Abstract

Hashtag segmentation is the task of breaking a hashtag into its constituent tokens. Hashtags often encode the essence of usergenerated posts, along with information like topic and sentiment, which are useful in downstream tasks. Hashtags prioritize brevity and are written in unique ways - transliterating and mixing languages, spelling variations, creative named entities. Benchmark datasets used for the hashtag segmentation task - STAN, BOUN - are small in size and extracted from a single set of tweets. However, datasets should reflect the variations in writing styles of hashtags and also account for domain and language specificity, failing which the results will misrepresent model performance. We argue that model performance should be assessed on a wider variety of hashtags, and datasets should be carefully curated. To this end, we propose HashSet, a dataset comprising of: a) 1.9k manually annotated dataset; b) 3.3M loosely supervised dataset. HashSet dataset is sampled from a different set of tweets when compared to existing datasets and provides an alternate distribution of hashtags to build and validate hashtag segmentation models. We show that the performance of SOTA models for Hashtag Segmentation drops substantially on proposed dataset, indicating that the proposed dataset provides an alternate set of hashtags to train and assess models. Datasets and results are released publicly and can be accessed from https://github.com/prashantkodali/HashSet

Abstract

Targeted image retrieval has long been a challenging problem since each person has a different perception of different features leading to inconsistency among users in describing the details of a particular image. Due to this, each user needs a system personalized according to the way they have structured the image in their mind. One important application of this task is suspect identifcation in forensic investigations where a witness needs to identify the suspect from an existing criminal database. Existing methods require the attributes for each image or suffer from poor latency during training and inference. We propose a new approach to tackle this problem through explicit relevance feedback by introducing a novel loss function and a corresponding scoring function. For this, we leverage contrastive learning on the user feedback to generate the next set of suggested images while improving the level of personalization with each user feedback iteration.

Abstract

NFT or Non-Fungible Token is a token that certifies a digital asset to be unique. A wide range of assets including, digital art, music, tweets, memes, are being sold as NFTs. NFT-related content has been widely shared on social media sites such as Twitter. We aim to understand the dominant factors that influence NFT asset valuation. Towards this objective, we create a first-of-its-kind dataset linking Twitter and OpenSea (the largest NFT marketplace) to capture social media profiles and linked NFT assets. Our dataset contains 245,159 tweets posted by 17,155 unique users, directly linking 62,997 NFT assets on OpenSea worth 19 Million USD. We have made the dataset public. We analyze the growth of NFTs, characterize the Twitter users promoting NFT assets, and gauge the impact of Twitter features on the virality of an NFT. Further, we investigate the effectiveness of different social media and NFT platform features by experimenting with multiple machine learning and deep learning models to predict an asset's value. Our results show that social media features improve the accuracy by 6% over baseline models that use only NFT platform features. Among social media features, count of user membership lists, number of likes and retweets are important features.

Abstract

Recent years have witnessed a massive growth in the proliferation of fake news online. User-generated content is a blend of text and visual information leading to producing different variants of fake news. As a result, researchers started targeting multimodal methods for fake news detection. Existing methods capture high-level information from different modalities and jointly model them to decide. Given multiple input modalities, we hypothesize that not all modalities may be equally responsible for decision-making. Hence, this paper presents a novel architecture that effectively identifies and suppresses information from weaker modalities and extracts relevant information from the strong modality on a per-sample basis. We also establish intra-modality relationship by extracting fine-grained image and text features. We conduct extensive experiments on real-world datasets to show that our approach outperforms the state-of-the-art by an average of 3.05% and 4.525% on accuracy and F1-score, respectively. We also release the code, implementation details, and model checkpoints for the community’s interest.1

Abstract

Music sharing trends have been shown to change during times of socio-economic crises. Studies have also shown that music can act as a social surrogate, helping to significantly reduce loneliness by acting as an empathetic friend. We explored these phenomena through a novel study of online music sharing during the Covid-19 pandemic in India. We collected tweets from the popular social media platform Twitter during India’s first and second wave of the pandemic (n = 1,364). We examined the different ways in which music was able to accomplish the role of a social surrogate via analyzing tweet text using Natural Language Processing techniques. Additionally, we analyzed the emotional connotations of the music shared through the acoustic features and lyrical content and compared the results between pandemic and pre-pandemic times. It was observed that the role of music shifted to a more community focused function rather than tending to a more self-serving utility. Results demonstrated that people shared music during the Covid-19 pandemic which had lower valence and shared songs with topics that reflected turbulent times such as Hardship and Exclusion when compared to songs shared during pre-Covid times. The results are further discussed in the context of individualistic versus collectivistic cultures.

