spb/wp30_uqo

1{2  "document_id": "lb_qasper_db4afd55783aaf6d069c5228152492cf0804e9cf310cb238",3  "family": "public_qasper",4  "split": "test",5  "text": "Introduction\nWith the increasing popularity of the Internet, online texts provided by social media platform (e.g. Twitter) and news media sites (e.g. Google news) have become important sources of real-world events. Therefore, it is crucial to automatically extract events from online texts.\nDue to the high variety of events discussed online and the difficulty in obtaining annotated data for training, traditional template-based or supervised learning approaches for event extraction are no longer applicable in dealing with online texts. Nevertheless, newsworthy events are often discussed by many tweets or online news articles. Therefore, the same event could be mentioned by a high volume of redundant tweets or news articles. This property inspires the research community to devise clustering-based models BIBREF0 , BIBREF1 , BIBREF2 to discover new or previously unidentified events without extracting structured representations.\nTo extract structured representations of events such as who did what, when, where and why, Bayesian approaches have made some progress. Assuming that each document is assigned to a single event, which is modeled as a joint distribution over the named entities, the date and the location of the event, and the event-related keywords, Zhou et al. zhou2014simple proposed an unsupervised Latent Event Model (LEM) for open-domain event extraction. To address the limitation that LEM requires the number of events to be pre-set, Zhou et al. zhou2017event further proposed the Dirichlet Process Event Mixture Model (DPEMM) in which the number of events can be learned automatically from data. However, both LEM and DPEMM have two limitations: (1) they assume that all words in a document are generated from a single event which can be represented by a quadruple INLINEFORM0 entity, location, keyword, date INLINEFORM1 . However, long texts such news articles often describe multiple events which clearly violates this assumption; (2) During the inference process of both approaches, the Gibbs sampler needs to compute the conditional posterior distribution and assigns an event for each document. This is time consuming and takes long time to converge.\nTo deal with these limitations, in this paper, we propose the Adversarial-neural Event Model (AEM) based on adversarial training for open-domain event extraction. The principle idea is to use a generator network to learn the projection function between the document-event distribution and four event related word distributions (entity distribution, location distribution, keyword distribution and date distribution). Instead of providing an analytic approximation, AEM uses a discriminator network to discriminate between the reconstructed documents from latent events and the original input documents. This essentially helps the generator to construct a more realistic document from a random noise drawn from a Dirichlet distribution. Due to the flexibility of neural networks, the generator is capable of learning complicated nonlinear distributions. And the supervision signal provided by the discriminator will help generator to capture the event-related patterns. Furthermore, the discriminator also provides low-dimensional discriminative features which can be used to visualize documents and events.\nThe main contributions of the paper are summarized below:\nRelated Work\nOur work is related to two lines of research, event extraction and Generative Adversarial Nets.\nEvent Extraction\nRecently there has been much interest in event extraction from online texts, and approaches could be categorized as domain-specific and open-domain event extraction.\nDomain-specific event extraction often focuses on the specific types of events (e.g. sports events or city events). Panem et al. panem2014structured devised a novel algorithm to extract attribute-value pairs and mapped them to manually generated schemes for extracting the natural disaster events. Similarly, to extract the city-traffic related event, Anantharam et al. anantharam2015extracting viewed the task as a sequential tagging problem and proposed an approach based on the conditional random fields. Zhang zhang2018event proposed an event extraction approach based on imitation learning, especially on inverse reinforcement learning.