In this study, we explore the concept of anchoring bias in the context of sequential sentiment analysis of review corpora. We introduce a novel approach that involves using a carefully selected, limited group of reviews at the beginning of the annotation process for calibration, aiming to reduce this bias. Through an extensive set of experiments we confirm the existence of sentiment bias and demonstrate that indeed its impact can be moderated through initial calibration. We also demonstrate that the composition of the calibration set is critical, underscoring the importance of establishing sound criteria for selecting these initial reviews. By comparing the accuracy of annotators who utilized our calibration method against those who did not calibrate or used a randomly chosen set for calibration, we found that our method significantly decreases the overall annotation error. Moreover, the guidelines we develop for selecting the calibration set prove to be highly effective and adaptable, even when applied to a domain other than the one they were originally developed for. Acknowledging the overhead of labeling calibration reviews, we demonstrate that this approach is more efficient compared to eliciting multiple sentiment scores per review, a common strategy to reduce Mean Absolute Error (MAE). Our findings reveal that the proposed calibration method significantly reduces resource expenditure, compared to relying on parallel labeling, while maintaining accuracy.

Purpose This study investigates how public engagement in political discourse on Facebook evolves during periods of continuous political instability. While prior research has examined engagement drivers during routine and election periods, this study focuses on how political communication and public responsiveness change across repeated election cycles under sustained crisis conditions. Design/methodology/approach Using computational methods, we examined more than 8,000 posts published by leading Israeli politicians and millions of public responses across four election cycles (2019–2021). The analysis tracks changes in sentiment, content features and communication patterns and examines their associations with public engagement as the political instability persisted. Findings Results reveal several notable patterns. Although negative sentiment dominated across all four campaigns, negativity consistently declined over time. Contrary to expectations, longer posts were associated with higher engagement. Overall, public engagement increased during the first three election cycles and then stabilized slightly in the fourth, suggesting competing dynamics of heightened political attention and gradual fatigue during prolonged instability. Originality/value This study provides a longitudinal examination of political discourse and engagement during an extended period of political instability. By jointly examining sentiment, content characteristics and engagement indicators, it contributes to understanding how online political communication evolves under prolonged crisis conditions and how competing dynamics of emotional amplification, heightened public attention and electoral fatigue may shape engagement over time. Peer review The peer review history for this article is available at: Link to the website.

This article presents a novel approach for the citation network construction from Jewish Responsa literature based on automatic extraction of references from texts. Jewish Responsa literature contains thousands of answers to questions related to Jewish law (Halachah), spanning over 1,300 years by authors from all over the world. This literature is abundant with references, but because of their high lexical and format variability their automatic identification and extraction is very challenging. In this article we present a novel, multi layered approach that splits the reference extraction task into two main subtasks: i) reference boundaries’ identification; ii) reference internal components’ identification. We experimented with several different machine learning models: CRF (Conditional Random Field) model, BERT (Bidirectional Encoder Representations from Transformers) model, and a combined approach, BERT-CRF. Additionally, we examined the influence of the training corpus on the model’s accuracy by comparing the performance of the models trained on modern Hebrew vs. Rabbinic Hebrew. We found that the best results were achieved by a BERT-CRF model trained on Rabbinic Hebrew. The constructed network can be utilized to build various tools for analyzing trends and influences in the Jewish Halachic corpus, such as the most influencing authors, the authors’ sources of authority, and their evolution over time and place.

The narratives expressed by patients during psychotherapy sessions provide a cognitive framework through which they process their experiences. While these narratives offer therapists valuable insights, automatically detecting them presents unique challenges due to their unstructured nature, parallel storytelling patterns, and frequent interruptions - characteristics distinct from narrative detection in traditional domains like news or social media. This paper first establishes a formal definition of therapeutic narratives as continuous segments containing characters, actions, and consequent changes, situated in specific temporal-spatial contexts. Working with domain experts, we developed and validated annotation guidelines achieving high inter-annotator agreement on a comprehensive corpus of Hebrew psychotherapy sessions. Building on this foundation, we introduce APhyND, a novel layered framework specifically designed for automatic narrative detection in psychotherapy session transcripts. Our framework makes several key innovations: (1) a clinically-informed architecture incorporating therapist-validated narrative criteria and specialized feature extraction for therapeutic discourse (2) integration of deep contextual understanding through BERT with sequence modeling via Conditional Random Fields, specifically adapted for therapeutic discourse, and (3) the first comprehensive solution for narrative detection in Hebrew psychotherapy transcripts, addressing the challenges of morphologically rich languages in therapeutic contexts. The framework employs a clinically-informed architecture incorporating therapist-validated narrative criteria, specialized feature extraction for therapeutic discourse, and adaptive threshold adjustment to handle inherent class imbalance. To evaluate the framework’s effectiveness, we conducted experiments on a dataset of 38,434 sentences from 79 psychotherapy sessions in Hebrew. Our results demonstrate significant improvements over existing approaches, achieving an f1-score of 0.804, with particularly strong performance in handling interrupted narratives and speaker transitions. The framework’s ability to process full therapy sessions in real-time while maintaining high accuracy makes it particularly valuable for clinical applications, addressing a critical gap in automated psychotherapy analysis tools.

