MOTIVATION: Accurate TNM staging from lung cancer radiology reports is crucial for treatment planning and prognosis assessment. Manual staging processes are time-consuming and subject to inter-observer variability. Large language models (LLMs) offer opportunities to automate TNM staging with enhanced interpretability and clinical reasoning. RESULTS: We developed two complementary systems for automated TNM staging from English radiology reports. System I employs GPT-4o with reasoning-based few-shot learning and multi-step voting. System II integrates multiple LLMs (GPT-4o and Gemini-2) using DSPy framework with MIPROv2 optimization. In NTCIR-18 RadNLP 2024 English main task, our approaches achieved first (joint accuracy: 0.6543) and second place (joint accuracy: 0.6296), demonstrating superior performance in T, N, and M classification with accuracies of 0.7037/0.9136/0.8889 and 0.7284/0.9383/0.8395, respectively. AVAILABILITY AND IMPLEMENTATION: Source code freely available at https://github.com/nlptmu/multi-expert-tnm-staging under MIT license. An archival snapshot of the version used in this study is deposited on Zenodo at https://doi.org/10.5281/zenodo.20338561. Implemented in Python 3.12+ with PyTorch 2.6 and DSPY 3.0, supporting Linux.

Timeline summarization involves condensing events from news articles to illustrate the temporal development of a specific topic. Traditional methods often extract events based on the number of related reports but tend to overlook the movement of protagonists, the leading actors participating in events that shape the progression of the topic. This oversight can result in the extraction of sensationalized events unrelated to the topic's progression, distracting readers from tracking the topic's development. To address this limitation, we propose a novel strategy that identifies protagonists through dependency relations and tracks changes in the context surrounding them over time using a multi-faceted temporal graph. This temporal graph is a sequence of graphs that effectively captures information progression and shifts over time. Our approach aims to build a biographical timeline with accurate chronology by identifying and following the movement of protagonists. Our experiments demonstrate that our method, PIECE, outperforms previous approaches in date assignment for timeline summarization across different language datasets.

Che-Wei Tsai, Yen-Hao Huang, Tsu-Keng Liao, Didier Fernando Salazar Estrada, Retnani Latifah, Yi-Shin Chen. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Recent pandemics, such as H1N1 and COVID-19, have had extensive negative effects on the social and economic well-being of communities. Despite efforts to prevent and control their spread, governments have turned to a strategy of Living With the Virus to manage, rather than eliminate, the impact of these pandemics. However, group activities such as collaborative spatial crowdsourcing can still lead to the significant spread of infection due to the correlation between individuals’ mobility, interactions, and infection spread. In this paper, we address the problem of spatial crowdsourcing-induced infection spread and propose Epidemic-aware Maximum Task Assignment (EMTA). EMTA aims to form and assign collaborative worker groups to spatial crowdsourcing tasks while taking into consideration the control of epidemic spread. We prove that EMTA is NP-hard and inapproximable. We then propose the Epidemic-aware Task Assignment Algorithm (ETAA) that leverages epidemic characteristics to fully address EMTA. The experimental results from real LBSN and real epidemic datasets demonstrate that the proposed algorithm outperforms the state-of-the-art baselines in terms of effectiveness and efficiency.

Real-timesocial interactions andmulti-streamingare two critical features oflive streaming services. In this paper, we formulate a new fundamental service query,Social-aware Diverse and Preferred Organization Query (SDSQ), that jointly selects a set of diverse and preferred live streaming channels and a group of socially tight viewers for organization of a live multi-streaming soiree. We prove that SDSQ is NP-hard and inapproximable within any factor, and designSDSSel, a 2-approximation algorithm with a guaranteed error bound. Moreover, we study SDSQ-T, a special case of SDSQ, where the social graph is a threshold graph, and proposeTDSSel, a 2-approximation algorithm without any error to SDSQ-T. We propose two pruning strategies,PCPandCDPto boost SDSSel and TDSSel. We further propose a more challenging but practical service query,Generalized Social-aware Maximum Preferred and Diverse Query (GSPQ), a generalization of SDSQ. We designGPDSel, a 4-approximation algorithm for GSPQ with a guaranteed error bound. We propose a strategy to improve the approximation ratios of the proposed algorithms. A user study on Twitch validates SDSQ, and the large-scale experiments on real datasets demonstrate the superiority of the proposed algorithms over several baselines for live-streaming services.