Abstract

—Stack Overflow is a Community Question Answering service that attracts millions of users to seek answers to their questions. Maintaining high-quality content is necessary for relevant question retrieval, question recommendation, and enhancing the user experience. Manually removing low-quality content from the platform is time-consuming and challenging for site moderators. Thus, it is imperative to assess the content quality by automatically detecting and ‘closing’ the low-quality questions. Previous works have explored lexical, communitybased, vote-based, and style-based features to detect low-quality questions. These approaches are limited to writing styles, textual, and handcrafted features. However, these features fall short in understanding semantic features and capturing the implicit relationships between tags and questions. In contrast, we propose LQuaD (Low-Quality Question Detection), a multi-tier hybrid framework that, a) incorporates semantic information of questions associated with each post using transformers, b) includes the question and tag information that enables learning via a graph convolutional network. LQuaD outperforms the state-of-the-art methods by a 21% higher F1-score on the dataset of 2.8 million questions. Furthermore, we apply survival analysis which acts as a proactive intervention to reduce the number of questions closed by informing users to take appropriate action. We find that the timeframe between the stages from the question’s creation till it gets ‘closed’ varies significantly for tags and different ‘closing’ reasons for these questions. Index Terms—Stack Overflow, Community Question Answering, Low-quality questions

Abstract

The early diagnosis of stress symptoms is essential for preventing various mental disorder such as depression. Electroencephalography (EEG) signals are frequently employed in stress detection research and are both inexpensive and noninvasive modality. This paper proposes a stress classification system by utilizing an EEG signal. EEG signals from thirty-five volunteers were analysed which were acquired using four EEG sensors using a commercially available 4-electrode Muse EEG headband. Four movie clips were chosen as stress elicitation material. Two clips were selected to induce stress as it contains emotionally inductive scenes. The other two clips were chosen that do not induce stress as it has many comedy scenes. The recorded signals were then used to build the stress classification model. We compared the Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) for classifying stress and nonstress group. The maximum classification accuracy of 93.17% was achieved using two-layer LSTM architecture.

Abstract

Stress, either physical or mental, is experienced by almost every person at some point in his lifetime. Stress is one of the leading causes of various diseases and burdens society globally. Stress badly affects an individual's well-being. Thus, stress-related study is an emerging field, and in the past decade, a lot of attention has been given to the detection and classification of stress. The estimation of stress in the individual helps in stress management before it invades the human mind and body. In this paper, we proposed a system for the detection and classification of stress. We compared the various machine learning algorithms for stress classification using EEG signal recordings. Interaxon Muse device having four dry electrodes

Abstract

Code-mixed languages are very popular in multilingual societies around the world, yet the resources lag behind to enable robust systems on such languages. A major contributing factor is the informal nature of these languages which makes it difficult to collect code-mixed data. In this paper, we propose our system for Task 1 of CACLS 2021 to generate a machine translation system for English to Hinglish in a supervised setting. Translating in the given direction can help expand the set of resources for several tasks by translating valuable datasets from high resource languages. We propose to use mBART, a pre-trained multilingual sequence-to-sequence model, and fully utilize the pre-training of the model by transliterating the roman Hindi words in the code-mixed sentences to Devanagri script. We evaluate how expanding the input by concatenating Hindi translations of the English sentences improves mBART’s performance. Our system gives a BLEU score of 12.22 on test set. Further, we perform a detailed error analysis of our proposed systems and explore the limitations of the provided dataset and metrics.

Abstract

The ACM Hypertext and Social Media Conference is a premium venue for high quality peerreviewed research on hypertext theory, systems and applications. It is concerned with all aspects of modern hypertext research including social media, semantic web, dynamic and computed hypertext and hypermedia as well as narrative systems and applications. ACM Hypertext 2021, the 32nd ACM Conference on Hypertext and Social Media was organized as a virtual event, hosted by the ADAPT Centre and Trinity College Dublin, Ireland. The theme of HT'21 is: "Hypertext in a Multimodal World". The hypertext paradigm has transformed the way we store and transfer knowledge and how we think about information and access it. We access the same sources via multiple devices, ranging from smart watches and smart phones to laptops. Information and transactions have become ever more visual, with video increasingly replacing traditional text-based web pages. Particularly, we use interactive hypertext, news feeds and videos for getting things done: planning a trip, managing our finances, interacting with colleagues, friends and family, being productive and creative. The Covid-19 situation has put a sharp focus on the interwoven disciplines needed to realise such rich applications. The research of the Hypertext community has never been more relevant. The main tracks for ACM Hypertext 2021 are: Adaptive Web and Recommender Systems, Social Web, Semantic Web and NLP, Human-information Interaction, Search and Retrieval, and Digital Humanities, Games, and Culture. In addition, the conference features a Late Breaking Results track, in which visionary Blue Sky papers were also invited, and a Doctoral Consortium.