In e-commerce, customers often struggle to find relevant items when their needs involve subjective properties characterized by personal or collective perception, tastes, and opinions, which are typically not captured in catalog data. This challenge is particularly pronounced in event-based scenarios like gifting, where selecting the right product involves complex subjective reasoning. Customer reviews can be a valuable source of subjective information to bridge this gap. Consequently, customers often spend significant amount of time navigating multiple products and reading numerous reviews to find suitable gifts that meet their needs. In order to reduce the effort involved, we propose an agentic approach driven by large language models to streamline this process by autonomously executing various user actions. These include computational tasks like vagueness detection and subjective product needs extraction, conversational interactions to gather missing user information, and web browsing actions that search for product details, reviews, and review images. Additionally, the agent employs generative actions to synthesize gifting ideas and explanations, helping users discover suitable products more efficiently. The proposed approach not only reduces the cognitive burden on users but also facilitates the exploration of a wider range of products. Our solution highlights the potential of autonomous agents to handle subjective queries in e-commerce, enhancing personalization, product exploration, and selection in a user-centric manner.

This article presents an exploration of the use of pre-trained Large Language Models (LLMs) for crisis classification to address labeled data dependency issues. We present a methodology that enhances open LLMs through fine-tuning, creating zero-shot and few-shot classifiers that approach traditional supervised models in classifying crisis-related messages. A comparative study evaluates crisis classification tasks using general domain pre-trained LLMs, crisis-specific LLMs, and traditional supervised learning methods, establishing a benchmark in the field. Our task-specific fine-tuned Llama model achieved a 69% macro F1 score in classifying humanitarian information–a remarkable 26% improvement compared to the Llama baseline, even with limited training data. Moreover, it outperformed ChatGPT4 by 3% in macro F1. This improvement increased to 71% macro F1 when fine-tuning Llama with multitask data. For the binary classification of messages as related vs. not related to crises, we observed that pre-trained LLMs, such as Llama 2 and ChatGPT4, performed well without fine-tuning, achieving an 87% macro F1 score with ChatGPT4. This research expands our knowledge of how to exploit the potential of LLMs for crisis classification, representing a great opportunity for crisis scenarios that lack labeled data. The findings emphasize the potential of LLMs in crisis informatics to address cold start challenges, especially critical in the initial phases of a disaster, while also showcasing their capacity to attain high accuracy even with limited training data.

Many public pre-trained word embeddings have been shown to encode different types of biases. Embeddings are often obtained from training on large pre-existing corpora, and therefore resulting biases can be a reflection of unfair representations in the original data. Bias, in this scenario, is a challenging problem since current mitigation techniques require knowing and understanding existing biases in the embedding, which is not always possible. In this work, we propose to improve word embedding fairness by borrowing methods from the field of data privacy. The idea behind this approach is to treat bias as if it were a special type of training data leakage. This has the unique advantage of not requiring prior knowledge of potential biases in word embeddings. We investigated two types of privacy algorithms, and measured their effect on bias using four different metrics. To investigate techniques from differential privacy, we applied Gaussian perturbation to public pre-trained word embeddings. To investigate noiseless privacy, we applied vector quantization during training. Experiments show that both approaches improve fairness for commonly used embeddings, and additionally, noiseless privacy techniques reduce the size of the resulting embedding representation.

Social media data has emerged as a useful source of timely information about real-world crisis events. One of the main tasks related to the use of social media for disaster management is the automatic identification of crisis-related messages. Most of the studies on this topic have focused on the analysis of data for a particular type of event in a specific language. This limits the possibility of generalizing existing approaches because models cannot be directly applied to new types of events or other languages. In this work, we study the task of automatically classifying messages that are related to crisis events by leveraging cross-language and cross-domain labeled data. Our goal is to make use of labeled data from high-resource languages to classify messages from other (low-resource) languages and/or of new (previously unseen) types of crisis situations. For our study we consolidated from the literature a large unified dataset containing multiple crisis events and languages. Our empirical findings show that it is indeed possible to leverage data from crisis events in English to classify the same type of event in other languages, such as Spanish and Italian (80.0% F1-score). Furthermore, we achieve good performance for the cross-domain task (80.0% F1-score) in a cross-lingual setting. Overall, our work contributes to improving the data scarcity problem that is so important for multilingual crisis classification. In particular, mitigating cold-start situations in emergency events, when time is of essence.