Graph neural networks (GNNs) have been recently applied in natural language processing. Various GNN research studies are proposed to learn node interactions within the local graph of each document that contains words, sentences, or topics for inductive text classification. However, most inductive GNNs that are built on a word graph generally take global word embeddings as node features, without referring to document-wise contextual information. Consequently, we find that BERT models can perform better than inductive GNNs. An intuitive follow-up approach is used to enrich GNNs with contextual embeddings from BERT, yet there is a lack of related research. In this work, we propose a simple yet effective unified model, coined ConTextING, with a joint training mechanism to learn from both document embeddings and contextual word interactions simultaneously. Our experiments show that ConTextING outperforms pure inductive GNNs and BERT-style models. The analyses also highlight the benefits of the sub-word graph and joint training with separated classifiers.

Epidemics such as COVID-19, SARS, H1N1 have highly transmissible viruses and spread wildly through the population with negative consequences. Multiple studies have shown the correlation between the contact networks between individuals and the transmission of infections due to contact between colocated individuals. To mitigate the transmission of the virus, intervention measures have been applied without decisive success. Therefore, reducing transmissions through suitable epidemicaware POI recommendations to users is necessary to cope with user mobility. Current POI recommendation approaches do not take into consideration the transmission of infections between co-located users. In this paper, we formulate a new query named Epidemic-aware POI Recommendation Query (EPQ), to timely recommend a set of POIs to users at different time steps, while considering the spread of infection between co-located users, their social friendships, and their preference. We prove that EPQ is NP-hard and propose an effective and efficient algorithm, Epidemic-aware POI Recommendation (EpRec) to tackle EPQ. We evaluate EpRec on existing location-based social networks and pandemic datasets against state-of-the-art algorithms. The experimental results show that EpRec outperforms the baselines in effectiveness and efficiency.

Competitive Influence Maximization (CIM) aims to maximize the influence of a party given the competition from other parties in the same social network, like companies find key users to promote their competitive products on the social network to achieve maximum profit. Recently, learning-based solutions are introduced to tackle the competitive influence maximization problem. However, such studies focus on the static nature of social networks. This paper proposes a deep reinforcement learning-based framework employing network embedding, termed as DRL-EMB, to tackle the CIM problem on evolving social networks. The DRL-EMB key objective is to find the best strategy to maximize the party's reward, considering budget and competition with information propagation and network evolving being run in parallel. We validate our proposed framework with the DRL-based model using hand-crafted state features (DRL-HCF) and heuristic-based methods. Experimental results show that our proposed framework, DRL-EMB, achieves better results than heuristic-based and DRL-HCF models while significantly outperforming the DRL-HCF model in terms of time efficiency.

Recent studies have considered the reinforcement and deep reinforcement learning models to address the competitive influence maximization (CIM) problem. However, these models assume complete network topology information is available to address the CIM problem. This assumption is unrealistic as it is difficult to obtain complete social network data and requires exhaustive efforts to obtain it. In this work, we propose a deep reinforcement learning-based (DRL) model to tackle the competitive influence maximization on unknown social networks. Our proposed model has a two-fold objective: the first is to identify the time when to explore the network to collect network information. The second is to determine key influential users from the explored network, using optimal seed-selection strategy considering the competition in the social network. Moreover, we integrate the transfer learning in DRL to improve the training efficiency of DRL models. Experimental results show that our proposed DRL and transfer learning-based DRL models achieve significantly better performance than heuristic-based methods.

False information on the Internet has caused severe damage to society. Researchers have proposed methods to determine the credibility of news and have obtained good results. As different media sources (publishers) have different content generators (writers) and may focus on different topics or aspects, the word/topic distribution for each media source is divergent from others. We expose a challenge in the generalizability of existing content-based methods to perform consistently when applied to news from media sources non-existing in the training set, namely the cross-source failure. A cross-source setting can cause a decrease beyond in accuracy for current methods; content-sensitive features are considered one of the major causes of cross-source failure for a content-based approach. To overcome this challenge, we propose a syntactic network for news credibility (SYNC), which focuses on function words and syntactic structure to learn generalizable representations for news credibility and further reinforce the cross-source robustness for different media. Experiments with cross-validation on 194 real-world media sources showed that the proposed method could learn the generalizable features and outperformed the state-of-the-art methods on unseen media sources. Extensive analysis on the embedding feature representation represents a strength of the proposed method compared to current content embedding feature approaches. We envision that the proposed method is more robust for real-life application with SYNC on account of its good generalizability.