\nOpen-domain event extraction aims to extract events without limiting the specific types of events. To analyze individual messages and induce a canonical value for each event, Benson et al. benson2011event proposed an approach based on a structured graphical model. By representing an event with a binary tuple which is constituted by a named entity and a date, Ritter et al. ritter2012open employed some statistic to measure the strength of associations between a named entity and a date. The proposed system relies on a supervised labeler trained on annotated data. In BIBREF1 , Abdelhaq et al. developed a real-time event extraction system called EvenTweet, and each event is represented as a triple constituted by time, location and keywords. To extract more information, Wang el al. wang2015seeft developed a system employing the links in tweets and combing tweets with linked articles to identify events. Xia el al. xia2015new combined texts with the location information to detect the events with low spatial and temporal deviations. Zhou et al. zhou2014simple,zhou2017event represented event as a quadruple and proposed two Bayesian models to extract events from tweets.\nGenerative Adversarial Nets\nAs a neural-based generative model, Generative Adversarial Nets BIBREF3 have been extensively researched in natural language processing (NLP) community.\nFor text generation, the sequence generative adversarial network (SeqGAN) proposed in BIBREF4 incorporated a policy gradient strategy to optimize the generation process. Based on the policy gradient, Lin et al. lin2017adversarial proposed RankGAN to capture the rich structures of language by ranking and analyzing a collection of human-written and machine-written sentences. To overcome mode collapse when dealing with discrete data, Fedus et al. fedus2018maskgan proposed MaskGAN which used an actor-critic conditional GAN to fill in missing text conditioned on the surrounding context. Along this line, Wang et al. wang2018sentigan proposed SentiGAN to generate texts of different sentiment labels. Besides, Li et al. li2018learning improved the performance of semi-supervised text classification using adversarial training, BIBREF5 , BIBREF6 designed GAN-based models for distance supervision relation extraction.\nAlthough various GAN based approaches have been explored for many applications, none of these approaches tackles open-domain event extraction from online texts. We propose a novel GAN-based event extraction model called AEM. Compared with the previous models, AEM has the following differences: (1) Unlike most GAN-based text generation approaches, a generator network is employed in AEM to learn the projection function between an event distribution and the event-related word distributions (entity, location, keyword, date). The learned generator captures event-related patterns rather than generating text sequence; (2) Different from LEM and DPEMM, AEM uses a generator network to capture the event-related patterns and is able to mine events from different text sources (short and long). Moreover, unlike traditional inference procedure, such as Gibbs sampling used in LEM and DPEMM, AEM could extract the events more efficiently due to the CUDA acceleration; (3) The discriminative features learned by the discriminator of AEM provide a straightforward way to visualize the extracted events.\nMethodology\nWe describe Adversarial-neural Event Model (AEM) in this section. An event is represented as a quadruple < INLINEFORM0 >, where INLINEFORM1 stands for non-location named entities, INLINEFORM2 for a location, INLINEFORM3 for event-related keywords, INLINEFORM4 for a date, and each component in the tuple is represented by component-specific representative words.\nAEM is constituted by three components: (1) The document representation module, as shown at the top of Figure FIGREF4 , defines a document representation approach which converts an input document from the online text corpus into INLINEFORM0 which captures the key event elements; (2) The generator INLINEFORM1 , as shown in the lower-left part of Figure FIGREF4 , generates a fake document INLINEFORM2 which is constituted by four multinomial distributions using an event distribution INLINEFORM3 drawn from a Dirichlet distribution as input; (3) The discriminator INLINEFORM4 , as shown in the lower-right part of Figure FIGREF4 , distinguishes the real documents from the fake ones and its output is subsequently employed as a learning signal to update the INLINEFORM5 and INLINEFORM6 . The details of each component are presented below.