The objective of high-recall information retrieval (HRIR) is to retrieve as many relevant documents as possible for a given search topic. One type approach to HRIR is Technology-Assisted Review (TAR), which uses information retrieval and machine learning techniques to aid the review of large document collections. TAR systems are commonly used in legal eDiscovery and systematic literature reviews. Successful TAR systems are able to find the majority of relevant documents using the least number of assessments. Commonly used retrospective evaluation assumes that the system achieves a specific, fixed recall level first, and then measures the precision or work saved (e.g., precision at r% recall). This approach can cause problems related to understanding the behaviour of evaluation measures in a fixed recall setting. It is also problematic when estimating time and money savings during technology-assisted reviews.
This paper presents a new visual analytics tool to explore the dynamics of evaluation measures depending on recall level. We implemented 18 evaluation measures based on the confusion matrix terms, both from general IR tasks and specific to TAR. The tool allows for a comparison of the behaviour of these measures in a fixed recall evaluation setting. It can also simulate savings in time and money and a count of manual vs automatic assessments for different datasets depending on the model quality. The tool is open-source, and the demo is available under the following URL: https://vombat.streamlit.app

VoMBaT: A Tool for Visualising Evaluation Measure Behaviour in High-Recall Search Tasks

Transformer-based deep learning is currently the state-of-the-art in many NLP and IR tasks. However, fine-tuning such Transformers for specific tasks, especially in scenarios of ever-expanding volumes of data with constant re-training requirements and budget constraints, is costly (computationally and financially) and energy-consuming. In this paper, we focus on Instance Selection (IS) – a set of methods focused on selecting the most representative documents for training,  aimed at maintaining (or improving) classification effectiveness while reducing total time for training (or fine-tuning). We propose E2SC-IS -- Effective, Efficient, and Scalable Confidence-Based IS -- a two-step framework with a particular focus on Transformers and large datasets. E2SC-IS estimates the probability of each instance being removed from the training set based on scalable, fast, and calibrated weak classifiers. E2SC-IS  also exploits iterative heuristics to estimate a near-optimal reduction rate. Our solution can reduce the training sets by 29% on average while maintaining the effectiveness in all datasets, with speedup gains up to 70%,  scaling for very large datasets (something that the baselines cannot do).

An Effective, Efficient, and Scalable Confidence-Based Instance Selection Framework for Transformer-Based Text Classification

Conversational systems can be particularly effective in supporting complex information seeking scenarios with evolving information needs.
Finding the right products on an e-commerce platform is one such scenario, where a conversational agent would need to be able to provide search capabilities over the item catalog, understand and make recommendations based on the user's preferences, and answer a range of questions related to items and their usage.
Yet, existing conversational datasets do not fully support the idea of mixing different conversational goals (i.e., search, recommendation, and question answering) and instead focus on a single goal.
To address this, we introduce MG-ShopDial: a dataset of conversations mixing different goals in the domain of e-commerce.
Specifically, we make the following contributions.
First, we develop a coached human-human data collection protocol where each dialogue participant is given a set of instructions, instead of a specific script or answers to choose from.
Second, we implement a data collection tool to facilitate the collection of multi-goal conversations via a web chat interface, using the above protocol.
Third, we create the MG-ShopDial collection, which contains 64 high-quality dialogues with a total of 2,196 utterances for e-commerce scenarios of varying complexity. The dataset is additionally annotated with both intents and goals on the utterance level.
Finally, we present an analysis of this dataset and identify multi-goal conversational patterns.

MG-ShopDial: A Multi-Goal Conversational Dataset for e-Commerce

This paper presents the solution proposed by team FAST Machine Translation to the shared tasks of JOKER CLEF 2022 Automatic pun and humour translation. State-of-the-art Transformer-based models are used to solve the three tasks introduced in the JOKER CLEF workshop. The Transformer model is a kind of neural network that tries to learn the contextual information from the sequential data by implicitly comprehending the existing relationships. In task 1, given a piece of text, we need to classify/explain any instance of wordplay is present in it or not. The proposed solution to task 1 combines the pipeline of token classification, text classification, and text generation. In task 2, we need to translate single words (nouns) containing a wordplay. This task is mapped to the problem of question answering (Q/A) on programmatically extracted texts from the OPUS parallel corpus. In task 3, contestants are required to translate the entire phrase containing the wordplay.  Sequence-to sequence translation models are used to solve this task. The team has adopted different strategies for each task as they suited to the requirements therein. The paper reports proposed solutions, implementation details, experimental studies, and results obtained in JOKER CLEF 2022 automatic pun and humour translation tasks.

