Universal Sentence Representation Learning with Conditional Masked Language Model

Developing the utilized intelligent systems is increasingly important to learn effective text representations, especially extract the sentence features. Numerous previous studies have been concentrated on the task of sentence representation learning based on deep learning approaches. However, the present approaches are mostly proposed with the single task or replied on the labeled corpus when learning the embedding of the sentences. In this paper, we assess the factors in learning sentence representation and propose an efficient unsupervised learning framework with multi-task learning (USR-MTL), in which various text learning tasks are merged into the unitized framework. With the syntactic and semantic features of sentences, three different factors to some extent are reflected in the task of the sentence representation learning that is the wording, or the ordering of the neighbored sentences of a target sentence in other words. Hence, we integrate the word-order learning task, word prediction task, and the sentence-order learning task into the proposed framework to attain meaningful sentence embeddings. Here, the process of sentence embedding learning is reformulated as a multi-task learning framework of the sentence-level task and the two word-level tasks. Moreover, the proposed framework is motivated by an unsupervised learning algorithm utilizing the unlabeled corpus. Based on the experimental results, our approach achieves the state-of-the-art performances on the downstream natural language processing tasks compared to the popular unsupervised representation learning techniques. The experiments on representation visualization and task analysis demonstrate the effectiveness of the tasks in the proposed framework in creating reasonable sentence representations proving the capacity of the proposed unsupervised multi-task framework for the sentence representation learning.

Natural Language Inference (NLI) is a crucial task in natural language processing that involves determining the relationship between two sentences, typically referred to as the premise and the hypothesis. However, traditional NLI models solely rely on the semantic information inherent in independent sentences and lack relevant situational visual information, which can hinder a complete understanding of the intended meaning of the sentences due to the ambiguity and vagueness of language. To address this challenge, we propose an innovative ScenaFuse adapter that simultaneously integrates large-scale pre-trained linguistic knowledge and relevant visual information for NLI tasks. Specifically, we first design an image-sentence interaction module to incorporate visuals into the attention mechanism of the pre-trained model, allowing the two modalities to interact comprehensively. Furthermore, we introduce an image-sentence fusion module that can adaptively integrate visual information from images and semantic information from sentences. By incorporating relevant visual information and leveraging linguistic knowledge, our approach bridges the gap between language and vision, leading to improved understanding and inference capabilities in NLI tasks. Extensive benchmark experiments demonstrate that our proposed ScenaFuse, a scenario-guided approach, consistently boosts NLI performance.

In this paper, we propose a simple, tricky method to improve sentence representation of unsupervised contrastive learning.Even though contrastive learning has achieved great performances in both visual representation learning (VRL) and sentence representation learning (SRL) fields, we focus on the fact that there is a gap between the characteristics and training dynamics of VRL and SRL.We first examine the role of temperature to bridge the gap between VRL and SRL, and find some temperaturedependent elements in SRL; i.e., a higher temperature causes overfitting of the uniformity while improving the alignment in the earlier phase of training.Then, we design a temperature cool-down technique based on this observation, which helps PLMs to be more suitable for contrastive learning via the preparation of uniform representation space.Our experimental results on widely-utilized benchmarks demonstrate the effectiveness and an extensibility of our method.Our code is publicly available at https://github.com/myngsooo/Cooldown.

Large pretrained masked language models have become state-of-the-art solutions for many NLP problems. While studies have shown that monolingual models produce better results than multilingual models, the training datasets must be sufficiently large. We trained a trilingual LitLat BERT-like model for Lithuanian, Latvian, and English, and a monolingual Est-RoBERTa model for Estonian. We evaluate their performance on four downstream tasks: named entity recognition, dependency parsing, part-of-speech tagging, and word analogy. To analyze the importance of focusing on a single language and the importance of a large training set, we compare created models with existing monolingual and multilingual BERT models for Estonian, Latvian, and Lithuanian. The results show that the newly created LitLat BERT and Est-RoBERTa models improve the results of existing models on all tested tasks in most situations.KeywordsNatural language processingBERTTransformersEstonianLatvianLithuanian

This paper presents UniBERT, a compact multilingual language model that uses an innovative training framework that integrates three components: masked language modeling, adversarial training, and knowledge distillation. Pre-trained on a meticulously curated Wikipedia corpus spanning 107 languages, UniBERT is designed to reduce the computational demands of large-scale models while maintaining competitive performance across various natural language processing tasks. Comprehensive evaluations on four tasks, named entity recognition, natural language inference, question answering, and semantic textual similarity, demonstrate that our multilingual training strategy, enhanced by an adversarial objective, significantly improves cross-lingual generalization. Specifically, UniBERT models show an average relative improvement of 7.72% over traditional baselines, which achieved an average relative improvement of only 1.12%, and statistical analysis confirms the significance of these gains (p value = 0.0184). This work highlights the benefits of combining adversarial training and knowledge distillation to build robust and scalable language models, thus advancing the field of multilingual and cross-lingual natural language processing.