Online e-commerce product reviews can be highly influential in a customer's decision-making processes. Reviews often describe personal experiences with a product and provide candid opinions about a product's pros and cons. In some cases, reviewers choose to share information about themselves, just as they might do in social platforms. These descriptions are a valuable source of information about who finds a product most helpful. Customers benefit from key insights about a product from people with their same interests and sellers might use the information to better serve their customers needs. In this work, we present a comprehensive look into voluntary self-descriptive information found in public customer reviews. We analyzed what people share about themselves and how this contributes to their product opinions. We developed a taxonomy of types of self-descriptions, and a machine-learned classification model of reviews according to this taxonomy. We present new quantitative findings, and a thematic study of the perceived purpose descriptions in reviews.

The Web has become a resource that allows us to make sense of social phenomena around the world. This started the moment users became content creators, and has grown with the emergence of social platforms tailored to our need to connect and share with others. Throughout my work, I’ve come to appreciate how social media has democratized access to real-world news and social sentiment, while also witnessing the loss of trust created by fake information. As a computer scientist from Chile in Latin America, I have worked on a range of problems that were driven by local needs. Many times, I have tried to apply state of the art solutions to well-known problems, only to find that these don’t work outside of their initial evaluation dataset. In this talk, I’ll discuss how geographical, language, and social diversity have opened new avenues for innovation and better understanding the social Web. I’ll also show that to truly create useful technological solutions, we must develop inclusive research and resources.

Henry Rosales-Méndez, Aidan Hogan, Barbara Poblete. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Hate speech is an important problem that is seriously affecting the dynamics and usefulness of online social communities. Large scale social platforms are currently investing important resources into automatically detecting and classifying hateful content, without much success. On the other hand, the results reported by state-of-the-art systems indicate that supervised approaches achieve almost perfect performance but only within specific datasets. In this work, we analyze this apparent contradiction between existing literature and actual applications. We study closely the experimental methodology used in prior work and their generalizability to other datasets. Our findings evidence methodological issues, as well as an important dataset bias. As a consequence, performance claims of the current state-of-the-art have become significantly overestimated. The problems that we have found are mostly related to data overfitting and sampling issues. We discuss the implications for current research and re-conduct experiments to give a more accurate picture of the current state-of-the art methods.

The sheer amount of newsworthy information published by users in social media platforms makes it necessary to have efficient and effective methods to filter and organize content. In this scenario, off-the-shelf methods fail to process large amounts of data, which is usually approached by adding more computational resources. Simple data aggregations can help to cope with space and time constraints, while at the same time improve the effectiveness of certain applications, such as topic detection or summarization. We propose a lightweight representation of newsworthy social media data. The proposed representation leverages microblog features, such as redundancy and re-sharing capabilities, by using surrogate texts from shared URLs and word embeddings. Our representation allows us to achieve comparable clustering results to those obtained by using the complete data, while reducing running time and required memory. This is useful when dealing with noisy and raw user-generated social media data.

Timely detection and accurate description of extreme events, such as natural disasters and other crisis situations, are crucial for emergency management and mitigation. Extreme-event detection is challenging, since one has to rely upon reports from human observers appointed to specific geographical areas, or on an expensive and sophisticated infrastructure. In the case of earthquakes, geographically dense sensor networks are expensive to deploy and maintain. Therefore, only some regions-or even countries-are able to acquire useful information about the effects of earthquakes in their own territory. An inexpensive and viable alternative to this problem is to detect extreme real-world events through people's reactions in online social networks. In particular, Twitter has gained popularity within the scientific community for providing access to real-time “citizen sensor” activity. Nevertheless, the massive amount of messages in the Twitter stream, along with the noise it contains, underpin a number of difficulties when it comes to Twitter-based event detection. We contribute to address these challenges by proposing an online method for detecting unusual bursts in discrete-time signals extracted from Twitter. This method only requires a one-off semisupervised initialization and can be scaled to track multiple signals in a robust manner. We also show empirically how our proposed approach, which was envisioned for generic event detection, can be adapted for worldwide earthquake detection, where we compare the proposed model to the state of the art for earthquake tracking using social media. Experimental results validate our approach as a competitive alternative in terms of precision and recall to leading solutions, with the advantage of implementation simplicity and worldwide scalability.

