File size: 5,687 Bytes
801a2fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
[
  {
    "title": "Indirect abuse: baseline misses, A and B catch",
    "language": "English",
    "text": "I didn't know the right wing loved islam",
    "true_label": "harmful",
    "sarcasm_score": "0.75",
    "source": "cad",
    "models": [
      {
        "name": "Baseline",
        "prob": "0.45",
        "decision": "non-harmful (MISS)"
      },
      {
        "name": "Mechanism A",
        "prob": "--",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism B",
        "prob": "0.61",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism D",
        "prob": "0.60",
        "decision": "HARMFUL"
      }
    ],
    "story": "CAD is where the mechanisms' recall gains concentrate (baseline recall 0.389 -> B 0.593, p<1e-5): contextual, indirect abuse benefits most from the auxiliary non-literality signal."
  },
  {
    "title": "Tunisian dialect: the worry that turned out backwards",
    "language": "Arabic",
    "text": "اتقوا انفسكم ياتوانسه",
    "true_label": "harmful",
    "sarcasm_score": "0.21",
    "source": "thsab",
    "models": [
      {
        "name": "Baseline",
        "prob": "0.26",
        "decision": "non-harmful (MISS)"
      },
      {
        "name": "Mechanism A",
        "prob": "--",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism B",
        "prob": "0.80",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism D",
        "prob": "0.18",
        "decision": "non-harmful (MISS)"
      }
    ],
    "story": "Despite zero Tunisian data in the sarcasm layer's training pool, A and B are significantly effective on T-HSAB specifically (McNemar p=0.015 / p=0.029) -- the suspected weak point was verified to be a strength."
  },
  {
    "title": "Sarcastic camouflage: high sarcasm score, hostile intent",
    "language": "Arabic",
    "text": "يلعن صباحك نزعتلي نهاري",
    "true_label": "harmful",
    "sarcasm_score": "0.85",
    "source": "lhsab",
    "models": [
      {
        "name": "Baseline",
        "prob": "0.03",
        "decision": "non-harmful (MISS)"
      },
      {
        "name": "Mechanism A",
        "prob": "--",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism B",
        "prob": "0.98",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism D",
        "prob": "0.93",
        "decision": "HARMFUL"
      }
    ],
    "story": "The motivating case of the thesis: superficially non-hostile or joking wording carrying real hostility, recovered when the sarcasm signal is available."
  },
  {
    "title": "The honest trade-off: a false positive the mechanisms introduce",
    "language": "English",
    "text": "@user “ lets go gays ”",
    "true_label": "non-harmful",
    "sarcasm_score": "0.48",
    "source": "hatexplain",
    "models": [
      {
        "name": "Baseline",
        "prob": "0.43",
        "decision": "non-harmful"
      },
      {
        "name": "Mechanism A",
        "prob": "--",
        "decision": "HARMFUL (false positive)"
      },
      {
        "name": "Mechanism B",
        "prob": "0.77",
        "decision": "HARMFUL (false positive)"
      },
      {
        "name": "Mechanism D",
        "prob": "0.51",
        "decision": "HARMFUL (false positive)"
      }
    ],
    "story": "Every mechanism trades precision for recall. Showing a failure case demonstrates the evaluation was honest: the operating point shifts toward recall; the mechanisms do not dominate the baseline on both axes."
  },
  {
    "title": "Mechanism D's cross-lingual cost: Arabic harmful D misses",
    "language": "Arabic",
    "text": "صحيح انك انسان وضيع",
    "true_label": "harmful",
    "sarcasm_score": "0.83",
    "source": "lhsab",
    "models": [
      {
        "name": "Baseline",
        "prob": "0.99",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism A",
        "prob": "--",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism B",
        "prob": "0.98",
        "decision": "HARMFUL"
      },
      {
        "name": "Mechanism D",
        "prob": "0.05",
        "decision": "non-harmful (MISS)"
      }
    ],
    "story": "Mechanism D helps English but degrades Arabic (recall 0.8155 -> 0.7729, p=0.0009) -- a cost the constant-score ablation showed is independent of score values, and which generalized to French in zero-shot transfer (recall collapse to 0.15-0.18)."
  },
  {
    "title": "Mechanism D2: what reasoning injection looks like (pilot)",
    "language": "Arabic",
    "text": "صحيح انك انسان وضيع",
    "true_label": "harmful",
    "sarcasm_score": "0.83",
    "source": "lhsab",
    "models": [
      {
        "name": "Matched baseline",
        "prob": "--",
        "decision": "non-harmful (MISS)"
      },
      {
        "name": "Matched D (score)",
        "prob": "--",
        "decision": "HARMFUL"
      },
      {
        "name": "D2 (reasoning)",
        "prob": "--",
        "decision": "HARMFUL"
      }
    ],
    "story": "Generated reasoning for this example: “Literal meaning: It is true that you are a worthless person. Implied meaning: The speaker is insulting the addressee's worthiness and dignity. Speaker intention: To demean and criticize the addressee's value and status. Indirect language: No -- the statement is made directly without any implied irony or sarcasm.” — D2 significantly beats its matched baseline (p=0.0019, replicated with a second generator, p=0.0038) but does not exceed bare-score injection: the channel, not the content, is the binding constraint."
  }
]