Generating knowledge aware explanation for natural language inference

Natural language inference (NLI) is an increasingly important task of natural language processing, and the explainable NLI generates natural language explanations (NLEs) in addition to label prediction, to make NLI explainable and acceptable. However, NLEs generated by current models often present problems that disobey of commonsense or lack of informativeness. In this paper, we propose a knowledge enhanced explainable NLI framework (KxNLI) by leveraging Knowledge Graph (KG) to address these problems. The subgraphs from KG are constructed based on the concept set of the input sequence. Contextual embedding of input and the graph embedding of subgraphs, is used to guide the NLE generation by using a copy mechanism. Furthermore, the generated NLEs are used to augment the original data. Experimental results show that the performance of KxNLI can achieve state-of-the-art (SOTA) results on the SNLI dataset when the pretrained model is fine-tuned on the augmented data. Besides, the proposed mechanism of knowledge enhancement and rationales utilization can achieve ideal performance on vanilla seq2seq model, and obtain better transfer ability when transferred to the MultiNLI dataset. In order to comprehensively evaluate generated NLEs, we design two metrics from the perspectives of the accuracy and informativeness, to measure the quality of NLEs, respectively. The results show that KxNLI can provide high quality NLEs while making accurate prediction.     

Measuring and mitigating language model biases in abusive language detection

Warning: This paper contains abusive samples that may cause discomfort to readers.Abusive language on social media reinforces prejudice against an individual or a specific group of people, which greatly hampers freedom of expression. With the rise of large-scale pre-trained language models, classification based on pre-trained language models has gradually become a paradigm for automatic abusive language detection. However, the effect of stereotypes inherent in language models on the detection of abusive language remains unknown, although this may further reinforce biases against the minorities. To this end, in this paper, we use multiple metrics to measure the presence of bias in language models and analyze the impact of these inherent biases in automatic abusive language detection. On the basis of this quantitative analysis, we propose two different debiasing strategies, token debiasing and sentence debiasing, which are jointly applied to reduce the bias of language models in abusive language detection without degrading the classification performance. Specifically, for the token debiasing strategy, we reduce the discrimination of the language model against protected attribute terms of a certain group by random probability estimation. For the sentence debiasing strategy, we replace protected attribute terms and augment the original text by counterfactual augmentation to obtain debiased samples, and use the consistency regularization between the original data and the augmented samples to eliminate the bias at the sentence level of the language model. The experimental results confirm that our method can not only reduce the bias of the language model in the abusive language detection task, but also effectively improve the performance of abusive language detection.     

Question classification using limited labelled data

Question classification (QC) involves classifying given question based on the expected answer type and is an important task in the Question Answering(QA) system. Existing approaches for question classification use full training dataset to fine-tune the models. It is expensive and requires more time to develop labelled datasets in huge size. Hence, there is a need to develop approaches that can achieve comparable or state of the art performance using limited training instances. In this paper, we propose an approach that uses data augmentation as a tool to generate additional training instances. We evaluate our proposed approach on two question classification datasets namely TREC and ICHI datasets. Experimental results show that our proposed approach reduces the requirement of labelled instances (a) up to 81.7% and achieves new state of the art accuracy of 98.11 on TREC dataset and (b) up to 75% and achieves 67.9 on ICHI dataset.     

Transfer language selection for zero-shot cross-lingual abusive language detection

We study the selection of transfer languages for automatic abusive language detection. Instead of preparing a dataset for every language, we demonstrate the effectiveness of cross-lingual transfer learning for zero-shot abusive language detection. This way we can use existing data from higher-resource languages to build better detection systems for low-resource languages. Our datasets are from seven different languages from three language families. We measure the distance between the languages using several language similarity measures, especially by quantifying the World Atlas of Language Structures. We show that there is a correlation between linguistic similarity and classifier performance. This discovery allows us to choose an optimal transfer language for zero shot abusive language detection.     