We propose an algorithm and system that detects earthquakes worldwide in real time based on reports of social media users, or "citizen-sensors." Earthquake detections are based on user postings in any language and from any region. This approach is unsupervised, adapting automatically to changes in the input data stream, and only requires a general list of keywords for each language. Our method is noise tolerant and simple, providing good results both in terms of precision and recall. This complements prior work that mostly consists of supervised approaches that focus on performing detections in a specific geographical area and are difficult to generalize to a global scope. We demonstrate the effectiveness of this approach by using it within a real-time on-line system, which is publicly available and currently in use at National Seismology Center in Chile and Oceanographic and Hydrological Office of the Chilean Army. The quantitative evaluation of our system, performed during a 9-month period, shows that our solution is competitive to the best state-of-the-art methods. Overall, our findings indicate that our approach is an effective low-cost alternative for earthquake monitoring at a global scale.

In this talk I will describe "Twicalli", a real-time earthquake detection system based on citizen sensors. This system is publicly available for over a year, at http://twicalli.cl, and is currently in use as a decision support tool by the National Seismology Office and by the Hydrographic and the Oceanographic Service in Chile. The novelty of our system relies on the fact that it has a very good precision and recall tradeoff for earthquakes of all magnitude ranges that were reported on Twitter. Our earthquake detection methodology is simple, efficient, unsupervised, and it can detect earthquakes reported globally in any language and any region. This complements existing approaches that are either: i) supervised and customized to a particular geographical region, which makes them very expensive to scale geographically and keep up-to-date, or ii) unsupervised with low earthquake recall.The evaluation of our system, performed during a 9-month period, shows that our solution is competitive to the best state-of-the-art methods, providing very good precision and recall performance for a wide range of earthquake magnitudes.

Multimedia searches often return items that can be categorized into several "topics", allowing users to disambiguate and explore answers more efficiently. In this paper we investigate methods for clustering tags associated with multimedia search results, where each resulting cluster represents a topic computed online for that particular search. We specifically investigate the applicability of community detection algorithms to the tag graph induced from the search results. This type of approach allows us to exploit tag similarity and create ad-hoc topics for each search, without specify the number and sizes of clusters a priori. In this work we experiment with well-known algorithms in this field and propose two new methods based on adaptive island cuts. Using the Social20 dataset (a collection gathered from Flickr) we evaluate several community detection methods, with quantitative analysis of each algorithm in terms of the relative number of communities (which we interpret as topics) that they produce and their sizes, as well as qualitative analysis of topics per human judgement. Our evaluation shows that it is possible to extract ad-hoc topics for search results using community detection, but that different community detection methods produce very different results. In particular, our proposed methods produce more compact and less noisy clusters as well as less relative recall when compared to methods that produce much larger clusters.

Presents the introductory editorial for this issue of the publication.

Online Social Networks (OSN) have changed the way information is produced and consumed. Organizing and retrieving unstructured data extracted from these platforms is not an easy task. Galean is a visual and interactive tool that aims to help journalists and historians, among others, analyze news events discussed on Twitter. In this tool, news events are visually represented by the very countries from where the news originated, the date when they happened and their impact in the OSN. Galean considers countries as entities, as opposed to mere geographical locations as most of the tools in the state of the art. As a consequence, it allows users to explore and retrieve news not only by their geographical and temporal context, but also by the relationship among countries. With this tool users can search for behavioral patterns of news events and observe how countries are associated in specific events. We expect our work to become a public tool that helps conduct historical analyses of social media news coverage over time.

Nowadays, social media services are being used extensively as news sources and for spreading information on real-world events. Several studies have focused on detecting those events and locating them geographically. However, in order to study real-world events, for example, finding relationships between locations or detecting high impact events based on their coverage, we need more suitable models to represent events. In this work we propose a simple model to represent real-world news events using two sources of information: the locations that are mentioned in the event (where the event occurs), and the locations of users that discuss or comment on it. We then characterize a country based on the amount of events in which that country is mentioned and also participates on the event. We show some applications of the model: we find clusters of news events based on the level of participation of countries, identifying global and impactful events in certain areas. Also, we show groups of similar countries, finding promising insights about their relationships. This model can be useful at finding unsuspected relations among countries based on the news coverage and country participation, identifying different levels of news coverage in the world, and finding bias in international news sources.

Toolbar navigation logs provide rich data for enhancing information discovery on the Web. The value of this data resides in its scope, which goes beyond that of traditional query-mining data sources, such as search-engine logs. In this paper we present a methodology for extracting relevant association rules for queries, based on historic user navigational data. In addition, we propose a graph-based approach for extracting related queries and URLs for a given query.