"Echo chamber" is a metaphorical description of a situation in which beliefs are amplified inside a closed network, and social media platforms provide an environment that is well-suited to this phenomenon. Depending on the scale of the echo chamber, a user's judgment of different opinions may be restricted. The current study focuses on detecting echoing interaction between a post and its related comments to then quantify the predominating degree of echo chamber behavior on Facebook pages. To enable such detection, two content-based features are designed; the first aids stance representation of comments on a particular discussion topic, and the second focuses on the type and intensity of emotion elicited by a subject. This work also introduces data-driven semi-supervised approaches to extract such features from social media data.

The popularity of live streaming has led to the explosive growth in new video contents and social communities on emerging platforms such as Facebook Live and Twitch. Viewers on these platforms are able to follow multiple streams of live events simultaneously, while engaging discussions with friends. However, existing approaches for selecting live streaming channels still focus on satisfying individual preferences of users, without considering the need to accommodate real-time social interactions among viewers and to diversify the content of streams. In this paper, therefore, we formulate a new Social-aware Diverse and Preferred Live Streaming Channel Query (SDSQ) that jointly selects a set of diverse and preferred live streaming channels and a group of socially tight viewers. We prove that SDSQ is NP-hard and inapproximable within any factor, and design SDSSel, a 2-approximation algorithm with a guaranteed error bound. We perform a user study on Twitch with 432 participants to validate the need of SDSQ and the usefulness of SDSSel. We also conduct large-scale experiments on real datasets to demonstrate the superiority of the proposed algorithm over several baselines in terms of solution quality and efficiency.

Emotions are expressed in nuanced ways, which varies by collective or individual experiences, knowledge, and beliefs.Therefore, to understand emotion, as conveyed through text, a robust mechanism capable of capturing and modeling different linguistic nuances and phenomena is needed.We propose a semisupervised, graph-based algorithm to produce rich structural descriptors which serve as the building blocks for constructing contextualized affect representations from text.The pattern-based representations are further enriched with word embeddings and evaluated through several emotion recognition tasks.Our experimental results demonstrate that the proposed method outperforms state-of-the-art techniques on emotion recognition tasks.

Companies aim to promote their products under competitions and try to gain more profit than other companies. This problem is formulated as a Competitive Influence Maximization (CIM). Recently, a reinforcement learning has been used to solve the CIM problem, that is, to find an optimal strategy against competitor in order to maximize the commutative reward under the competition from other agents. However, reinforcement learning agents require huge training time to find an optimal strategy whenever the settings of the agents or the networks change. To tackle this issue, we propose a transfer learning method in reinforcement learning to reduce the training time and utilize the knowledge gained on source network to target network. Our method relies on two ideas, the first one is the state representation of the source and target networks in order to efficiently utilize the knowledge gained on source network to target network. The second idea is to transfer the final Q-solution of source network while learning on the target network. We validate our transfer learning method in similar or different settings of source and target networks while competing against the competitor's known strategies. Experimental results show that our proposed transfer learning method achieves similar or better performance as a baseline model while significantly reducing training time in all settings.

The rapid growth of social networks have enabled users to instantly share what is happening around them. With the character-limitation and other feature constraints imposed by microblogs, users are obliged to express their intentions in implicit forms. This behavior poses many challenges for contextual approaches that aim to identify user intentions. Furthermore, users have the tendency to display different degree of preferences towards specific interests, simultaneously in time, making it difficult for models to rank the discovered interests. We propose a dynamic interest keyword model, a graph-based ranking mechanism, that identifies the different degrees of interests of a user. Our results show that the proposed system detects human-inferred interests, 94% of the time, showing that the model is feasible and contributes various insights that can be used to improve user intention identification systems.

