Advances in Information Retrieval

Different users may be interested in different clustering views underlying a given collection (e.g., topic and writing style in documents). Enabling them to provide constraints reflecting their needs can then help obtain tailored clustering results. For document clustering, constraints can be provided in the form of seed words, each cluster being characterized by a small set of words. This seed-guided constrained document clustering problem was recently addressed through topic modeling approaches. In this paper, we jointly learn deep representations and bias the clustering results through the seed words, leading to a Seed-guided Deep Document Clustering approach. Its effectiveness is demonstrated on five public datasets.

Knowledge graphs’ incompleteness has motivated many researchers to propose methods to automatically infer missing facts in knowledge graphs. Knowledge graph embedding has been an active research area for knowledge graph completion, with great improvement from the early TransE to the current state-of-the-art ConvKB. ConvKB considers a knowledge graph as a set of triples, and employs a convolutional neural network to capture global relationships and transitional characteristics between entities and relations in the knowledge graph. However, it only utilizes the triple information, and ignores the rich information contained in relation paths. In fact, a path of one relation describes the relation from some aspect in a fine-grained way. Therefore, it is beneficial to take relation paths into consideration for knowledge graph embedding. In this paper, we present a novel convolutional neural network-based embedding model PConvKB, which improves knowledge graph embedding by incorporating relation paths locally and globally. Specifically, we introduce attention mechanism to measure the local importance of relation paths. Moreover, we propose a simple yet effective measure DIPF to compute the global importance of relation paths. Experimental results show that our model achieves substantial improvements against state-of-the-art methods.

Analyzing the readability of articles has been an important sociolinguistic task. Addressing this task is necessary to the automatic recommendation of appropriate articles to readers with different comprehension abilities, and it further benefits education systems, web information systems, and digital libraries. Current methods for assessing readability employ empirical measures or statistical learning techniques that are limited by their ability to characterize complex patterns such as article structures and semantic meanings of sentences. In this paper, we propose a new and comprehensive framework which uses a hierarchical self-attention model to analyze document readability. In this model, measurements of sentence-level difficulty are captured along with the semantic meanings of each sentence. Additionally, the sentence-level features are incorporated to characterize the overall readability of an article with consideration of article structures. We evaluate our proposed approach on three widely-used benchmark datasets against several strong baseline approaches. Experimental results show that our proposed method achieves the state-of-the-art performance on estimating the readability for various web articles and literature.

We address and formalise the task of sequence-to-sequence (seq2seq) cross-modal retrieval. Given a sequence of text passages as query, the goal is to retrieve a sequence of images that best describes and aligns with the query. This new task extends the traditional cross-modal retrieval, where each image-text pair is treated independently ignoring broader context. We propose a novel variational recurrent seq2seq (VRSS) retrieval model for this seq2seq task. Unlike most cross-modal methods, we generate an image vector corresponding to the latent topic obtained from combining the text semantics and context. This synthetic image embedding point associated with every text embedding point can then be employed for either image generation or image retrieval as desired. We evaluate the model for the application of stepwise illustration of recipes, where a sequence of relevant images are retrieved to best match the steps described in the text. To this end, we build and release a new Stepwise Recipe dataset for research purposes, containing 10K recipes (sequences of image-text pairs) having a total of 67K image-text pairs. To our knowledge, it is the first publicly available dataset to offer rich semantic descriptions in a focused category such as food or recipes. Our model is shown to outperform several competitive and relevant baselines in the experiments. We also provide qualitative analysis of how semantically meaningful the results produced by our model are through human evaluation and comparison with relevant existing methods.

Transcribing structured data into natural language descriptions has emerged as a challenging task, referred to as “data-to-text”. These structures generally regroup multiple elements, as well as their attributes. Most attempts rely on translation encoder-decoder methods which linearize elements into a sequence. This however loses most of the structure contained in the data. In this work, we propose to overpass this limitation with a hierarchical model that encodes the data-structure at the element-level and the structure level. Evaluations on RotoWire show the effectiveness of our model w.r.t. qualitative and quantitative metrics.

We propose a method for learning entity orders, for example, safety, popularity, and livability orders of countries. We train linear functions by using samples of ordered entities as training data, and attributes of entities as features. An example of such functions is f(Entity) $$= +0.5$$ (Police budget) $$-0.8$$ (Crime rate), for ordering countries in terms of safety. As the size of training data is typically small in this task, we propose a machine learning method referred to as context-guided learning (CGL) to overcome the over-fitting problem. Exploiting a large amount of contexts regarding relations between the labeling criteria (e.g. safety) and attributes, CGL guides learning in the correct direction by estimating a roughly appropriate weight for each attribute by the contexts. This idea was implemented by a regularization approach similar to support vector machines. Experiments were conducted with 158 kinds of orders in three datasets. The experimental results showed high effectiveness of the contextual guidance over existing ranking methods.

In this research, we improve upon the current state of the art in entity retrieval by re-ranking the result list using graph embeddings. The paper shows that graph embeddings are useful for entity-oriented search tasks. We demonstrate empirically that encoding information from the knowledge graph into (graph) embeddings contributes to a higher increase in effectiveness of entity retrieval results than using plain word embeddings. We analyze the impact of the accuracy of the entity linker on the overall retrieval effectiveness. Our analysis further deploys the cluster hypothesis to explain the observed advantages of graph embeddings over the more widely used word embeddings, for user tasks involving ranking entities.