Social media services have spread throughout the world in just a few years. They have become not only a new source of information, but also new mechanisms for societies world-wide to organize themselves and communicate. Therefore, social media has a very strong impact in many aspects -- at personal level, in business, and in politics, among many others. In spite of its fast adoption, little is known about social media usage in different countries, and whether patterns of behavior remain the same or not. To provide deep understanding of differences between countries can be useful in many ways, e.g.: to improve the design of social media systems (which features work best for which country?), and influence marketing and political campaigns. Moreover, this type of analysis can provide relevant insight into how societies might differ. In this paper we present a summary of a large-scale analysis of Twitter for an extended period of time. We analyze in detail various aspects of social media for the ten countries we identified as most active. We collected one year's worth of data and report differences and similarities in terms of activity, sentiment, use of languages, and network structure. To the best of our knowledge, this is the first on-line social network study of such characteristics.

Annotating or tagging multimedia objects is an important task for enhancing multimedia information retrieval processes. In the context of the Web, automatic tagging deals with many issues, such as loosely tagged images and huge collections of images with no textual data at all. Recently, graph representations have been shown useful for modeling relationships between images and their associated semantics. Using these types of graphs, it is possible to represent images and their textual labels as nodes, and the relationship between them as edges, under the assumption that visual similarity implies semantic similarity. In this work, we present an algorithm for automatic tag propagation in such a graph structure, called the visual-semantic graph. This graph has been used in prior work only for the task of image retrieval re-ranking. The goal of our work, is to show how the visual-semantic graph can be used for efficient tag propagation to unlabeled images. More specifically, our contributions are: (1) An algorithm to propagate tags automatically based on the breadth-first traversal and (2) A set of heuristics for pruning this approach for large size collections.

We analyze the information credibility of news propagated through Twitter, a popular microblogging service. Previous research has shown that most of the messages posted on Twitter are truthful, but the service is also used to spread misinformation and false rumors, often unintentionally.

We explore the application of a graph representation to model similarity relationships that exist among images found on the Web. The resulting similarity-induced graph allows us to model in a unified way different types of content-based similarities, as well as semantic relationships. Content-based similarities include different image descriptors, and semantic similarities can include relevance user feedback from search engines. The goal of our representation is to provide an experimental framework for combining apparently unrelated metrics into a unique graph structure, which allows us to enhance the results of Web image retrieval. We evaluate our approach by re-ranking Web image search results.

We introduce the concern of confidentiality protection of business information for the publication of search engine query logs and derived data. We study business confidentiality, as the protection of nonpublic data from institutions, such as companies and people in the public eye. In particular, we relate this concern to the involuntary exposure of confidential Web site information, and we transfer this problem into the field of privacy-preserving data mining. We characterize the possible adversaries interested in disclosing Web site confidential data and the attack strategies that they could use. These attacks are based on different vulnerabilities found in query log for which we present several anonymization heuristics to prevent them. We perform an experimental evaluation to estimate the remaining utility of the log after the application of our anonymization techniques. Our experimental results show that a query log can be anonymized against these specific attacks while retaining a significant volume of useful data.

We introduce a unified graph representation of the Web, which includes both structural and usage information. We model this graph using a simple union of the Web's hyperlink and click graphs. The hyperlink graph expresses link structure among Web pages, while the click graph is a bipartite graph of queries and documents denoting users' searching behavior extracted from a search engine's query log.

In this paper we present a new document representation based on implicit user feedback obtained from search engine queries. The main objective of this is to achieve better results in non-supervised tasks, such as clustering and labeling, through the incorporation of usage data obtained from search engine queries. This type of allows us to discover the motivations of users when visiting a certain document. The terms used in queries can provide a better choice of features, from the user's point of view, for summarizing the Web pages that were clicked from these queries. In this work we extend and formalize as model an existing but not very well known idea of view for document representation. Furthermore, we create a novel based on frequent query patterns called the model. Our evaluation shows that both query-based models outperform the vector-space when used for clustering and labeling documents in a website. In our experiments, the query-set reduces by more than 90% the number of features needed to represent a set of documents and improves by over 90% the quality of the results. We believe that this can be explained because our chooses better features and provides more accurate labels according to the user's expectations.

We present a novel model for validating and improving the content and structure organization of a website. This model studies the website as a graph and evaluates its interconnectivity in relation to the similarity of its documents. The aim of this model is to provide a simple way for improving the overall structure, contents and interconnectivity of a website. This model has been implemented as a prototype and applied to several websites, showing very interesting results. Our model is complementary to other methods of website personalization and improvement.