A deep network model for paraphrase detection in short text messages

This paper is concerned with paraphrase detection, i.e., identifying sentences that are semantically identical. The ability to detect similar sentences written in natural language is crucial for several applications, such as text mining, text summarization, plagiarism detection, authorship authentication and question answering. Recognizing this importance, we study in particular how to address the challenges with detecting paraphrases in user generated short texts, such as Twitter, which often contain language irregularity and noise, and do not necessarily contain as much semantic information as longer clean texts. We propose a novel deep neural network-based approach that relies on coarse-grained sentence modelling using a convolutional neural network (CNN) and a recurrent neural network (RNN) model, combined with a specific fine-grained word-level similarity matching model. More specifically, we develop a new architecture, called DeepParaphrase, which enables to create an informative semantic representation of each sentence by (1) using CNN to extract the local region information in form of important n-grams from the sentence, and (2) applying RNN to capture the long-term dependency information. In addition, we perform a comparative study on state-of-the-art approaches within paraphrase detection. An important insight from this study is that existing paraphrase approaches perform well when applied on clean texts, but they do not necessarily deliver good performance against noisy texts, and vice versa. In contrast, our evaluation has shown that the proposed DeepParaphrase-based approach achieves good results in both types of texts, thus making it more robust and generic than the existing approaches.     

Improving Abusive Language Detection with online interaction network

The rapid development of online social media makes Abusive Language Detection (ALD) a hot topic in the field of affective computing. However, most methods for ALD in social networks do not take into account the interactive relationships among user posts, which simply regard ALD as a task of text context representation learning. To solve this problem, we propose a pipeline approach that considers both the context of a post and the characteristics of interaction network in which it is posted. Specifically, our method is divided into pre-training and downstream tasks. First, to capture fine contextual features of the posts, we use Bidirectional Encoder Representation from Transformers (BERT) as Encoder to generate sentence representations. Later, we build a Relation-Special Network according to the semantic similarity between posts as well as the interaction network structural information. On this basis, we design a Relation-Special Graph Neural Network (RSGNN) to spread effective information in the interaction network and learn the classification of texts. The experiment proves that our method can effectively improve the detection effect of abusive posts over three public datasets. The results demonstrate that injecting interaction network structure into the abusive language detection task can significantly improve the detection results.     

基于词同现网络与支持向量机的论文甄别

单词在句子中的交互不是随机的,而是基于一定的规则,这种规则可以通过语言网络进行研究。词同现网络是人类语言网络的一种表现形式,它利用单词在句子中的相邻关系来确定一个连接。文中采用语言网络分析的方法对论文进行甄别:将论文用词同现网络表示,计算网络的特征参数并输出一个向量来表征论文,然后运用支持向量机对论文进行分类。结果表明,使用该方法对高水平的论文和文本发生器产生的论文具有很好的甄别效果,对领域差别大的论文甄别效果也较显著。     

A comparative study of automated legal text classification using random forests and deep learning

Automated legal text classification is a prominent research topic in the legal field. It lays the foundation for building an intelligent legal system. Current literature focuses on international legal texts, such as Chinese cases, European cases, and Australian cases. Little attention is paid to text classification for U.S. legal texts. Deep learning has been applied to improving text classification performance. Its effectiveness needs further exploration in domains such as the legal field. This paper investigates legal text classification with a large collection of labeled U.S. case documents through comparing the effectiveness of different text classification techniques. We propose a machine learning algorithm using domain concepts as features and random forests as the classifier. Our experiment results on 30,000 full U.S. case documents in 50 categories demonstrated that our approach significantly outperforms a deep learning system built on multiple pre-trained word embeddings and deep neural networks. In addition, applying only the top 400 domain concepts as features for building the random forests could achieve the best performance. This study provides a reference to select machine learning techniques for building high-performance text classification systems in the legal domain or other fields.     

Automatic,semantics-based indexing of natural language texts for information retrieval systems

The fundamental idea of the work reported here is to extract index phrases from texts with the help of a single word concept dictionary and a thesaurus containing relations among concepts. The work is based on the fact, that, within every phrase, the single words the phrase is composed of are related in a certain well denned manner, the type of relations holding between concepts depending only on the concepts themselves. Therefore relations can be stored in a semantic network. The algorithm described extracts single word concepts from texts and combines them to phrases using the semantic relations between these concepts, which are stored in the network. The results obtained show that phrase extraction from texts by this semantic method is possible and offers many advantages over other (purely syntactic or statistic) methods concerning preciseness and completeness of the meaning representation of the text. But the results show, too, that some syntactic and morphologic “filtering” should be included for effectivity reasons.     