Interlinking of spatio-textual entities is an open and quite challenging research problem, with application in several commercial fields, including geomarketing, navigation and social networks. It comprises the process of identifying, between different data sources, entity descriptions that refer to the same real-world entity. In this work, we focus on toponym interlinking, that is we handle spatio-textual entities that are exclusively represented by their name; additional properties, such as categories, coordinates, etc. are considered as either absent or of too low quality to be exploited in this setting. Toponyms are inherently heterogeneous entities; quite often several alternative names exist for the same toponym, with varying degrees of similarity between these names. State of the art approaches adopt mostly generic, domain-agnostic similarity functions and use them as is, or incorporate them as training features within classifiers for performing toponym interlinking. We claim that capturing the specificities of toponyms and exploiting them into elaborate meta-similarity functions and derived training features can significantly increase the effectiveness of interlinking methods. To this end, we propose the LGM-Sim meta-similarity function and a series of novel, similarity-based and statistical training features that can be utilized in similarity-based and classification-based interlinking settings respectively. We demonstrate that the proposed methods achieve large increases in accuracy, in both settings, compared to several methods from the literature in the widely used Geonames toponym dataset.

The identification of lexical semantic relations is of the utmost importance to enhance reasoning capacities of Natural Language Processing and Information Retrieval systems. Within this context, successful results have been achieved based on the distributional hypothesis and/or the paradigmatic assumption. However, both strategies solely rely on the input words to predict the lexical semantic relation. In this paper, we make the hypothesis that the decision process should not only rely on the input words but also on their K closest neighbors in some semantic space. For that purpose, we present different binary and multi-task classification strategies that include two distinct attention mechanisms based on PageRank. Evaluation results over four gold-standard datasets show that average improvements of 10.6% for binary and 8% for multi-task classification can be achieved over baseline approaches in terms of F$$_1$$. The code and the datasets are available upon demand.

Recent years have witnessed the emergence of novel models for ad-hoc entity search in knowledge graphs of varying complexity. Since these models are based on direct term matching, their accuracy can suffer from a mismatch between vocabularies used in queries and entity descriptions. Although successful applications of word embeddings and knowledge graph entity embeddings to address the issues of vocabulary mismatch in ad-hoc document retrieval and knowledge graph noisiness and incompleteness, respectively, have been reported in recent literature, the utility of joint word and entity embeddings for entity search in knowledge graphs has been relatively unexplored. In this paper, we propose Knowledge graph Entity and Word Embedding for Retrieval (KEWER), a novel method to embed entities and words into the same low-dimensional vector space, which takes into account a knowledge graph’s local structure and structural components, such as entities, attributes, and categories, and is designed specifically for entity search. KEWER is based on random walks over the knowledge graph and can be considered as a hybrid of word and network embedding methods. Similar to word embedding methods, KEWER utilizes contextual co-occurrences as training data, however, it treats words and entities as different objects. Similar to network embedding methods, KEWER takes into account knowledge graph’s local structure, however, it also differentiates between structural components. Experiments on publicly available entity search benchmarks and state-of-the-art word and joint word and entity embedding methods indicate that a combination of KEWER and BM25F results in a consistent improvement in retrieval accuracy over BM25F alone.

Extracting insights from data obtained from the web in order to identify people’s views and opinions on various topics is a growing practice. The standard insights extraction pipeline is typically an unsupervised machine learning task composed of processes that preprocess the text, visualize it, cluster and identify the topics and sentiment in each cluster, and then graph the network. Given the increasing amount of data being generated on the internet in Africa today, and the multilingual state of African countries, we evaluated how well the standard pipeline works when applied to text wholly or partially written in indigenous African languages, specifically Bantu languages. We carried out an exploratory investigation using Twitter data and compared the outputs from each step of the pipeline for an English dataset and a mixed Bantu language dataset. We found that for Bantu languages, due to their complex grammatical structure, extra preprocessing steps such as part-of-speech tagging and morphological analysis are required during data cleaning, threshold values should be adjusted during topic modeling, and semantic analysis should be performed before completing text preprocessing.

Contact recommendation is an important functionality in many social network scenarios including Twitter and Facebook, since they can help grow the social networks of users by suggesting, to a given user, people they might wish to follow. Recently, it has been shown that classical information retrieval (IR) weighting models – such as BM25 – can be adapted to effectively recommend new social contacts to a given user. However, the exact properties that make such adapted contact recommendation models effective at the task are as yet unknown. In this paper, inspired by new advances in the axiomatic theory of IR, we study the existing IR axioms for the contact recommendation task. Our theoretical analysis and empirical findings show that while the classical axioms related to term frequencies and term discrimination seem to have a positive impact on the recommendation effectiveness, those related to length normalization tend to be not desirable for the task.

We propose a framework for personalized music curator recommendation to connect users with curators who have matching curation style. Three unique features of the proposed framework are: (i) models of curation style to capture the coverage of music and curator’s individual style in assigning tracks to playlists; (ii) a curation-based embedding approach to capture inter-track agreement, beyond the audio features, resulting in models of music tracks that pair well together; and (iii) a novel neural pairwise ranking model for personalized music curator recommendation that naturally incorporates both curator style models and track embeddings. Experiments over a Spotify dataset show significant improvements in precision, recall, and F1 versus state-of-the-art.

