A Discriminative Semantic Ranker for Question Retrieval

Abstract

Similar question retrieval is a core task in community-based question answering (CQA) services. To balance the effectiveness and efficiency, the question retrieval system is typically implemented as multi-stage rankers: The first-stage ranker aims to recall potentially relevant questions from a large repository, and the latter stages attempt to re-rank the retrieved results. Most existing works on question retrieval mainly focused on the re-ranking stages, leaving the first-stage ranker to some traditional term-based methods. However, term-based methods often suffer from the vocabulary mismatch problem, especially on short texts, which may block the re-rankers from relevant questions at the very beginning. An alternative is to employ embedding-based methods for the first-stage ranker, which compress texts into dense vectors to enhance the semantic matching. However, these methods often lose the discriminative power as term-based methods, thus introduce noise during retrieval and hurt the recall performance. In this work, we aim to tackle the dilemma of the first-stage ranker, and propose a discriminative semantic ranker, namely DenseTrans, for highrecall retrieval. Specifically, DenseTrans is a densely connected Transformer, which learns semantic embeddings for texts based on Transformer layers. Meanwhile, DenseTrans promotes low-level features through dense connections to keep the discriminative power of the learned representations. DenseTrans is inspired by DenseNet in computer vision (CV), but poses a new way to use the dense connectivity which is totally different from its original design purpose. Experimental results over two question retrieval benchmark datasets show that our model can obtain significant gain on recall against strong term-based methods as well as stateof-the-art embedding-based methods.

Information Retrieval: Recent Advances and Beyond

“…Following this, the paper [115] propose a new method for training dense retrieval models by using pseudo-relevance feedback and multiple representations, which allows the model to learn more robust representations of queries. To further improve performance, [116] proposed a new method for training a discriminative semantic ranker for question retrieval. This approach focuses on training the model to differentiate between relevant and non-relevant documents, which is important for accurate retrieval.…”

Information Retrieval: Recent Advances and Beyond

“…As researchers strive to improve information retrieval, they have developed methods to train models that can effectively distinguish between relevant and non-relevant documents in dense retrieval settings. Cai et al [109] suggested a method for training a discriminative semantic ranker for question retrieval, focusing on this crucial aspect of accurate retrieval. To further refine the understanding of the relationship between passages and queries, Wu et al [110] proposed a representation decoupling method that improves open-domain passage retrieval by separating the encoding of passages and queries.…”

An Empirical Study of Uniform-Architecture Knowledge Distillation in Document Ranking

“…There are multiple architectures for BERT-based ranking models, including cross-encoder [12], dual encoders [2] and ColBERT [8]. Comparing with other architectures, cross-encoder can achieve the best performance, while its inference latency is the highest comparing with the other ones.…”

“…Most of existing approaches [7,[15][16][17]22] for language model distillation aim at improving the performance of student model in general natural language understanding (NLU) tasks [21]. Specially for document retrieval and ranking tasks, some latest approaches are focusing on cross-architecture distillation approaches [5,10] using cross-encoder [12] teachers and ColBERT [8] or dual-encoder [2] based students. For those cross-architecture approaches, both teachers and students are using pre-trained models in the same scale (i.e.…”