\nDocument Representation\nEach document INLINEFORM0 in a given corpus INLINEFORM1 is represented as a concatenation of 4 multinomial distributions which are entity distribution ( INLINEFORM2 ), location distribution ( INLINEFORM3 ), keyword distribution ( INLINEFORM4 ) and date distribution ( INLINEFORM5 ) of the document. As four distributions are calculated in a similar way, we only describe the computation of the entity distribution below as an example.\nThe entity distribution INLINEFORM0 is represented by a normalized INLINEFORM1 -dimensional vector weighted by TF-IDF, and it is calculated as: INLINEFORM2\nwhere INLINEFORM0 is the pseudo corpus constructed by removing all non-entity words from INLINEFORM1 , INLINEFORM2 is the total number of distinct entities in a corpus, INLINEFORM3 denotes the number of INLINEFORM4 -th entity appeared in document INLINEFORM5 , INLINEFORM6 represents the number of documents in the corpus, and INLINEFORM7 is the number of documents that contain INLINEFORM8 -th entity, and the obtained INLINEFORM9 denotes the relevance between INLINEFORM10 -th entity and document INLINEFORM11 .\nSimilarly, location distribution INLINEFORM0 , keyword distribution INLINEFORM1 and date distribution INLINEFORM2 of INLINEFORM3 could be calculated in the same way, and the dimensions of these distributions are denoted as INLINEFORM4 , INLINEFORM5 and INLINEFORM6 , respectively. Finally, each document INLINEFORM7 in the corpus is represented by a INLINEFORM8 -dimensional ( INLINEFORM9 = INLINEFORM10 + INLINEFORM11 + INLINEFORM12 + INLINEFORM13 ) vector INLINEFORM14 by concatenating four computed distributions.\nNetwork Architecture\nThe generator network INLINEFORM0 is designed to learn the projection function between the document-event distribution INLINEFORM1 and the four document-level word distributions (entity distribution, location distribution, keyword distribution and date distribution).\nMore concretely, INLINEFORM0 consists of a INLINEFORM1 -dimensional document-event distribution layer, INLINEFORM2 -dimensional hidden layer and INLINEFORM3 -dimensional event-related word distribution layer. Here, INLINEFORM4 denotes the event number, INLINEFORM5 is the number of units in the hidden layer, INLINEFORM6 is the vocabulary size and equals to INLINEFORM7 + INLINEFORM8 + INLINEFORM9 + INLINEFORM10 . As shown in Figure FIGREF4 , INLINEFORM11 firstly employs a random document-event distribution INLINEFORM12 as an input. To model the multinomial property of the document-event distribution, INLINEFORM13 is drawn from a Dirichlet distribution parameterized with INLINEFORM14 which is formulated as: DISPLAYFORM0\nwhere INLINEFORM0 is the hyper-parameter of the dirichlet distribution, INLINEFORM1 is the number of events which should be set in AEM, INLINEFORM2 , INLINEFORM3 represents the proportion of event INLINEFORM4 in the document and INLINEFORM5 .\nSubsequently, INLINEFORM0 transforms INLINEFORM1 into a INLINEFORM2 -dimensional hidden space using a linear layer followed by layer normalization, and the transformation is defined as: DISPLAYFORM0\nwhere INLINEFORM0 represents the weight matrix of hidden layer, and INLINEFORM1 denotes the bias term, INLINEFORM2 is the parameter of LeakyReLU activation and is set to 0.1, INLINEFORM3 and INLINEFORM4 denote the normalized hidden states and the outputs of the hidden layer, and INLINEFORM5 represents the layer normalization.\nThen, to project INLINEFORM0 into four document-level event related word distributions ( INLINEFORM1 , INLINEFORM2 , INLINEFORM3 and INLINEFORM4 shown in Figure FIGREF4 ), four subnets (each contains a linear layer, a batch normalization layer and a softmax layer) are employed in INLINEFORM5 . And the exact transformation is based on the formulas below: DISPLAYFORM0\nwhere INLINEFORM0 means softmax layer, INLINEFORM1 , INLINEFORM2 , INLINEFORM3 and INLINEFORM4 denote the weight matrices of the linear layers in subnets, INLINEFORM5 , INLINEFORM6 , INLINEFORM7 and INLINEFORM8 represent the corresponding bias terms, INLINEFORM9 , INLINEFORM10 , INLINEFORM11 and INLINEFORM12 are state vectors. INLINEFORM13 , INLINEFORM14 , INLINEFORM15 and INLINEFORM16 denote the generated entity distribution, location distribution, keyword distribution and date distribution, respectively, that correspond to the given event distribution INLINEFORM17 . And each dimension represents the relevance between corresponding entity/location/keyword/date term and the input event distribution.