With the popularity of Location-based Social Networks, Point-of-Interest (POI) recommendation has become an important task, which learns the users’ preferences and mobility patterns to recommend POIs. Previous studies show that incorporating contextual information such as geographical and temporal influences is necessary to improve POI recommendation by addressing the data sparsity problem. However, existing methods model the geographical influence based on the physical distance between POIs and users, while ignoring the temporal characteristics of such geographical influences. In this paper, we perform a study on the user mobility patterns where we find out that users’ check-ins happen around several centers depending on their current temporal state. Next, we propose a spatio-temporal activity-centers algorithm to model users’ behavior more accurately. Finally, we demonstrate the effectiveness of our proposed contextual model by incorporating it into the matrix factorization model under two different settings: (i) static and (ii) temporal. To show the effectiveness of our proposed method, which we refer to as STACP, we conduct experiments on two well-known real-world datasets acquired from Gowalla and Foursquare LBSNs. Experimental results show that the STACP model achieves a statistically significant performance improvement, compared to the state-of-the-art techniques. Also, we demonstrate the effectiveness of capturing geographical and temporal information for modeling users’ activity centers and the importance of modeling them jointly.

New research is being published at a rate, at which it is infeasible for many scholars to read and assess everything possibly relevant to their work. In pursuit of a remedy, efforts towards automated processing of publications, like semantic modelling of papers to facilitate their digital handling, and the development of information filtering systems, are an active area of research. In this paper, we investigate the benefits of semantically modelling citation contexts for the purpose of citation recommendation. For this, we develop semantic models of citation contexts based on entities and claim structures. To assess the effectiveness and conceptual soundness of our models, we perform a large offline evaluation on several data sets and furthermore conduct a user study. Our findings show that the models can outperform a non-semantic baseline model and do, indeed, capture the kind of information they’re conceptualized for.

The text of a review expresses the sentiment a customer has towards a particular product. This is exploited in sentiment analysis where machine learning models are used to predict the review score from the text of the review. Furthermore, the products costumers have purchased in the past are indicative of the products they will purchase in the future. This is what recommender systems exploit by learning models from purchase information to predict the items a customer might be interested in. The underlying structure of this problem setting is a bipartite graph, wherein customer nodes are connected to product nodes via ‘review’ links. This is reminiscent of knowledge bases, with ‘review’ links replacing relation types. We propose TransRev, an approach to the product recommendation problem that integrates ideas from recommender systems, sentiment analysis, and multi-relational learning into a joint learning objective.TransRev learns vector representations for users, items, and reviews. The embedding of a review is learned such that (a) it performs well as input feature of a regression model for sentiment prediction; and (b) it always translates the reviewer embedding to the embedding of the reviewed item. This is reminiscent of TransE [5], a popular embedding method for link prediction in knowledge bases. This allows TransRev to approximate a review embedding at test time as the difference of the embedding of each item and the user embedding. The approximated review embedding is then used with the regression model to predict the review score for each item. TransRev outperforms state of the art recommender systems on a large number of benchmark data sets. Moreover, it is able to retrieve, for each user and item, the review text from the training set whose embedding is most similar to the approximated review embedding.

We examine the novel task of domain-independent scientific concept extraction from abstracts of scholarly articles and present two contributions. First, we suggest a set of generic scientific concepts that have been identified in a systematic annotation process. This set of concepts is utilised to annotate a corpus of scientific abstracts from 10 domains of Science, Technology and Medicine at the phrasal level in a joint effort with domain experts. The resulting dataset is used in a set of benchmark experiments to (a) provide baseline performance for this task, (b) examine the transferability of concepts between domains. Second, we present a state-of-the-art deep learning baseline. Further, we propose the active learning strategy for an optimal selection of instances from among the various domains in our data. The experimental results show that (1) a substantial agreement is achievable by non-experts after consultation with domain experts, (2) the baseline system achieves a fairly high F1 score, (3) active learning enables us to nearly halve the amount of required training data.

A search engine’s ability to retrieve desirable datasets is important for data sharing and reuse. Existing dataset search engines typically rely on matching queries to dataset descriptions. However, a user may not have enough prior knowledge to write a query using terms that match with description text. We propose a novel schema label generation model which generates possible schema labels based on dataset table content. We incorporate the generated schema labels into a mixed ranking model which not only considers the relevance between the query and dataset metadata but also the similarity between the query and generated schema labels. To evaluate our method on real-world datasets, we create a new benchmark specifically for the dataset retrieval task. Experiments show that our approach can effectively improve the precision and NDCG scores of the dataset retrieval task compared with baseline methods. We also test on a collection of Wikipedia tables to show that the features generated from schema labels can improve the unsupervised and supervised web table retrieval task as well.

Formal concept analysis has been largely applied to explore taxonomic relationships and derive ontologies from text collections. Despite its recognized relevance, it generally misses relevant concept associations and suffers from the need to learn from Boolean space models. Biclustering, the discovery of coherent concept associations (subsets of documents correlated on subsets of terms and topics), is here suggested to address the aforementioned problems. This work proposes a structured view on why, when and how to apply biclustering for concept analysis, a subject remaining largely unexplored up to date. Gathered results from a large text collection confirm the relevance of biclustering to find less-trivial, yet actionable and statistically significant concept associations.

Mental disorders are a major concern in societies all over the world, and in spite of the improved diagnosis rates of such disorders in recent years, many cases still go undetected. Nowadays, many people are increasingly utilising online social media platforms to share their feelings and moods. Despite the collective efforts in the community to develop models for identifying potential cases of mental disorders, not much work has been done to provide insights that could be used by a predictive system or a health practitioner in the elaboration of a diagnosis.In this paper, we present our research towards better visualising and understanding the factors that characterise and differentiate social media users who are affected by mental disorders from those who are not. Furthermore, we study to which extent various mental disorders, such as depression and anorexia, differ in terms of language use. We conduct different experiments considering various dimensions of language such as vocabulary, psychometric attributes and emotional indicators. Our findings reveal that positive instances of mental disorders show significant differences from control individuals in the way they write and express emotions in social media. However, there are not quantifiable differences that could be used to distinguish one mental disorder from each other.

