Social text understanding is prone to fail when opinions are conveyed implicitly or sarcastically.It is therefore desired to model users' contexts in processing the texts authored by them.In this work, we represent users within a social embedding space that was learned from the Twitter network at large-scale.Similar to word embeddings that encode lexical semantics, the network embeddings encode latent dimensions of social semantics.We perform extensive experiments on author-informed stance prediction, demonstrating improved generalization through inductive social user modeling, both within and across topics.Similar results were obtained for author-informed toxicity and incivility detection.The proposed approach may pave way to social NLP that considers user embeddings as contextual modality.However, our investigation also reveals that user stances are correlated with the personal socio-demographic traits encoded in their embeddings.Hence, author-informed NLP approaches may inadvertently model and reinforce socio-demographic and other social biases.

Toxic online political discourse has become prevalent, where scholars debate about its impact on democratic processes.This work presents a large-scale study of political incivility on Twitter.In line with theories of political communication, we differentiate between harsh impolite style and intolerant substance.We present a dataset of 13K political tweets in the U.S. context, which we collected and labeled by this multidimensional distinction using crowd sourcing.The evaluation of state-of-the-art classifiers illustrates the challenges involved in political incivility detection, which often requires high-level semantic and social understanding.Nevertheless, performing incivility detection at scale, we are able to characterise its distribution across individual users and geopolitical regions.Our findings align with and extend existing theories of political communication.In particular, we find that roughly 80% of the uncivil tweets are authored by 20% of the users, where users who are politically engaged are more inclined to use uncivil language.We further find that political incivility exhibits network homophily, and that incivility is more prominent in highly competitive geopolitical regions.Our results apply to both uncivil style and substance.Warning: The paper contains examples that readers might find offensive.

Although large neural language models (LMs) like BERT can be finetuned to yield state-of-the-art results on many NLP tasks, it is often unclear what these models actually learn. Here we study using such LMs to fill in entities in human-authored comparative questions, like ``Which country is older, India or _____?''---i.e., we study the ability of neural LMs to ask (not answer) reasonable questions. We show that accuracy in this fill-in-the-blank task is well-correlated with human judgements of whether a question is reasonable, and that these models can be trained to achieve nearly human-level performance in completing comparative questions in three different subdomains. However, analysis shows that what they learn fails to model any sort of broad notion of which entities are semantically comparable or similar---instead the trained models are very domain-specific, and performance is highly correlated with co-occurrences between specific entities observed in the training set. This is true both for models that are pretrained on general text corpora, as well as models trained on a large corpus of comparison questions. Our study thus reinforces recent results on the difficulty of making claims about a deep model's world knowledge or linguistic competence based on performance on specific benchmark problems. We make our evaluation datasets publicly available to foster future research on complex understanding and reasoning in such models at standards of human interaction.

When we consider our CV, it is full of entities - where we studied, where we worked, who we collaborated with on a project or on a paper. Entities we are linked to are part of our profile and as such they help to understand who we are and what are we interested in. Hence, we adapt the typed entity-relation graph (profile) concept and based on this presentation we propose a personalized multifaceted graph visualization for the entity profile. In the context of an academic conference, we allow scholars to explore a graph of related entities and a word cloud representing the links, providing the user a comprehensive, compact and structured overview about the explored scholar. We evaluated our proposed personalized multifaceted visualization in a user study with encouraging results which showed that this visualization is engaging, easy to use and helpful.

Graph-based recommendation approaches can model associations between users and items alongside additional contextual information. Recent studies demonstrated that representing features extracted from social media (SM) auxiliary data, like friendships, jointly with traditional users/items ratings in the graph, contribute to recommendation accuracy. In this work, we take a step further and propose an extended graph representation that includes socio-demographic and personal traits extracted from the content posted by the user on SM. Empirical results demonstrate that processing unstructured textual information collected from Twitter and representing it in structured form in the graph improves recommendation performance, especially in cold start conditions.

We describe a graph-based entity profiling system (GBEP) that extracts information about persons of interest from the Web and uses this information to construct a joint social graph. GBEP then employs graph-based measures to assess inter-personal relatedness, performing social recommendation. Importantly, GBEP provides detailed explanations for its suggestions in the form of relational connecting paths. Initial positive results were obtained for recommending related conference participants to each other using a joint social graph constructed for this purpose.

