Christina Lioma

Sara Vera Marjanović, Haeun Yu, Pepa Atanasova et al.

Knowledge-intensive language understanding tasks require Language Models (LMs) to integrate relevant context, mitigating their inherent weaknesses, such as incomplete or outdated knowledge. However, conflicting knowledge can be present in the LM's parameters, termed intra-memory conflict, which can affect a model's propensity to accept contextual knowledge. To study the effect of intra-memory conflict on an LM's ability to accept relevant context, we utilize two knowledge conflict measures and a novel dataset containing inherently conflicting data, DynamicQA. This dataset includes facts with a temporal dynamic nature where facts can change over time and disputable dynamic facts, which can change depending on the viewpoint. DynamicQA is the first to include real-world knowledge conflicts and provide context to study the link between the different types of knowledge conflicts. We also evaluate several measures on their ability to reflect the presence of intra-memory conflict: semantic entropy and a novel coherent persuasion score. With our extensive experiments, we verify that LMs exhibit a greater degree of intra-memory conflict with dynamic facts compared to facts that have a single truth value. Furthermore, we reveal that facts with intra-memory conflict are harder to update with context, suggesting that retrieval-augmented generation will struggle with the most commonly adapted facts.

Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma et al.

Automating the fact checking (FC) process relies on information obtained from external sources. In this work, we posit that it is crucial for FC models to make veracity predictions only when there is sufficient evidence and otherwise indicate when it is not enough. To this end, we are the first to study what information FC models consider sufficient by introducing a novel task and advancing it with three main contributions. First, we conduct an in-depth empirical analysis of the task with a new fluency-preserving method for omitting information from the evidence at the constituent and sentence level. We identify when models consider the remaining evidence (in)sufficient for FC, based on three trained models with different Transformer architectures and three FC datasets. Second, we ask annotators whether the omitted evidence was important for FC, resulting in a novel diagnostic dataset, SufficientFacts, for FC with omitted evidence. We find that models are least successful in detecting missing evidence when adverbial modifiers are omitted (21% accuracy), whereas it is easiest for omitted date modifiers (63% accuracy). Finally, we propose a novel data augmentation strategy for contrastive self-learning of missing evidence by employing the proposed omission method combined with tri-training. It improves performance for Evidence Sufficiency Prediction by up to 17.8 F1 score, which in turn improves FC performance by up to 2.6 F1 score.

Qiuchi Li, Benyou Wang, Yudong Zhu et al.

The emerging classical-quantum transfer learning paradigm has brought a decent performance to quantum computational models in many tasks, such as computer vision, by enabling a combination of quantum models and classical pre-trained neural networks. However, using quantum computing with pre-trained models has yet to be explored in natural language processing (NLP). Due to the high linearity constraints of the underlying quantum computing infrastructures, existing Quantum NLP models are limited in performance on real tasks. We fill this gap by pre-training a sentence state with complex-valued BERT-like architecture, and adapting it to the classical-quantum transfer learning scheme for sentence classification. On quantum simulation experiments, the pre-trained representation can bring 50\% to 60\% increases to the capacity of end-to-end quantum models.

Shivam Adarsh, Maria Maistro, Christina Lioma

Large Language Models (LLMs) often encode whether a statement is true as a vector in their residual stream activations. These vectors, also known as truth vectors, have been studied in prior work, however how they change when context is introduced remains unexplored. We study this question by measuring (1) the directional change ($θ$) between the truth vectors with and without context and (2) the relative magnitude of the truth vectors upon adding context. Across four LLMs and four datasets, we find that (1) truth vectors are roughly orthogonal in early layers, converge in middle layers, and may stabilize or continue increasing in later layers; (2) adding context generally increases the truth vector magnitude, i.e., the separation between true and false representations in the activation space is amplified; (3) larger models distinguish relevant from irrelevant context mainly through directional change ($θ$), while smaller models show this distinction through magnitude differences. We also find that context conflicting with parametric knowledge produces larger geometric changes than parametrically aligned context. To the best of our knowledge, this is the first work that provides a geometric characterization of how context transforms the truth vector in the activation space of LLMs.

Qiuchi Li, Christina Lioma

Text generation has long been a popular research topic in NLP. However, the task of generating recruitment emails from recruiters to candidates in the job recommendation scenario has received little attention by the research community. This work aims at defining the topic of automatic email generation for job recommendation, identifying the challenges, and providing a baseline template-based solution for Danish jobs. Evaluation by human experts shows that our method is effective. We wrap up by discussing the future research directions for better solving this task.

Theresia Veronika Rampisela, Maria Maistro, Tuukka Ruotsalo et al.

Fairness in recommender systems (RSs) is commonly categorised into group fairness and individual fairness. However, there is no established scientific understanding of the relationship between the two fairness types, as prior work on both types has used different evaluation measures or evaluation objectives for each fairness type, thereby not allowing for a proper comparison of the two. As a result, it is currently not known how increasing one type of fairness may affect the other. To fill this gap, we study the relationship of group and individual fairness through a comprehensive comparison of evaluation measures that can be used for both fairness types. Our experiments with 8 runs across 3 datasets show that recommendations that are highly fair for groups can be very unfair for individuals. Our finding is novel and useful for RS practitioners aiming to improve the fairness of their systems. Our code is available at: https://github.com/theresiavr/stairway-to-fairness.

Pietro Tropeano, Maria Maistro, Tuukka Ruotsalo et al.

Language Model (LM) pruning compresses the model by removing weights, nodes, or other parts of its architecture. Typically, pruning focuses on the resulting efficiency gains at the cost of effectiveness. However, when looking at how individual data points are affected by pruning, it turns out that a particular subset of data points always bears most of the brunt (in terms of reduced accuracy) when pruning, but this effect goes unnoticed when reporting the mean accuracy of all data points. These data points are called PIEs and have been studied in image processing, but not in NLP. In a study of various NLP datasets, pruning methods, and levels of compression, we find that PIEs impact inference quality considerably, regardless of class frequency, and that BERT is more prone to this than BiLSTM. We also find that PIEs contain a high amount of data points that have the largest influence on how well the model generalises to unseen data. This means that when pruning, with seemingly moderate loss to accuracy across all data points, we in fact hurt tremendously those data points that matter the most. We trace what makes PIEs both hard and impactful to inference to their overall longer and more semantically complex text. These findings are novel and contribute to understanding how LMs are affected by pruning. The code is available at: https://github.com/pietrotrope/AsEasyAsPIE

Benyou Wang, Donghao Zhao, Christina Lioma et al.