Predicting the runtime complexity of a programming code is an arduous task. In fact, even for humans, it requires a subtle analysis and comprehensive knowledge of algorithms to predict time complexity with high fidelity, given any code. As per Turing’s Halting problem proof, estimating code complexity is mathematically impossible. Nevertheless, an approximate solution to such a task can help developers to get real-time feedback for the efficiency of their code. In this work, we model this problem as a machine learning task and check its feasibility with thorough analysis. Due to the lack of any open source dataset for this task, we propose our own annotated dataset, (The complete dataset is available for use at https://github.com/midas-research/corcod-dataset/blob/master/README.md ) CoRCoD: Code Runtime Complexity Dataset, extracted from online coding platforms. We establish baselines using two different approaches: feature engineering and code embeddings, to achieve state of the art results and compare their performances. Such solutions can be highly useful in potential applications like automatically grading coding assignments, IDE-integrated tools for static code analysis, and others.

Document network embedding aims at learning representations for a structured text corpus i.e. when documents are linked to each other. Recent algorithms extend network embedding approaches by incorporating the text content associated with the nodes in their formulations. In most cases, it is hard to interpret the learned representations. Moreover, little importance is given to the generalization to new documents that are not observed within the network. In this paper, we propose an interpretable and inductive document network embedding method. We introduce a novel mechanism, the Topic-Word Attention (TWA), that generates document representations based on the interplay between word and topic representations. We train these word and topic vectors through our general model, Inductive Document Network Embedding (IDNE), by leveraging the connections in the document network. Quantitative evaluations show that our approach achieves state-of-the-art performance on various networks and we qualitatively show that our model produces meaningful and interpretable representations of the words, topics and documents.

In this paper, we propose an approach to tackle the problem of the automatic restoration of Arabic diacritics that includes three components stacked in a pipeline: a deep learning model which is a multi-layer recurrent neural network with LSTM and Dense layers, a character-level rule-based corrector which applies deterministic operations to prevent some errors, and a word-level statistical corrector which uses the context and the distance information to fix some diacritization issues. This approach is novel in a way that combines methods of different types and adds edit distance based corrections.We used a large public dataset containing raw diacritized Arabic text (Tashkeela) for training and testing our system after cleaning and normalizing it. On a newly-released benchmark test set, our system outperformed all the tested systems by achieving DER of 3.39% and WER of 9.94% when taking all Arabic letters into account, DER of 2.61% and WER of 5.83% when ignoring the diacritization of the last letter of every word.

Named Entity Recognition (NER) is an essential part of many natural language processing (NLP) tasks. The existing Chinese NER methods are mostly based on word segmentation, or use the character sequences as input. However, using a single granularity representation would suffer from the problems of out-of-vocabulary and word segmentation errors, and the semantic content is relatively simple. In this paper, we introduce the self-attention mechanism into the BiLSTM-CRF neural network structure for Chinese named entity recognition with two embedding. Different from other models, our method combines character and word features at the sequence level, and the attention mechanism computes similarity on the total sequence consisted of characters and words. The character semantic information and the structure of words work together to improve the accuracy of word boundary segmentation and solve the problem of long-phrase combination. We validate our model on MSRA and Weibo corpora, and experiments demonstrate that our model can significantly improve the performance of the Chinese NER task.

Much progress has been made recently on text classification with methods based on neural networks. In particular, models using attention mechanism such as BERT have shown to have the capability of capturing the contextual information within a sentence or document. However, their ability of capturing the global information about the vocabulary of a language is more limited. This latter is the strength of Graph Convolutional Networks (GCN). In this paper, we propose VGCN-BERT model which combines the capability of BERT with a Vocabulary Graph Convolutional Network (VGCN). Local information and global information interact through different layers of BERT, allowing them to influence mutually and to build together a final representation for classification. In our experiments on several text classification datasets, our approach outperforms BERT and GCN alone, and achieve higher effectiveness than that reported in previous studies.

Searching literature for a systematic review begins with a manually constructed search strategy by an expert information specialist. The typical process of constructing search strategies is often undocumented, ad-hoc, and subject to individual expertise, which may introduce bias in the systematic review. A new method for objectively deriving search strategies has arisen from information specialists attempting to address these shortcomings. However, this proposed method still presents a number of manual, ad-hoc interventions, and trial-and-error processes, potentially still introducing bias into systematic reviews. Moreover, this method has not been rigorously evaluated on a large set of systematic review cases, thus its generalisability is unknown. In this work, we present a computational adaptation of this proposed objective method. Our adaptation removes the human-in-the-loop processes involved in the initial steps of creating a search strategy for a systematic review; reducing bias due to human factors and increasing the objectivity of the originally proposed method. Our proposed computational adaptation further enables a formal and rigorous evaluation over a large set of systematic reviews. We find that our computational adaptation of the original objective method provides an effective starting point for information specialists to continue refining. We also identify a number of avenues for extending and improving our adaptation to further promote supporting information specialists.

Coordination level matching is a ranking method originally proposed to rank documents given Boolean queries that is now several decades old. Rank fusion is a relatively recent method for combining runs from multiple systems into a single ranking, and has been shown to significantly improve the ranking. This paper presents a novel extension to coordination level matching, by applying rank fusion to each sub-clause of a Boolean query. We show that, for the tasks of systematic review screening prioritisation and stopping estimation, our method significantly outperforms the state-of-the-art learning to rank and bag-of-words-based systems for this domain. Our fully automatic, unsupervised method has (i) the potential for significant real-world cost savings (ii) does not rely on any intervention from the user, and (iii) is significantly better at ranking documents given only a Boolean query in the context of systematic reviews when compared to other approaches.

