对话是一个顺序交互的过程,回应选择旨在根据已有对话上文选择合适的回应,是自然语言处理领域的研究热点。已有研究取得了一定的成功,但仍然存在两个突出的问题。一是现有的编码器在挖掘对话文本语义信息上尚存在不足;二是只考虑每一回合对话与备选回应之间的关系,忽视了对话上文的整体语义信息。针对问题一,本文借助多头自注意力机制有效捕捉对话文本的语义信息;针对问题二,整合对话上文的整体语义信息,分别从单词、句子以及整体对话上文三个层次与备选回应进行匹配,充分保证匹配信息的完整。在Ubuntu Corpus V1和Douban Conversation Corpus数据集上的对比实验表明了本文给出方法的有效性。
Recent advances of multilingual word representations weaken the input divergences across languages, making cross-lingual transfer similar to the monolingual cross-domain and semi-supervised settings. Thus self-training, which is effective for these settings, could be possibly beneficial to cross-lingual as well. This paper presents the first comprehensive study for self-training in cross-lingual dependency parsing. Three instance selection strategies are investigated, where two of which are based on the baseline dependency parsing model, and the third one adopts an auxiliary cross-lingual POS tagging model as evidence. We conduct experiments on the universal dependencies for eleven languages. Results show that self-training can boost the dependency parsing performances on the target languages. In addition, the POS tagger assistant instance selection can achieve further improvements consistently. Detailed analysis is conducted to examine the potentiality of self-training in-depth.
In Chinese dependency parsing, the joint model of word segmentation, POS tagging and dependency parsing has become the mainstream framework because it can eliminate error propagation and share knowledge, where the transition-based model with feature templates maintains the best performance. Recently, the graph-based joint model (Yan et al., 2019) on word segmentation and dependency parsing has achieved better performance, demonstrating the advantages of the graph-based models. However, this work can not provide POS information for downstream tasks, and the POS tagging task was proved to be helpful to the dependency parsing according to the research of the transition-based model. Therefore, we propose a graph-based joint model for Chinese word segmentation, POS tagging and dependency parsing. We designed a charater-level POS tagging task, and then train it jointly with the model of Yan et al. (2019). We adopt two methods of joint POS tagging task, one is by sharing parameters, the other is by using tag attention mechanism, which enables the three tasks to better share intermediate information and improve each other’s performance. The experimental results on the Penn Chinese treebank (CTB5) show that our proposed joint model improved by 0.38% on dependency parsing than the model of Yan et al. (2019). Compared with the best transition-based joint model, our model improved by 0.18%, 0.35% and 5.99% respectively in terms of word segmentation, POS tagging and dependency parsing.
Deep learning-based Chinese zero pronoun resolution model has achieved better performance than traditional machine learning-based model. However, the existing work related to Chinese zero pronoun resolution has not yet well integrated linguistic information into the deep learningbased Chinese zero pronoun resolution model. This paper adopts the idea based on the pre-trained model, and integrates the semantic representations in the pre-trained Chinese semantic dependency graph parser into the Chinese zero pronoun resolution model. The experimental results on OntoNotes-5.0 dataset show that our proposed Chinese zero pronoun resolution model with pretrained Chinese semantic dependency parser improves the F-score by 0.4% compared with our baseline model, and obtains better results than other deep learning-based Chinese zero pronoun resolution models. In addition, we integrate the BERT representations into our model so that the performance of our model was improved by 0.7% compared with our baseline model.
Neural network based models have achieved impressive results on the sentence classification task. However, most of previous work focuses on designing more sophisticated network or effective learning paradigms on monolingual data, which often suffers from insufficient discriminative knowledge for classification. In this paper, we investigate to improve sentence classification by multilingual data augmentation and consensus learning. Comparing to previous methods, our model can make use of multilingual data generated by machine translation and mine their language-share and language-specific knowledge for better representation and classification. We evaluate our model using English (i.e., source language) and Chinese (i.e., target language) data on several sentence classification tasks. Very positive classification performance can be achieved by our proposed model.
