Reading Wikipedia to Answer Open-Domain Questions (2017)
Document Retriever
Document Reader
個人總結
Dense Passage Retrieval for Open-Domain Question Answering (2020)
Encoders
Training
End-to-end QA System
個人總結
Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering (2020)
Encoder
Decoder
result
個人總結
DISTILLING KNOWLEDGE FROM READER TO RETRIEVER FOR QUESTION ANSWERING (2021)
Reader
CROSS-ATTENTION SCORE
Retriever
Training
Result
個人總結
題外話
前言
這篇文章是年前欠下來的,當時在選擇比賽的項目時,SDP @NAACL 的第二項任務SCIVER: Verifying Scientific Claims with Evidence,和開放域問答系統類型非常相似,因此調研了一些開放域問答的經典文獻和SOTA方法,在這里做一個總結。通過這4篇論文,可以對開放域問答任務的研究方向有一個基本的了解,從統計特征到可訓練特征再到無樣本學習,從span抽取到直接生成。
涉及論文:
Reading Wikipedia to Answer Open-Domain Questions
Dense Passage Retrieval for Open-Domain Question Answering
Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering
DISTILLING KNOWLEDGE FROM READER TO RETRIEVER FOR QUESTION ANSWERING
SCIVER: Verifying Scientific Claims with Evidence
任務介紹
Due to the rapid growth in scientific literature, it is difficult for scientists to stay up-to-date on the latest findings. This challenge is especially acute during pandemics due to the risk of making decisions based on outdated or incomplete information. There is a need for AI systems that can help scientists with information overload and support scientific fact checking and evidence synthesis.
{"id": 52,"claim": "ALDH1 expression is associated with poorer prognosis for breast cancer primary tumors.","evidence": {"11": [ // 2 evidence sets in document 11 support the claim.{"sentences": [0, 1], // Sentences 0 and 1, taken together, support the claim."label": "SUPPORT"},{"sentences": [11], // Sentence 11, on its own, supports the claim."label": "SUPPORT"}],"15": [ // A single evidence set in document 15 supports the claim.{"sentences": [4], "label": "SUPPORT"}]},"cited_doc_ids": [11, 15]}
Reading Wikipedia to Answer Open-Domain Questions (2017)
作為開放域問答的經典論文,Danqi Chen提出的DrQA,搭建了開放域問答解決系統的基本框架: (1) the Document Retriever module for finding relevant articles and (2) a machine comprehension model, Document Reader, for extracting answers from a single document or a small collection of documents.
Word embeddings: 300-dimensional Glove word embeddings trained from 840B Web crawl data. 作者固定了詞向量的embedding在模型訓練時只fine-tune前1000高頻的詞向量,原因在于認為像what, how, which 這種高頻的提問詞可能對于QA systems來說更加關鍵。
與傳統的transformer的decoder相似,每一個block中先做一次輸出的self-attention,之后的通過Q=WQH,K=WKX,V=WVXQ = W_QH, K = W_KX, V = W_VXQ=WQ?H,K=WK?X,V=WV?X來計算cross-attention,具體計算公式與transformers一致,不再展開敘述。其中X為encoder輸出的最后的特征表示,H為decoder中上一個self-attention的輸出。
根據Fusion-in-Decoder論文提到的,decoder中的cross-attention計算公式如下:Q=WQH,K=WKX,V=WVXQ = W_QH, K = W_KX, V = W_VXQ=WQ?H,K=WK?X,V=WV?X
作者假設α:,jα_{:,j}α:,j?可以用來度量第 j 個key token對于模型通過value計算下一個特征表示的重要性,并以此作為與該key token對應文檔的重要性的代表——the more the tokens in a text segment are attended to, the more relevant the text segment is to answer the question.
DISTILLING KNOWLEDGE FROM READER TO RETRIEVER FOR QUESTION ANSWERING這篇文章的方法可拓展性強嗎?由于其Reader模塊采用的是Seq2Seq結構配合問答,是否存在特殊性?適用于其他的召回 + X 系統嗎(排序/打分等)
其實最近也比較流行用NLG完成NLU任務的論文,All NLP Tasks Are Generation Tasks: A General Pretraining Framework,GPT Understands, Too等。這種召回+生成框架遷移的根本難點在于我們如何針對不同的任務類型進行設計不同的“Decoder”。值得研究~