Exploiting users’ implicit feedback, such as clicks, to learn rankers is attractive as it does not require editorial labelling effort, and adapts to users’ changing preferences, among other benefits. However, directly learning a ranker from implicit data is challenging, as users’ implicit feedback usually contains bias (e.g., position bias, selection bias) and noise (e.g., clicking on irrelevant but attractive snippets, adversarial clicks). Two main methods have arisen for optimizing rankers based on implicit feedback: counterfactual learning to rank (CLTR), which learns a ranker from the historical click-through data collected from a deployed, logging ranker; and online learning to rank (OLTR), where a ranker is updated by recording user interaction with a result list produced by multiple rankers (usually via interleaving).In this paper, we propose a counterfactual online learning to rank algorithm (COLTR) that combines the key components of both CLTR and OLTR. It does so by replacing the online evaluation required by traditional OLTR methods with the counterfactual evaluation common in CLTR. Compared to traditional OLTR approaches based on interleaving, COLTR can evaluate a large number of candidate rankers in a more efficient manner. Our empirical results show that COLTR significantly outperforms traditional OLTR methods. Furthermore, COLTR can reach the same effectiveness of the current state-of-the-art, under noisy click settings, and has room for future extensions.

Computational argumentation has recently become a fast growing field of research. An argument consists of a claim, such as “We should abandon fossil fuels”, which is supported or attacked by at least one premise, for example “Burning fossil fuels is one cause for global warming”. From an information retrieval perspective, an interesting task within this setting is finding the best supporting and attacking premises for a given query claim from a large corpus of arguments. Since the same logical premise can be formulated differently, the system needs to avoid retrieving duplicate results and thus needs to use some form of clustering. In this paper we propose a principled probabilistic ranking framework for premises based on the idea of tf-idf that, given a query claim, first identifies highly similar claims in the corpus, and then clusters and ranks their premises, taking clusters of claims as well as the stances of query and premises into account. We compare our approach to a baseline system that uses BM25F which we outperform even with a primitive implementation of our framework utilising BERT.

Word mismatch between queries and documents is a long-standing challenge in information retrieval. Recent advances in distributed word representations address the word mismatch problem by enabling semantic matching. However, most existing models rank documents based on semantic matching between query and document terms without an explicit understanding of the relationship of the match to relevance. To consider semantic matching between query and document, we propose an unsupervised semantic matching model by simulating a user who makes relevance decisions. The primary goal of the proposed model is to combine the exact and semantic matching between query and document terms, which has been shown to produce effective performance in information retrieval. As semantic matching between queries and entire documents is computationally expensive, we propose to use local contexts of query terms in documents for semantic matching. Matching with smaller query-related contexts of documents stems from the relevance judgment process recorded by human observers. The most relevant part of a document is then recognized and used to rank documents with respect to the query. Experimental results on several representative retrieval models and standard datasets show that our proposed semantic matching model significantly outperforms competitive baselines in all measures.

In many information extraction applications, entity linking (EL) has emerged as a crucial task that allows leveraging information about named entities from a knowledge base. In this paper, we address the task of multimodal entity linking (MEL), an emerging research field in which textual and visual information is used to map an ambiguous mention to an entity in a knowledge base (KB). First, we propose a method for building a fully annotated Twitter dataset for MEL, where entities are defined in a Twitter KB. Then, we propose a model for jointly learning a representation of both mentions and entities from their textual and visual contexts. We demonstrate the effectiveness of the proposed model by evaluating it on the proposed dataset and highlight the importance of leveraging visual information when it is available.

The recent upsurge in the usage of social media and the multimedia data generated therein has attracted many researchers for analyzing and decoding the information to automate decision-making in several fields. This work focuses on one such application: disaster management in times of crises and calamities. The existing research on disaster damage analysis has primarily taken only unimodal information in the form of text or image into account. These unimodal systems, although useful, fail to model the relationship between the various modalities. Different modalities often present supporting facts about the task, and therefore, learning them together can enhance performance. We present MEMIS, a system that can be used in emergencies like disasters to identify and analyze the damage indicated by user-generated multimodal social media posts, thereby helping the disaster management groups in making informed decisions. Our leave-one-disaster-out experiments on a multimodal dataset suggest that not only does fusing information in different media forms improves performance, but that our system can also generalize well to new disaster categories. Further qualitative analysis reveals that the system is responsive and computationally efficient.

Interactive learning has been suggested as a key method for addressing analytic multimedia tasks arising in several domains. Until recently, however, methods to maintain interactive performance at the scale of today’s media collections have not been addressed. We propose an interactive learning approach that builds on and extends the state of the art in user relevance feedback systems and high-dimensional indexing for multimedia. We report on a detailed experimental study using the ImageNet and YFCC100M collections, containing 14 million and 100 million images respectively. The proposed approach outperforms the relevant state-of-the-art approaches in terms of interactive performance, while improving suggestion relevance in some cases. In particular, even on YFCC100M, our approach requires less than 0.3 s per interaction round to generate suggestions, using a single computing core and less than 7 GB of main memory.