Building for tomorrow: Assessing the temporal persistence of text classifiers

Performance of text classification models tends to drop over time due to changes in data, which limits the lifetime of a pretrained model. Therefore an ability to predict a model’s ability to persist over time can help design models that can be effectively used over a longer period of time. In this paper, we provide a thorough discussion into the problem, establish an evaluation setup for the task. We look at this problem from a practical perspective by assessing the ability of a wide range of language models and classification algorithms to persist over time, as well as how dataset characteristics can help predict the temporal stability of different models. We perform longitudinal classification experiments on three datasets spanning between 6 and 19 years, and involving diverse tasks and types of data. By splitting the longitudinal datasets into years, we perform a comprehensive set of experiments by training and testing across data that are different numbers of years apart from each other, both in the past and in the future. This enables a gradual investigation into the impact of the temporal gap between training and test sets on the classification performance, as well as measuring the extent of the persistence over time. Through experimenting with a range of language models and algorithms, we observe a consistent trend of performance drop over time, which however differs significantly across datasets; indeed, datasets whose domain is more closed and language is more stable, such as with book reviews, exhibit a less pronounced performance drop than open-domain social media datasets where language varies significantly more. We find that one can estimate how a model will retain its performance over time based on (i) how well the model performs over a restricted time period and its extrapolation to a longer time period, and (ii) the linguistic characteristics of the dataset, such as the familiarity score between subsets from different years. Findings from these experiments have important implications for the design of text classification models with the aim of preserving performance over time.     

Dealing with textual noise for robust and effective BERT re-ranking

The pre-trained language models (PLMs), such as BERT, have been successfully employed in two-phases ranking pipeline for information retrieval (IR). Meanwhile, recent studies have reported that BERT model is vulnerable to imperceptible textual perturbations on quite a few natural language processing (NLP) tasks. As for IR tasks, current established BERT re-ranker is mainly trained on large-scale and relatively clean dataset, such as MS MARCO, but actually noisy text is more common in real-world scenarios, such as web search. In addition, the impact of within-document textual noises (perturbations) on retrieval effectiveness remains to be investigated, especially on the ranking quality of BERT re-ranker, considering its contextualized nature. To mitigate this gap, we carry out exploratory experiments on the MS MARCO dataset in this work to examine whether BERT re-ranker can still perform well when ranking text with noise. Unfortunately, we observe non-negligible effectiveness degradation of BERT re-ranker over a total of ten different types of synthetic within-document textual noise. Furthermore, to address the effectiveness losses over textual noise, we propose a novel noise-tolerant model, De-Ranker, which is learned by minimizing the distance between the noisy text and its original clean version. Our evaluation on the MS MARCO and TREC 2019–2020 DL datasets demonstrates that De-Ranker can deal with synthetic textual noise more effectively, with 3%–4% performance improvement over vanilla BERT re-ranker. Meanwhile, extensive zero-shot transfer experiments on a total of 18 widely-used IR datasets show that De-Ranker can not only tackle natural noise in real-world text, but also achieve 1.32% improvement on average in terms of cross-domain generalization ability on the BEIR benchmark.     

Named entity recognition in Turkish: A comparative study with detailed error analysis

Named entity recognition aims to detect pre-determined entity types in unstructured text. There is a limited number of studies on this task for low-resource languages such as Turkish. We provide a comprehensive study for Turkish named entity recognition by comparing the performances of existing state-of-the-art models on the datasets with varying domains to understand their generalization capability and further analyze why such models fail or succeed in this task. Our experimental results, supported by statistical tests, show that the highest weighted F1 scores are obtained by Transformer-based language models, varying from 80.8% in tweets to 96.1% in news articles. We find that Transformer-based language models are more robust to entity types with a small sample size and longer named entities compared to traditional models, yet all models have poor performance for longer named entities in social media. Moreover, when we shuffle 80% of words in a sentence to imitate flexible word order in Turkish, we observe more performance deterioration, 12% in well-written texts, compared to 7% in noisy text.     

A lemmatization method for Mongolian and its application to indexing for information retrieval

In Mongolian, two different alphabets are used, Cyrillic and Mongolian. In this paper, we focus solely on the Mongolian language using the Cyrillic alphabet, in which a content word can be inflected when concatenated with one or more suffixes. Identifying the original form of content words is crucial for natural language processing and information retrieval. We propose a lemmatization method for Mongolian. The advantage of our lemmatization method is that it does not rely on noun dictionaries, enabling us to lemmatize out-of-dictionary words. We also apply our method to indexing for information retrieval. We use newspaper articles and technical abstracts in experiments that show the effectiveness of our method. Our research is the first significant exploration of the effectiveness of lemmatization for information retrieval in Mongolian.     