Information extraction from documents such as receipts or invoices is a fundamental and crucial step for office automation. Many approaches focus on extracting entities and relationships from plain texts, however, when it comes to document images, such demand becomes quite challenging since visual and layout information are also of great significance to help tackle this problem. In this work, we propose the attention-based graph neural network to combine textual and visual information from document images. Moreover, the global node is introduced in our graph construction algorithm which is used as a virtual hub to collect the information from all the nodes and edges to help improve the performance. Extensive experiments on real-world datasets show that our method outperforms baseline methods by significant margins.
Multi-turn conversational Question Answering (ConvQA) is a practical task that requires the understanding of conversation history, such as previous QA pairs, the passage context, and current question. It can be applied to a variety of scenarios with human-machine dialogue. The major challenge of this task is to require the model to consider the relevant conversation history while understanding the passage. Existing methods usually simply prepend the history to the current question, or use the complicated mechanism to model the history. This article proposes an impression feature, which use the word-level inter attention mechanism to learn multi-oriented information from conversation history to the input sequence, including attention from history tokens to each token of the input sequence, and history turn inter attention from different history turns to each token of the input sequence, and self-attention within input sequence, where the input sequence contains a current question and a passage. Then a feature selection method is designed to enhance the useful history turns of conversation and weaken the unnecessary information. Finally, we demonstrate the effectiveness of the proposed method on the QuAC dataset, analyze the impact of different feature selection methods, and verify the validity and reliability of the proposed features through visualization and human evaluation.
This paper presents our work in long and short form choice, a significant question of lexical choice, which plays an important role in many Natural Language Understanding tasks. Long and short form sharing at least one identical word meaning but with different number of syllables is a highly frequent linguistic phenomenon in Chinese like 老虎-虎(laohu-hu, tiger)
Recent work on data-to-text generation has made progress under the neural encoder-decoder architectures. However, the data input size is often enormous, while not all data records are important for text generation and inappropriate input may bring noise into the final output. To solve this problem, we propose a two-step approach which first selects and orders the important data records and then generates text from the noise-reduced data. Here we propose a learning to rank model to rank the importance of each record which is supervised by a relation extractor. With the noise-reduced data as input, we implement a text generator which sequentially models the input data records and emits a summary. Experiments on the ROTOWIRE dataset verifies the effectiveness of our proposed method in both performance and efficiency.
Story generation is a challenging task of automatically creating natural languages to describe a sequence of events, which requires outputting text with not only a consistent topic but also novel wordings. Although many approaches have been proposed and obvious progress has been made on this task, there is still a large room for improvement, especially for improving thematic consistency and wording diversity. To mitigate the gap between generated stories and those written by human writers, in this paper, we propose a planning-based conditional variational autoencoder, namely Plan-CVAE, which first plans a keyword sequence and then generates a story based on the keyword sequence. In our method, the keywords planning strategy is used to improve thematic consistency while the CVAE module allows enhancing wording diversity. Experimental results on a benchmark dataset confirm that our proposed method can generate stories with both thematic consistency and wording novelty, and outperforms state-of-the-art methods on both automatic metrics and human evaluations.
Causal explanation analysis (CEA) can assist us to understand the reasons behind daily events, which has been found very helpful for understanding the coherence of messages. In this paper, we focus on Causal Explanation Detection, an important subtask of causal explanation analysis, which determines whether a causal explanation exists in one message. We design a Pyramid Salient-Aware Network (PSAN) to detect causal explanations on messages. PSAN can assist in causal explanation detection via capturing the salient semantics of discourses contained in their keywords with a bottom graph-based word-level salient network. Furthermore, PSAN can modify the dominance of discourses via a top attention-based discourse-level salient network to enhance explanatory semantics of messages. The experiments on the commonly used dataset of CEA shows that the PSAN outperforms the state-of-the-art method by 1.8% F1 value on the Causal Explanation Detection task.