Visual re-ranking has received considerable attention in recent years. It aims to enhance the performance of text-based image retrieval by boosting the rank of relevant images using visual information. Hypergraph has been widely used for relevance estimation, where textual results are taken as vertices and the re-ranking problem is formulated as a transductive learning on the hypergraph. The potential of the hypergraph learning is essentially determined by the hypergraph construction scheme. To this end, in this paper, we introduce a novel data representation technique named adaptive collaborative representation for hypergraph learning. Compared to the conventional collaborative representation, we consider the data locality to adaptively select relevant and close samples for a test sample and discard irrelevant and faraway ones. Moreover, at the feature level, we impose a weight matrix on the representation errors to adaptively highlight the important features and reduce the effect of redundant/noisy ones. Finally, we also add a nonnegativity constraint on the representation coefficients to enhance the hypergraph interpretability. These attractive properties allow constructing a more informative and quality hypergraph, thereby achieving better retrieval performance than other hypergraph models. Extensive experiments on the public MediaEval benchmarks demonstrate that our re-ranking method achieves consistently superior results, compared to state-of-the-art methods.

There is a growing amount of human motion data captured as a continuous 3D skeleton sequence without any information about its semantic partitioning. To make such unsegmented and unlabeled data efficiently accessible, we propose to transform them into a text-like representation and employ well-known text retrieval models. Specifically, we partition each motion synthetically into a sequence of short segments and quantize the segments into motion words, i.e. compact features with similar characteristics as words in text documents. We introduce several quantization techniques for building motion-word vocabularies and propose application-independent criteria for assessing the vocabulary quality. We verify these criteria on two real-life application scenarios.

Style transfer deals with the algorithms to transfer the stylistic properties of a piece of text into that of another while ensuring that the core content is preserved. There has been a lot of interest in the field of text style transfer due to its wide application to tailored text generation. Existing works evaluate the style transfer models based on content preservation and transfer strength. In this work, we propose a reinforcement learning based framework that directly rewards the framework on these target metrics yielding a better transfer of the target style. We show the improved performance of our proposed framework based on automatic and human evaluation on three independent tasks: wherein we transfer the style of text from formal to informal, high excitement to low excitement, modern English to Shakespearean English, and vice-versa in all the three cases. Improved performance of the proposed framework over existing state-of-the-art frameworks indicates the viability of the approach.

Automatically recognizing an existing semantic relation (such as “is a”, “part of”, “property of”, “opposite of” etc.) between two arbitrary words (phrases, concepts, etc.) is an important task affecting many information retrieval and artificial intelligence tasks including query expansion, common-sense reasoning, question answering, and database federation. Currently, two classes of approaches exist to classify a relation between words (concepts) X and Y: (1) path-based and (2) distributional. While the path-based approaches look at word-paths connecting X and Y in text, the distributional approaches look at statistical properties of X and Y separately, not necessary in the proximity of each other. Here, we suggest how both types can be improved and empirically compare them using several standard benchmarking datasets. For our distributional approach, we are suggesting using an attention-based transformer. While they are known to be capable of supporting knowledge transfer between different tasks, and recently set a number of benchmarking records in various applications, we are the first to successfully apply them to the task of recognizing semantic relations. To improve a path-based approach, we are suggesting our original neural word path model that combines useful properties of convolutional and recurrent networks, and thus addressing several shortcomings from the prior path-based models. Both our models significantly outperforms the state-of-the-art within its type accordingly. Our transformer-based approach outperforms current state-of-the-art by 1–12% points on 4 out of 6 standard benchmarking datasets. This results in 15–40% error reduction and is closing the gap between the automated and human performance by up to 50%. It also needs much less training data than prior approaches. For the ease of re-producing our results, we make our source code and trained models publicly available.

Rumour detection on Twitter is an important problem. Existing studies mainly focus on high detection accuracy, which often requires large volumes of data on contents, source credibility or propagation. In this paper we focus on early detection of rumours when data for information sources or propagation is scarce. We observe that tweets attract immediate comments from the public who often express uncertain and questioning attitudes towards rumour tweets. We therefore propose to learn user attitude distribution for Twitter posts from their comments, and then combine it with content analysis for early detection of rumours. Specifically we propose convolutional neural network (CNN) CNN and BERT neural network language models to learn attitude representation for user comments without human annotation via transfer learning based on external data sources for stance classification. We further propose CNN-BiLSTM- and BERT-based deep neural models to combine attitude representation and content representation for early rumour detection. Experiments on real-world rumour datasets show that our BERT-based model can achieve effective early rumour detection and significantly outperform start-of-the-art rumour detection models.

We are interested in the problem of cross-modal retrieval for web image search, where the goal is to retrieve images relevant to a text query. While most of the current approaches for cross-modal retrieval revolve around learning how to represent text and images in a shared latent space, we take a different direction: we propose to generalize the cross-modal relevance feedback mechanism, a simple yet effective unsupervised method, that relies on standard information retrieval heuristics and the choice of a few hyper-parameters. We show that we can cast it as a supervised representation learning problem on graphs, using graph convolutions operating jointly over text and image features, namely cross-modal graph convolutions. The proposed architecture directly learns how to combine image and text features for the ranking task, while taking into account the context given by all the other elements in the set of images to be (re-)ranked. We validate our approach on two datasets: a public dataset from a MediaEval challenge, and a small sample of proprietary image search query logs, referred as WebQ. Our experiments demonstrate that our model improves over standard baselines.

Word embeddings, made widely popular in 2013 with the release of word2vec, have become a mainstay of NLP engineering pipelines. Recently, with the release of BERT, word embeddings have moved from the term-based embedding space to the contextual embedding space—each term is no longer represented by a single low-dimensional vector but instead each term and its context determine the vector weights. BERT’s setup and architecture have been shown to be general enough to be applicable to many natural language tasks. Importantly for Information Retrieval (IR), in contrast to prior deep learning solutions to IR problems which required significant tuning of neural net architectures and training regimes, “vanilla BERT” has been shown to outperform existing retrieval algorithms by a wide margin, including on tasks and corpora that have long resisted retrieval effectiveness gains over traditional IR baselines (such as Robust04). In this paper, we employ the recently proposed axiomatic dataset analysis technique—that is, we create diagnostic datasets that each fulfil a retrieval heuristic (both term matching and semantic-based)—to explore what BERT is able to learn. In contrast to our expectations, we find BERT, when applied to a recently released large-scale web corpus with ad-hoc topics, to not adhere to any of the explored axioms. At the same time, BERT outperforms the traditional query likelihood retrieval model by 40%. This means that the axiomatic approach to IR (and its extension of diagnostic datasets created for retrieval heuristics) may in its current form not be applicable to large-scale corpora. Additional—different—axioms are needed.