Arabic abstractive text summarization using RNN-based and transformer-based architectures

Recently, the Transformer model architecture and the pre-trained Transformer-based language models have shown impressive performance when used in solving both natural language understanding and text generation tasks. Nevertheless, there is little research done on using these models for text generation in Arabic. This research aims at leveraging and comparing the performance of different model architectures, including RNN-based and Transformer-based ones, and different pre-trained language models, including mBERT, AraBERT, AraGPT2, and AraT5 for Arabic abstractive summarization. We first built an Arabic summarization dataset of 84,764 high-quality text-summary pairs. To use mBERT and AraBERT in the context of text summarization, we employed a BERT2BERT-based encoder-decoder model where we initialized both the encoder and decoder with the respective model weights. The proposed models have been tested using ROUGE metrics and manual human evaluation. We also compared their performance on out-of-domain data. Our pre-trained Transformer-based models give a large improvement in performance with ～79% less data. We found that AraT5 scores ～3 ROUGE higher than a BERT2BERT-based model that is initialized with AraBERT, indicating that an encoder-decoder pre-trained Transformer is more suitable for summarizing Arabic text. Also, both of these two models perform better than AraGPT2 by a clear margin, which we found to produce summaries with high readability but with relatively lesser quality. On the other hand, we found that both AraT5 and AraGPT2 are better at summarizing out-of-domain text. We released our models and dataset publicly^1,.²     

LDA单词图像表示的蒙古文古籍图像关键词检索方法

[目的]为了克服传统视觉词袋方法（Bag-of-Visual-Words）中忽略视觉单词间的空间关系和语义信息等问题。[方法]本文提出一种与视觉语言模型相结合的基于LDA主题模型,并采用查询似然模型实现检索。[结果]实验数据表明,本文所提出的基于LDA的表示方法可以高效、准确地解决蒙古文古籍的关键词检索问题。[结论]同时,该方法的性能比BoVW方法有显著提高。     

Improving topic disentanglement via contrastive learning

With the emergence and development of deep generative models, such as the variational auto-encoders (VAEs), the research on topic modeling successfully extends to a new area: neural topic modeling, which aims to learn disentangled topics to understand the data better. However, the original VAE framework had been shown to be limited in disentanglement performance, bringing their inherent defects to a neural topic model (NTM). In this paper, we put forward that the optimization objectives of contrastive learning are consistent with two important goals (alignment and uniformity) of well-disentangled topic learning. Also, the optimization objectives of contrastive learning are consistent with two key evaluation measures for topic models, topic coherence and topic diversity. So, we come to the important conclusion that alignment and uniformity of disentangled topic learning can be quantified with topic coherence and topic diversity. Accordingly, we are inspired to propose the Contrastive Disentangled Neural Topic Model (CNTM). By representing both words and topics as low-dimensional vectors in the same embedding space, we apply contrastive learning to neural topic modeling to produce factorized and disentangled topics in an interpretable manner. We compare our proposed CNTM with strong baseline models on widely-used metrics. Our model achieves the best topic coherence scores under the most general evaluation setting (100% proportion topic selected) with 25.0%, 10.9%, 24.6%, and 51.3% improvements above the second-best models’ scores reported on four datasets of 20 Newsgroups, Web Snippets, Tag My News, and Reuters, respectively. Our method also gets the second-best topic diversity scores on the dataset of 20Newsgroups and Web Snippets. Our experimental results show that CNTM can effectively leverage the disentanglement ability from contrastive learning to solve the inherent defect of neural topic modeling and obtain better topic quality.     

Fake news detection via knowledgeable prompt learning

The spread of fake news has become a significant social problem, drawing great concern for fake news detection (FND). Pretrained language models (PLMs), such as BERT and RoBERTa can benefit this task much, leading to state-of-the-art performance. The common paradigm of utilizing these PLMs is fine-tuning, in which a linear classification layer is built upon the well-initialized PLM network, resulting in an FND mode, and then the full model is tuned on a training corpus. Although great successes have been achieved, this paradigm still involves a significant gap between the language model pretraining and target task fine-tuning processes. Fortunately, prompt learning, a new alternative to PLM exploration, can handle the issue naturally, showing the potential for further performance improvements. To this end, we propose knowledgeable prompt learning (KPL) for this task. First, we apply prompt learning to FND, through designing one sophisticated prompt template and the corresponding verbal words carefully for the task. Second, we incorporate external knowledge into the prompt representation, making the representation more expressive to predict the verbal words. Experimental results on two benchmark datasets demonstrate that prompt learning is better than the baseline fine-tuning PLM utilization for FND and can outperform all previous representative methods. Our final knowledgeable model (i.e, KPL) can provide further improvements. In particular, it achieves an average increase of 3.28% in F1 score under low-resource conditions compared with fine-tuning.     

