（完成）SIGIR2020

• （TKL）Local Self-Attention over Long Text for Efficient Document Retrieval

  对Transformer-kernel的改进，主要是针对Transformer-kernel用于长文本时。之前的基于Transformer的模型用于长文本时通常的做法是直接截断。本文在Transformer-kernel模型的基础上提出了local self-attention。并且计算出每个区域的相关度之后再使用top-local-max机制组合得到全局相关度。 数据集：MARCO的document retrieval数据集

• Training Curricula for Open Domain Answer Re-Ranking

  for passage re-ranking 阶段。

  提出课程学习策略，改进BERT和Conv-KNRM的训练过程，刚开始给容易的passage更大的权重，给难的passage更小的权重，后期给相同的权重。

• ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT

• 【PreTTR】Efficient Document Re-Ranking for Transformers by Precomputing Term Representations

• SetRank: Learning a Permutation-Invariant Ranking Model for Information Retrieval

• Open-Retrieval Conversational Question Answering

  提出了一个开放域对话式QA数据集：OR-QuAC。

  提出了一个baseline，包括retriever、re-ranker，reader。

• Match$^2$: A Matching over Matching Model for Similar Question Identification

• MarkedBERT: Integrating Traditional IR cues in pre-trained language models for passage retrieval

  We proposed MarkedBERT that incorporates Exact Matching signals via a simple yet effective marking technique that only modifies the model input.

• Context-Aware Term Weighting For First-Stage Passage Retrieval

• Learning Term Discrimination

• Improving Contextual Language Models for Response Retrieval in Multi-Turn Conversation

  对BERT类的上下文预训练模型提出针对多轮对话任务的两种改进策略：1）Speaker Segmentation；2）Dialogue Augmentation

• Large-scale Image Retrieval with Sparse Binary Projections

• 【EPIC】Expansion via Prediction of Importance with Contextualization

  把query表示成稀疏向量，document表示成dense向量，计算点积。

• Efficiency Implications of Term Re-Weighting for Passage Retrieval

• DC-BERT: Decoupling Question and Document for Efficient Contextual Encoding

• Improving Matching Models with Hierarchical Contextualized Representations for Multi-turn Response Selection

• Distilling Knowledge for fast retrieval-based chat-bots

  we introduced an enhanced BERT cross-encoder architecture modified for the task of response retrieval. Alongside that, we utilized knowledge distillation to compress the complex BERT cross-encoder network as a teacher model into the student BERT bi-encoder model. This increases the BERT bi-encoders prediction quality without affecting its inference speed.

• Having Your Cake and Eating it Too: Training Neural Retrieval for Language Inference without Losing Lexical Match

  We presented a simple approach to infuse lexical matching using unsupervised IR methods into a state-of-the-art transformer method - RoBERTa. We show that infusing lexical-matching improves the performance on simpler retrieval based question and the (justification) retrieval task itself。

• An analysis of BERT in document ranking

  What Does BERT Look At? An Analysis of BERT’s Attention.

  How Contextual are Contextualized Word Representations.

  Understanding the Behaviors of BERT in Ranking（刘知远）

• Read, Attend, and Exclude: Multi-Choice Reading Comprehension by Mimicking Human Reasoning Process

  提出了一个针对多选型阅读理解任务的模型。

• Unsupervised Text Summarization with Sentence Graph Compression

  对多篇文档生成摘要。

---