Privacy concerns can prohibit research access to large-scale commercial query logs. Here we focus on generation of a synthetic log from a publicly available dataset, suitable for evaluation of query auto completion (QAC) systems. The synthetic log contains plausible string sequences reflecting how users enter their queries in a QAC interface. Properties that would influence experimental outcomes are compared between a synthetic log and a real QAC log through a set of side-by-side experiments, and confirm the applicability of the generated log for benchmarking the performance of QAC methods.

A significant amount of prior research has been devoted to understanding query reformulations. The majority of these works rely on time-based sessions which are sequences of contiguous queries segmented using time threshold on users’ activities. However, queries are generally issued by users having in mind a particular task, and time-based sessions unfortunately fail in revealing such tasks. In this paper, we are interested in revealing in which extent time-based sessions vs. task-based sessions represent significantly different background contexts to be used in the perspective of better understanding users’ query reformulations. Using insights from large-scale search logs, our findings clearly show that task is an additional relevant search unit that helps better understanding user’s query reformulation patterns and predicting the next user’s query. The findings from our analyses provide potential implications for model design of task-based search engines.

A substantial portion of the query volume for e-commerce search engines consists of infrequent queries and identifying user intent in such tail queries is critical in retrieving relevant products. The intent of a query is defined as a labelling of its tokens with the product attributes whose values are matched against the query tokens during retrieval. Tail queries in e-commerce search tend to have multiple correct attribute labels for their tokens due to multiple valid matches in the product catalog. In this paper, we propose a latent variable generative model along with a novel data dependent regularisation technique for identifying multiple intents in such queries. We demonstrate the superior performance of our proposed model against several strong baseline models on an editorially labelled data set as well as in a large scale online A/B experiment at Flipkart, a major Indian e-commerce company.

Query Auto-completion (QAC) is a prominently used feature in search engines, where user interaction with such explicit feature is facilitated by the possible automatic suggestion of queries based on a prefix typed by the user. Existing QAC models have pursued a little on user interaction and cannot capture a user’s information need (IN) context. In this work, we devise a new task of QAC applied on an image for estimating patch (one of the key components of Information Foraging Theory) probabilities for query suggestion. Our work supports query completion by extending a user query prefix (one or two characters) to a complete query utilising a foraging-based probabilistic patch selection model. We present iBERT, to fine-tune the BERT (Bidirectional Encoder Representations from Transformers) model, which leverages combined textual-image queries for a solution to image QAC by computing probabilities of a large set of image patches. The reflected patch probabilities are used for selection while being agnostic to changing information need or contextual mechanisms. Experimental results show that query auto-completion using both natural language queries and images is more effective than using only language-level queries. Also, our fine-tuned iBERT model allows to efficiently rank patches in the image.

Interactive search sessions often contain multiple queries, where the user submits a reformulated version of the previous query in response to the original results. We aim to enhance the query recommendation experience for a commercial image search engine. Our proposed methodology incorporates current state-of-the-art practices from relevant literature – the use of generation-based sequence-to-sequence models that capture session context, and a multitask architecture that simultaneously optimizes the ranking of results. We extend this setup by driving the learning of such a model with captions of clicked images as the target, instead of using the subsequent query within the session. Since these captions tend to be linguistically richer, the reformulation mechanism can be seen as assistance to construct more descriptive queries. In addition, via the use of a pairwise loss for the secondary ranking task, we show that the generated reformulations are more diverse.

Neural ranking models are traditionally trained on a series of random batches, sampled uniformly from the entire training set. Curriculum learning has recently been shown to improve neural models’ effectiveness by sampling batches non-uniformly, going from easy to difficult instances during training. In the context of neural Information Retrieval (IR) curriculum learning has not been explored yet, and so it remains unclear (1) how to measure the difficulty of training instances and (2) how to transition from easy to difficult instances during training. To address both challenges and determine whether curriculum learning is beneficial for neural ranking models, we need large-scale datasets and a retrieval task that allows us to conduct a wide range of experiments. For this purpose, we resort to the task of conversation response ranking: ranking responses given the conversation history. In order to deal with challenge (1), we explore scoring functions to measure the difficulty of conversations based on different input spaces. To address challenge (2) we evaluate different pacing functions, which determine the velocity in which we go from easy to difficult instances. We find that, overall, by just intelligently sorting the training data (i.e., by performing curriculum learning) we can improve the retrieval effectiveness by up to 2% (The source code is available at https://github.com/Guzpenha/transformers_cl .).

Formula retrieval systems using substructure matching are effective, but suffer from slow retrieval times caused by the complexity of structure matching. We present a specialized inverted index and rank-safe dynamic pruning algorithm for faster substructure retrieval. Formulas are indexed from their Operator Tree (OPT) representations. Our model is evaluated using the NTCIR-12 Wikipedia Formula Browsing Task and a new formula corpus produced from Math StackExchange posts. Our approach preserves the effectiveness of structure matching while allowing queries to be executed in real-time.

