504. Feedback Prize - Predicting Effective Arguments | feedback-prize-effectiveness
恭喜大家,感谢佐治亚州立大学、Kaggle 和所有组织者举办了如此精彩的比赛。我们很遗憾在两个赛道的比赛中都以微弱差距与奖项失之交臂,但我们将带着这些收获继续前行!
我们的解决方案基于 deberta-v3、MLM(掩码语言模型)、伪标签和 LGB 堆叠。我们发现结合多种 NLP 改进技术可以显著提高分数。
我们解决方案的核心是按顺序组合所有语篇 ID,然后在开头和结尾添加语篇类型 ID 和特殊标记。
使用了一个 cls_position 作为语篇的开始或结束标记。Python 代码如下:
input_ids = []
essay_df = = trndf.loc[essay_id]
for i in range(len(essay_df)):
text = f'{discourse_types[i]}: {discourse_texts[i]} {self.tokenizer.sep_token} :{discourse_types[i]}'
i_ids = self.tokenizer.encode(text, add_special_tokens=False)
input_ids += i_ids
cls_pos += [0] * (len(i_ids) - 1) + [1]
输出示例如下:
Lead: Computers can do many things for us. The idea that a computer can read your emotions is amazing [SEP] :Lead Position: I do not belive it is true. I will believe it when I see it. [SEP] :Position Evidence: In paragraph 8 it says " But in a false smile, the mouth is stretched sideways ising the zygomatic major and different muscle, the risorius." however this may be true for most people; there has to be someone out there in our world of 7 billion that smiles there smiles with their zygomatic muscle or their risorius muscle [SEP] :Evidence Claim: Everyone has diffrent emotions and everyone shows them diffrently [SEP] :Claim Counterclaim: The muscles in our face does say a lot about the emotions we are feeling. [SEP] :Counterclaim Concluding Statement: This is why I believe computures can not read your emotions [SEP] :Concluding Statement
我们还使用了辅助标签,如排名(Ineffective -> Adequate -> Effective)和文章主题聚类。
我们还将未包含在标注语篇中的文章文本添加到了末尾,有时也会穿插在语篇之间。
我们发现,在部分模型中用特殊标记替换换行符并清理文本,可以显著增加整体融合模型的多样性。
我们使用了三种不同的模型流程,在数据预处理、辅助标签和模型头部方面做了一些细微调整。我们仅使用 deberta-v3 作为骨干网络。其他骨干网络在交叉验证(CV)中未能提升我们的融合效果。
重要的是要在所有模型中移除 deberta 隐藏层的 dropout