Named entity recognition (NER) is an important task in the natural language processing field. Existing NER methods heavily rely on labeled data for model training, and their performance on rare entities is usually unsatisfactory. Entity dictionaries can cover many entities including both popular ones and rare ones, and are useful for NER. However, many entity names are context-dependent and it is not optimal to directly apply dictionaries without considering the context. In this paper, we propose a neural NER approach which can exploit dictionary knowledge with contextual information. We propose to learn context-aware dictionary knowledge by modeling the interactions between the entities in dictionaries and their contexts via context-dictionary attention. In addition, we propose an auxiliary term classification task to predict the types of the matched entity names, and jointly train it with the NER model to fuse both contexts and dictionary knowledge into NER. Extensive experiments on the CoNLL-2003 benchmark dataset validate the effectiveness of our approach in exploiting entity dictionaries to improve the performance of various NER models.
Named entity recognition (NER) aims to identify text spans that mention named entities and classify them into pre-defined categories. For Chinese NER task, most of the existing methods are character-based sequence labeling models and achieve great success. However, these methods usually ignore lexical knowledge, which leads to false prediction of entity boundaries. Moreover, these methods have difficulties in capturing tag dependencies. In this paper, we propose an Adaptive Multi-pass Memory Network with Hierarchical Tagging Mechanism (AMMNHT) to address all above problems. Specifically, to reduce the errors of predicting entity boundaries, we propose an adaptive multi-pass memory network to exploit lexical knowledge. In addition, we propose a hierarchical tagging layer to learn tag dependencies. Experimental results on three widely used Chinese NER datasets demonstrate that our proposed model significantly outperforms other state-of-the-art methods.
The increasing amount of semi-structured and unstructured data on tourism websites brings a need for information extraction (IE) so as to construct a Tourism-domain Knowledge Graph (TKG), which is helpful to manage tourism information and develop downstream applications such as tourism search engine, recommendation and Q & A. However, the existing TKG is deficient, and there are few open methods to promote the construction and widespread application of TKG. In this paper, we present a systematic framework to build a TKG for Hainan, collecting data from popular tourism websites and structuring it into triples. The data is multi-source and heterogeneous, which raises a great challenge for processing it. So we develop two pipelines of processing methods for semi-structured data and unstructured data respectively. We refer to tourism InfoBox for semi-structured knowledge extraction and leverage deep learning algorithms to extract entities and relations from unstructured travel notes, which are colloquial and high-noise, and then we fuse the extracted knowledge from two sources. Finally, a TKG with 13 entity types and 46 relation types is established, which totally contains 34,079 entities and 441,371 triples. The systematic procedure proposed by this paper can construct a TKG from tourism websites, which can further applied to many scenarios and provide detailed reference for the construction of other domain-specific knowledge graphs.
Event extraction is an essential yet challenging task in information extraction. Previous approaches have paid little attention to the problem of roles overlap which is a common phenomenon in practice. To solve this problem, this paper defines event relation triple to explicitly represent relations among triggers, arguments and roles which are incorporated into the model to learn their inter-dependencies. The task of argument extraction is converted to event relation triple extraction. A novel joint framework for multiple Chinese event extraction is proposed which jointly performs predictions for event triggers and arguments based on shared feature representations from pre-trained language model. Experimental comparison with state-of-the-art baselines on ACE 2005 dataset shows the superiority of the proposed method in both trigger classification and argument classification.
Joint entity and relation extraction has received increasing interests recently, due to the capability of utilizing the interactions between both steps. Among existing studies, the Multi-Head Selection (MHS) framework is efficient in extracting entities and relations simultaneously. However, the method is weak for its limited performance. In this paper, we propose several effective insights to address this problem. First, we propose an entity-specific Relative Position Representation (eRPR) to allow the model to fully leverage the distance information between entities and context tokens. Second, we introduce an auxiliary Global Relation Classification (GRC) to enhance the learning of local contextual features. Moreover, we improve the semantic representation by adopting a pre-trained language model BERT as the feature encoder. Finally, these new keypoints are closely integrated with the multi-head selection framework and optimized jointly. Extensive experiments on two benchmark datasets demonstrate that our approach overwhelmingly outperforms previous works in terms of all evaluation metrics, achieving significant improvements for relation F1 by +2.40% on CoNLL04 and +1.90% on ACE05, respectively.