实体-属性抽取的GRU+CRF方法

[目的/意义]研究利用深度学习的循环神经网络GRU结合条件随机场CRF对标注的中文文本序列进行预测,来抽取在线评论文本中的实体-属性。[方法/过程]首先根据设计好的文本序列标注规范,对评论语料分词后进行实体及其属性的命名实体标注,得到单词序列、词性序列和标注序列;然后将单词序列、词性序列转为分布式词向量表示并用于GRU循环神经网络的输入;最后输出层采用条件随机场CRF,输出标签即是实体或属性。[结果/结论]实验结果表明,本文的方法将实体-属性抽取简化为命名实体标注,并利用深度学习的GRU捕获输入数据的上下文语义以及条件随机场CRF获取输出标签的前后关系,比传统的基于规则或一般的机器学习方法具有较大的应用优势。     

Automated estimation of item difficulty for multiple-choice tests: An application of word embedding techniques

Pretesting is the most commonly used method for estimating test item difficulty because it provides highly accurate results that can be applied to assessment development activities. However, pretesting is inefficient, and it can lead to item exposure. Hence, an increasing number of studies have invested considerable effort in researching the automated estimation of item difficulty. Language proficiency tests constitute the majority of researched test topics, while comparatively less research has focused on content subjects. This paper introduces a novel method for the automated estimation of item difficulty for social studies tests. In this study, we explore the difficulty of multiple-choice items, which consist of the following item elements: a question and alternative options. We use learning materials to construct a semantic space using word embedding techniques and project an item's texts into the semantic space to obtain corresponding vectors. Semantic features are obtained by calculating the cosine similarity between the vectors of item elements. Subsequently, these semantic features are sent to a classifier for training and testing. Based on the output of the classifier, an estimation model is created and item difficulty is estimated. Our findings suggest that the semantic similarity between a stem and the options has the strongest impact on item difficulty. Furthermore, the results indicate that the proposed estimation method outperforms pretesting, and therefore, we expect that the proposed approach will complement and partially replace pretesting in future.     

Word sense disambiguation based on context selection using knowledge-based word similarity

In this paper, we introduce a novel knowledge-based word-sense disambiguation (WSD) system. In particular, the main goal of our research is to find an effective way to filter out unnecessary information by using word similarity. For this, we adopt two methods in our WSD system. First, we propose a novel encoding method for word vector representation by considering the graphical semantic relationships from the lexical knowledge bases, and the word vector representation is utilized to determine the word similarity in our WSD system. Second, we present an effective method for extracting the contextual words from a text for analyzing an ambiguous word based on word similarity. The results demonstrate that the suggested methods significantly enhance the baseline WSD performance in all corpora. In particular, the performance on nouns is similar to those of the state-of-the-art knowledge-based WSD models, and the performance on verbs surpasses that of the existing knowledge-based WSD models.     

Cognitive-inspired domain adaptation of sentiment lexicons

Sentiment lexicons are essential tools for polarity classification and opinion mining. In contrast to machine learning methods that only leverage text features or raw text for sentiment analysis, methods that use sentiment lexicons embrace higher interpretability. Although a number of domain-specific sentiment lexicons are made available, it is impractical to build an ex ante lexicon that fully reflects the characteristics of the language usage in endless domains. In this article, we propose a novel approach to simultaneously train a vanilla sentiment classifier and adapt word polarities to the target domain. Specifically, we sequentially track the wrongly predicted sentences and use them as the supervision instead of addressing the gold standard as a whole to emulate the life-long cognitive process of lexicon learning. An exploration-exploitation mechanism is designed to trade off between searching for new sentiment words and updating the polarity score of one word. Experimental results on several popular datasets show that our approach significantly improves the sentiment classification performance for a variety of domains by means of improving the quality of sentiment lexicons. Case-studies also illustrate how polarity scores of the same words are discovered for different domains.     