\nFinally, four generated distributions are concatenated to represent the generated document INLINEFORM0 corresponding to the input INLINEFORM1 : DISPLAYFORM0\nThe discriminator network INLINEFORM0 is designed as a fully-connected network which contains an input layer, a discriminative feature layer (discriminative features are employed for event visualization) and an output layer. In AEM, INLINEFORM1 uses fake document INLINEFORM2 and real document INLINEFORM3 as input and outputs the signal INLINEFORM4 to indicate the source of the input data (lower value denotes that INLINEFORM5 is prone to predict the input data as a fake document and vice versa).\nAs have previously been discussed in BIBREF7 , BIBREF8 , lipschitz continuity of INLINEFORM0 network is crucial to the training of the GAN-based approaches. To ensure the lipschitz continuity of INLINEFORM1 , we employ the spectral normalization technique BIBREF9 . More concretely, for each linear layer INLINEFORM2 (bias term is omitted for simplicity) in INLINEFORM3 , the weight matrix INLINEFORM4 is normalized by INLINEFORM5 . Here, INLINEFORM6 is the spectral norm of the weight matrix INLINEFORM7 with the definition below: DISPLAYFORM0\nwhich is equivalent to the largest singular value of INLINEFORM0 . The weight matrix INLINEFORM1 is then normalized using: DISPLAYFORM0\nObviously, the normalized weight matrix INLINEFORM0 satisfies that INLINEFORM1 and further ensures the lipschitz continuity of the INLINEFORM2 network BIBREF9 . To reduce the high cost of computing spectral norm INLINEFORM3 using singular value decomposition at each iteration, we follow BIBREF10 and employ the power iteration method to estimate INLINEFORM4 instead. With this substitution, the spectral norm can be estimated with very small additional computational time.\nObjective and Training Procedure\nThe real document INLINEFORM0 and fake document INLINEFORM1 shown in Figure FIGREF4 could be viewed as random samples from two distributions INLINEFORM2 and INLINEFORM3 , and each of them is a joint distribution constituted by four Dirichlet distributions (corresponding to entity distribution, location distribution, keyword distribution and date distribution). The training objective of AEM is to let the distribution INLINEFORM4 (produced by INLINEFORM5 network) to approximate the real data distribution INLINEFORM6 as much as possible.\nTo compare the different GAN losses, Kurach kurach2018gan takes a sober view of the current state of GAN and suggests that the Jansen-Shannon divergence used in BIBREF3 performs more stable than variant objectives. Besides, Kurach also advocates that the gradient penalty (GP) regularization devised in BIBREF8 will further improve the stability of the model. Thus, the objective function of the proposed AEM is defined as: DISPLAYFORM0\nwhere INLINEFORM0 denotes the discriminator loss, INLINEFORM1 represents the gradient penalty regularization loss, INLINEFORM2 is the gradient penalty coefficient which trade-off the two components of objective, INLINEFORM3 could be obtained by sampling uniformly along a straight line between INLINEFORM4 and INLINEFORM5 , INLINEFORM6 denotes the corresponding distribution.\nThe training procedure of AEM is presented in Algorithm SECREF15 , where INLINEFORM0 is the event number, INLINEFORM1 denotes the number of discriminator iterations per generator iteration, INLINEFORM2 is the batch size, INLINEFORM3 represents the learning rate, INLINEFORM4 and INLINEFORM5 are hyper-parameters of Adam BIBREF11 , INLINEFORM6 denotes INLINEFORM7 . In this paper, we set INLINEFORM8 , INLINEFORM9 , INLINEFORM10 . Moreover, INLINEFORM11 , INLINEFORM12 and INLINEFORM13 are set as 0.0002, 0.5 and 0.999.\n[!h] Training procedure for AEM [1] INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 , INLINEFORM4 , INLINEFORM5 , INLINEFORM6 the trained INLINEFORM7 and INLINEFORM8 . Initial INLINEFORM9 parameters INLINEFORM10 and INLINEFORM11 parameter INLINEFORM12 INLINEFORM13 has not converged INLINEFORM14 INLINEFORM15 Sample INLINEFORM16 , Sample a random INLINEFORM17 Sample a random number INLINEFORM18 INLINEFORM19 INLINEFORM20 INLINEFORM21 INLINEFORM22 INLINEFORM23 INLINEFORM24 Sample INLINEFORM25 noise INLINEFORM26 INLINEFORM27\nEvent Generation\nAfter the model training, the generator INLINEFORM0 learns the mapping function between the document-event distribution and the document-level event-related word distributions (entity, location, keyword and date). In other words, with an event distribution INLINEFORM1 as input, INLINEFORM2 could generate the corresponding entity distribution, location distribution, keyword distribution and date distribution.