This paper focuses in the inherent anchoring bias present in sequential reviews-sentiment corpora annotation processes. It proposes employing a limited subset of meticulously chosen reviews at the outset of the process, as a means of calibration, effectively mitigating the phenomenon. Through extensive experimentation we validate the phenomenon of sentiment bias in the annotation process and show that its magnitude can be influenced by pre-calibration. Furthermore, we show that the choice of the calibration set matters, hence the need for effective guidelines for choosing the reviews to be included in it. A comparison of annotators performance with the proposed calibration to annotation processes that do not use calibration or use a randomly-picked calibration set, reveals that indeed the calibration set picked is highly effective---it manages to substantially reduce the average absolute error compared to the other cases. Furthermore, the proposed selection guidelines are found to be highly robust in picking an effective calibration set also for domains different than the one based on which these rules were extracted.

Citation networks enable analysis of author groups, defining in-group dynamics, and mapping out inter-group relationships. While intellectual diversity and inclusiveness is one of the important principles of modern scholarship, it is intriguing to explore the extent to which these principles apply to historical communities of leaders and intellectuals. This paper introduces a novel methodological framework aimed at assessing the degree of viewpoint plurality and diversity of historical scholarship communities, through an in-depth analysis of the citations used in their literature, which has become possible due to the recently developed advanced computational analysis techniques. To achieve this goal, we have devised a set of new network-based indicators grounded in standard network metrics. These indicators can be applied at both the individual author and community levels. The developed methodology was applied to a citation network automatically constructed from a corpus of Rabbinic Halachic literature spanning the 10th to 15th centuries. This corpus includes over 5,000 citations from hundreds of books authored by approximately 140 Rabbinic scholars from six diverse geographic communities. We found that most of the authors and communities cite many more external resources from other communities than their own reflecting a willingness to engage with a diverse range of viewpoints. A more in-depth analysis based on the novel proportional diversity measures unveils more intriguing insights. Contrary to expectations, communities with the greatest number of external citations, such as Spain and Ashkenaz, surprisingly exhibit lower levels of viewpoint plurality compared to others, such as Italy and North Africa, elucidating a key finding of the study.

The practice of online astroturfing has become increasingly pervasive in recent years, with the growth in popularity of social media. Astroturfing consists of promoting social, political, or other agendas in a non-transparent or deceitful way, where the promoters masquerade as normative users while acting behind a mask that conceals their true identity, and at times that they are not human. In politics, astroturfing is currently considered one of the most severe online threats to democracy. The ability to automatically identify astroturfers thus constitutes a first step in eradicating this threat. We present a complete framework for handling a dataset of profiles, from data collection and efficient labeling, through feature extraction, and finally, to the identification of astroturfers lurking in the dataset. The data were collected over a period of 15 months, during which three consecutive elections were held in Israel. These raw data are unique in scope and size, consisting of several million public comments and reactions to posts on political candidates’ pages. For the manual labeling stage, we present a technique that can zoom in on a sufficiently large subset of astroturfer profiles, thus making the procedure highly efficient. The feature extraction stage consists of a temporal layer of features, which proves useful for identifying astroturfers. We then applied and compared several algorithms in the classification stage, and achieved improved results, with an F1 score of 77% and accuracy of 92%.