Visitors to museums and other cultural heritage sites encounter a wealth of exhibits in a variety of subject areas, but can explore only a small number of them. Moreover, there typically exists rich complementary information that can be delivered to the visitor about exhibits of interest, but only a fraction of this information can be consumed during the limited time of the visit. Recommender systems may help visitors to cope with this information overload. Ideally, the recommender system of choice should model user preferences, as well as background knowledge about the museum's environment, considering aspects of physical and thematic relevancy. We propose a personalized graph‐based recommender framework, representing rating history and background multi‐facet information jointly as a relational graph. A random walk measure is applied to rank available complementary multimedia presentations by their relevancy to a visitor's profile, integrating the various dimensions. We report the results of experiments conducted using authentic data collected at the Hecht museum. An evaluation of multiple graph variants, compared with several popular and state‐of‐the‐art recommendation methods, indicates on advantages of the graph‐based approach.

We consider the task of record extraction from text documents, where the goal is to automatically populate the fields of target relations, such as scientific seminars or corporate acquisition events. There are various inferences involved in the record-extraction process, including mention detection, unification, and field assignments. We use structured learning to find the appropriate field-value assignments. Unlike previous works, the proposed approach generates feature-rich models that enable the modeling of domain semantics and structural coherence at all levels and across fields. Given labeled examples, such an approach can, for instance, learn likely event durations and the fact that start times should come before end times. While the inference space is large, effective learning is achieved using a perceptron-style method and simple, greedy beam decoding. A main focus of this article is on practical aspects involved in implementing the proposed framework for real-world applications. We argue and demonstrate that this approach is favorable in conditions of data shift, a real-world setting in which models learned using a limited set of labeled examples are applied to examples drawn from a different data distribution. Much of the framework’s robustness is attributed to the modeling of domain knowledge. We describe design and implementation details for the case study of seminar event extraction from email announcements, and discuss design adaptations across different domains and text genres.

Ni Lao, Einat Minkov, William Cohen. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

We outline a learning framework that aims at identifying useful contextual cues for knowledge-based word sense disambiguation. The usefulness of individual context words is evaluated based on diverse lexico-statistical and syntactic information, as well as simple word distance. Experiments using two dif-ferent knowledge-based methods and bench-mark datasets show significant improvements due to context modeling, beating the conven-tional window-based approach. 1

While there has been much research on automatically constructing structured Knowledge Bases (KBs), most of it has focused on generating facts to populate a KB. However, a useful KB must go beyond facts. For example, glosses (short natural language definitions) have been found to be very useful in tasks such as Word Sense Disambiguation. However, the important problem of Automatic Gloss Finding, i.e., assigning glosses to entities in an initially gloss-free KB, is relatively unexplored. We address that gap in this paper. In particular, we propose GLOFIN, a hierarchical semi-supervised learning algorithm for this problem which makes effective use of limited amounts of supervision and available ontological constraints. To the best of our knowledge, GLOFIN is the first system for this task. Through extensive experiments on real-world datasets, we demonstrate GLOFIN's effectiveness. It is encouraging to see that GLOFIN outperforms other state-of-the-art SSL algorithms, especially in low supervision settings. We also demonstrate GLOFIN's robustness to noise through experiments on a wide variety of KBs, ranging from user contributed (e.g., Freebase) to automatically constructed (e.g., NELL). To facilitate further research in this area, we have made the datasets and code used in this paper publicly available.

A plethora of tools exist for extracting and visualizing key sentiment information from a corpus of text documents. Often, however, there is a need for quickly assessing the sentiment and feelings that arise from an individual document. We describe an interactive tool that visualizes the sentiment of a specific document such as an online opinion, blog, or transcript, by visually highlighting the sentiment features while leaving the document text intact.

Abstract We consider a dependency-parsed text corpus as an instance of a labeled directed graph, where nodes represent words and weighted directed edges represent the syntactic relations between them. We show that graph walks, combined with existing techniques of supervised learning that model local and global information about the graph walk process, can be used to derive a task-specific word similarity measure in this graph. We also propose and evaluate a new learning method in this framework, a path-constrained graph walk variant, in which the walk process is guided by high-level knowledge about meaningful edge sequences (paths) in the graph. Empirical evaluation on the tasks of named entity coordinate term extraction and general word synonym extraction show that this framework is preferable to, or competitive with, vector-based models when learning is applied, and using small to moderate size text corpora.