\nIn AEM, we employ event seed INLINEFORM0 , an INLINEFORM1 -dimensional vector with one-hot encoding, to generate the event related word distributions. For example, in ten event setting, INLINEFORM2 represents the event seed of the first event. With the event seed INLINEFORM3 as input, the corresponding distributions could be generated by INLINEFORM4 based on the equation below: DISPLAYFORM0\nwhere INLINEFORM0 , INLINEFORM1 , INLINEFORM2 and INLINEFORM3 denote the entity distribution, location distribution, keyword distribution and date distribution of the first event respectively.\nExperiments\nIn this section, we firstly describe the datasets and baseline approaches used in our experiments and then present the experimental results.\nExperimental Setup\nTo validate the effectiveness of AEM for extracting events from social media (e.g. Twitter) and news media sites (e.g. Google news), three datasets (FSD BIBREF12 , Twitter, and Google datasets) are employed. Details are summarized below:\nFSD dataset (social media) is the first story detection dataset containing 2,499 tweets. We filter out events mentioned in less than 15 tweets since events mentioned in very few tweets are less likely to be significant. The final dataset contains 2,453 tweets annotated with 20 events.\nTwitter dataset (social media) is collected from tweets published in the month of December in 2010 using Twitter streaming API. It contains 1,000 tweets annotated with 20 events.\nGoogle dataset (news article) is a subset of GDELT Event Database INLINEFORM0 , documents are retrieved by event related words. For example, documents which contain `malaysia', `airline', `search' and `plane' are retrieved for event MH370. By combining 30 events related documents, the dataset contains 11,909 news articles.\nWe choose the following three models as the baselines:\nK-means is a well known data clustering algorithm, we implement the algorithm using sklearn toolbox, and represent documents using bag-of-words weighted by TF-IDF.\nLEM BIBREF13 is a Bayesian modeling approach for open-domain event extraction. It treats an event as a latent variable and models the generation of an event as a joint distribution of its individual event elements. We implement the algorithm with the default configuration.\nDPEMM BIBREF14 is a non-parametric mixture model for event extraction. It addresses the limitation of LEM that the number of events should be known beforehand. We implement the model with the default configuration.\nFor social media text corpus (FSD and Twitter), a named entity tagger specifically built for Twitter is used to extract named entities including locations from tweets. A Twitter Part-of-Speech (POS) tagger BIBREF15 is used for POS tagging and only words tagged with nouns, verbs and adjectives are retained as keywords. For the Google dataset, we use the Stanford Named Entity Recognizer to identify the named entities (organization, location and person). Due to the `date' information not being provided in the Google dataset, we further divide the non-location named entities into two categories (`person' and `organization') and employ a quadruple <organization, location, person, keyword> to denote an event in news articles. We also remove common stopwords and only keep the recognized named entities and the tokens which are verbs, nouns or adjectives.\nExperimental Results\nTo evaluate the performance of the proposed approach, we use the evaluation metrics such as precision, recall and F-measure. Precision is defined as the proportion of the correctly identified events out of the model generated events. Recall is defined as the proportion of correctly identified true events. For calculating the precision of the 4-tuple, we use following criteria:\n(1) Do the entity/organization, location, date/person and keyword that we have extracted refer to the same event?\n(2) If the extracted representation contains keywords, are they informative enough to tell us what happened?\nTable TABREF35 shows the event extraction results on the three datasets. The statistics are obtained with the default parameter setting that INLINEFORM0 is set to 5, number of hidden units INLINEFORM1 is set to 200, and INLINEFORM2 contains three fully-connected layers. The event number INLINEFORM3 for three datasets are set to 25, 25 and 35, respectively. The examples of extracted events are shown in Table. TABREF36 .