FAST-MT Participation for the JOKER CLEF-2022 Automatic Pun and Humour Translation Tasks

Conversational question answering (ConvQA) tackles sequential information needs where contexts in follow-up questions are left implicit. Current ConvQA systems operate over homogeneous sources of information: either a knowledge base (KB), or a text corpus, or a collection of tables. This paper addresses the novel issue of jointly tapping into all of these together, this way boosting answer coverage and confidence. We present CONVINSE, an end-to-end pipeline for ConvQA over heterogeneous sources, operating in three stages: i) learning an explicit structured representation of an incoming question and its conversational context, ii) harnessing this frame-like representation to uniformly capture relevant evidences from KB, text, and tables, and iii) running a fusion-in-decoder model to generate the answer. We construct and release the first benchmark, ConvMix, for ConvQA over heterogeneous sources, comprising $3000$ real-user conversations with $16000$ questions, along with entity annotations, completed question utterances, and question paraphrases. Experiments demonstrate the viability and advantages of our method, compared to state-of-the-art baselines.

Conversational Question Answering on Heterogeneous Sources

The rise of personal assistants has made conversational question answering (ConvQA) a very popular mechanism for user-system interaction. State-of-the-art methods for ConvQA over knowledge graphs (KGs) can only learn from crisp question-answer pairs found in popular benchmarks. In reality, however, such training data is hard to come by: users would rarely mark answers explicitly as correct or wrong. In this work, we take a step towards a more natural learning paradigm - from noisy and implicit feedback via question reformulations. A reformulation is likely to be triggered by an incorrect system response, whereas a new follow-up question could be a positive signal on the previous turn's answer. We present a reinforcement learning model, termed CONQUER, that can learn from a conversational stream of questions and reformulations. CONQUER models the answering process as multiple agents walking in parallel on the KG, where the walks are determined by actions sampled using a policy network. This policy network takes the question along with the conversational context as inputs and is trained via noisy rewards obtained from the reformulation likelihood. To evaluate CONQUER, we create and release ConvRef, a benchmark with about 11k natural conversations containing around 205k reformulations. Experiments show that CONQUER successfully learns to answer conversational questions from noisy reward signals, significantly improving over a state-of-the-art baseline.

Reinforcement Learning from Reformulations in Conversational Question Answering over Knowledge Graphs

Understanding why specific items are recommended to users can significantly increase their trust and satisfaction in the system. While neural recommenders have become the state-of-the-art in recent years, the complexity of deep models still makes the generation of tangible explanations for end users a challenging problem. Existing methods are usually based on attention distributions over a variety of features, which are still questionable regarding their suitability as explanations, and rather unwieldy to grasp for an end user. Counterfactual explanations based on a small set of the user's own actions have been shown to be an acceptable solution to the tangibility problem. However, current work on such counterfactuals cannot be readily applied to neural models. In this work, we propose ACCENT, the first general framework for finding counterfactual explanations for neural recommenders. It extends recently-proposed influence functions for identifying training points most relevant to a recommendation, from a single to a pair of items, while deducing a counterfactual set in an iterative process. We use ACCENT to generate counterfactual explanations for two popular neural models, Neural Collaborative Filtering (NCF) and Relational Collaborative Filtering (RCF), and demonstrate its feasibility on a sample of the popular MovieLens 100K dataset.

Counterfactual Explanations for Neural Recommenders

Existing fair ranking systems, especially those designed to be demographically fair, assume that accurate demographic information about individuals is available to the ranking algorithm. In practice, however, this assumption may not hold -- in real-world contexts like ranking job applicants or credit seekers, social and legal barriers may prevent algorithm operators from collecting peoples' demographic information. In these cases, algorithm operators may attempt to infer peoples' demographics and then supply these inferences as inputs to the ranking algorithm.
In this study, we investigate how uncertainty and errors in demographic inference impact the fairness offered by fair ranking algorithms. Using simulations and three case studies with real datasets, we show how demographic inferences drawn from real systems can lead to unfair rankings. Our results suggest that developers should not use inferred demographic data as input to fair ranking algorithms, unless the inferences are extremely accurate.

When Fair Ranking Meets Uncertain Inference

In web search on debated topics, algorithmic and cognitive biases strongly influence how users consume and process information. Recent research has shown that this can lead to a search engine manipulation effect (SEME): when search result rankings are biased towards a particular viewpoint, users tend to adopt this favored viewpoint. To better understand the mechanisms underlying SEME, we present a pre-registered, 5 × 3 factorial user study investigating whether order effects (i.e., users adopting the viewpoint pertaining to higher-ranked documents) can cause SEME. For five different debated topics, we evaluated attitude change after exposing participants with mild pre-existing attitudes to search results that were overall viewpoint-balanced but reflected one of three levels of algorithmic ranking bias. We found that attitude change did not differ across levels of ranking bias and did not vary based on individual user differences. Our results thus suggest that order effects may not be an underlying mechanism of SEME. Exploratory analyses lend support to the presence of exposure effects (i.e., users adopting the majority viewpoint among the results they examine) as a contributing factor to users’ attitude change. We discuss how our findings can inform the design of user bias mitigation strategies.