Evaluation discrepancy and overcorrection phenomenon are two common problems in neural machine translation (NMT). NMT models are generally trained with word-level learning objective, but evaluated by sentence-level metrics. Moreover, the cross-entropy loss function discourages model to generate synonymous predictions and overcorrect them to ground truth words. To address these two drawbacks, we adopt multi-task learning and propose a mixed learning objective (MLO) which combines the strength of word-level and sentence-level evaluation without modifying model structure. At word-level, it calculates semantic similarity between predicted and ground truth words. At sentence-level, it computes probabilistic n-gram matching scores of generated translations. We also combine a loss-sensitive scheduled sampling decoding strategy with MLO to explore its extensibility. Experimental results on IWSLT 2016 German-English and WMT 2019 English-Chinese datasets demonstrate that our methodology can significantly promote translation quality. The ablation study shows that both word-level and sentence-level learning objectives can improve BLEU scores. Furthermore, MLO is consistent with state-of-the-art scheduled sampling methods and can achieve further promotion.
Reinforcement learning (RL) has made remarkable progress in neural machine translation (NMT). However, it exists the problems with uneven sampling distribution, sparse rewards and high variance in training phase. Therefore, we propose a multi-reward reinforcement learning training strategy to decouple action selection and value estimation. Meanwhile, our method combines with language model rewards to jointly optimize model parameters. In addition, we add Gumbel noise in sampling to obtain more effective semantic information. To verify the robustness of our method, we not only conducted experiments on large corpora, but also performed on low-resource languages. Experimental results show that our work is superior to the baselines in WMT14 English-German, LDC2014 Chinese-English and CWMT2018 Mongolian-Chinese tasks, which fully certificates the effectiveness of our method.
Text classification tends to be difficult when data are inadequate considering the amount of manually labeled text corpora. For low-resource agglutinative languages including Uyghur, Kazakh, and Kyrgyz (UKK languages), in which words are manufactured via stems concatenated with several suffixes and stems are used as the representation of text content, this feature allows infinite derivatives vocabulary that leads to high uncertainty of writing forms and huge redundant features. There are major challenges of low-resource agglutinative text classification the lack of labeled data in a target domain and morphologic diversity of derivations in language structures. It is an effective solution which fine-tuning a pre-trained language model to provide meaningful and favorable-to-use feature extractors for downstream text classification tasks. To this end, we propose a low-resource agglutinative language model fine-tuning AgglutiFiT, specifically, we build a low-noise fine-tuning dataset by morphological analysis and stem extraction, then fine-tune the cross-lingual pre-training model on this dataset. Moreover, we propose an attention-based fine-tuning strategy that better selects relevant semantic and syntactic information from the pre-trained language model and uses those features on downstream text classification tasks. We evaluate our methods on nine Uyghur, Kazakh, and Kyrgyz classification datasets, where they have significantly better performance compared with several strong baselines.
Although named entity recognition achieved great success by introducing the neural networks, it is challenging to apply these models to low resource languages including Uyghur while it depends on a large amount of annotated training data. Constructing a well-annotated named entity corpus manually is very time-consuming and labor-intensive. Most existing methods based on the parallel corpus combined with the word alignment tools. However, word alignment methods introduce alignment errors inevitably. In this paper, we address this problem by a named entity tag transfer method based on the common neural machine translation. The proposed method marks the entity boundaries in Chinese sentence and translates the sentences to Uyghur by neural machine translation system, hope that neural machine translation will align the source and target entity by the self-attention mechanism. The experimental results show that the Uyghur named entity recognition system trained by the constructed corpus achieve good performance on the test set, with 73.80% F1 score(3.79% improvement by baseline)
The manual labeling work for constructing the Korean corpus is too time-consuming and laborious. It is difficult for low-minority languages to integrate resources. As a result, the research progress of Korean language information processing is slow. From the perspective of representation learning, reinforcement learning was combined with traditional deep learning methods. Based on the Korean text classification effect as a benchmark, and studied how to extract important Korean words in sentences. A structured model Information Distilled of Korean (IDK) was proposed. The model recognizes the words in Korean sentences and retains important words and deletes non-important words. Thereby transforming the reconstruction of the sentence into a sequential decision problem. So you can introduce the Policy Gradient method in reinforcement learning to solve the conversion problem. The results show that the model can identify the important words in Korean instead of manual annotation for representation learning. Furthermore, compared with traditional text classification methods, the model also improves the effect of Korean text classification.
