Constructing and analyzing domain-specific language model for financial text mining

The application of natural language processing (NLP) to financial fields is advancing with an increase in the number of available financial documents. Transformer-based models such as Bidirectional Encoder Representations from Transformers (BERT) have been successful in NLP in recent years. These cutting-edge models have been adapted to the financial domain by applying financial corpora to existing pre-trained models and by pre-training with the financial corpora from scratch. In Japanese, by contrast, financial terminology cannot be applied from a general vocabulary without further processing. In this study, we construct language models suitable for the financial domain. Furthermore, we compare methods for adapting language models to the financial domain, such as pre-training methods and vocabulary adaptation. We confirm that the adaptation of a pre-training corpus and tokenizer vocabulary based on a corpus of financial text is effective in several downstream financial tasks. No significant difference is observed between pre-training with the financial corpus and continuous pre-training from the general language model with the financial corpus. We have released our source code and pre-trained models.     

BERT-SMAP: Paying attention to Essential Terms in passage ranking beyond BERT

Passage ranking has attracted considerable attention due to its importance in information retrieval (IR) and question answering (QA). Prior works have shown that pre-trained language models (e.g. BERT) can improve ranking performance. However, these simple BERT-based methods tend to focus on passage terms that exactly match the question, which makes them easily fooled by the overlapping but irrelevant (distracting) passages. To solve this problem, we propose a self-matching attention-pooling mechanism (SMAP) to highlight the Essential Terms in the question-passage pairs. Further, we propose a hybrid passage ranking architecture, called BERT-SMAP, which combines SMAP with BERT to more effectively identify distracting passages and downplay their influence. BERT-SMAP uses the representations obtained through SMAP to enhance BERT’s classification mechanism as an interaction-focused neural ranker, and as the inputs of a matching function. Experimental results on three evaluation datasets show that our model outperforms the previous best BERT${}_{base}$-based approaches, and is comparable to the state-of-the-art method that utilizes a much stronger pre-trained language model.     

Unified benchmark for zero-shot Turkish text classification

Effective learning schemes such as fine-tuning, zero-shot, and few-shot learning, have been widely used to obtain considerable performance with only a handful of annotated training data. In this paper, we presented a unified benchmark to facilitate the problem of zero-shot text classification in Turkish. For this purpose, we evaluated three methods, namely, Natural Language Inference, Next Sentence Prediction and our proposed model that is based on Masked Language Modeling and pre-trained word embeddings on nine Turkish datasets for three main categories: topic, sentiment, and emotion. We used pre-trained Turkish monolingual and multilingual transformer models which can be listed as BERT, ConvBERT, DistilBERT and mBERT. The results showed that ConvBERT with the NLI method yields the best results with 79% and outperforms previously used multilingual XLM-RoBERTa model by 19.6%. The study contributes to the literature using different and unattempted transformer models for Turkish and showing improvement of zero-shot text classification performance for monolingual models over multilingual models.     

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.     

面向网络虚假医疗信息的识别模型构建研究——一种基于预训练的BERT模型

【 目的/意义】解决获取虚假网络医疗信息数据集时专业知识不足的问题,帮助在小样本领域构建虚假网络 医疗信息识别模型。【方法/过程】本文提出一种基于权威辟谣信息转化提取构建网络虚假医疗信息数据集的思路, 并依次构建传统机器学习模型、CNN模型和BERT模型进行分类识别。【结果/结论】结果表明,基于辟谣信息能够 实现以较低成本、不依赖专家标注构建虚假医疗信息数据集。通过对比实验发现,基于微博数据预训练的 BERT 模型准确率为 95.91%,F1值为 94.57%,相比于传统机器学习模型和 CNN模型提升分别接近 6%和 4%,表明本文构 建的基于预训练的BERT模型在网络虚假医疗信息识别任务上取得了更好的效果。【创新/局限】本文提出的方法能 以较低成本建立专业领域的虚假信息数据集,所构建的BERT虚假医疗信息识别模型在小样本领域也具有实用价 值,但在数据集规模、深度学习模型对比、模型性能评价指标等方面还有待拓展与延伸。     

AsU-OSum: Aspect-augmented unsupervised opinion summarization

Opinion summarization can facilitate user’s decision-making by mining the salient review information. However, due to the lack of sufficient annotated data, most of the early works are based on extractive methods, which restricts the performance of opinion summarization. In this work, we aim to improve the informativeness of opinion summarization to provide better guidance to users. We consider the setting with only reviews without corresponding summaries, and propose an aspect-augmented model for unsupervised abstractive opinion summarization, denoted as AsU-OSum. We first employ an aspect-based sentiment analysis system to extract opinion phrases from reviews. Then, we construct a heterogeneous graph consisting of reviews and opinion clusters as nodes, which is used to enhance the Transformer-based encoder–decoder framework. Furthermore, we design a novel cascaded attention mechanism to prompt the decoder to pay more attention to the aspects that are more likely to appear in summary. During training, we introduce a sentiment accuracy reward that further enhances the learning ability of our model. We conduct comprehensive experiments on the Yelp, Amazon, and Rotten Tomatoes datasets. Automatic evaluation results show that our model is competitive and performs better than the state-of-the-art (SOTA) models on some ROUGE metrics. Human evaluation results further verify that our model can generate more informative summaries and reduce redundancy.     