Sequential word order is important when processing text. Currently, neural networks (NNs) address this by modeling word position using position embeddings. The problem is that position embeddings capture the position of individual words, but not the ordered relationship (e.g., adjacency or precedence) between individual word positions. We present a novel and principled solution for modeling both the global absolute positions of words and their order relationships. Our solution generalizes word embeddings, previously defined as independent vectors, to continuous word functions over a variable (position). The benefit of continuous functions over variable positions is that word representations shift smoothly with increasing positions. Hence, word representations in different positions can correlate with each other in a continuous function. The general solution of these functions is extended to complex-valued domain due to richer representations. We extend CNN, RNN and Transformer NNs to complex-valued versions to incorporate our complex embedding (we make all code available). Experiments on text classification, machine translation and language modeling show gains over both classical word embeddings and position-enriched word embeddings. To our knowledge, this is the first work in NLP to link imaginary numbers in complex-valued representations to concrete meanings (i.e., word order).

Christian Hansen, Casper Hansen, Stephen Alstrup et al.

Modelling sequential music skips provides streaming companies the ability to better understand the needs of the user base, resulting in a better user experience by reducing the need to manually skip certain music tracks. This paper describes the solution of the University of Copenhagen DIKU-IR team in the 'Spotify Sequential Skip Prediction Challenge', where the task was to predict the skip behaviour of the second half in a music listening session conditioned on the first half. We model this task using a Multi-RNN approach consisting of two distinct stacked recurrent neural networks, where one network focuses on encoding the first half of the session and the other network focuses on utilizing the encoding to make sequential skip predictions. The encoder network is initialized by a learned session-wide music encoding, and both of them utilize a learned track embedding. Our final model consists of a majority voted ensemble of individually trained models, and ranked 2nd out of 45 participating teams in the competition with a mean average accuracy of 0.641 and an accuracy on the first skip prediction of 0.807. Our code is released at https://github.com/Varyn/WSDM-challenge-2019-spotify.

Radu Dragusin, Paula Petcu, Christina Lioma et al.

Background: The web has become a primary information resource about illnesses and treatments for both medical and non-medical users. Standard web search is by far the most common interface for such information. It is therefore of interest to find out how well web search engines work for diagnostic queries and what factors contribute to successes and failures. Among diseases, rare (or orphan) diseases represent an especially challenging and thus interesting class to diagnose as each is rare, diverse in symptoms and usually has scattered resources associated with it. Methods: We use an evaluation approach for web search engines for rare disease diagnosis which includes 56 real life diagnostic cases, state-of-the-art evaluation measures, and curated information resources. In addition, we introduce FindZebra, a specialized (vertical) rare disease search engine. FindZebra is powered by open source search technology and uses curated freely available online medical information. Results: FindZebra outperforms Google Search in both default setup and customised to the resources used by FindZebra. We extend FindZebra with specialized functionalities exploiting medical ontological information and UMLS medical concepts to demonstrate different ways of displaying the retrieved results to medical experts. Conclusions: Our results indicate that a specialized search engine can improve the diagnostic quality without compromising the ease of use of the currently widely popular web search engines. The proposed evaluation approach can be valuable for future development and benchmarking. The FindZebra search engine is available at http://www.findzebra.com/.

Lovisa Hagström, Sara Vera Marjanović, Haeun Yu et al.

Retrieval-augmented generation (RAG) helps address the limitations of parametric knowledge embedded within a language model (LM). In real world settings, retrieved information can vary in complexity, yet most investigations of LM utilisation of context has been limited to synthetic text. We introduce DRUID (Dataset of Retrieved Unreliable, Insufficient and Difficult-to-understand contexts) with real-world queries and contexts manually annotated for stance. The dataset is based on the prototypical task of automated claim verification, for which automated retrieval of real-world evidence is crucial. We compare DRUID to synthetic datasets (CounterFact, ConflictQA) and find that artificial datasets often fail to represent the complexity and diversity of realistically retrieved context. We show that synthetic datasets exaggerate context characteristics rare in real retrieved data, which leads to inflated context utilisation results, as measured by our novel ACU score. Moreover, while previous work has mainly focused on singleton context characteristics to explain context utilisation, correlations between singleton context properties and ACU on DRUID are surprisingly small compared to other properties related to context source. Overall, our work underscores the need for real-world aligned context utilisation studies to represent and improve performance in real-world RAG settings.

Ziyi Ye, Jingtao Zhan, Qingyao Ai et al. · tsinghua

Query augmentation is a crucial technique for refining semantically imprecise queries. Traditionally, query augmentation relies on extracting information from initially retrieved, potentially relevant documents. If the quality of the initially retrieved documents is low, then the effectiveness of query augmentation would be limited as well. We propose Brain-Aug, which enhances a query by incorporating semantic information decoded from brain signals. BrainAug generates the continuation of the original query with a prompt constructed with brain signal information and a ranking-oriented inference approach. Experimental results on fMRI (functional magnetic resonance imaging) datasets show that Brain-Aug produces semantically more accurate queries, leading to improved document ranking performance. Such improvement brought by brain signals is particularly notable for ambiguous queries.

