Benyou Wang

IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models

This paper provides a unified account of two schools of thinking in information retrieval modelling: the generative retrieval focusing on predicting relevant documents given a query, and the discriminative retrieval focusing on predicting relevancy given a query-document pair. We propose a game theoretical minimax game to iteratively optimise both models. On one hand, the discriminative model, aiming to mine signals from labelled and unlabelled data, provides guidance to train the generative model towards fitting the underlying relevance distribution over documents given the query. On the other hand, the generative model, acting as an attacker to the current discriminative model, generates difficult examples for the discriminative model in an adversarial way by minimising its discrimination objective. With the competition between these two models, we show that the unified framework takes advantage of both schools of thinking: (i) the generative model learns to fit the relevance distribution over documents via the signals from the discriminative model, and (ii) the discriminative model is able to exploit the unlabelled data selected by the generative model to achieve a better estimation for document ranking. Our experimental results have demonstrated significant performance gains as much as 23.96% on Precision@5 and 15.50% on MAP over strong baselines in a variety of applications including web search, item recommendation, and question answering.

Exploration of Quantum Interference in Document Relevance Judgement Discrepancy

Quantum theory has been applied in a number of fields outside physics, e.g., cognitive science and information retrieval (IR). Recently, it has been shown that quantum theory can subsume various key IR models into a single mathematical formalism of Hilbert vector spaces. While a series of quantum-inspired IR models has been proposed, limited effort has been devoted to verify the existence of the quantum-like phenomenon in real users’ information retrieval processes, from a real user study perspective. In this paper, we aim to explore and model the quantum interference in users’ relevance judgement about documents, caused by the presentation order of documents. A user study in the context of IR tasks have been carried out. The existence of the quantum interference is tested by the violation of the law of total probability and the validity of the order effect. Our main findings are: (1) there is an apparent judging discrepancy across different users and document presentation orders, and empirical data have violated the law of total probability; (2) most search trials recorded in the user study show the existence of the order effect, and the incompatible decision perspectives in the quantum question (QQ) model are valid in some trials. We further explain the judgement discrepancy in more depth, in terms of four effects (comparison, unfamiliarity, attraction and repulsion) and also analyse the dynamics of document relevance judgement in terms of the evolution of the information need subspace.

A Quantum Query Expansion Approach for Session Search

Recently, Quantum Theory (QT) has been employed to advance the theory of Information Retrieval (IR). Various analogies between QT and IR have been established. Among them, a typical one is applying the idea of photon polarization in IR tasks, e.g., for document ranking and query expansion. In this paper, we aim to further extend this work by constructing a new superposed state of each document in the information need space, based on which we can incorporate the quantum interference idea in query expansion. We then apply the new quantum query expansion model to session search, which is a typical Web search task. Empirical evaluation on the large-scale Clueweb12 dataset has shown that the proposed model is effective in the session search tasks, demonstrating the potential of developing novel and effective IR models based on intuitions and formalisms of QT.

A Chinese Question Answering Approach Integrating Count-Based and Embedding-Based Features

Document-based Question Answering system, which needs to match semantically the short text pairs, has gradually become an important topic in the fields of natural language processing and information retrieval. Question Answering system based on English corpus has developed rapidly with the utilization of the deep learning technology, whereas an effective Chinese-customized system needs to be paid more attention. Thus, we explore a Question Answering system which is characterized in Chinese for the QA task of NLPCC. In our approach, the ordered sequential information of text and deep matching of semantics of Chinese textual pairs have been captured by our count-based traditional methods and embedding-based neural network. The ensemble strategy has achieved a good performance which is much stronger than the provided baselines.

Learning to Diversify Web Search Results with a Document Repulsion Model

Search diversification (also called diversity search), is an important approach to tackling the query ambiguity problem in information retrieval. It aims to diversify the search results that are originally ranked according to their probabilities of relevance to a given query, by re-ranking them to cover as many as possible different aspects (or subtopics) of the query. Most existing diversity search models heuristically balance the relevance ranking and the diversity ranking, yet lacking an efficient learning mechanism to reach an optimized parameter setting. To address this problem, we propose a learning-to-diversify approach which can directly optimize the search diversification performance (in term of any effectiveness metric). We first extend the ranking function of a widely used learning-to-rank framework, i.e., LambdaMART, so that the extended ranking function can correlate relevance and diversity indicators. Furthermore, we develop an effective learning algorithm, namely Document Repulsion Model (DRM), to train the ranking function based on a Document Repulsion Theory (DRT). DRT assumes that two result documents covering similar query aspects (i.e., subtopics) should be mutually repulsive, for the purpose of search diversification. Accordingly, the proposed DRM exerts a repulsion force between each pair of similar documents in the learning process, and includes the diversity effectiveness metric to be optimized as part of the loss function. Although there have been existing learning based diversity search methods, they often involve an iterative sequential selection process in the ranking process, which is computationally complex and time consuming for training, while our proposed learning strategy can largely reduce the time cost. Extensive experiments are conducted on the TREC diversity track data (2009, 2010 and 2011). The results demonstrate that our model significantly outperforms a number of baselines in terms of effectiveness and robustness. Further, an efficiency analysis shows that the proposed DRM has a lower computational complexity than the state of the art learning-to-diversify methods.