Text classification from unlabeled documents with bootstrapping and feature projection techniques

Many machine learning algorithms have been applied to text classification tasks. In the machine learning paradigm, a general inductive process automatically builds a text classifier by learning, generally known as supervised learning. However, the supervised learning approaches have some problems. The most notable problem is that they require a large number of labeled training documents for accurate learning. While unlabeled documents are easily collected and plentiful, labeled documents are difficultly generated because a labeling task must be done by human developers. In this paper, we propose a new text classification method based on unsupervised or semi-supervised learning. The proposed method launches text classification tasks with only unlabeled documents and the title word of each category for learning, and then it automatically learns text classifier by using bootstrapping and feature projection techniques. The results of experiments showed that the proposed method achieved reasonably useful performance compared to a supervised method. If the proposed method is used in a text classification task, building text classification systems will become significantly faster and less expensive.     

Legal case document similarity: You need both network and text

Estimating the similarity between two legal case documents is an important and challenging problem, having various downstream applications such as prior-case retrieval and citation recommendation. There are two broad approaches for the task — citation network-based and text-based. Prior citation network-based approaches consider citations only to prior-cases (also called precedents) (PCNet). This approach misses important signals inherent in Statutes (written laws of a jurisdiction). In this work, we propose Hier-SPCNet that augments PCNet with a heterogeneous network of Statutes. We incorporate domain knowledge for legal document similarity into Hier-SPCNet, thereby obtaining state-of-the-art results for network-based legal document similarity.Both textual and network similarity provide important signals for legal case similarity; but till now, only trivial attempts have been made to unify the two signals. In this work, we apply several methods for combining textual and network information for estimating legal case similarity. We perform extensive experiments over legal case documents from the Indian judiciary, where the gold standard similarity between document-pairs is judged by law experts from two reputed Law institutes in India. Our experiments establish that our proposed network-based methods significantly improve the correlation with domain experts’ opinion when compared to the existing methods for network-based legal document similarity. Our best-performing combination method (that combines network-based and text-based similarity) improves the correlation with domain experts’ opinion by 11.8% over the best text-based method and 20.6% over the best network-based method. We also establish that our best-performing method can be used to recommend/retrieve citable and similar cases for a source (query) case, which are well appreciated by legal experts.     

Sememe knowledge and auxiliary information enhanced approach for sarcasm detection

Sarcasm expression is a pervasive literary technique in which people intentionally express the opposite of what is implied. Accurate detection of sarcasm in a text can facilitate the understanding of speakers’ true intentions and promote other natural language processing tasks, especially sentiment analysis tasks. Since sarcasm is a kind of implicit sentiment expression and speakers deliberately confuse the audience, it is challenging to detect sarcasm only by text. Existing approaches based on machine learning and deep learning achieved unsatisfactory performance when handling sarcasm text with complex expression or needing specific background knowledge to understand. Especially, due to the characteristics of the Chinese language itself, sarcasm detection in Chinese is more difficult. To alleviate this dilemma on Chinese sarcasm detection, we propose a sememe and auxiliary enhanced attention neural model, SAAG. At the word level, we introduce sememe knowledge to enhance the representation learning of Chinese words. Sememe is the minimum unit of meaning, which is a fine-grained portrayal of a word. At the sentence level, we leverage some auxiliary information, such as the news title, to learning the representation of the context and background of sarcasm expression. Then, we construct the representation of text expression progressively and dynamically. The evaluation on a sarcasm dateset, consisting of comments on news text, reveals that our proposed approach is effective and outperforms the state-of-the-art models.     

A benchmark for evaluating Arabic contextualized word embedding models

Word embeddings, which represent words as numerical vectors in a high-dimensional space, are contextualized by generating a unique vector representation for each sense of a word based on the surrounding words and sentence structure. They are typically generated using such deep learning models as BERT and trained on large amounts of text data and using self-supervised learning techniques. Resulting embeddings are highly effective at capturing the nuances of language, and have been shown to significantly improve the performance of numerous NLP tasks. Word embeddings represent textual records of human thinking, with all the mental relations that we utilize to produce the succession of sentences that make up texts and discourses. Consequently, the distributed representation of words within embeddings ought to capture the reasoning relations that hold texts together. This paper makes its contribution to the field by proposing a benchmark for the assessment of contextualized word embeddings that probes into their capability for true contextualization by inspecting how well they capture resemblance, contrariety, comparability, identity, relations in time and space, causation, analogy, and sense disambiguation. The proposed metrics adopt a triangulation approach, so they use (1) Hume’s reasoning relations, (2) standard analogy, and (3) sense disambiguation. The benchmark has been evaluated against 22 Arabic contextualized embeddings and has proven to be capable of quantifying their differential performance in terms of these reasoning relations. Results of evaluation of the target embeddings revealed that they do take context into account and that they do reasonably well in sense disambiguation but have weakness in their identification of converseness, synonymy, complementarity, and analogy. Results also show that size of an embedding has diminishing returns because the highly frequent language patterns swamp low frequency patterns. Furthermore, the suggest that future research endeavors should not be concerned with the quantity of data as much as its quality, and that it should focus more on the representativeness of data, and on model architecture, design, and training.     