Sara Vera Marjanović, Isabelle Augenstein, Christina Lioma

Explainable AI methods facilitate the understanding of model behaviour, yet, small, imperceptible perturbations to inputs can vastly distort explanations. As these explanations are typically evaluated holistically, before model deployment, it is difficult to assess when a particular explanation is trustworthy. Some studies have tried to create confidence estimators for explanations, but none have investigated an existing link between uncertainty and explanation quality. We artificially simulate epistemic uncertainty in text input by introducing noise at inference time. In this large-scale empirical study, we insert different levels of noise perturbations and measure the effect on the output of pre-trained language models and different uncertainty metrics. Realistic perturbations have minimal effect on performance and explanations, yet masking has a drastic effect. We find that high uncertainty doesn't necessarily imply low explanation plausibility; the correlation between the two metrics can be moderately positive when noise is exposed during the training process. This suggests that noise-augmented models may be better at identifying salient tokens when uncertain. Furthermore, when predictive and epistemic uncertainty measures are over-confident, the robustness of a saliency map to perturbation can indicate model stability issues. Integrated Gradients shows the overall greatest robustness to perturbation, while still showing model-specific patterns in performance; however, this phenomenon is limited to smaller Transformer-based language models.

Pietro Tropeano, Maria Maistro, Tuukka Ruotsalo et al.

Pruning Large Language Models (LLMs) reduces memory and inference costs by removing parts of the network, producing smaller models that retain most of their accuracy. As attention layers are the most resource-intensive parts of LLMs, pruning them is a promising compression strategy. Prior work shows that up to 33% of attention layers can be pruned with minimal accuracy loss. Nevertheless, the impact of attention pruning on model interpretability, specifically faithfulness and confidence calibration, remains unstudied. To address this gap, we study how pruning attention layers affects explanation faithfulness and confidence calibration across five LLMs and eight datasets. While the pruned models often maintain high accuracy, we find that their faithfulness and calibration often degrade. Notably, faithfulness and calibration can fluctuate significantly, even when accuracy remains stable, highlighting a misalignment between model confidence, interpretability, and accuracy. Our findings suggest that layer pruning can affect LLMs' interpretability and reliability in ways not captured by accuracy and efficiency measures alone. We recommend including explainability and calibration metrics when evaluating pruned models.

Qiuchi Li, Christina Lioma

The matching of competences, such as skills, occupations or knowledges, is a key desiderata for candidates to be fit for jobs. Automatic extraction of competences from CVs and Jobs can greatly promote recruiters' productivity in locating relevant candidates for job vacancies. This work presents the first model that jointly extracts and classifies competence from Danish job postings. Different from existing works on skill extraction and skill classification, our model is trained on a large volume of annotated Danish corpora and is capable of extracting a wide range of Danish competences, including skills, occupations and knowledges of different categories. More importantly, as a single BERT-like architecture for joint extraction and classification, our model is lightweight and efficient at inference. On a real-scenario job matching dataset, our model beats the state-of-the-art models in the overall performance of Danish competence extraction and classification, and saves over 50% time at inference.

Ziyi Ye, Qingyao Ai, Yiqun Liu et al.

Generating human language through non-invasive brain-computer interfaces (BCIs) has the potential to unlock many applications, such as serving disabled patients and improving communication. Currently, however, generating language via BCIs has been previously successful only within a classification setup for selecting pre-generated sentence continuation candidates with the most likely cortical semantic representation. Inspired by recent research that revealed associations between the brain and the large computational language models, we propose a generative language BCI that utilizes the capacity of a large language model (LLM) jointly with a semantic brain decoder to directly generate language from functional magnetic resonance imaging (fMRI) input. The proposed model can generate coherent language sequences aligned with the semantic content of visual or auditory language stimuli perceived, without prior knowledge of any pre-generated candidates. We compare the language generated from the presented model with a random control, pre-generated language selection approach, and a standard LLM, which generates common coherent text solely based on the next word likelihood according to statistical language training data. The proposed model is found to generate language that is more aligned with semantic stimulus in response to which brain input is sampled. Our findings demonstrate the potential and feasibility of employing BCIs in direct language generation.

Pepa Atanasova, Oana-Maria Camburu, Christina Lioma et al.

Explanations of neural models aim to reveal a model's decision-making process for its predictions. However, recent work shows that current methods giving explanations such as saliency maps or counterfactuals can be misleading, as they are prone to present reasons that are unfaithful to the model's inner workings. This work explores the challenging question of evaluating the faithfulness of natural language explanations (NLEs). To this end, we present two tests. First, we propose a counterfactual input editor for inserting reasons that lead to counterfactual predictions but are not reflected by the NLEs. Second, we reconstruct inputs from the reasons stated in the generated NLEs and check how often they lead to the same predictions. Our tests can evaluate emerging NLE models, proving a fundamental tool in the development of faithful NLEs.

Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma et al.

Explanations shed light on a machine learning model's rationales and can aid in identifying deficiencies in its reasoning process. Explanation generation models are typically trained in a supervised way given human explanations. When such annotations are not available, explanations are often selected as those portions of the input that maximise a downstream task's performance, which corresponds to optimising an explanation's Faithfulness to a given model. Faithfulness is one of several so-called diagnostic properties, which prior work has identified as useful for gauging the quality of an explanation without requiring annotations. Other diagnostic properties are Data Consistency, which measures how similar explanations are for similar input instances, and Confidence Indication, which shows whether the explanation reflects the confidence of the model. In this work, we show how to directly optimise for these diagnostic properties when training a model to generate sentence-level explanations, which markedly improves explanation quality, agreement with human rationales, and downstream task performance on three complex reasoning tasks.

Qiuchi Li, Dimitris Gkoumas, Christina Lioma et al.

We tackle the crucial challenge of fusing different modalities of features for multimodal sentiment analysis. Mainly based on neural networks, existing approaches largely model multimodal interactions in an implicit and hard-to-understand manner. We address this limitation with inspirations from quantum theory, which contains principled methods for modeling complicated interactions and correlations. In our quantum-inspired framework, the word interaction within a single modality and the interaction across modalities are formulated with superposition and entanglement respectively at different stages. The complex-valued neural network implementation of the framework achieves comparable results to state-of-the-art systems on two benchmarking video sentiment analysis datasets. In the meantime, we produce the unimodal and bimodal sentiment directly from the model to interpret the entangled decision.