Question classification is a crucial subtask in question answering system. Mongolian is a kind of few resource language. It lacks public labeled corpus. And the complex morphological structure of Mongolian vocabulary makes the data-sparse problem. This paper proposes a classification model, which combines the Bi-LSTM model with the Multi-Head Attention mechanism. The Multi-Head Attention mechanism extracts relevant information from different dimensions and representation subspace. According to the characteristics of Mongolian word-formation, this paper introduces Mongolian morphemes representation in the embedding layer. Morpheme vector focuses on the semantics of the Mongolian word. In this paper, character vector and morpheme vector are concatenated to get word vector, which sends to the Bi-LSTM getting context representation. Finally, the Multi-Head Attention obtains global information for classification. The model experimented on the Mongolian corpus. Experimental results show that our proposed model significantly outperforms baseline systems.
A clause complex consists of clauses, which are connected by component sharing relations and logic-semantic relations. Hence, clause-complex level structural transformations in translation are concerned with the expression adjustment of these two types of relations. In this paper, a formal scheme for tagging structural transformations in English-Chinese translation is designed. The annotation scheme include 3 steps operated on two grammatical levels: parsing an English clause complex into constructs and assembling construct translations on the clause complex level; translating constructs independently on the clause level. The assembling step involves 2 operations: performing operation functions and inserting Chinese words. The corpus annotation shows that it is feasible to divide structural transformations in English-Chinese translation into 2 levels. The corpus, which unfolds formally the operations of clause-complex level structural transformations, would help to improve the end-to-end translation of complicated sentences.
Recently, more and more data have been generated in the online world, filled with offensive language such as threats, swear words or straightforward insults. It is disgraceful for a progressive society, and then the question arises on how language resources and technologies can cope with this challenge. However, previous work only analyzes the problem as a whole but fails to detect particular types of offensive content in a more fine-grained way, mainly because of the lack of annotated data. In this work, we present a densely annotated data-set COLA
In this paper, we introduce LiveQA, a new question answering dataset constructed from play-by-play live broadcast. It contains 117k multiple-choice questions written by human commentators for over 1,670 NBA games, which are collected from the Chinese Hupu1 website. Derived from the characteristics of sports games, LiveQA can potentially test the reasoning ability across timeline-based live broadcasts, which is challenging compared to the existing datasets. In LiveQA, the questions require understanding the timeline, tracking events or doing mathematical computations. Our preliminary experiments show that the dataset introduces a challenging problem for question answering models, and a strong baseline model only achieves the accuracy of 53.1% and cannot beat the dominant option rule. We release the code and data of this paper for future research.
Elementary Discourse Unit (EDU) recognition is the basic task of discourse analysis, and the Chinese and English discourse alignment corpus is helpful to the studies of EDU recognition. This paper first builds Chinese-English parallel discourse corpus, in which EDUs are annotated and aligned. Then, we present the framework of Bi-LSTM-CRF EDUs recognition model using word embedding, POS and syntactic features, which can combine the advantage of CRF and Bi-LSTM. The results show that F1 is about 2% higher than the traditional method. Compared with CRF and Bi-LSTM, the Bi-LSTM-CRF model can combine the advantages of them and obtains satisfactory results for Chinese and English EDUs recognition. The experiment of feature contribution shows that using all features together can get best result, the syntactic feature outperforms than other features.
Aspect-category sentiment classification (ACSC) aims to identify the sentiment polarities towards the aspect categories mentioned in a sentence. Because a sentence often mentions more than one aspect category and expresses different sentiment polarities to them, finding aspect category-related information from the sentence is the key challenge to accurately recognize the sentiment polarity. Most previous models take both sentence and aspect category as input and query aspect category-related information based on the aspect category. However, these models represent the aspect category as a context-independent vector called aspect embedding, which may not be effective enough as a query. In this paper, we propose two contextualized aspect category representations, Contextualized Aspect Vector (CAV) and Contextualized Aspect Matrix (CAM). Specifically, we use the coarse aspect category-related information found by the aspect category detection task to generate CAV or CAM. Then the CAV or CAM as queries are used to search for fine-grained aspect category-related information like aspect embedding by aspect-category sentiment classification models. In experiments, we integrate the proposed CAV and CAM into several representative aspect embedding-based aspect-category sentiment classification models. Experimental results on the SemEval-2014 Restaurant Review dataset and the Multi-Aspect Multi-Sentiment dataset demonstrate the effectiveness of CAV and CAM.
