Information Retrieval has a long history of applying either discriminative or generative modeling to retrieval and ranking tasks. Recent developments in transformer architectures and multi-task learning techniques have dramatically improved our ability to train effective neural models capable of resolving a wide variety of tasks using either of these paradigms. In this paper, we propose a novel multi-task learning approach which can be used to produce more effective neural ranking models. The key idea is to improve the quality of the underlying transformer model by cross-training a retrieval task and one or more complementary language generation tasks. By targeting the training on the encoding layer in the transformer architecture, our experimental results show that the proposed multi-task learning approach consistently improves retrieval effectiveness on the targeted collection and can easily be re-targeted to new ranking tasks. We provide an in-depth analysis showing how multi-task learning modifies model behaviors, resulting in more general models.

As one of the most popular services over online platforms, social recommendation has attracted increasing research efforts recently. Among all the recommendation tasks, an important one is item recommendation over high speed social media streams. Existing stream recommendation techniques are not effective for handling social users with diverse interests. Meanwhile, approaches for recommending items to a particular user are not efficient when applied to a huge number of users over high speed streams. In this paper, we propose a novel framework for the social recommendation over streams. Specifically, we first propose a novel Bi-Layer Hidden Markov Model (BiHMM) that adaptively captures the users' behaviors and their interactions with influential official accounts to predict their long-term and short-term interests. Then, we design a new probabilistic entity matching scheme for identifying the relevance score of a streaming item to a user. Moreover, we propose a novel index scheme called CPPse-index for improving the efficiency of our solution. Extensive tests are conducted to prove the superiority of our approach in terms of the recommendation quality and time cost.

Allocating resources to customers in the customer service is a difficult problem, because designing an optimal strategy to achieve an optimal trade-off between available resources and customers' satisfaction is non-trivial. In this paper, we formalize the customer routing problem, and propose a novel framework based on deep reinforcement learning (RL) to address this problem. To make it more practical, a demo is provided to show and compare different models, which visualizes all decision process, and in particular, the system shows how the optimal strategy is reached. Besides, our demo system also ships with a variety of models that users can choose based on their needs.

Direct answering of questions that involve multiple entities and relations is a challenge for text-based QA. This problem is most pronounced when answers can be found only by joining evidence from multiple documents. Curated knowledge graphs (KGs) may yield good answers, but are limited by their inherent incompleteness and potential staleness. This paper presents QUEST, a method that can answer complex questions directly from textual sources on-the-fly, by computing similarity joins over partial results from different documents. Our method is completely unsupervised, avoiding training-data bottlenecks and being able to cope with rapidly evolving ad hoc topics and formulation style in user questions. QUEST builds a noisy quasi KG with node and edge weights, consisting of dynamically retrieved entity names and relational phrases. It augments this graph with types and semantic alignments, and computes the best answers by an algorithm for Group Steiner Trees. We evaluate QUEST on benchmarks of complex questions, and show that it substantially outperforms state-of-the-art baselines.

Increasing test collection sizes and limited judgment budgets create measurement challenges for IR batch evaluations, challenges that are greater when using deep effectiveness metrics than when using shallow metrics, because of the increased likelihood that unjudged documents will be encountered. Here we study the problem of metric score adjustment, with the goal of accurately estimating system performance when using deep metrics and limited judgment sets, assuming that dynamic score adjustment is required per topic due to the variability in the number of relevant documents. We seek to induce system orderings that are as close as is possible to the orderings that would arise if full judgments were available. Starting with depth-based pooling, and no prior knowledge of sampling probabilities, the first phase of our two-stage process computes a background gain for each document based on rank-level statistics. The second stage then accounts for the distributional variance of relevant documents. We also exploit the frequency statistics of pooled relevant documents in order to determine a threshold for dynamically determining the set of topics to be adjusted. Taken together, our results show that: (i) better score estimates can be achieved when compared to previous work; (ii) by setting a global threshold, we are able to adapt our methods to different collections; and (iii) the proposed estimation methods reliably approximate the system orderings achieved when many more relevance judgments are available. We also consider pools generated by a two-strata sampling approach.

Sophisticated ranking mechanisms make use of term dependency features in order to compute similarity scores for documents. These features often include exact phrase occurrences, and term proximity estimates. Both cases build on the intuition that if multiple query terms appear near each other, the document is more likely to be relevant to the query. In this paper we examine the processes used to compute these statistics. Two distinct input structures can be used -- inverted files and direct files. Inverted files must store the position offsets of the terms, while "direct" files represent each document as a sequence of preprocessed term identifiers. Based on these two input modalities, a number of algorithms can be used to compute proximity statistics. Until now, these algorithms have been described in terms of a single set of query terms. But similarity computations such as the Full Dependency Model compute proximity statistics for a collection of related term sets. We present a new approach in which such collections are processed holistically in time that is much less than would be the case if each subquery were to be evaluated independently. The benefits of the new method are demonstrated by a comprehensive experimental study.

No abstract available.