This Is Not What We Ordered: Exploring Why Biased Search Result Rankings Affect User Attitudes on Debated Topics

Disinformation and fake news have posed detrimental effects on individuals and society in recent years, attracting broad attention to fake news detection. The majority of existing fake news detection algorithms focus on mining news content and/or the surrounding exogenous context for discovering deceptive signals; while the endogenous preference of a user when he/she decides to spread a piece of fake news or not is ignored. The confirmation bias theory has indicated that a user is more likely to spread a piece of fake news when it confirms his/her existing beliefs/preferences. Users' historical, social engagements such as posts provide rich information about users' preferences toward news and have great potential to advance fake news detection. However, the work on exploring user preference for fake news detection is somewhat limited. Therefore, in this paper, we study the novel problem of exploiting user preference for fake news detection. We propose a new framework, UPFD, which simultaneously captures various signals from user preferences by joint content and graph modeling. Experimental results on real-world datasets demonstrate the effectiveness of the proposed framework. We release our code and data as a  benchmark for  GNN-based fake news detection:https://github.com/safe-graph/GNN-FakeNews.

User Preference-aware Fake News Detection

The use of IR methodology in the evaluation of recommender systems has become common practice in recent years. IR metrics have been found however to be strongly biased towards rewarding algorithms that recommend popular items –the same bias that state of the art recommendation algorithms display. Recent research has confirmed and measured such biases, and proposed methods to avoid them. The fundamental question remains open though whether popularity is really a bias we should avoid or not; whether it could be a useful and reliable signal in recommendation, or it may be unfairly rewarded by the experimental biases. We address this question at a formal level by identifying and modeling the conditions that can determine the answer, in terms of dependencies between key random variables, involving item rating, discovery and relevance. We find conditions that guarantee popularity to be effective or quite the opposite, and for the measured metric values to reflect a true effectiveness, or qualitatively deviate from it. We exemplify and confirm the theoretical findings with empirical results. We build a crowdsourced dataset devoid of the usual biases displayed by common publicly available data, in which we illustrate contradictions between the accuracy that would be measured in a common biased offline experimental setting, and the actual accuracy that can be measured with unbiased observations.

Should I Follow the Crowd? A Probabilistic Analysis of the Effectiveness of Popularity in Recommender Systems

We conducted a laboratory-based observational study where pairs of people performed search tasks communicating verbally. Examination of the discourse allowed commonly used interactions to be identified for Spoken Conversational Search (SCS). We compared the interactions to existing models of search behaviour. We find that SCS is more complex and interactive than traditional search. This work enhances our understanding of different search behaviours and proposes research opportunities for an audio-only search system. Future work will focus on creating models of search behaviour for SCS and evaluating these against actual SCS systems.

Informing the Design of Spoken Conversational Search

Many applications require semantic understanding of short texts, and inferring discriminative and coherent latent topics is a critical and fundamental task in these applications. Conventional topic models largely rely on word co-occurrences to derive topics from a collection of documents. However, due to the length of each document, short texts are much more sparse in terms of word co-occurrences. Recent studies show that the Dirichlet Multinomial Mixture (DMM) model is effective for topic inference over short texts by assuming that each piece of short text is generated by a single topic. However, DMM has two main limitations. First, even though it seems reasonable to assume that each short text has only one topic because of its shortness, the definition of “shortness” is subjective and the length of the short texts is dataset dependent. That is, the single-topic assumption may be too strong for some datasets. To address this limitation, we propose to model the topic number as a Poisson distribution, allowing each short text to be associated with a small number of topics (e.g., one to three topics). This model is named PDMM. Second, DMM (and also PDMM) does not have access to background knowledge (e.g., semantic relations between words) when modeling short texts. When a human being interprets a piece of short text, the understanding is not solely based on its content words, but also their semantic relations. Recent advances in word embeddings offer effective learning of word semantic relations from a large corpus. Such auxiliary word embeddings enable us to address the second limitation. To this end, we propose to promote the semantically related words under the same topic during the sampling process, by using the generalized Pólya urn (GPU) model. Through the GPU model, background knowledge about word semantic relations learned from millions of external documents can be easily exploited to improve topic modeling for short texts. By directly extending the PDMM model with the GPU model, we propose two more effective topic models for short texts, named GPU-DMM and GPU-PDMM. Through extensive experiments on two real-world short text collections in two languages, we demonstrate that PDMM achieves better topic representations than state-of-the-art models, measured by topic coherence.
The learned topic representation leads to better accuracy in a text classification task, as an indirect evaluation. Both GPU-DMM and GPU-PDMM further improve topic coherence and text classification accuracy. GPU-PDMM outperforms GPU-DMM at the price of higher computational costs.

Enhancing Topic Modeling for Short Texts with Auxiliary Word Embeddings

Deep learning models have been employed to perform IR tasks and have shown competitive results. Depending on the structure of the models, previous deep IR models could be roughly divided into: representation-based models and interaction-based models. A number of experiments have been conducted to test these models, but often under different conditions, making it difficult to draw a clear conclusion on their comparison. In order to compare the two learning schemas for ad hoc search under the same condition, we build similar convolution networks to learn either representations or interaction patterns between document and query and test them on the same test collection. In addition, we also propose multi- level matching models to cope with various types of query, rather than the existing single-level matching. Our experiments show that interaction-based approach generally performs better than representation-based approach, and multi-level matching performs better than single-level matching. We will provide some possible explanations to these observations.

ACM SIGIR Badging Submissions