Dongsheng Wang, Casper Hansen, Lucas Chaves Lima et al.

The state of the art in learning meaningful semantic representations of words is the Transformer model and its attention mechanisms. Simply put, the attention mechanisms learn to attend to specific parts of the input dispensing recurrence and convolutions. While some of the learned attention heads have been found to play linguistically interpretable roles, they can be redundant or prone to errors. We propose a method to guide the attention heads towards roles identified in prior work as important. We do this by defining role-specific masks to constrain the heads to attend to specific parts of the input, such that different heads are designed to play different roles. Experiments on text classification and machine translation using 7 different datasets show that our method outperforms competitive attention-based, CNN, and RNN baselines.

Lucas Chaves Lima, Casper Hansen, Christian Hansen et al.

This report describes the participation of two Danish universities, University of Copenhagen and Aalborg University, in the international search engine competition on COVID-19 (the 2020 TREC-COVID Challenge) organised by the U.S. National Institute of Standards and Technology (NIST) and its Text Retrieval Conference (TREC) division. The aim of the competition was to find the best search engine strategy for retrieving precise biomedical scientific information on COVID-19 from the largest, at that point in time, dataset of curated scientific literature on COVID-19 -- the COVID-19 Open Research Dataset (CORD-19). CORD-19 was the result of a call to action to the tech community by the U.S. White House in March 2020, and was shortly thereafter posted on Kaggle as an AI competition by the Allen Institute for AI, the Chan Zuckerberg Initiative, Georgetown University's Center for Security and Emerging Technology, Microsoft, and the National Library of Medicine at the US National Institutes of Health. CORD-19 contained over 200,000 scholarly articles (of which more than 100,000 were with full text) about COVID-19, SARS-CoV-2, and related coronaviruses, gathered from curated biomedical sources. The TREC-COVID challenge asked for the best way to (a) retrieve accurate and precise scientific information, in response to some queries formulated by biomedical experts, and (b) rank this information decreasingly by its relevance to the query. In this document, we describe the TREC-COVID competition setup, our participation to it, and our resulting reflections and lessons learned about the state-of-art technology when faced with the acute task of retrieving precise scientific information from a rapidly growing corpus of literature, in response to highly specialised queries, in the middle of a pandemic.

Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma et al.

Recent developments in machine learning have introduced models that approach human performance at the cost of increased architectural complexity. Efforts to make the rationales behind the models' predictions transparent have inspired an abundance of new explainability techniques. Provided with an already trained model, they compute saliency scores for the words of an input instance. However, there exists no definitive guide on (i) how to choose such a technique given a particular application task and model architecture, and (ii) the benefits and drawbacks of using each such technique. In this paper, we develop a comprehensive list of diagnostic properties for evaluating existing explainability techniques. We then employ the proposed list to compare a set of diverse explainability techniques on downstream text classification tasks and neural network architectures. We also compare the saliency scores assigned by the explainability techniques with human annotations of salient input regions to find relations between a model's performance and the agreement of its rationales with human ones. Overall, we find that the gradient-based explanations perform best across tasks and model architectures, and we present further insights into the properties of the reviewed explainability techniques.

Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma et al.

Most existing work on automated fact checking is concerned with predicting the veracity of claims based on metadata, social network spread, language used in claims, and, more recently, evidence supporting or denying claims. A crucial piece of the puzzle that is still missing is to understand how to automate the most elaborate part of the process -- generating justifications for verdicts on claims. This paper provides the first study of how these explanations can be generated automatically based on available claim context, and how this task can be modelled jointly with veracity prediction. Our results indicate that optimising both objectives at the same time, rather than training them separately, improves the performance of a fact checking system. The results of a manual evaluation further suggest that the informativeness, coverage and overall quality of the generated explanations are also improved in the multi-task model.

Isabelle Augenstein, Christina Lioma, Dongsheng Wang et al.

We contribute the largest publicly available dataset of naturally occurring factual claims for the purpose of automatic claim verification. It is collected from 26 fact checking websites in English, paired with textual sources and rich metadata, and labelled for veracity by human expert journalists. We present an in-depth analysis of the dataset, highlighting characteristics and challenges. Further, we present results for automatic veracity prediction, both with established baselines and with a novel method for joint ranking of evidence pages and predicting veracity that outperforms all baselines. Significant performance increases are achieved by encoding evidence, and by modelling metadata. Our best-performing model achieves a Macro F1 of 49.2%, showing that this is a challenging testbed for claim veracity prediction.

Casper Hansen, Christian Hansen, Stephen Alstrup et al.

Word embeddings predict a word from its neighbours by learning small, dense embedding vectors. In practice, this prediction corresponds to a semantic score given to the predicted word (or term weight). We present a novel model that, given a target word, redistributes part of that word's weight (that has been computed with word embeddings) across words occurring in similar contexts as the target word. Thus, our model aims to simulate how semantic meaning is shared by words occurring in similar contexts, which is incorporated into bag-of-words document representations. Experimental evaluation in an unsupervised setting against 8 state of the art baselines shows that our model yields the best micro and macro F1 scores across datasets of increasing difficulty.

Casper Hansen, Christian Hansen, Jakob Grue Simonsen et al.

Fast similarity search is a key component in large-scale information retrieval, where semantic hashing has become a popular strategy for representing documents as binary hash codes. Recent advances in this area have been obtained through neural network based models: generative models trained by learning to reconstruct the original documents. We present a novel unsupervised generative semantic hashing approach, \textit{Ranking based Semantic Hashing} (RBSH) that consists of both a variational and a ranking based component. Similarly to variational autoencoders, the variational component is trained to reconstruct the original document conditioned on its generated hash code, and as in prior work, it only considers documents individually. The ranking component solves this limitation by incorporating inter-document similarity into the hash code generation, modelling document ranking through a hinge loss. To circumvent the need for labelled data to compute the hinge loss, we use a weak labeller and thus keep the approach fully unsupervised. Extensive experimental evaluation on four publicly available datasets against traditional baselines and recent state-of-the-art methods for semantic hashing shows that RBSH significantly outperforms all other methods across all evaluated hash code lengths. In fact, RBSH hash codes are able to perform similarly to state-of-the-art hash codes while using 2-4x fewer bits.