Looking deep in the eyes: Investigating interpretation methods for neural models on reading tasks using human eye-movement behaviour

This paper provides the first broad overview of the relation between different interpretation methods and human eye-movement behaviour across different tasks and architectures. The interpretation methods of neural networks provide the information the machine considers important, while the human eye-gaze has been believed to be a proxy of the human cognitive process. Thus, comparing them explains machine behaviour in terms of human behaviour, leading to improvement in machine performance through minimising their difference. We consider three types of natural language processing (NLP) tasks: sentiment analysis, relation classification and question answering, and four interpretation methods based on: simple gradient, integrated gradient, input-perturbation and attention, and three architectures: LSTM, CNN and Transformer. We leverage two corpora annotated with eye-gaze information: the Zuco dataset and the MQA-RC dataset. This research sets up two research questions. First, we investigate whether the saliency (importance) of input-words conform with those from human eye-gaze features. To this end, we compute a saliency distance (SD) between input words (by an interpretation method) and an eye-gaze feature. SD is defined as the KL-divergence between the saliency distribution over input words and an eye-gaze feature. We found that the SD scores vary depending on the combinations of tasks, interpretation methods and architectures. Second, we investigate whether the models with good saliency conformity to human eye-gaze behaviour have better prediction performances. To this end, we propose a novel evaluation device called “SD-performance curve” (SDPC) which represents the cumulative model performance against the SD scores. SDPC enables us to analyse the underlying phenomena that were overlooked using only the macroscopic metrics, such as average SD scores and rank correlations, that are typically used in the past studies. We observe that the impact of good saliency conformity between humans and machines on task performance varies among the combinations of tasks, interpretation methods and architectures. Our findings should be considered when introducing eye-gaze information for model training to improve the model performance.     

Is my stance the same as your stance? A cross validation study of stance detection datasets

Stance detection identifies a person’s evaluation of a subject, and is a crucial component for many downstream applications. In application, stance detection requires training a machine learning model on an annotated dataset and applying the model on another to predict stances of text snippets. This cross-dataset model generalization poses three central questions, which we investigate using stance classification models on 7 publicly available English Twitter datasets ranging from 297 to 48,284 instances. (1) Are stance classification models generalizable across datasets? We construct a single dataset model to train/test dataset-against-dataset, finding models do not generalize well (avg F1=0.33). (2) Can we improve the generalizability by aggregating datasets? We find a multi dataset model built on the aggregation of datasets has an improved performance (avg F1=0.69). (3) Given a model built on multiple datasets, how much additional data is required to fine-tune it? We find it challenging to ascertain a minimum number of data points due to the lack of pattern in performance. Investigating possible reasons for the choppy model performance we find that texts are not easily differentiable by stances, nor are annotations consistent within and across datasets. Our observations emphasize the need for an aggregated dataset as well as consistent labels for the generalizability of models.     

Selecting optimal training data for learning to rank

This paper is concerned with the quality of training data in learning to rank for information retrieval. While many data selection techniques have been proposed to improve the quality of training data for classification, the study on the same issue for ranking appears to be insufficient. As pointed out in this paper, it is inappropriate to extend technologies for classification to ranking, and the development of novel technologies is sorely needed. In this paper, we study the development of such technologies. To begin with, we propose the concept of “pairwise preference consistency” (PPC) to describe the quality of a training data collection from the ranking point of view. PPC takes into consideration the ordinal relationship between documents as well as the hierarchical structure on queries and documents, which are both unique properties of ranking. Then we select a subset of the original training documents, by maximizing the PPC of the selected subset. We further propose an efficient solution to the maximization problem. Empirical results on the LETOR benchmark datasets and a web search engine dataset show that with the subset of training data selected by our approach, the performance of the learned ranking model can be significantly improved.     

