Real Time Event Detection Adopting Incremental TF-IDF based LSH and Event Summary Generation

Abstract

Recently, twitter users are leveraged to detect social and physical events such as festivals and traffic jam at real time. Real time event detection and summarization from Cricket sports is the process of detecting events such as boundary at real time from live Cricket tweet stream as soon as event happens and generating a quick game summary. This is an interesting, yet a complex problem. Because of the need for rapid detection of sports events and for the generation of a concise summary from huge volume of tweets for Cricket enthusiasts. In this paper, a novel framework is proposed for detecting key events from live Cricket tweets and for generating a game summary using the crawled tweets. Feature vectors of live tweets are created using incremental TF-IDF representation and tweet clusters are discovered using Locality Sensitive Hashing (LSH) where the post rate of each cluster determines the key event. A key event is recognized from that cluster using our domain specific event lexicon. Then, important moments from the crawled tweets are computed by identifying the spikes in the tweets volume. Topk tweets from each moment are selected by ranking tweets on top-k words. Representative tweets from top-k tweets are identified using Jaccard similarity. The evaluation on 2017 IPL T20 Cricket live tweets using ROC measure shows that the proposed incremental TF-IDF based LSH approach detects key events with nearly 95% true positive rate and around 5% false positive rate. The proposed game summarization algorithm generates summaries which are readable and competitive to human tailored summaries.

Update Frequency and Background Corpus Selection in Dynamic TF-IDF Models for First Story Detection

“…Accordingly, their novelty scores are defined as the distance of the incoming data to: an existing data point for P2P models, a cluster of existing data points for P2C models, and all the existing data points for P2A models. The P2P models are normally nearest neighbour-based [2] [19] [3] or approximate nearest neighbour-based [12] [7] that aim at finding the most similar existing story to the incoming story. The P2C models use clusters of existing stories to represent previous events and evaluate the incoming story by comparing it with these clustered events [2] [19].…”

“…The most well-known term vector model is TF-IDF, short for term frequencyinverse document frequency, in which the weight of each term in a specific document is calculated as the product of the TF (term frequency) component and the IDF (inverse document frequency) component. There are many schemes of calculating these two components, but a widely-applied scheme is shown as follows [3] [7]:…”

“…In this way, the terms that are unseen in the background corpus but have been seen up to a certain point in the target data stream are also taken into account. Because of its online nature, a dynamic model should usually be adopted for FSD [12][11] [7]. Given this, in the next section, we will analyse how a dynamic TF-IDF model works for FSD and its difference from a static model in this context.…”

“…There is no standard update frequency for a dynamic TF-IDF model. Typically, updates are implemented so as to be less frequent than every 100 documents [7], as the update process is very expensive if the update frequency is higher than every 100 documents. In our experiments, we evaluate a range of update frequencies -specifically, every 100, 500, 1000, 10000 and 100000 documents, and also implement a static TF-IDF model as a baseline.…”

“…Nearest neighbour-based models with the traditional term vector document representations currently achieve the state of the art in FSD [13] [11] [16]. Because of its online characteristic, a dynamic term vector model that is incrementally updated during detection is usually adopted for FSD instead of a traditional static model [12][11] [7]. However, very little previous research has investigated how a dynamic term vector model works in practice for FSD, or has investigated how to select hyper-parameters (such as the model update frequency) and background corpora for such dynamic models.…”