Casper Hansen, Christian Hansen, Stephen Alstrup et al.

Automatic fact-checking systems detect misinformation, such as fake news, by (i) selecting check-worthy sentences for fact-checking, (ii) gathering related information to the sentences, and (iii) inferring the factuality of the sentences. Most prior research on (i) uses hand-crafted features to select check-worthy sentences, and does not explicitly account for the recent finding that the top weighted terms in both check-worthy and non-check-worthy sentences are actually overlapping [15]. Motivated by this, we present a neural check-worthiness sentence ranking model that represents each word in a sentence by \textit{both} its embedding (aiming to capture its semantics) and its syntactic dependencies (aiming to capture its role in modifying the semantics of other terms in the sentence). Our model is an end-to-end trainable neural network for check-worthiness ranking, which is trained on large amounts of unlabelled data through weak supervision. Thorough experimental evaluation against state of the art baselines, with and without weak supervision, shows our model to be superior at all times (+13% in MAP and +28% at various Precision cut-offs from the best baseline with statistical significance). Empirical analysis of the use of weak supervision, word embedding pretraining on domain-specific data, and the use of syntactic dependencies of our model reveals that check-worthy sentences contain notably more identical syntactic dependencies than non-check-worthy sentences.

Dongsheng Wang, Quichi Li, Lucas Chaves Lima et al.

When the meaning of a phrase cannot be inferred from the individual meanings of its words (e.g., hot dog), that phrase is said to be non-compositional. Automatic compositionality detection in multi-word phrases is critical in any application of semantic processing, such as search engines; failing to detect non-compositional phrases can hurt system effectiveness notably. Existing research treats phrases as either compositional or non-compositional in a deterministic manner. In this paper, we operationalize the viewpoint that compositionality is contextual rather than deterministic, i.e., that whether a phrase is compositional or non-compositional depends on its context. For example, the phrase `green card' is compositional when referring to a green colored card, whereas it is non-compositional when meaning permanent residence authorization. We address the challenge of detecting this type of contextual compositionality as follows: given a multi-word phrase, we enrich the word embedding representing its semantics with evidence about its global context (terms it often collocates with) as well as its local context (narratives where that phrase is used, which we call usage scenarios). We further extend this representation with information extracted from external knowledge bases. The resulting representation incorporates both localized context and more general usage of the phrase and allows to detect its compositionality in a non-deterministic and contextual way. Empirical evaluation of our model on a dataset of phrase compositionality, manually collected by crowdsourcing contextual compositionality assessments, shows that our model outperforms state-of-the-art baselines notably on detecting phrase compositionality.

Niels Dalum Hansen, Kåre Mølbak, Ingemar Cox et al.

Consumption of antimicrobial drugs, such as antibiotics, is linked with antimicrobial resistance. Surveillance of antimicrobial drug consumption is therefore an important element in dealing with antimicrobial resistance. Many countries lack sufficient surveillance systems. Usage of web mined data therefore has the potential to improve current surveillance methods. To this end, we study how well antimicrobial drug consumption can be predicted based on web search queries, compared to historical purchase data of antimicrobial drugs. We present two prediction models (linear Elastic Net, and non-linear Gaussian Processes), which we train and evaluate on almost 6 years of weekly antimicrobial drug consumption data from Denmark and web search data from Google Health Trends. We present a novel method of selecting web search queries by considering diseases and drugs linked to antimicrobials, as well as professional and layman descriptions of antimicrobial drugs, all of which we mine from the open web. We find that predictions based on web search data are marginally more erroneous but overall on a par with predictions based on purchases of antimicrobial drugs. This marginal difference corresponds to $<1$\% point mean absolute error in weekly usage. Best predictions are reported when combining both web search and purchase data. This study contributes a novel alternative solution to the real-life problem of predicting (and hence monitoring) antimicrobial drug consumption, which is particularly valuable in countries/states lacking centralised and timely surveillance systems.

Niels Dalum Hansen, Kåre Mølbak, Ingemar J. Cox et al.

Influenza-like illness (ILI) estimation from web search data is an important web analytics task. The basic idea is to use the frequencies of queries in web search logs that are correlated with past ILI activity as features when estimating current ILI activity. It has been noted that since influenza is seasonal, this approach can lead to spurious correlations with features/queries that also exhibit seasonality, but have no relationship with ILI. Spurious correlations can, in turn, degrade performance. To address this issue, we propose modeling the seasonal variation in ILI activity and selecting queries that are correlated with the residual of the seasonal model and the observed ILI signal. Experimental results show that re-ranking queries obtained by Google Correlate based on their correlation with the residual strongly favours ILI-related queries.

Christina Lioma, Birger Larsen, Peter Ingwersen

When submitting queries to information retrieval (IR) systems, users often have the option of specifying which, if any, of the query terms are heavily dependent on each other and should be treated as a fixed phrase, for instance by placing them between quotes. In addition to such cases where users specify term dependence, automatic ways also exist for IR systems to detect dependent terms in queries. Most IR systems use both user and algorithmic approaches. It is not however clear whether and to what extent user-defined term dependence agrees with algorithmic estimates of term dependence, nor which of the two may fetch higher performance gains. Simply put, is it better to trust users or the system to detect term dependence in queries? To answer this question, we experiment with 101 crowdsourced search engine users and 334 queries (52 train and 282 test TREC queries) and we record 10 assessments per query. We find that (i) user assessments of term dependence differ significantly from algorithmic assessments of term dependence (their overlap is approximately 30%); (ii) there is little agreement among users about term dependence in queries, and this disagreement increases as queries become longer; (iii) the potential retrieval gain that can be fetched by treating term dependence (both user- and system-defined) over a bag of words baseline is reserved to a small subset (approxi-mately 8%) of the queries, and is much higher for low-depth than deep preci-sion measures. Points (ii) and (iii) constitute novel insights into term dependence.