Speciesist language and nonhuman animal bias in English Masked Language Models

Warning: This paper contains examples of offensive language, including insulting or objectifying expressions.Various existing studies have analyzed what social biases are inherited by NLP models. These biases may directly or indirectly harm people, therefore previous studies have focused only on human attributes. However, until recently no research on social biases in NLP regarding nonhumans existed. In this paper,¹ we analyze biases to nonhuman animals, i.e. speciesist bias, inherent in English Masked Language Models such as BERT. We analyzed speciesist bias against 46 animal names using template-based and corpus-extracted sentences containing speciesist (or non-speciesist) language. We found that pre-trained masked language models tend to associate harmful words with nonhuman animals and have a bias toward using speciesist language for some nonhuman animal names. Our code for reproducing the experiments will be made available on GitHub.²     

Text summarization using Wikipedia

Automatic text summarization has been an active field of research for many years. Several approaches have been proposed, ranging from simple position and word-frequency methods, to learning and graph based algorithms. The advent of human-generated knowledge bases like Wikipedia offer a further possibility in text summarization – they can be used to understand the input text in terms of salient concepts from the knowledge base. In this paper, we study a novel approach that leverages Wikipedia in conjunction with graph-based ranking. Our approach is to first construct a bipartite sentence–concept graph, and then rank the input sentences using iterative updates on this graph. We consider several models for the bipartite graph, and derive convergence properties under each model. Then, we take up personalized and query-focused summarization, where the sentence ranks additionally depend on user interests and queries, respectively. Finally, we present a Wikipedia-based multi-document summarization algorithm. An important feature of the proposed algorithms is that they enable real-time incremental summarization – users can first view an initial summary, and then request additional content if interested. We evaluate the performance of our proposed summarizer using the ROUGE metric, and the results show that leveraging Wikipedia can significantly improve summary quality. We also present results from a user study, which suggests that using incremental summarization can help in better understanding news articles.     

Meta-prompt based learning for low-resource false information detection

The wide spread of false information has detrimental effects on society, and false information detection has received wide attention. When new domains appear, the relevant labeled data is scarce, which brings severe challenges to the detection. Previous work mainly leverages additional data or domain adaptation technology to assist detection. The former would lead to a severe data burden; the latter underutilizes the pre-trained language model because there is a gap between the downstream task and the pre-training task, which is also inefficient for model storage because it needs to store a set of parameters for each domain. To this end, we propose a meta-prompt based learning (MAP) framework for low-resource false information detection. We excavate the potential of pre-trained language models by transforming the detection tasks into pre-training tasks by constructing template. To solve the problem of the randomly initialized template hindering excavation performance, we learn optimal initialized parameters by borrowing the benefit of meta learning in fast parameter training. The combination of meta learning and prompt learning for the detection is non-trivial: Constructing meta tasks to get initialized parameters suitable for different domains and setting up the prompt model’s verbalizer for classification in the noisy low-resource scenario are challenging. For the former, we propose a multi-domain meta task construction method to learn domain-invariant meta knowledge. For the latter, we propose a prototype verbalizer to summarize category information and design a noise-resistant prototyping strategy to reduce the influence of noise data. Extensive experiments on real-world data demonstrate the superiority of the MAP in new domains of false information detection.     

Graph-based Arabic text semantic representation

Semantic representation reflects the meaning of the text as it may be understood by humans. Thus, it contributes to facilitating various automated language processing applications. Although semantic representation is very useful for several applications, a few models were proposed for the Arabic language. In that context, this paper proposes a graph-based semantic representation model for Arabic text. The proposed model aims to extract the semantic relations between Arabic words. Several tools and concepts have been employed such as dependency relations, part-of-speech tags, name entities, patterns, and Arabic language predefined linguistic rules. The core idea of the proposed model is to represent the meaning of Arabic sentences as a rooted acyclic graph. Textual entailment recognition challenge is considered in order to evaluate the ability of the proposed model to enhance other Arabic NLP applications. The experiments have been conducted using a benchmark Arabic textual entailment dataset, namely, ArbTED. The results proved that the proposed graph-based model is able to enhance the performance of the textual entailment recognition task in comparison to other baseline models. On average, the proposed model achieved 8.6%, 30.2%, 5.3% and 16.2% improvement in terms of accuracy, recall, precision, and F-score results, respectively.     