Indexing and stemming approaches for the Czech language

This paper describes and evaluates various stemming and indexing strategies for the Czech language. Based on Czech test-collection, we have designed and evaluated two stemming approaches, a light and a more aggressive one. We have compared them with a no stemming scheme as well as a language-independent approach (n-gram). To evaluate the suggested solutions we used various IR models, including Okapi, Divergence from Randomness (DFR), a statistical language model (LM) as well as the classical tf idf vector-space approach. We found that the Divergence from Randomness paradigm tend to propose better retrieval effectiveness than the Okapi, LM or tf idf models, the performance differences were however statistically significant only with the last two IR approaches. Ignoring the stemming reduces generally the MAP by more than 40%, and these differences are always significant. Finally, if our more aggressive stemmer tends to show the best performance, the differences in performance with a light stemmer are not statistically significant.     

Entrepreneurship and subjective vs objective institutional performance: A decade of US hospital data

Regional entrepreneurial activity can importantly affect the performance of local public service institutions. Yet, the literature explaining these relationships suffers from five methodological challenges: 1) inferred direction of influence; 2) unavailability of representative data; 3) blurring of objective and subjective performance; 4) a lack of longitudinal data; 5) and a lack of fine-grained regional data. This paper relies on a rich dataset from the ubiquitous institution of hospitals to explore these effects and overcome these challenges. We discriminate between objective and subjective institutional performance, suggesting that both performance categories deserve empirical attention, and may react differently to entrepreneurship. Our empirical approach applies econometric, mixed-effects regression models to a novel longitudinal dataset representing the entire hospital population in over 3000 U.S. counties between 2006 and 2018 merged with two sources of entrepreneurial activity at the county level. Interestingly, the results suggest divergent relationships: regional entrepreneurial activity positively affects objective institutional performance and also negatively affects subjective performance. Further, an institution's research designation attenuates the effect on subjective performance. These findings suggest that institutional performance is an often-overlooked byproduct of regional entrepreneurial activity and offer significant theoretical and policy implications.     

A joint learning approach with knowledge injection for zero-shot cross-lingual hate speech detection

Hate speech is an increasingly important societal issue in the era of digital communication. Hateful expressions often make use of figurative language and, although they represent, in some sense, the dark side of language, they are also often prime examples of creative use of language. While hate speech is a global phenomenon, current studies on automatic hate speech detection are typically framed in a monolingual setting. In this work, we explore hate speech detection in low-resource languages by transferring knowledge from a resource-rich language, English, in a zero-shot learning fashion. We experiment with traditional and recent neural architectures, and propose two joint-learning models, using different multilingual language representations to transfer knowledge between pairs of languages. We also evaluate the impact of additional knowledge in our experiment, by incorporating information from a multilingual lexicon of abusive words. The results show that our joint-learning models achieve the best performance on most languages. However, a simple approach that uses machine translation and a pre-trained English language model achieves a robust performance. In contrast, Multilingual BERT fails to obtain a good performance in cross-lingual hate speech detection. We also experimentally found that the external knowledge from a multilingual abusive lexicon is able to improve the models’ performance, specifically in detecting the positive class. The results of our experimental evaluation highlight a number of challenges and issues in this particular task. One of the main challenges is related to the issue of current benchmarks for hate speech detection, in particular how bias related to the topical focus in the datasets influences the classification performance. The insufficient ability of current multilingual language models to transfer knowledge between languages in the specific hate speech detection task also remain an open problem. However, our experimental evaluation and our qualitative analysis show how the explicit integration of linguistic knowledge from a structured abusive language lexicon helps to alleviate this issue.     

商业模式匹配跨界搜索战略对创新绩效的影响

商业模式关系到企业的成败及利润的高低,与其他战略的匹配有利于企业持续获得竞争优势。本文以123家企业为研究样本,采用多元回归技术检验了商业模式匹配跨界搜索战略对企业创新绩效的影响。实证结果表明,效率型商业模式匹配技术知识跨界搜索战略、新颖型商业模式匹配市场知识跨界搜索战略对创新绩效有积极地促进作用。同时,我们还发现,技术知识和市场知识分别积极调节效率型和新颖型商业模式对创新绩效的作用,进而促成匹配。