Social network analysis of characters in historical works is a popular research methodology in the study of historical literature. This article proposes using this methodology to characterize and comparatively analyze editing styles of similar historical literary works to determine whether they were edited by the same hand. To that end, the study proposes constructing a network of characters for each of the works being studied and to compute standard statistical measures for these networks, thus producing a network-based profile for the editing style of each work, which can be compared to the profiles of various other works. To determine the effectiveness of this new approach, it was tested on two similar works from the realm of Rabbinic literature—the Mishnah and the Tosefta. Our findings show that despite the abundant structural, thematic, and linguistic similarities of the works, their network-based profiles demonstrated clear differences between them with respect to various parameters, like the degree of connectivity, density, and centrality of the networks and their communities, and also with respect to the usage of different types of relationships in each network. These differences are reflected in the network features of the works, rather than in their texts, and so it would be difficult to identify them using direct stylometric analysis on the texts of the works, especially given the stylistic and thematic similarity between them. The approach presented in this article forms a basis for developing automatic classifiers to identify different editors and editing styles based on works’ network-based profiles.

Billions of ads are displayed on a daily basis, making it a multi-billion industry. Most of web pages contain multiple ads, which are largely served in real time using a bidding process where buyers (advertisers) offer a price tag to the seller (publishers) for each given possible ad on the page. There are multiple factors that impact an ad price, one of the primary ones is the ad-location’s viewability likelihood. Due to the length of many web pages, certain ad locations are invisible to the visiting user, as he may not scroll far enough on the page to where the ads are placed. According to recent industry metrics, less than 60% of ads are viewable. This poses a challenge to both: buyers and sellers. Buyers want to optimize the likelihood they buy an ad that will be viewed, while sellers want to maximize ad prices (by setting higher floor prices) by providing as many possible ad placements with high viewability probability. This paper addresses the viewability prediction from the publisher’s side, and proposes a novel algorithm based on cascading gradient boosting. The algorithm enables sellers to predict an accurate viewability probability for ad impressions, which is optimized to match the actual viewability rate that will be measured for the served ads. Unlike other algorithms that optimize these problems to an average minimal difference from a central mean error, we propose an algorithm that increases the amount of extreme cases - which are the most valuable ones, thus expanding the premium ad inventory. We evaluate the algorithm on two datasets with a total of over 500 million impressions. We found that the algorithm outperforms other viewability prediction algorithms, works well for publishers while providing a measurable fairness metric to advertisers.

. Astroturfing is the practice of using a fake online social media (OSM) profile in order to influence public opinion, while giving the impression that the profile belongs to an authentic human user. In at-tempting to train a classifier for discriminating between authentic users and astroturfers, a labeled dataset must first be arranged. The labeling is generally done manually, by human judges, on a collection of profiles garnered from the social media network. However, the fact that any randomly collected set of profiles will statistically contain a small proportion of astroturfers, renders this process inefficient: a lot of time and effort is invested on manually labeling lots of data, while producing only a small set of astroturfer profiles. We present here a method for quickly and ef-ficiently collecting a data set for manual labeling, with a high percent of astroturfers.

The international conference #DHJewish - Jewish Studies in the Digital Age will bring together more than 60 scholars and heritage practitioners to discuss how the digital turn affects the field of Jewish Studies. The conference is organised by the [Luxembourg Centre for Contemporary and Digital History (C²DH).](https://www.c2dh.uni.lu/) [For more information, and a PDF download of the program, please see also here.](https://www.c2dh.uni.lu/events/dhjewish-jewish-studies-digital-age) In order to meet other participants and for informal chats in the breaks, join us in our [wonder.me room](https://www.wonder.me/r?id=73a12756-8d93-4871-9374-143253bfbf59).

This paper suggests the use of automatic topic modeling for large-scale corpora of privacy policies using unsupervised learning techniques. The advantages of using unsupervised learning for this task are numerous. The primary advantages include the ability to analyze any new corpus with a fraction of the effort required by supervised learning, the ability to study changes in topics of interest along time, and the ability to identify finer-grained topics of interest in these privacy policies. Based on general principles of document analysis we synthesize a cohesive framework for privacy policy topic modeling and apply it over a corpus of 4,982 privacy policies of mobile applications crawled from the Google Play Store. The results demonstrate that even with this relatively moderate-size corpus quite comprehensive insights can be attained regarding the focus and scope of current privacy policy documents. The topics extracted, their structure and the applicability of the unsupervised approach for that matter are validated through an extensive comparison to similar findings reported in prior work that uses supervised learning (which heavily depends on manual annotation of experts). The comparison suggests a substantial overlap between the topics found and those reported in prior work, and also unveils some new topics of interest.