Multimodal sentiment analysis aims to learn a joint representation of multiple features. As demonstrated by previous studies, it is shown that the language modality may contain more semantic information than that of other modalities. Based on this observation, we propose a Multi-perspective Fusion Network(MPFN) focusing on Sense Attentive Language for multimodal sentiment analysis. Different from previous studies, we use the language modality as the main part of the final joint representation, and propose a multi-stage and uni-stage fusion strategy to get the fusion representation of the multiple modalities to assist the final language-dominated multimodal representation. In our model, a Sense-Level Attention Network is proposed to dynamically learn the word representation which is guided by the fusion of the multiple modalities. As in turn, the learned language representation can also help the multi-stage and uni-stage fusion of the different modalities. In this way, the model can jointly learn a well integrated final representation focusing on the language and the interactions between the multiple modalities both on multi-stage and uni-stage. Several experiments are carried on the CMU-MOSI, the CMU-MOSEI and the YouTube public datasets. The experiments show that our model performs better or competitive results compared with the baseline models.
Emotion recognition in dialogue systems has gained attention in the field of natural language processing recent years, because it can be applied in opinion mining from public conversational data on social media. In this paper, we propose a hierarchical model to recognize emotions in the dialogue. In the first layer, in order to extract textual features of utterances, we propose a convolutional self-attention network(CAN). Convolution is used to capture n-gram information and attention mechanism is used to obtain the relevant semantic information among words in the utterance. In the second layer, a GRU-based network helps to capture contextual information in the conversation. Furthermore, we discuss the effects of unidirectional and bidirectional networks. We conduct experiments on Friends dataset and EmotionPush dataset. The results show that our proposed model(CAN-GRU) and its variants achieve better performance than baselines.
Aspect-category sentiment analysis (ACSA) aims to predict the aspect categories mentioned in texts and their corresponding sentiment polarities. Some joint models have been proposed to address this task. Given a text, these joint models detect all the aspect categories mentioned in the text and predict the sentiment polarities toward them at once. Although these joint models obtain promising performances, they train separate parameters for each aspect category and therefore suffer from data deficiency of some aspect categories. To solve this problem, we propose a novel joint model which contains a shared sentiment prediction layer. The shared sentiment prediction layer transfers sentiment knowledge between aspect categories and alleviates the problem caused by data deficiency. Experiments conducted on SemEval-2016 Datasets demonstrate the effectiveness of our model.
It is well known that deep learning model has huge parameters and is computationally expensive, especially for embedded and mobile devices. Polyphone pronunciations selection is a basic function for Chinese Text-to-Speech (TTS) application. Recurrent neural network (RNN) is a good sequence labeling solution for polyphone pronunciation selection. However, huge parameters and computation make compression needed to alleviate its disadvantage. In contrast to existing quantization with low precision data format and projection layer, we propose a novel method based on shared labels, which focuses on compressing the fully-connected layer before Softmax for models with a huge number of labels in TTS polyphone selection. The basic idea is to compress large number of target labels into a few label clusters, which will share the parameters of fully-connected layer. Furthermore, we combine it with other methods to further compress the polyphone pronunciation selection model. The experimental result shows that for Bi-LSTM (Bidirectional Long Short Term Memory) based polyphone selection, shared labels model decreases about 52% of original model size and accelerates prediction by 44% almost without performance loss. It is worth mentioning that the proposed method can be applied for other tasks to compress the model and accelerate the calculation.