We demonstrate a method for collaborative ranking of future events. Previous work on recommender systems typically relies on feedback on a particular item, such as a movie, and generalizes this to other items or other people. In contrast, we examine a setting where no feedback exists on the particular item. Because direct feedback does not exist for events that have not taken place, we recommend them based on individuals' preferences for past events, combined collaboratively with other peoples' likes and dislikes. We examine the topic of unseen item recommendation through a user study of academic (scientific) talk recommendation, where we aim to correctly estimate a ranking function for each user, predicting which talks would be of most interest to them. Then by decomposing user parameters into shared and individual dimensions, we induce a similarity metric between users based on the degree to which they share these dimensions. We show that the collaborative ranking predictions of future events are more effective than pure content-based recommendation. Finally, to further reduce the need for explicit user feedback, we suggest an active learning approach for eliciting feedback and a method for incorporating available implicit user cues.

Most current methods for 802.11-based indoor localization depend on surveys conducted by experts or skilled technicians. Some recent systems have incorporated surveying by users. Structuring localization systems "organically," however, introduces its own set of challenges: conveying uncertainty, determining when user input is actually required, and discounting erroneous and stale data. Through deployment of an organic location system in our nine-story building, which contains nearly 1,400 distinct spaces, we evaluate new algorithms for addressing these challenges. We describe the use of Voronoi regions for conveying uncertainty and reasoning about gaps in coverage, and a clustering method for identifying potentially erroneous user data. Our algorithms facilitate rapid coverage while maintaining positioning accuracy comparable to that achievable with survey-driven indoor deployments.

Relational or semistructured data is naturally represented by a graph, where nodes denote entities and directed typed edges represent the relations between them. Such graphs are heterogeneous, describing different types of objects and links. We represent personal information as a graph that includes messages , terms , persons , dates , and other object types, and relations like sent-to and has-term . Given the graph, we apply finite random graph walks to induce a measure of entity similarity, which can be viewed as a tool for performing search in the graph. Experiments conducted using personal email collections derived from the Enron corpus and other corpora show how the different tasks of alias finding , threading , and person name disambiguation can be all addressed as search queries in this framework, where the graph-walk-based similarity metric is preferable to alternative approaches, and further improvements are achieved with learning. While researchers have suggested to tune edge weight parameters to optimize the graph walk performance per task, we apply reranking to improve the graph walk results, using features that describe high-level information such as the paths traversed in the walk. High performance, together with practical runtimes, suggest that the described framework is a useful search system in the PIM domain, as well as in other semistructured domains.

Similarity measures for text have historically been an important tool for solving information retrieval problems. In many interesting settings, however, documents are often closely connected to other documents, as well as other non-textual objects: for instance, email messages are connected to other messages via header information. In this paper we consider extended similarity metrics for documents and other objects embedded in graphs, facilitated via a lazy graph walk. We provide a detailed instantiation of this framework for email data, where content, social networks and a timeline are integrated in a structural graph. The suggested framework is evaluated for two email-related problems: disambiguating names in email documents, and threading. We show that reranking schemes based on the graph-walk similarity measures often outperform baseline methods, and that further improvements can be obtained by use of appropriate learning methods.

BACKGROUND: One step in the model organism database curation process is to find, for each article, the identifier of every gene discussed in the article. We consider a relaxation of this problem suitable for semi-automated systems, in which each article is associated with a ranked list of possible gene identifiers, and experimentally compare methods for solving this geneId ranking problem. In addition to baseline approaches based on combining named entity recognition (NER) systems with a "soft dictionary" of gene synonyms, we evaluate a graph-based method which combines the outputs of multiple NER systems, as well as other sources of information, and a learning method for reranking the output of the graph-based method. RESULTS: We show that named entity recognition (NER) systems with similar F-measure performance can have significantly different performance when used with a soft dictionary for geneId-ranking. The graph-based approach can outperform any of its component NER systems, even without learning, and learning can further improve the performance of the graph-based ranking approach. CONCLUSION: The utility of a named entity recognition (NER) system for geneId-finding may not be accurately predicted by its entity-level F1 performance, the most common performance measure. GeneId-ranking systems are best implemented by combining several NER systems. With appropriate combination methods, usefully accurate geneId-ranking systems can be constructed based on easily-available resources, without resorting to problem-specific, engineered components.

Decision-tree algorithms are known to be unstable: small variations in the training set can result in different trees and different predictions for the same validation examples. Both accuracy and stability can be improved by learning multiple models from bootstrap samples of training data, but the "meta-learner" approach makes the extracted knowledge hardly interpretable. In the following paper, we present the Info-Fuzzy Network (IFN), a novel information-theoretic method for building stable and comprehensible decision-tree models. The stability of the IFN algorithm is ensured by restricting the tree structure to using the same feature for all nodes of the same tree level and by the built-in statistical significance tests. The IFN method is shown empirically to produce more compact and stable models than the "meta-learner" techniques, while preserving a reasonable level of predictive accuracy.