Christina Lioma

Building machines that can understand text like humans is an AI-complete problem. A great deal of research has already gone into this, with astounding results, allowing everyday people to discuss with their telephones, or have their reading materials analysed and classified by computers. A prerequisite for processing text semantics, common to the above examples, is having some computational representation of text as an abstract object. Operations on this representation practically correspond to making semantic inferences, and by extension simulating understanding text. The complexity and granularity of semantic processing that can be realised is constrained by the mathematical and computational robustness, expressiveness, and rigour of the tools used. This dissertation contributes a series of such tools, diverse in their mathematical formulation, but common in their application to model semantic inferences when machines process text. These tools are principally expressed in nine distinct models that capture aspects of semantic dependence in highly interpretable and non-complex ways. This dissertation further reflects on present and future problems with the current research paradigm in this area, and makes recommendations on how to overcome them. The amalgamation of the body of work presented in this dissertation advances the complexity and granularity of semantic inferences that can be made automatically by machines.

Christina Lioma, Jakob Grue Simonsen, Birger Larsen

Recent discussions on alternative facts, fake news, and post truth politics have motivated research on creating technologies that allow people not only to access information, but also to assess the credibility of the information presented to them by information retrieval systems. Whereas technology is in place for filtering information according to relevance and/or credibility, no single measure currently exists for evaluating the accuracy or precision (and more generally effectiveness) of both the relevance and the credibility of retrieved results. One obvious way of doing so is to measure relevance and credibility effectiveness separately, and then consolidate the two measures into one. There at least two problems with such an approach: (I) it is not certain that the same criteria are applied to the evaluation of both relevance and credibility (and applying different criteria introduces bias to the evaluation); (II) many more and richer measures exist for assessing relevance effectiveness than for assessing credibility effectiveness (hence risking further bias). Motivated by the above, we present two novel types of evaluation measures that are designed to measure the effectiveness of both relevance and credibility in ranked lists of retrieval results. Experimental evaluation on a small human-annotated dataset (that we make freely available to the research community) shows that our measures are expressive and intuitive in their interpretation.

Christian Hansen, Casper Hansen, Niklas Hjuler et al.

The analysis of log data generated by online educational systems is an important task for improving the systems, and furthering our knowledge of how students learn. This paper uses previously unseen log data from Edulab, the largest provider of digital learning for mathematics in Denmark, to analyse the sessions of its users, where 1.08 million student sessions are extracted from a subset of their data. We propose to model students as a distribution of different underlying student behaviours, where the sequence of actions from each session belongs to an underlying student behaviour. We model student behaviour as Markov chains, such that a student is modelled as a distribution of Markov chains, which are estimated using a modified k-means clustering algorithm. The resulting Markov chains are readily interpretable, and in a qualitative analysis around 125,000 student sessions are identified as exhibiting unproductive student behaviour. Based on our results this student representation is promising, especially for educational systems offering many different learning usages, and offers an alternative to common approaches like modelling student behaviour as a single Markov chain often done in the literature.

Wei Lu, Qikai Cheng, Christina Lioma

TextRank is a variant of PageRank typically used in graphs that represent documents, and where vertices denote terms and edges denote relations between terms. Quite often the relation between terms is simple term co-occurrence within a fixed window of k terms. The output of TextRank when applied iteratively is a score for each vertex, i.e. a term weight, that can be used for information retrieval (IR) just like conventional term frequency based term weights. So far, when computing TextRank term weights over co- occurrence graphs, the window of term co-occurrence is al- ways ?xed. This work departs from this, and considers dy- namically adjusted windows of term co-occurrence that fol- low the document structure on a sentence- and paragraph- level. The resulting TextRank term weights are used in a ranking function that re-ranks 1000 initially returned search results in order to improve the precision of the ranking. Ex- periments with two IR collections show that adjusting the vicinity of term co-occurrence when computing TextRank term weights can lead to gains in early precision.

Birger Larsen, Christina Lioma, Arjen de Vries

The ECIR half-day workshop on Task-Based and Aggregated Search (TBAS) was held in Barcelona, Spain on 1 April 2012. The program included a keynote talk by Professor Jarvelin, six full paper presentations, two poster presentations, and an interactive discussion among the approximately 25 participants. This report overviews the aims and contents of the workshop and outlines the major outcomes.

Christina Lioma, Roi Blanco

Automatic language processing tools typically assign to terms so-called weights corresponding to the contribution of terms to information content. Traditionally, term weights are computed from lexical statistics, e.g., term frequencies. We propose a new type of term weight that is computed from part of speech (POS) n-gram statistics. The proposed POS-based term weight represents how informative a term is in general, based on the POS contexts in which it generally occurs in language. We suggest five different computations of POS-based term weights by extending existing statistical approximations of term information measures. We apply these POS-based term weights to information retrieval, by integrating them into the model that matches documents to queries. Experiments with two TREC collections and 300 queries, using TF-IDF & BM25 as baselines, show that integrating our POS-based term weights to retrieval always leads to gains (up to +33.7% from the baseline). Additional experiments with a different retrieval model as baseline (Language Model with Dirichlet priors smoothing) and our best performing POS-based term weight, show retrieval gains always and consistently across the whole smoothing range of the baseline.

Christina Lioma, Birger Larsen, Hinrich Schütze et al.

Interactive Information Retrieval refers to the branch of Information Retrieval that considers the retrieval process with respect to a wide range of contexts, which may affect the user's information seeking experience. The identification and representation of such contexts has been the object of the principle of Polyrepresentation, a theoretical framework for reasoning about different representations arising from interactive information retrieval in a given context. Although the principle of Polyrepresentation has received attention from many researchers, not much empirical work has been done based on it. One reason may be that it has not yet been formalised mathematically. In this paper we propose an up-to-date and exible mathematical formalisation of the principle of Polyrepresentation for information needs. Specifically, we apply Subjective Logic to model different representations of information needs as beliefs marked by degrees of uncertainty. We combine such beliefs using different logical operators, and we discuss these combinations with respect to different retrieval scenarios and situations. A formal model is introduced and discussed, with illustrative applications to the modelling of information needs.