A large number of studies in cognitive science have revealed that probabilistic outcomes of certain human decisions do not agree with the axioms of classical probability theory. The field of Quantum Cognition provides an alternative probabilistic model to explain such paradoxical findings. It posits that cognitive systems have an underlying quantum-like structure, especially in decision-making under uncertainty. In this paper, we hypothesise that relevance judgement, being a multidimensional, cognitive concept, can be used to probe the quantum-like structure for modelling users’ cognitive states in information seeking. Extending from an experiment protocol inspired by the Stern-Gerlach experiment in Quantum Physics, we design a crowd-sourced user study to show violation of the Kolmogorovian probability axioms as a proof of the quantum-like structure, and provide a comparison between a quantum probabilistic model and a Bayesian model for predictions of relevance.

We propose a temporal latent space model for user community prediction in social networks, whose goal is to predict future emerging user communities based on past history of users’ topics of interest. Our model assumes that each user lies within an unobserved latent space, and similar users in the latent space representation are more likely to be members of the same user community. The model allows each user to adjust its location in the latent space as her topics of interest evolve over time. Empirically, we demonstrate that our model, when evaluated on a Twitter dataset, outperforms existing approaches under two application scenarios, namely news recommendation and user prediction on a host of metrics such as mrr, ndcg as well as precision and f-measure.

Exploring a knowledge base is often an iterative process: initially vague information needs are refined by interaction. We propose a novel approach for such interaction that supports sequential question answering (SQA) on knowledge graphs. As opposed to previous work, we focus on exploratory settings, which we support with a visual representation of graph structures, helping users to better understand relationships. In addition, our approach keeps track of context – an important challenge in SQA – by allowing users to make their focus explicit via subgraph selection. Our results show that the interaction principle is either understood immediately or picked up very quickly – and that the possibility of exploring the information space iteratively is appreciated.

Long-term news article archives are valuable resources about our past, allowing people to know detailed information of events that occurred at specific time points. To make better use of such heritage collections, this work considers the task of large scale question answering on long-term news article archives. Questions on such archives are often event-related. In addition, they usually exhibit strong temporal aspects and can be roughly categorized into two types: (1) ones containing explicit temporal expressions, and (2) ones only implicitly associated with particular time periods. We focus on the latter type as such questions are more difficult to be answered, and we propose a retriever-reader model with an additional module for reranking articles by exploiting temporal information from different angles. Experimental results on carefully constructed test set show that our model outperforms the existing question answering systems, thanks to an additional module that finds more relevant documents.

Ethical issues, along with transparency, disinformation, and bias, are in the focus of our information society. In this work, we propose the bias goggles model, for computing the bias characteristics of web domains to user-defined concepts based on the structure of the web graph. For supporting the model, we exploit well-known propagation models and the newly introduced Biased-PR PageRank algorithm, that models various behaviours of biased surfers. An implementation discussion, along with a preliminary evaluation over a subset of the greek web graph, shows the applicability of the model even in real-time for small graphs, and showcases rather promising and interesting results. Finally, we pinpoint important directions for future work. A constantly evolving prototype of the bias goggles system is readily available.

In this paper, we present a conditional GAN with two generators and a common discriminator for multiview learning problems where observations have two views, but one of them may be missing for some of the training samples. This is for example the case for multilingual collections where documents are not available in all languages. Some studies tackled this problem by assuming the existence of view generation functions to approximately complete the missing views; for example Machine Translation to translate documents into the missing languages. These functions generally require an external resource to be set and their quality has a direct impact on the performance of the learned multiview classifier over the completed training set. Our proposed approach addresses this problem by jointly learning the missing views and the multiview classifier using a tripartite game with two generators and a discriminator. Each of the generators is associated to one of the views and tries to fool the discriminator by generating the other missing view conditionally on the corresponding observed view. The discriminator then tries to identify if for an observation, one of its views is completed by one of the generators or if both views are completed along with its class. Our results on a subset of Reuters RCV1/RCV2 collections show that the discriminator achieves significant classification performance; and that the generators learn the missing views with high quality without the need of any consequent external resource.

Graphs offer a natural abstraction for modeling complex real-world systems where entities are represented as nodes and edges encode relations between them. In such networks, entities may share common or similar attributes and may be connected by paths through multiple attribute modalities. In this work, we present an approach that uses semantically meaningful, bimodal random walks on real-world heterogeneous networks to extract correlations between nodes and bring together nodes with shared or similar attributes. An attention-based mechanism is used to combine multiple attribute-specific representations in a late fusion setup. We focus on a real-world network formed by restaurants and their shared attributes and evaluate performance on predicting the number of reviews a restaurant receives, a strong proxy for popularity. Our results demonstrate the rich expressiveness of such representations in predicting review volume and the ability of an attention-based model to selectively combine individual representations for maximum predictive power on the chosen downstream task.

Many people browse reviews online before making purchasing decisions. It is essential to identify the subset of helpful reviews from the large number of reviews of varying quality. This paper aims to build a model to predict review helpfulness automatically. Our work is inspired by the observation that a customer’s expectation of a review can be greatly affected by review sentiment and the degree to which the customer is aware of pertinent product information. Consequently, a customer may pay more attention to that specific content of a review which contributes more to its helpfulness from their perspective. To model such customer expectations and capture important information from a review text, we propose a novel neural network which leverages review sentiment and product information. Specifically, we encode the sentiment of a review through an attention module, to get sentiment-driven information from review text. We also introduce a product attention layer that fuses information from both the target product and related products, in order to capture the product related information from review text. Our experimental results show an AUC improvement of 5.4% and 1.5% over the previous state of the art model on Amazon and Yelp data sets, respectively.