Middle training methods aim to bridge the gap between the Masked Language Model (MLM) pre-training and the final finetuning for retrieval. Recent models such as CoCondenser, RetroMAE, and LexMAE argue that the MLM task is not sufficient enough to pre-train a transformer network for retrieval and hence propose various tasks to do so. Intrigued by those novel methods, we noticed that all these models used different finetuning protocols, making it hard to assess the benefits of middle training. We propose in this paper a benchmark of CoCondenser, RetroMAE, and LexMAE, under the same finetuning conditions. We compare both dense and sparse approaches under various finetuning protocols and middle training on different collections (MS MARCO, Wikipedia or Tripclick). We use additional middle training baselines, such as a standard MLM finetuning on the retrieval collection, optionally augmented by a CLS predicting the passage term frequency. For the sparse approach, our study reveals that there is almost no statistical difference between those methods: the more effective the finetuning procedure is, the less difference there is between those models. For the dense approach, RetroMAE using MS MARCO as middle-training collection shows excellent results in almost all the settings. Finally, we show that middle training on the retrieval collection, thus adapting the language model to it, is a critical factor. Overall, a better experimental setup should be adopted to evaluate middle training methods. Code available at https://github.com/naver/splade/tree/benchmarch-SIGIR23
Sparse neural retrievers, such as DeepImpact, uniCOIL and SPLADE, have been introduced recently as an efficient and effective way to perform retrieval with inverted indexes. They aim to learn term importance and, in some cases, document expansions, to provide a more effective document ranking compared to traditional bag-of-words retrieval models such as BM25. However, these sparse neural retrievers have been shown to increase the computational costs and latency of query processing compared to their classical counterparts. To mitigate this, we apply a well-known family of techniques for boosting the efficiency of query processing over inverted indexes: static pruning. We experiment with three static pruning strategies, namely document-centric, term-centric and agnostic pruning, and we assess, over diverse datasets, that these techniques still work with sparse neural retrievers. In particular, static pruning achieves $2\times$ speedup with negligible effectiveness loss ($\leq 2\%$ drop) and, depending on the use case, even $4\times$ speedup with minimal impact on the effectiveness ($\leq 8\%$ drop). Moreover, we show that neural rerankers are robust to candidates from statically pruned indexes.
Neural retrieval models have acquired significant effectiveness gains over the last few years compared to term-based methods. Nevertheless, those models may be brittle when faced to typos, distribution shifts or vulnerable to malicious attacks. For instance, several recent papers demonstrated that such variations severely impacted models performances, and then tried to train more resilient models. Usual approaches include synonyms replacements or typos injections -- as data-augmentation -- and the use of more robust tokenizers (characterBERT, BPE-dropout). To further complement the literature, we investigate in this paper adversarial training as another possible solution to this robustness issue. Our comparison includes the two main families of BERT-based neural retrievers, i.e. dense and sparse, with and without distillation techniques. We then demonstrate that one of the most simple adversarial training techniques -- the Fast Gradient Sign Method (FGSM) -- can improve first stage rankers robustness and effectiveness. In particular, FGSM increases models performances on both in-domain and out-of-domain distributions, and also on queries with typos, for multiple neural retrievers.
Recent IR approaches based on Pretrained Language Models (PLM) have now largely outperformed their predecessors on a variety of IR tasks. However, what happens to learned word representations with distribution shifts remains unclear. Recently, the BEIR benchmark was introduced to assess the performance of neural rankers in zero-shot settings and revealed deficiencies for several models. In complement to BEIR, we propose to control \textit{explicitly} distribution shifts. We selected different query subsets leading to different distribution shifts: short versus long queries, wh-words types of queries and 5 topic-based clusters. Then, we benchmarked state of the art neural rankers such as dense Bi-Encoder, SPLADE and ColBERT under these different training and test conditions. Our study demonstrates that it is possible to design distribution shift experiments within the MSMARCO collection, and that the query subsets we selected constitute an additional benchmark to better study factors of generalization for various models.
MeSH (Medical Subject Headings) is a large thesaurus created by the National Library of Medicine and used for fine-grained indexing of publications in the biomedical domain. In the context of the COVID-19 pandemic, MeSH descriptors have emerged in relation to articles published on the corresponding topic. Zero-shot classification is an adequate response for timely labeling of the stream of papers with MeSH categories. In this work, we hypothesise that rich semantic information available in MeSH has potential to improve BioBERT representations and make them more suitable for zero-shot/few-shot tasks. We frame the problem as determining if MeSH term definitions, concatenated with paper abstracts are valid instances or not, and leverage multi-task learning to induce the MeSH hierarchy in the representations thanks to a seq2seq task. Results establish a baseline on the MedLine and LitCovid datasets, and probing shows that the resulting representations convey the hierarchical relations present in MeSH.