Christina Lioma, Birger Larsen, Peter Ingwersen

According to the principle of polyrepresentation, retrieval accuracy may improve through the combination of multiple and diverse information object representations about e.g. the context of the user, the information sought, or the retrieval system. Recently, the principle of polyrepresentation was mathematically expressed using subjective logic, where the potential suitability of each representation for improving retrieval performance was formalised through degrees of belief and uncertainty. No experimental evidence or practical application has so far validated this model. We extend the work of Lioma et al. (2010), by providing a practical application and analysis of the model. We show how to map the abstract notions of belief and uncertainty to real-life evidence drawn from a retrieval dataset. We also show how to estimate two different types of polyrepresentation assuming either (a) independence or (b) dependence between the information objects that are combined. We focus on the polyrepresentation of different types of context relating to user information needs (i.e. work task, user background knowledge, ideal answer) and show that the subjective logic model can predict their optimal combination prior and independently to the retrieval process.

Christina Lioma, Birger Larsen, Wei Lu

Typically, every part in most coherent text has some plausible reason for its presence, some function that it performs to the overall semantics of the text. Rhetorical relations, e.g. contrast, cause, explanation, describe how the parts of a text are linked to each other. Knowledge about this socalled discourse structure has been applied successfully to several natural language processing tasks. This work studies the use of rhetorical relations for Information Retrieval (IR): Is there a correlation between certain rhetorical relations and retrieval performance? Can knowledge about a document's rhetorical relations be useful to IR? We present a language model modification that considers rhetorical relations when estimating the relevance of a document to a query. Empirical evaluation of different versions of our model on TREC settings shows that certain rhetorical relations can benefit retrieval effectiveness notably (> 10% in mean average precision over a state-of-the-art baseline).

Christina Lioma, Niels Dalum Hansen

Compositionality in language refers to how much the meaning of some phrase can be decomposed into the meaning of its constituents and the way these constituents are combined. Based on the premise that substitution by synonyms is meaning-preserving, compositionality can be approximated as the semantic similarity between a phrase and a version of that phrase where words have been replaced by their synonyms. Different ways of representing such phrases exist (e.g., vectors [1] or language models [2]), and the choice of representation affects the measurement of semantic similarity. We propose a new compositionality detection method that represents phrases as ranked lists of term weights. Our method approximates the semantic similarity between two ranked list representations using a range of well-known distance and correlation metrics. In contrast to most state-of-the-art approaches in compositionality detection, our method is completely unsupervised. Experiments with a publicly available dataset of 1048 human-annotated phrases shows that, compared to strong supervised baselines, our approach provides superior measurement of compositionality using any of the distance and correlation metrics considered.

Christina Lioma, Birger Larsen, Wei Lu et al.

Much of the information processed by Information Retrieval (IR) systems is unreliable, biased, and generally untrustworthy [1], [2], [3]. Yet, factuality & objectivity detection is not a standard component of IR systems, even though it has been possible in Natural Language Processing (NLP) in the last decade. Motivated by this, we ask if and how factuality & objectivity detection may benefit IR. We answer this in two parts. First, we use state-of-the-art NLP to compute the probability of document factuality & objectivity in two TREC collections, and analyse its relation to document relevance. We find that factuality is strongly and positively correlated to document relevance, but objectivity is not. Second, we study the impact of factuality & objectivity to retrieval effectiveness by treating them as query independent features that we combine with a competitive language modelling baseline. Experiments with 450 TREC queries show that factuality improves precision >10% over strong baselines, especially for uncurated data used in web search; objectivity gives mixed results. An overall clear trend is that document factuality & objectivity is much more beneficial to IR when searching uncurated (e.g. web) documents vs. curated (e.g. state documentation and newswire articles). To our knowledge, this is the first study of factuality & objectivity for back-end IR, contributing novel findings about the relation between relevance and factuality/objectivity, and statistically significant gains to retrieval effectiveness in the competitive web search task.

Casper Petersen, Jakob Grue Simonsen, Kalervo Jarvelin et al.

Divergence From Randomness (DFR) ranking models assume that informative terms are distributed in a corpus differently than non-informative terms. Different statistical models (e.g. Poisson, geometric) are used to model the distribution of non-informative terms, producing different DFR models. An informative term is then detected by measuring the divergence of its distribution from the distribution of non-informative terms. However, there is little empirical evidence that the distributions of non-informative terms used in DFR actually fit current datasets. Practically this risks providing a poor separation between informative and non-informative terms, thus compromising the discriminative power of the ranking model. We present a novel extension to DFR, which first detects the best-fitting distribution of non-informative terms in a collection, and then adapts the ranking computation to this best-fitting distribution. We call this model Adaptive Distributional Ranking (ADR) because it adapts the ranking to the statistics of the specific dataset being processed each time. Experiments on TREC data show ADR to outperform DFR models (and their extensions) and be comparable in performance to a query likelihood language model (LM).

Niels Dalum Hansen, Christina Lioma, Kåre Mølbak

We present a method that uses ensemble learning to combine clinical and web-mined time-series data in order to predict future vaccination uptake. The clinical data is official vaccination registries, and the web data is query frequencies collected from Google Trends. Experiments with official vaccine records show that our method predicts vaccination uptake effectively (4.7 Root Mean Squared Error). Whereas performance is best when combining clinical and web data, using solely web data yields comparative performance. To our knowledge, this is the first study to predict vaccination uptake using web data (with and without clinical data).

Brian Brost, Yevgeny Seldin, Ingemar J. Cox et al.