Mass media sources, specifically the news media, have traditionally informed us of daily events. In modern times, social media services such as Twitter provide an enormous amount of user-generated data, which have great potential to contain informative news-related content. For these resources to be useful, we must find a way to filter noise and only capture the content that, based on its similarity to the news media, is considered valuable. However, even after noise is removed, information overload may still exist in the remaining data-hence, it is convenient to prioritize it for consumption. To achieve prioritization, information must be ranked in order of estimated importance considering three factors. First, the temporal prevalence of a particular topic in the news media is a factor of importance, and can be considered the media focus (MF) of a topic. Second, the temporal prevalence of the topic in social media indicates its user attention (UA). Last, the interaction between the social media users who mention this topic indicates the strength of the community discussing it, and can be regarded as the user interaction (UI) toward the topic. We propose an unsupervised framework-SociRank-which identifies news topics prevalent in both social media and the news media, and then ranks them by relevance using their degrees of MF, UA, and UI. Our experiments show that SociRank improves the quality and variety of automatically identified news topics.

Image tagging approaches have gained popularity in recent years. Advances in social computing research have enabled a faster completion of this task which before was known to be tedious. Most of which focus on the image-label relationship, leading to a vast amount of image repositories with corresponding text descriptors. There are several other collections such as dictionaries, lexicons or ontologies that aid different tasks the domains of text mining, Natural Language Processing and the different applications that come with it. Many of the research in this area is oriented towards social network analysis. These collections are based on formal ways of expression and unfortunately social networks content is not always formal. There is an increasing use of informal, many times regional language popularly called slang. We propose a Game With a Purpose (GWAP) system to facilitate the collection of informal terms. Leveraging on social interactions we seek to obtain a lexicon that is more suitable for social media related analysis.

Mental disorders are currently affecting millions of people from different cultures, age groups and geographic regions. The challenge of mental disorders is that they are difficult to detect on suffering patients, thus presenting an alarming number of undetected cases and misdiagnosis. In this paper, we aim at building predictive models that leverage language and behavioral patterns, used particularly in social media, to determine whether a user is suffering from two cases of mental disorder. These predictive models are made possible by employing a novel data collection process, coined as Subconscious Crowdsourcing, which helps to collect a faster and more reliable dataset of patients. Our experiments suggest that extracting specific language patterns and social interaction features from reliable patient datasets can greatly contribute to further analysis and detection of mental disorders.

Mental illnesses rank as some of the most disabling conditions, affecting millions of people, across the globe. In general, the main challenge of mental disorders is that they remain difficult to detect on suffering patients. In an online environment, the challenge extends to the collection of patients data and the implementation of proper algorithms to assist in the detection of such illnesses. In this paper, we propose a novel data collection mechanism and build predictive models that leverage language and behavioral patterns, used particularly on Twitter, to determine whether a user is suffering from a mental disorder. After training the predictive models, they are further pre-trained to serve as the backend for our demonstration, MIDAS. MIDAS offers an analytics web-service to explore several characteristics pertaining to user's linguistic and behavioral patterns on social media, with respect to mental illnesses.

Expanding social networks have led to a collateral growth of user generated content on the web. Micro-blogs have positioned themselves as a very common and popular channel of expression. In recent years there has been an incremental understanding that if these opinions are analyzed and interpreted correctly they can provide useful information such as understanding how people feel or react towards a specific topic. A broad version of this task attempts to determine if a given text is an expression of positive or negative opinion. More detailed alternatives classify texts into specific emotion labels. This has made it crucial to devise algorithms that efficiently identify the emotions expressed within the opinionated content. This work proposes an unsupervised graph-based algorithm to extract emotion bearing patterns from micro-blog posts. Having the extracted patterns, a classifier is implemented to efficiently identify the emotions expressed in posts without depending on predefined emotional dictionaries, lexicons or ontologies. The system also considers that posts maybe written in multiple languages. It then takes advantage of the pattern extraction method to successfully perform in different languages, domains and data sets. Experimental results are shown for English, Spanish and French tweets and achieve a desired accuracy, generality, adaptability and minimal supervision.

Traditional classifiers require extracting high dimensional feature representations, which become computationally expensive to process and can misrepresent or deteriorate the accuracy of a classifier. By utilizing a more representative list of extracted patterns, we can improve the precision and recall of a classification task. In this paper, we propose an unsupervised graph-based approach for bootstrapping Twitter-specific emotion-bearing patterns. Due to its novel bootstrapping process, the full system is also adaptable to different domains and classification problems. Furthermore, we explore how emotion-bearing patterns can help boost an emotion classification task. The experimented results demonstrate that the extracted patterns are effective in identifying emotions for English, Spanish and French Twitter streams.