An ensemble model for classifying idioms and literal texts using BERT and RoBERTa

An idiom is a common phrase that means something other than its literal meaning. Detecting idioms automatically is a serious challenge in natural language processing (NLP) domain applications like information retrieval (IR), machine translation and chatbot. Automatic detection of Idioms plays an important role in all these applications. A fundamental NLP task is text classification, which categorizes text into structured categories known as text labeling or categorization. This paper deals with idiom identification as a text classification task. Pre-trained deep learning models have been used for several text classification tasks; though models like BERT and RoBERTa have not been exclusively used for idiom and literal classification. We propose a predictive ensemble model to classify idioms and literals using BERT and RoBERTa, fine-tuned with the TroFi dataset. The model is tested with a newly created in house dataset of idioms and literal expressions, numbering 1470 in all, and annotated by domain experts. Our model outperforms the baseline models in terms of the metrics considered, such as F-score and accuracy, with a 2% improvement in accuracy.     

基于深度学习的细粒度命名实体识别研究——以番茄病虫害为例

【 目的/意义】开展面向领域的细粒度命名实体识别研究对于提升文本挖掘精度具有重要的意义,本文以番 茄病虫害命名实体为例,探索采用深度学习技术实现面向领域的细粒度命名实体识别研究方法。【目的/意义】文章 以电子书、论文、网页作为数据源,选择品种、病虫害、症状、时间、部位、防治药剂六类实体进行标注,利用BERT和 CBOW 预训练字向量分别输入 BiLSTM-CRF 模型训练,并在识别后补充规则控制实体的边界。【结果/结论】 BERT预训练的字向量和BiLSTM-CRF结合,在补充规则控制后F值达到了81.03%,优于其它模型,在番茄病虫害 领域的实体识别中具有较好的效果。【创新/局限】BERT预训练的字向量可以有效降低番茄病虫害领域实体因分 词错误带来的影响,针对不同实体的特点,补充规则可以有效控制实体边界,提高识别准确率。但本文的规则补充 仅在测试阶段,并没有加入训练过程,整体的准确率还有待提高。     

Summarizing court decisions

In the field of law there is an absolute need for summarizing the texts of court decisions in order to make the content of the cases easily accessible for legal professionals. During the SALOMON and MOSAIC² projects we investigated the summarization and retrieval of legal cases. This article presents some of the main findings while integrating the research results of experiments on legal document summarization by other research groups. In addition, we propose novel avenues of research for automatic text summarization, which we currently exploit when summarizing court decisions in the ACILA³ project. Techniques for automated concept learning and argument recognition are here the most challenging.     

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.     

The use of domain-specific concepts in biomedical text summarization

Text summarization is a method for data reduction. The use of text summarization enables users to reduce the amount of text that must be read while still assimilating the core information. The data reduction offered by text summarization is particularly useful in the biomedical domain, where physicians must continuously find clinical trial study information to incorporate into their patient treatment efforts. Such efforts are often hampered by the high-volume of publications. This paper presents two independent methods (BioChain and FreqDist) for identifying salient sentences in biomedical texts using concepts derived from domain-specific resources. Our semantic-based method (BioChain) is effective at identifying thematic sentences, while our frequency-distribution method (FreqDist) removes information redundancy. The two methods are then combined to form a hybrid method (ChainFreq). An evaluation of each method is performed using the ROUGE system to compare system-generated summaries against a set of manually-generated summaries. The BioChain and FreqDist methods outperform some common summarization systems, while the ChainFreq method improves upon the base approaches. Our work shows that the best performance is achieved when the two methods are combined. The paper also presents a brief physician’s evaluation of three randomly-selected papers from an evaluation corpus to show that the author’s abstract does not always reflect the entire contents of the full-text.     

Candidate sentence selection for extractive text summarization

Text summarization is a process of generating a brief version of documents by preserving the fundamental information of documents as much as possible. Although most of the text summarization research has been focused on supervised learning solutions, there are a few datasets indeed generated for summarization tasks, and most of the existing summarization datasets do not have human-generated goal summaries which are vital for both summary generation and evaluation. Therefore, a new dataset was presented for abstractive and extractive summarization tasks in this study. This dataset contains academic publications, the abstracts written by the authors, and extracts in two sizes, which were generated by human readers in this research. Then, the resulting extracts were evaluated to ensure the validity of the human extract production process. Moreover, the extractive summarization problem was reinvestigated on the proposed summarization dataset. Here the main point taken into account was to analyze the feature vector to generate more informative summaries. To that end, a comprehensive syntactic feature space was generated for the proposed dataset, and the impact of these features on the informativeness of the resulting summary was investigated. Besides, the summarization capability of semantic features was experienced by using GloVe and word2vec embeddings. Finally, the use of ensembled feature space, which corresponds to the joint use of syntactic and semantic features, was proposed on a long short-term memory-based neural network model. ROUGE metrics evaluated the model summaries, and the results of these evaluations showed that the use of the proposed ensemble feature space remarkably improved the single-use of syntactic or semantic features. Additionally, the resulting summaries of the proposed approach on ensembled features prominently outperformed or provided comparable performance than summaries obtained by state-of-the-art models for extractive summarization.     

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.