New ranking algorithms are continually being developed and refined, necessitating the development of efficient methods for evaluating these rankers. Online ranker evaluation focuses on the challenge of efficiently determining, from implicit user feedback, which ranker out of a finite set of rankers is the best. Online ranker evaluation can be modeled by dueling ban- dits, a mathematical model for online learning under limited feedback from pairwise comparisons. Comparisons of pairs of rankers is performed by interleaving their result sets and examining which documents users click on. The dueling bandits model addresses the key issue of which pair of rankers to compare at each iteration, thereby providing a solution to the exploration-exploitation trade-off. Recently, methods for simultaneously comparing more than two rankers have been developed. However, the question of which rankers to compare at each iteration was left open. We address this question by proposing a generalization of the dueling bandits model that uses simultaneous comparisons of an unrestricted number of rankers. We evaluate our algorithm on synthetic data and several standard large-scale online ranker evaluation datasets. Our experimental results show that the algorithm yields orders of magnitude improvement in performance compared to stateof- the-art dueling bandit algorithms.

Brian Brost, Ingemar J. Cox, Yevgeny Seldin et al.

Online ranker evaluation is a key challenge in information retrieval. An important task in the online evaluation of rankers is using implicit user feedback for inferring preferences between rankers. Interleaving methods have been found to be efficient and sensitive, i.e. they can quickly detect even small differences in quality. It has recently been shown that multileaving methods exhibit similar sensitivity but can be more efficient than interleaving methods. This paper presents empirical results demonstrating that existing multileaving methods either do not scale well with the number of rankers, or, more problematically, can produce results which substantially differ from evaluation measures like NDCG. The latter problem is caused by the fact that they do not correctly account for the similarities that can occur between rankers being multileaved. We propose a new multileaving method for handling this problem and demonstrate that it substantially outperforms existing methods, in some cases reducing errors by as much as 50%.

Christina Lioma, Fabien Tarissan, Jakob Grue Simonsen et al.

Document coherence describes how much sense text makes in terms of its logical organisation and discourse flow. Even though coherence is a relatively difficult notion to quantify precisely, it can be approximated automatically. This type of coherence modelling is not only interesting in itself, but also useful for a number of other text processing tasks, including Information Retrieval (IR), where adjusting the ranking of documents according to both their relevance and their coherence has been shown to increase retrieval effectiveness [34,37]. The state of the art in unsupervised coherence modelling represents documents as bipartite graphs of sentences and discourse entities, and then projects these bipartite graphs into one-mode undirected graphs. However, one-mode projections may incur significant loss of the information present in the original bipartite structure. To address this we present three novel graph metrics that compute document coherence on the original bipartite graph of sentences and entities. Evaluation on standard settings shows that: (i) one of our coherence metrics beats the state of the art in terms of coherence accuracy; and (ii) all three of our coherence metrics improve retrieval effectiveness because, as closer analysis reveals, they capture aspects of document quality that go undetected by both keyword-based standard ranking and by spam filtering. This work contributes document coherence metrics that are theoretically principled, parameter-free, and useful to IR.

Christina Lioma, Birger Larsen, Casper Petersen et al.

What if Information Retrieval (IR) systems did not just retrieve relevant information that is stored in their indices, but could also "understand" it and synthesise it into a single document? We present a preliminary study that makes a first step towards answering this question. Given a query, we train a Recurrent Neural Network (RNN) on existing relevant information to that query. We then use the RNN to "deep learn" a single, synthetic, and we assume, relevant document for that query. We design a crowdsourcing experiment to assess how relevant the "deep learned" document is, compared to existing relevant documents. Users are shown a query and four wordclouds (of three existing relevant documents and our deep learned synthetic document). The synthetic document is ranked on average most relevant of all.

Casper Petersen, Christina Lioma, Jakob Grue Simonsen et al.

We present two novel models of document coherence and their application to information retrieval (IR). Both models approximate document coherence using discourse entities, e.g. the subject or object of a sentence. Our first model views text as a Markov process generating sequences of discourse entities (entity n-grams); we use the entropy of these entity n-grams to approximate the rate at which new information appears in text, reasoning that as more new words appear, the topic increasingly drifts and text coherence decreases. Our second model extends the work of Guinaudeau & Strube [28] that represents text as a graph of discourse entities, linked by different relations, such as their distance or adjacency in text. We use several graph topology metrics to approximate different aspects of the discourse flow that can indicate coherence, such as the average clustering or betweenness of discourse entities in text. Experiments with several instantiations of these models show that: (i) our models perform on a par with two other well-known models of text coherence even without any parameter tuning, and (ii) reranking retrieval results according to their coherence scores gives notable performance gains, confirming a relation between document coherence and relevance. This work contributes two novel models of document coherence, the application of which to IR complements recent work in the integration of document cohesiveness or comprehensibility to ranking [5, 56].

Christina Lioma, Jakob Grue Simonsen, Birger Larsen et al.

Modelling term dependence in IR aims to identify co-occurring terms that are too heavily dependent on each other to be treated as a bag of words, and to adapt the indexing and ranking accordingly. Dependent terms are predominantly identified using lexical frequency statistics, assuming that (a) if terms co-occur often enough in some corpus, they are semantically dependent; (b) the more often they co-occur, the more semantically dependent they are. This assumption is not always correct: the frequency of co-occurring terms can be separate from the strength of their semantic dependence. E.g. "red tape" might be overall less frequent than "tape measure" in some corpus, but this does not mean that "red"+"tape" are less dependent than "tape"+"measure". This is especially the case for non-compositional phrases, i.e. phrases whose meaning cannot be composed from the individual meanings of their terms (such as the phrase "red tape" meaning bureaucracy). Motivated by this lack of distinction between the frequency and strength of term dependence in IR, we present a principled approach for handling term dependence in queries, using both lexical frequency and semantic evidence. We focus on non-compositional phrases, extending a recent unsupervised model for their detection [21] to IR. Our approach, integrated into ranking using Markov Random Fields [31], yields effectiveness gains over competitive TREC baselines, showing that there is still room for improvement in the very well-studied area of term dependence in IR.