Opinionated user-generated content has been increasingly flooding the internet since the rise of the Web 2.0. Many of this content is generated by the occurrence of different events varying in time, scale and location. In recent years there has been a growing interest in having a deeper understanding of these events and how the public reacts to them. Towards this goal there is a constant development in areas such as opinion mining. Nevertheless these methods alone are insufficient to provide a greater insight regarding several event related behaviors. In this demonstration we present a time based visualization platform for analyzing events, focusing specifically on the emotional transition generated by their occurrence, to value the impact they have over society.

Today, most personalized and recommendation services are built around interest extraction models but the outputs of these algorithms are ambiguous in nature. This makes it difficult to understand what users are personally interested in and more importantly what they are feeling towards these interests and how their interests transition through time. By studying both users' interests and emotions, simultaneously, one can further investigate the motivation behind these interests. Such findings can be useful to build better interest extraction models and algorithms that leverage personalized and recommendation services (e.g., ads. targeting, e-commerce and dating sites). In this paper, we propose the demonstration of a web visualization tool - EmoViz - which facilitates the further exploration of users' interests and their emotions at a global scale. Such tool, through the use of various visual components, aims to alleviate the problem of understanding what users of the world are interested in and the motivations behind their interests and feelings.

Social networks, which have become extremely popular nowadays, contain a tremendous amount of user-generated content about real-world events. This user-generated content can naturally reflect the real-world event as they happen, and sometimes even ahead of the newswire. The goal of this work is to identify events from social streams. A model called "keyword-based evolving graph sequences" (kEGS) is proposed to capture the characteristics of information propagation in social streams. The experimental results show the usefulness of our approach in identifying real-world events in social streams.

A conducting system for conductors of all skill levels is in high demand. An interactive conducting system could mediate the interpretation of a conductor's gestures and the performance of music based on recognized results from each distributed player. Such systems would be a helpful training tool for students, an experiencing tool for professional conductors and composers to shape music at a low cost, or an entertainment tool for nonprofessional music lovers. In this paper, we propose a real-time interactive conducting system using Microsoft Kinect. The proposed system overcomes the limitation that Kinect is usually designed for large body movements; hence, delicate conducting signals can be correctly recognized without referencing any prior knowledge. The system was evaluated by conductors of all skill levels and had a high level of accuracy and a low latency.

Accurately estimating network size is essential in unstructured networks. In previous studies, proposed sampling mechanisms for estimating network sizes assumed the probability that a peer is sampled is proportional to the number of its neighbors. This assumption leads to a sampling bias in favor of peers with many neighbors - something that commonly occurs in power law networks. To reduce this sampling bias, we propose a pheromone mechanism, that calibrates sampling probability by the amount of pheromone. This mechanism can be adapted to existing size-estimation techniques. Our empirical studies show that by adapting the pheromone mechanism, most size-estimation techniques can be significantly improved (in some cases, by more than 100%).

While keyword queries have become the "standard" query language of web search and many other database applications, their brevity and unstructuredness make it difficult to detect what users really want. In this demonstration, we aim to detect such hidden query intents, which we define as the frequent phrases that users co-ask with the query term, by exploring query logs. Toward building an online search system AIDE, we offer users the function to detect general and unique intents using arbitrary ad-hoc queries at run time. We will also demonstrate the effectiveness of the system which achieves indexing and searching over 14M MSN query log records.

Although much has been written about various information integration technologies, little has been said regarding how to combine these technologies together to build an entire application. We demonstrate TheaterLoc, an information integration application that allows users to retrieve information about theatres and restaurants for various U.S. cities, including an interactive map depicting their relative locations. The data retrieved by TheaterLoc comes from five distinct heterogeneous and distributed sources. The enabling technology used to achieve the integration includes: the Ariadne information mediator, a Web site wrapper learning tool, the Theseus execution system, and a mechanism for distributed spatial query planning. Our system is novel because it demonstrates how "virtual applications" can be rapidly built from a set of integration tools and existing online data sources.