\nIt can be observed that K-means performs the worst over all three datasets. On the social media datasets, AEM outpoerforms both LEM and DPEMM by 6.5% and 1.7% respectively in F-measure on the FSD dataset, and 4.4% and 3.7% in F-measure on the Twitter dataset. We can also observe that apart from K-means, all the approaches perform worse on the Twitter dataset compared to FSD, possibly due to the limited size of the Twitter dataset. Moreover, on the Google dataset, the proposed AEM performs significantly better than LEM and DPEMM. It improves upon LEM by 15.5% and upon DPEMM by more than 30% in F-measure. This is because: (1) the assumption made by LEM and DPEMM that all words in a document are generated from a single event is not suitable for long text such as news articles; (2) DPEMM generates too many irrelevant events which leads to a very low precision score. Overall, we see the superior performance of AEM across all datasets, with more significant improvement on the for Google datasets (long text).\nWe next visualize the detected events based on the discriminative features learned by the trained INLINEFORM0 network in AEM. The t-SNE BIBREF16 visualization results on the datasets are shown in Figure FIGREF19 . For clarity, each subplot is plotted on a subset of the dataset containing ten randomly selected events. It can be observed that documents describing the same event have been grouped into the same cluster.\nTo further evaluate if a variation of the parameters INLINEFORM0 (the number of discriminator iterations per generator iteration), INLINEFORM1 (the number of units in hidden layer) and the structure of generator INLINEFORM2 will impact the extraction performance, additional experiments have been conducted on the Google dataset, with INLINEFORM3 set to 5, 7 and 10, INLINEFORM4 set to 100, 150 and 200, and three INLINEFORM5 structures (3, 4 and 5 layers). The comparison results on precision, recall and F-measure are shown in Figure FIGREF20 . From the results, it could be observed that AEM with the 5-layer generator performs the best and achieves 96.7% in F-measure, and the worst F-measure obtained by AEM is 85.7%. Overall, the AEM outperforms all compared approaches acorss various parameter settings, showing relatively stable performance.\nFinally, we compare in Figure FIGREF37 the training time required for each model, excluding the constant time required by each model to load the data. We could observe that K-means runs fastest among all four approaches. Both LEM and DPEMM need to sample the event allocation for each document and update the relevant counts during Gibbs sampling which are time consuming. AEM only requires a fraction of the training time compared to LEM and DPEMM. Moreover, on a larger dataset such as the Google dataset, AEM appears to be far more efficient compared to LEM and DPEMM.\nConclusions and Future Work\nIn this paper, we have proposed a novel approach based on adversarial training to extract the structured representation of events from online text. The experimental comparison with the state-of-the-art methods shows that AEM achieves improved extraction performance, especially on long text corpora with an improvement of 15% observed in F-measure. AEM only requires a fraction of training time compared to existing Bayesian graphical modeling approaches. In future work, we will explore incorporating external knowledge (e.g. word relatedness contained in word embeddings) into the learning framework for event extraction. Besides, exploring nonparametric neural event extraction approaches and detecting the evolution of events over time from news articles are other promising future directions.\nAcknowledgments\nWe would like to thank anonymous reviewers for their valuable comments and helpful suggestions. This work was funded by the National Key Research and Development Program of China (2016YFC1306704), the National Natural Science Foundation of China (61772132), the Natural Science Foundation of Jiangsu Province of China (BK20161430).",6  "char_len": 26496,7  "document_hash": "c21b83bb359fc7dfed8eb94716974d6b6c0ab6f746221ee35714838688c2eca5",8  "questions": [9    {10      "question_id": "q00",11      "qtype": "qasper",12      "question": "Do they report results only on English data?"13    }14  ]15}