CNN-BiLSTM-Attention: A multi-label neural classifier for short texts with a small set of labels

We propose a CNN-BiLSTM-Attention classifier to classify online short messages in Chinese posted by users on government web portals, so that a message can be directed to one or more government offices. Our model leverages every bit of information to carry out multi-label classification, to make use of different hierarchical text features and the labels information. In particular, our designed method extracts label meaning, the CNN layer extracts local semantic features of the texts, the BiLSTM layer fuses the contextual features of the texts and the local semantic features, and the attention layer selects the most relevant features for each label. We evaluate our model on two public large corpuses, and our high-quality handcraft e-government multi-label dataset, which is constructed by the text annotation tool doccano and consists of 29920 data points. Experimental results show that our proposed method is effective under common multi-label evaluation metrics, achieving micro-f1 of 77.22%, 84.42%, 87.52%, and marco-f1 of 77.68%, 73.37%, 83.57% on these three datasets respectively, confirming that our classifier is robust. We conduct ablation study to evaluate our label embedding method and attention mechanism. Moreover, case study on our handcraft e-government multi-label dataset verifies that our model integrates all types of semantic information of short messages based on different labels to achieve text classification.     

Learning to recognize webpage genres

Webpages are mainly distinguished by their topic (e.g., politics, sports etc.) and genre (e.g., blogs, homepages, e-shops, etc.). Automatic detection of webpage genre could considerably enhance the ability of modern search engines to focus on the requirements of the user’s information need. In this paper, we present an approach to webpage genre detection based on a fully-automated extraction of the feature set that represents the style of webpages. The features we propose (character n-grams of variable length and HTML tags) are language-independent and easily-extracted while they can be adapted to the properties of the still evolving web genres and the noisy environment of the web. Experiments based on two publicly-available corpora show that the performance of the proposed approach is superior in comparison to previously reported results. It is also shown that character n-grams are better features than words when the dimensionality increases while the binary representation is more effective than the term-frequency representation for both feature types. Moreover, we perform a series of cross-check experiments (e.g., training using a genre palette and testing using a different genre palette as well as using the features extracted from one corpus to discriminate the genres of the other corpus) to illustrate the robustness of our approach and its ability to capture the general stylistic properties of genre categories even when the feature set is not optimized for the given corpus.     

Network text analysis: A two-way classification approach

Text clustering is a well-known method for information retrieval and numerous methods for classifying words, documents or both together have been proposed. Frequently, textual data are encoded using vector models so the corpus is transformed in to a matrix of terms by documents; using this representation text clustering generates groups of similar objects on the basis of the presence/absence of the words in the documents. An alternative way to work on texts is to represent them as a network where nodes are entities connected by the presence and distribution of the words in the documents. In this work, after summarising the state of the art of text clustering we will present a new network approach to textual data. We undertake text co-clustering using methods developed for social network analysis. Several experimental results will be presented to demonstrate the validity of the approach and the advantages of this technique compared to existing methods.     

Sentence modeling via multiple word embeddings and multi-level comparison for semantic textual similarity

Recently, using a pretrained word embedding to represent words achieves success in many natural language processing tasks. According to objective functions, different word embedding models capture different aspects of linguistic properties. However, the Semantic Textual Similarity task, which evaluates similarity/relation between two sentences, requires to take into account of these linguistic aspects. Therefore, this research aims to encode various characteristics from multiple sets of word embeddings into one embedding and then learn similarity/relation between sentences via this novel embedding. Representing each word by multiple word embeddings, the proposed MaxLSTM-CNN encoder generates a novel sentence embedding. We then learn the similarity/relation between our sentence embeddings via Multi-level comparison. Our method M-MaxLSTM-CNN consistently shows strong performances in several tasks (i.e., measure textual similarity, identify paraphrase, recognize textual entailment). Our model does not use hand-crafted features (e.g., alignment features, Ngram overlaps, dependency features) as well as does not require pre-trained word embeddings to have the same dimension.