The softmax operator is one of the most important functions in machine learning models. When applying neural networks to multi-category classification, the correlations among different categories are often ignored. For example, in text generation, a language model makes a choice of each new word based only on the former selection of its context. In this scenario, the link statistics information of concurrent words based on a corpus (an analogy of the natural way of expression) is also valuable in choosing the next word, which can help to improve the sentence's fluency and smoothness. To fully explore such important information, we propose a graph softmax function for text generation. It is expected that the final classification result would be dominated by both the language model and graphical text relationships among words. We use a graph total variation term to regularize softmax so as to incorporate the concurrent relationship into the language model. The total variation of the generated words should be small locally. We apply the proposed graph softmax to GPT2 for the text generation task. Experimental results demonstrate that the proposed graph softmax achieves better BLEU and perplexity than softmax. Human testers can also easily distinguish the text generated by the graph softmax or softmax.
In this paper, we present CopyCat2 (CC2), a novel model capable of: a) synthesizing speech with different speaker identities, b) generating speech with expressive and contextually appropriate prosody, and c) transferring prosody at fine-grained level between any pair of seen speakers. We do this by activating distinct parts of the network for different tasks. We train our model using a novel approach to two-stage training. In Stage I, the model learns speaker-independent word-level prosody representations from speech which it uses for many-to-many fine-grained prosody transfer. In Stage II, we learn to predict these prosody representations using the contextual information available in text, thereby, enabling multi-speaker TTS with contextually appropriate prosody. We compare CC2 to two strong baselines, one in TTS with contextually appropriate prosody, and one in fine-grained prosody transfer. CC2 reduces the gap in naturalness between our baseline and copy-synthesised speech by $22.79\%$. In fine-grained prosody transfer evaluations, it obtains a relative improvement of $33.15\%$ in target speaker similarity.
Key information extraction (KIE) from document images requires understanding the contextual and spatial semantics of texts in two-dimensional (2D) space. Many recent studies try to solve the task by developing pre-training language models focusing on combining visual features from document images with texts and their layout. On the other hand, this paper tackles the problem by going back to the basic: effective combination of text and layout. Specifically, we propose a pre-trained language model, named BROS (BERT Relying On Spatiality), that encodes relative positions of texts in 2D space and learns from unlabeled documents with area-masking strategy. With this optimized training scheme for understanding texts in 2D space, BROS shows comparable or better performance compared to previous methods on four KIE benchmarks (FUNSD, SROIE*, CORD, and SciTSR) without relying on visual features. This paper also reveals two real-world challenges in KIE tasks--(1) minimizing the error from incorrect text ordering and (2) efficient learning from fewer downstream examples--and demonstrates the superiority of BROS over previous methods. Our code will be open to the public.
Stylistic analysis of text is a key task in research areas ranging from authorship attribution to forensic analysis and personality profiling. The existing approaches for stylistic analysis are plagued by issues like topic influence, lack of discriminability for large number of authors and the requirement for large amounts of diverse data. In this paper, the source of these issues are identified along with the necessity for a cognitive perspective on authorial style in addressing them. A novel feature representation, called Trajectory-based Style Estimation (TraSE), is introduced to support this purpose. Authorship attribution experiments with over 27,000 authors and 1.4 million samples in a cross-domain scenario resulted in 90% attribution accuracy suggesting that the feature representation is immune to such negative influences and an excellent candidate for stylistic analysis. Finally, a qualitative analysis is performed on TraSE using physical human characteristics, like age, to validate its claim on capturing cognitive traits.
Drawing causal conclusions from observational data requires making assumptions about the true data-generating process. Causal inference research typically considers low-dimensional data, such as categorical or numerical fields in structured medical records. High-dimensional and unstructured data such as natural language complicates the evaluation of causal inference methods; such evaluations rely on synthetic datasets with known causal effects. Models for natural language generation have been widely studied and perform well empirically. However, existing methods not immediately applicable to producing synthetic datasets for causal evaluations, as they do not allow for quantifying a causal effect on the text itself. In this work, we develop a framework for adapting existing generation models to produce synthetic text datasets with known causal effects. We use this framework to perform an empirical comparison of four recently-proposed methods for estimating causal effects from text data. We release our code and synthetic datasets.