Legal Judgement Prediction has attracted more and more attention in recent years. One of the challenges is how to design a model with better interpretable prediction results. Previous studies have proposed different interpretable models based on the generation of court views and the extraction of charge keywords. Different from previous work, we propose a multi-task legal judgement prediction model which combines a subtask of the seriousness of charges. By introducing this subtask, our model can capture the attention weights of different terms of penalty corresponding to the charges and give more attention to the correct terms of penalty in the fact descriptions. Meanwhile, our model also incorporates the position of defendant making it capable of giving attention to the contextual information of the defendant. We carry several experiments on the public CAIL2018 dataset. Experimental results show that our model achieves better or comparable performance on three subtasks compared with the baseline models. Moreover, we also analyze the interpretable contribution of our model.
Clickbait is a form of web content designed to attract attention and entice users to click on specific hyperlinks. The detection of clickbaits is an important task for online platforms to improve the quality of web content and the satisfaction of users. Clickbait detection is typically formed as a binary classification task based on the title and body of a webpage, and existing methods are mainly based on the content of title and the relevance between title and body. However, these methods ignore the stylistic patterns of titles, which can provide important clues on identifying clickbaits. In addition, they do not consider the interactions between the contexts within title and body, which are very important for measuring their relevance for clickbait detection. In this paper, we propose a clickbait detection approach with style-aware title modeling and co-attention. Specifically, we use Transformers to learn content representations of title and body, and respectively compute two content-based clickbait scores for title and body based on their representations. In addition, we propose to use a character-level Transformer to learn a style-aware title representation by capturing the stylistic patterns of title, and we compute a title stylistic score based on this representation. Besides, we propose to use a co-attention network to model the relatedness between the contexts within title and body, and further enhance their representations by encoding the interaction information. We compute a title-body matching score based on the representations of title and body enhanced by their interactions. The final clickbait score is predicted by a weighted summation of the aforementioned four kinds of scores. Extensive experiments on two benchmark datasets show that our approach can effectively improve the performance of clickbait detection and consistently outperform many baseline methods.
The obstetric Electronic Medical Record (EMR) contains a large amount of medical data and health information. It plays a vital role in improving the quality of the diagnosis assistant service. In this paper, we treat the diagnosis assistant as a multi-label classification task and propose a Knowledge-Enabled Diagnosis Assistant (KEDA) model for the obstetric diagnosis assistant. We utilize the numerical information in EMRs and the external knowledge from Chinese Obstetric Knowledge Graph (COKG) to enhance the text representation of EMRs. Specifically, the bidirectional maximum matching method and similarity-based approach are used to obtain the entities set contained in EMRs and linked to the COKG. The final knowledge representation is obtained by a weight-based disease prediction algorithm, and it is fused with the text representation through a linear weighting method. Experiment results show that our approach can bring about +3.53 F1 score improvements upon the strong BERT baseline in the diagnosis assistant task.
Academic Phrasebank is an important resource for academic writers. Student writers use the phrases of Academic Phrasebank organizing their research article to improve their writing ability. Due to the limited size of Academic Phrasebank, it can not meet all the academic writing needs. There are still a large number of academic phraseology in the authentic research article. In this paper, we proposed an academic phraseology extraction model based on constituency parsing and dependency parsing, which can automatically extract the academic phraseology similar to phrases of Academic Phrasebank from an unlabelled research article. We divided the proposed model into three main components including an academic phraseology corpus module, a sentence simplification module, and a syntactic parsing module. We created a corpus of academic phraseology of 2,129 words to help judge whether a word is neutral and general, and created two datasets under two scenarios to verify the feasibility of the proposed model.
One challenge in Natural Language Processing (NLP) area is to learn semantic representation in different contexts. Recent works on pre-trained language model have received great attentions and have been proven as an effective technique. In spite of the success of pre-trained language model in many NLP tasks, the learned text representation only contains the correlation among the words in the sentence itself and ignores the implicit relationship between arbitrary tokens in the sequence. To address this problem, we focus on how to make our model effectively learn word representations that contain the relational information between any tokens of text sequences. In this paper, we propose to integrate the relational network(RN) into a Wasserstein autoencoder(WAE). Specifically, WAE and RN are used to better keep the semantic structurse and capture the relational information, respectively. Extensive experiments demonstrate that our proposed model achieves significant improvements over the traditional Seq2Seq baselines.