Corpus-based automatic thesaurus construction uses linguistic methods, such as Part-of-Speech taggers and parsers, which often perform poorly on MRLs. Therefore, in this paper, we focused on the complex task of adapting corpus-based thesaurus construction methods for MRLs. We investigated two statistical approaches for thesaurus construction; a) a first-order co-occurrence-based approach and b) a second-order distributional-based approach. We explored alternative levels of morphological term representations complemented by grouping the morphological variants. We then introduced and adopted a generic algorithmic scheme for thesaurus construction in MRLs for both first-order and second-order approaches. Our scheme investigated alternative representation levels and offered alternative configurations. We demonstrated the empirical benefits of our methodology for a diachronic Hebrew thesaurus construction. We used morphological analysis tools, defined and applied a new annotation scheme, and demonstrated its optimal configuration, which outperforms the baseline for both first and second order corpus-based thesaurus construction approaches.

Automatic thesaurus construction for Modern Hebrew is a complicated task, due to its high degree of inflectional ambiguity. Linguistics tools, including morphological analyzers, part-of-speech taggers and parsers often have limited in performance on Morphologically Rich Languages (MRLs) such as Hebrew. In this paper, we adopted a schematic methodology for generating a cooccurrence based thesaurus in a MRL and extended the methodology to create distributional similarity thesaurus. We explored three alternative levels of morphological term representations, surface form, lemma, and multiple lemmas, all complemented by the clustering of morphological variants. First, we evaluated both the co-occurrence based method and the distributional similarity method using Hebrew WordNet as our gold standard. However, due to Hebrew WordNet's low coverage, we completed our analysis with a manual evaluation. The results showed that for Modern Hebrew corpus-based thesaurus construction, the most directly applied statistical collection, using linguistics tools at the lemma level, is not optimal.

This paper presents a methodology for tackling the authorship verification problem. The approach is based on comparing the similarity between a given unknown document against the known documents using a graph representation that captures the syntactic sequence of texts and a graph similarity measure. An unknown document can be classified as having been written by the same author if the majority of the comparisons surpass a predefined threshold. The best results were obtained on the Clef PAN 2014 dataset: 79% for the Spanish and 68% for English, showing that the proposed methodology could be a way for determining a document authorship.

Online advertising is a huge, rapidly growing advertising market in today's world. One common form of online advertising is using image ads. A decision is made (often in real time) every time a user sees an ad, and the advertiser is eager to determine the best ad to display. Consequently, many algorithms have been developed that calculate the optimal ad to show to the current user at the present time. Typically, these algorithms focus on variations of the ad, optimizing among different properties such as background color, image size, or set of images. However, there is a more fundamental layer. Our study looks at new qualities of ads that can be determined before an ad is shown (rather than online optimization) and defines which ads are most likely to be successful. We present a set of novel algorithms that utilize deep-learning image processing, machine learning, and graph theory to investigate online advertising and to construct prediction models which can foresee an image ad's success. We evaluated our algorithms on a dataset with over 260,000 ad images, as well as a smaller dataset specifically related to the automotive industry, and we succeeded in constructing regression models for ad image click rate prediction. The obtained results emphasize the great potential of using deep-learning algorithms to effectively and efficiently analyze image ads and to create better and more innovative online ads. Moreover, the algorithms presented in this paper can help predict ad success and can be applied to analyze other large-scale image corpora.

Kate Niederhoffer, Jonathan Schler, Patrick Crutchley, Kate Loveys, Glen Coppersmith. Proceedings of the Fourth Workshop on Computational Linguistics and Clinical Psychology — From Linguistic Signal to Clinical Reality. 2017.

A cross-period (diachronic) thesaurus enables users to search for information using modern terminology and obtain semantically related terms from earlier historical periods. The complex task of supporting the construction of a diachronic thesaurus by a domain expert lexicographer has hardly been addressed computationally until now. In this article, we introduce a semiautomatic iterative Query Expansion (QE) scheme for supporting diachronic thesaurus construction, which identifies candidate related terms based on statistical corpus-based measures. We use ancient-modern period classification to increase the performance of the statistical cooccurrence measures and extend our methods to deal with Multi-Word Expressions (MWEs). We demonstrate the empirical benefit of our scheme for a Jewish cross-period thesaurus and evaluate its impact on recall and on the effectiveness of the lexicographer’s manual efforts.