Data is published on the web over time in great volumes, but majority of the data is unstructured, making it hard to understand and difficult to interpret. Information Extraction (IE) methods extract structured information from unstructured data. One of the challenging IE tasks is Event Extraction (EE) which seeks to derive information about specific incidents and their actors from the text. EE is useful in many domains such as building a knowledge base, information retrieval, summarization and online monitoring systems. In the past decades, some event ontologies like ACE, CAMEO and ICEWS were developed to define event forms, actors and dimensions of events observed in the text. These event ontologies still have some shortcomings such as covering only a few topics like political events, having inflexible structure in defining argument roles, lack of analytical dimensions, and complexity in choosing event sub-types. To address these concerns, we propose an event ontology, namely COfEE, that incorporates both expert domain knowledge, previous ontologies and a data-driven approach for identifying events from text. COfEE consists of two hierarchy levels (event types and event sub-types) that include new categories relating to environmental issues, cyberspace, criminal activity and natural disasters which need to be monitored instantly. Also, dynamic roles according to each event sub-type are defined to capture various dimensions of events. In a follow-up experiment, the proposed ontology is evaluated on Wikipedia events, and it is shown to be general and comprehensive. Moreover, in order to facilitate the preparation of gold-standard data for event extraction, a language-independent online tool is presented based on COfEE.
This work introduces Focused-Variation Network (FVN), a novel model to control language generation. The main problems in previous controlled language generation models range from the difficulty of generating text according to the given attributes, to the lack of diversity of the generated texts. FVN addresses these issues by learning disjoint discrete latent spaces for each attribute inside codebooks, which allows for both controllability and diversity, while at the same time generating fluent text. We evaluate FVN on two text generation datasets with annotated content and style, and show state-of-the-art performance as assessed by automatic and human evaluations.
In this work we evaluate the applicability of an ensemble of population models and machine learning models to predict the near future evolution of the COVID-19 pandemic, with a particular use case in Spain. We rely solely in open and public datasets, fusing incidence, vaccination, human mobility and weather data to feed our machine learning models (Random Forest, Gradient Boosting, k-Nearest Neighbours and Kernel Ridge Regression). We use the incidence data to adjust classic population models (Gompertz, Logistic, Richards, Bertalanffy) in order to be able to better capture the trend of the data. We then ensemble these two families of models in order to obtain a more robust and accurate prediction. Furthermore, we have observed an improvement in the predictions obtained with machine learning models as we add new features (vaccines, mobility, climatic conditions), analyzing the importance of each of them using Shapley Additive Explanation values. As in any other modelling work, data and predictions quality have several limitations and therefore they must be seen from a critical standpoint, as we discuss in the text. Our work concludes that the ensemble use of these models improves the individual predictions (using only machine learning models or only population models) and can be applied, with caution, in cases when compartmental models cannot be utilized due to the lack of relevant data.
In the current IT world, developers write code while system operators run the code mostly as a black box. The connection between both worlds is typically established with log messages: the developer provides hints to the (unknown) operator, where the cause of an occurred issue is, and vice versa, the operator can report bugs during operation. To fulfil this purpose, developers write log instructions that are structured text commonly composed of a log level (e.g., "info", "error"), static text ("IP {} cannot be reached"), and dynamic variables (e.g. IP {}). However, as opposed to well-adopted coding practices, there are no widely adopted guidelines on how to write log instructions with good quality properties. For example, a developer may assign a high log level (e.g., "error") for a trivial event that can confuse the operator and increase maintenance costs. Or the static text can be insufficient to hint at a specific issue. In this paper, we address the problem of log quality assessment and provide the first step towards its automation. We start with an in-depth analysis of quality log instruction properties in nine software systems and identify two quality properties: 1) correct log level assignment assessing the correctness of the log level, and 2) sufficient linguistic structure assessing the minimal richness of the static text necessary for verbose event description. Based on these findings, we developed a data-driven approach that adapts deep learning methods for each of the two properties. An extensive evaluation on large-scale open-source systems shows that our approach correctly assesses log level assignments with an accuracy of 0.88, and the sufficient linguistic structure with an F1 score of 0.99, outperforming the baselines. Our study shows the potential of the data-driven methods in assessing instructions quality and aid developers in comprehending and writing better code.
The academic literature of social sciences is the literature that records human civilization and studies human social problems. With the large-scale growth of this literature, ways to quickly find existing research on relevant issues have become an urgent demand for researchers. Previous studies, such as SciBERT, have shown that pre-training using domain-specific texts can improve the performance of natural language processing tasks in those fields. However, there is no pre-trained language model for social sciences, so this paper proposes a pre-trained model on many abstracts published in the Social Science Citation Index (SSCI) journals. The models, which are available on Github (https://github.com/S-T-Full-Text-Knowledge-Mining/SSCI-BERT), show excellent performance on discipline classification and abstract structure-function recognition tasks with the social sciences literature.