}
|
||
}
|
||
|
||
if (feature.Contains(namedEntity) && !ps.Contains(feature)) ps.Add(feature);
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
//如果主语==动作发出者
|
||
//if (!string.IsNullOrEmpty(person) && !ps.Contains(person) && !ps.Contains(string.Format("[{0}]", person))) ps.Add(string.Format("[{0}]", person));
|
||
//动作接受者
|
||
//if (!string.IsNullOrEmpty(arg1) && !ps.Contains(arg1)) ps.Add(arg1);
|
||
|
||
person = QuotedString.ToDelimiteredList(ps.ToArray());
|
||
|
||
return person;
|
||
}
|
||
|
||
public string GetAgr0(Hanlp2 hanlp, string text)
|
||
{
|
||
string depRoot = hanlp.GetDepRoot(text);
|
||
|
||
List<NamedEntity> ners = hanlp.GetHanlpNer(text);
|
||
|
||
List<List<MeaningRepresentationParsingEntity>> srls = hanlp.GetHanlpSrl(text);
|
||
|
||
List<HanlpConstituencyNode> cons = hanlp.GetHanlpCon(text);
|
||
|
||
SpeakerAnalysisHelper speakerFeatureLearner = new SpeakerAnalysisHelper();
|
||
|
||
List<string> ps = new List<string>();
|
||
|
||
string arg1 = string.Empty;
|
||
|
||
return GetArg0(hanlp, depRoot, ners, srls, cons, ref arg1);
|
||
}
|
||
|
||
public string GetArg0(Hanlp2 hanlp, string depRoot, List<NamedEntity> ners, List<List<MeaningRepresentationParsingEntity>> srls, List<HanlpConstituencyNode> cons, ref string arg1)
|
||
{
|
||
string arg0 = string.Empty;
|
||
|
||
if (!string.IsNullOrEmpty(depRoot) && srls != null && srls.Count() > 0)
|
||
{
|
||
foreach (var srl in srls)
|
||
{
|
||
if (srl.Where(x => x.Text == depRoot).Any())
|
||
{
|
||
foreach (var mrp in srl)
|
||
{
|
||
if (mrp.Type.ToUpper() != "ARG0") continue;
|
||
|
||
if (ners != null && ners.Where(x => mrp.Text.Contains(x.Text)).Any() &&
|
||
ners.Where(x => mrp.Text.Contains(x.Text)).FirstOrDefault().Text.Length > 1)
|
||
return ners.Where(x => mrp.Text.Contains(x.Text)).FirstOrDefault().Text;
|
||
|
||
if (mrp.Text.Length > 4)
|
||
{
|
||
arg0 = GetAgr0(hanlp, mrp.Text);
|
||
}
|
||
else if (mrp.Text.Length > 1)
|
||
{
|
||
if (ners == null || ners.Where(x => x.Type.Contains("PERSON")).Count() == 0)
|
||
return mrp.Text;
|
||
}
|
||
}
|
||
}
|
||
|
||
foreach (var mrp in srl)
|
||
{
|
||
if (mrp.Type.ToUpper() != "ARG0") continue;
|
||
if (mrp.Text.Length > 4)
|
||
{
|
||
var temp = GetAgr0(hanlp, mrp.Text);
|
||
if (!string.IsNullOrEmpty(temp))
|
||
{
|
||
arg1 = temp;
|
||
return arg0;
|
||
}
|
||
}
|
||
else
|
||
{
|
||
arg1 = mrp.Text;
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
return arg0;
|
||
}
|
||
|
||
//public AiSentenceEntity[] AiParagraphs { get; set; }
|
||
|
||
public AiSentenceEntity[] AiSentences { get; set; }
|
||
|
||
public AiChapterPersonEntity[] AiChapterPersons { get; set; }
|
||
}
|
||
}
|
||
/*
|
||
“文本解析→人物提取→上下文匹配→结果输出”
|
||
|
||
graph TD
|
||
A[对话行] --> B{规则1：文本特征匹配};
|
||
B -->|匹配成功| C[直接赋值说话者];
|
||
B -->|匹配失败| D{规则2：紧邻旁白匹配};
|
||
D -->|匹配成功| C;
|
||
D -->|匹配失败| E{规则3：段落内最近人物};
|
||
E -->|匹配成功| C;
|
||
E -->|匹配失败| F[赋值默认值+标记异常];
|
||
|
||
规则 1：文本特征精准匹配（最高优先级）
|
||
基于对话内容中的 “专属特征” 直接匹配说话者，避免上下文干扰：
|
||
示例规则：
|
||
对话内容特征 匹配的说话者
|
||
包含 “第二元婴” 韩立
|
||
包含 “银月所化” 银月
|
||
包含 “本座 / 本尊” 反派角色
|
||
|
||
落地：维护一个 “特征 - 人物” 映射字典，匹配时优先使用：
|
||
规则 2：紧邻旁白匹配（次高优先级）只找对话行 “紧邻的前 1-2 行旁白”，而非整个段落的旁白，避免跨越多个人物导致匹配错误：
|
||
规则：仅检查对话行的前 1 行、前 2 行旁白（最多前 3 行），超过则停止；
|
||
落地：修改向前查找的逻辑，限制查找范围
|
||
|
||
规则 3：排除无效人物（避免干扰）从旁白中提取人物时，过滤掉 “非核心人物”（如 “店小二”“路人”），只保留主要人物：
|
||
落地：维护 “核心人物白名单”，提取人物时只保留白名单内的名称
|
||
|
||
3. 增加异常处理和人工校验（兜底保障）
|
||
即使逻辑再完善，也会有边缘场景匹配错误，需通过 “异常标记 + 人工复核” 兜底：
|
||
|
||
步骤 1：标记可疑匹配结果对以下情况标记为 “可疑”，存入日志或单独字段：
|
||
匹配到的人物在旁白中出现次数≥2（比如同一段落有韩立和银月，无法确定）；
|
||
对话行前后 3 行都没有找到人物，最终赋值为 “未知人物”；
|
||
对话内容无任何特征，且段落内人物数量≥3；
|
||
落地：为实体增加扩展字段（如IsSuspected），或生成异常日志
|
||
|
||
步骤 2：人工复核可疑数据
|
||
规则：所有标记为 “可疑” 的匹配结果，必须经过人工复核后再更新到数据库；
|
||
落地：生成 “可疑数据清单”（包含Guid、Content、PersonName），供运营 / 编辑人员校验，修正错误后重新填充。
|
||
|
||
|
||
收集错误案例：记录每次匹配错误的场景（如 “银月的对话匹配给韩立”），分析错误原因；
|
||
迭代特征字典：将错误案例转化为新的特征规则（如新增 “乌云”→银月，“元婴”→韩立）；
|
||
A/B 测试：对优化后的规则做小范围测试，对比匹配准确率，确认有效后全量上线。
|
||
*/ |