PLASTIC: Prioritize Long and Short-term Information in Top-n Recommendation using Adversarial Training

Recommender systems provide users with ranked lists of items based on individual’s preferences and constraints. Two types of models are commonly used to generate ranking results: long-term models and session-based models. While long-term models represent the interactions between users and items that are supposed to change slowly across time, session-based models encode the information of users’ interests and changing dynamics of items’ attributes in short terms. In this paper, we propose a PLASTIC model, Prioritizing Long And ShortTerm Information in top-n reCommendation using adversarial training. In the adversarial process, we train a generator as an agent of reinforcement learning which recommends the next item to a user sequentially. We also train a discriminator which attempts to distinguish the generated list of items from the real list recorded. Extensive experiments show that our model exhibits significantly better performances on two widely used datasets.1

A Multi-task Learning Approach for Image Captioning

In this paper, we propose a Multi-task Learning Approach for Image Captioning (MLAIC), motivated by the fact that humans have no difficulty performing such task because they have the capabilities of multiple domains. Specifically, MLAIC consists of three key components: (i) A multi-object classification model that learns rich category-aware image representations using a CNN image encoder; (ii) A syntax generation model that learns better syntaxaware LSTM based decoder; (iii) An image captioning model that generates image descriptions in text, sharing its CNN encoder and LSTM decoder with the object classification task and the syntax generation task, respectively. In particular, the image captioning model can benefit from the additional object categorization and syntax knowledge. The experimental results on MS-COCO dataset demonstrate that our model achieves impressive results compared to other strong competitors.1

A quantum-inspired multimodal sentiment analysis framework

Abstract Multimodal sentiment analysis aims to capture diversified sentiment information implied in data that are of different modalities (e.g., an image that is associated with a textual description or a set of textual labels). The key challenge is rooted on the “semantic gap” between different low-level content features and high-level semantic information. Existing approaches generally utilize a combination of multimodal features in a somehow heuristic way. However, how to employ and combine multiple information from different sources effectively is still an important yet largely unsolved problem. To address the problem, in this paper, we propose a Quantum-inspired Multimodal Sentiment Analysis (QMSA) framework. The framework consists of a Quantum-inspired Multimodal Representation (QMR) model (which aims to fill the “semantic gap” and model the correlations between different modalities via density matrix), and a Multimodal decision Fusion strategy inspired by Quantum Interference (QIMF) in the double-slit experiment (in which the sentiment label is analogous to a photon, and the data modalities are analogous to slits). Extensive experiments are conducted on two large scale datasets, which are collected from the Getty Images and Flickr photo sharing platform. The experimental results show that our approach significantly outperforms a wide range of baselines and state-of-the-art methods.

Enhanced Embedding Based Attentive Pooling Network for Answer Selection

Document-based Question Answering tries to rank the candidate answers for given questions, which needs to evaluate matching score between the question sentence and answer sentence. Existing works usually utilize convolution neural network (CNN) to adaptively learn the latent matching pattern between the question/answer pair. However, CNN can only perceive the order of a word in a local windows, while the global order of the windows is ignored due to the window-sliding operation. In this report, we design an enhanced CNN (https://github.com/shuishen112/pairwise-deep-qa) with extended order information (e.g. overlapping position and global order) into inputting embedding, such rich representation makes it possible to learn an order-aware matching in CNN. Combining with standard convolutional paradigm like attentive pooling, pair-wise training and dynamic negative sample, this end-to-end CNN achieve a good performance on the DBQA task of NLPCC 2017 without any other extra features.

Regularizing Neural Networks via Retaining Confident Connections

Regularization of neural networks can alleviate overfitting in the training phase. Current regularization methods, such as Dropout and DropConnect, randomly drop neural nodes or connections based on a uniform prior. Such a data-independent strategy does not take into consideration of the quality of individual unit or connection. In this paper, we aim to develop a data-dependent approach to regularizing neural network in the framework of Information Geometry. A measurement for the quality of connections is proposed, namely confidence. Specifically, the confidence of a connection is derived from its contribution to the Fisher information distance. The network is adjusted by retaining the confident connections and discarding the less confident ones. The adjusted network, named as ConfNet, would carry the majority of variations in the sample data. The relationships among confidence estimation, Maximum Likelihood Estimation and classical model selection criteria (like Akaike information criterion) is investigated and discussed theoretically. Furthermore, a Stochastic ConfNet is designed by adding a self-adaptive probabilistic sampling strategy. The proposed data-dependent regularization methods achieve promising experimental results on three data collections including MNIST, CIFAR-10 and CIFAR-100.