Recent work has shown that predicate-argument label representations from semantic role labeling (SRL) can be concatenated with BERT representations to improve natural language understanding tasks such as natural language inference (NLI) and reading comprehension. Two natural questions that arise are whether infusing SRL representations with BERT 1) improves model performance and 2) increases the model’s linguistic awareness. This paper aims at answering both questions with a case study on the NLI task. We start by analyzing whether and how infusing SRL information helps BERT learn linguistic knowledge. We compare model performance on two benchmark datasets, SNLI and MNLI. We also conduct in-depth analysis on two probing datasets, Breaking NLI and HANS, which contain abundant examples where SRL information is expected to be helpful. We found that combining SRL representations with BERT representations does outperform BERT-only representations in general, with better awareness in lexical meaning and world knowledge but not in logic knowledge. We also found that infusing SRL information via predicate-wise concatenation with BERT word representations followed by an interaction layer is more effective than sentence-wise concatenation.

Natural Language Inference (NLI) datasets contain examples with highly ambiguous labels. While many research works do not pay much attention to this fact, several recent efforts have been made to acknowledge and embrace the existence of ambiguity, such as UNLI and ChaosNLI. In this paper, we explore the option of training directly on the estimated label distribution of the annotators in the NLI task, using a learning loss based on this ambiguity distribution instead of the gold-labels. We prepare AmbiNLI, a trial dataset obtained from readily available sources, and show it is possible to reduce ChaosNLI divergence scores when finetuning on this data, a promising first step towards learning how to capture linguistic ambiguity. Additionally, we show that training on the same amount of data but targeting the ambiguity distribution instead of gold-labels can result in models that achieve higher performance and learn better representations for downstream tasks.

Natural language inference (NLI) is an essential subtask in many natural language processing applications. It is a directional relationship from premise to hypothesis. A pair of texts is defined as entailed if a text infers its meaning from the other text. The NLI is also known as textual entailment recognition, and it recognizes entailed and contradictory sentences in various NLP systems like Question Answering, Summarization and Information retrieval systems. This paper describes the NLI problem attempted for a low resource Indian language Malayalam, the regional language of Kerala. More than 30 million people speak this language. The paper is about the Malayalam NLI dataset, named MaNLI dataset, and its application of NLI in Malayalam language using different models, namely Doc2Vec (paragraph vector), fastText, BERT (Bidirectional Encoder Representation from Transformers), and LASER (Language Agnostic Sentence Representation). Our work attempts NLI in two ways, as binary classification and as multiclass classification. For both the classifications, LASER outperformed the other techniques. For multiclass classification, NLI using LASER based sentence embedding technique outperformed the other techniques by a significant margin of 12% accuracy. There was also an accuracy improvement of 9% for LASER based NLI system for binary classification over the other techniques.

We examine the effect of domain-specific external knowledge variations on deep large scale language model performance. Recent work in enhancing BERT with external knowledge has been very popular, resulting in models such as ERNIE (Zhang et al., 2019a). Using the ERNIE architecture, we provide a detailed analysis on the types of knowledge that result in a performance increase on the Natural Language Inference (NLI) task, specifically on the Multi-Genre Natural Language Inference Corpus (MNLI). While ERNIE uses general TransE embeddings, we instead train domain-specific knowledge embeddings and insert this knowledge via an information fusion layer in the ERNIE architecture, allowing us to directly control and analyze knowledge input. Using several different knowledge training objectives, sources of knowledge, and knowledge ablations, we find a strong correlation between knowledge and classification labels within the same polarity, illustrating that knowledge polarity is an important feature in predicting entailment. We also perform classification change analysis across different knowledge variations to illustrate the importance of selecting appropriate knowledge input regarding content and polarity, and show representative examples of these changes.

Understanding the biology underpinning the natural regeneration of plant species in order to make plans for effective reforestation is a complex task. This can be aided by providing access to databases that contain long-term and wide-scale geographical information on species distribution, habitat, and reproduction. Although there exists widely-used biodiversity databases that contain structured information on species and their occurrences, such as the Global Biodiversity Information Facility (GBIF) and the Atlas of Living Australia (ALA), the bulk of knowledge about biodiversity still remains embedded in textual documents. Unstructured information can be made more accessible and useful for large-scale studies if there are tools and services that automatically extract meaningful information from text and store it in structured formats, e.g., open biodiversity databases, ready to be consumed for analysis (Thessen et al. 2022). We aim to enrich biodiversity occurrence databases with information on species reproductive condition and habitat, derived from text. In previous work, we developed unsupervised approaches to extract related habitats and their locations, and related reproductive condition and temporal expressions (Gabud and Batista-Navarro 2018). We built a new unsupervised hybrid approach for relation extraction (RE), which is a combination of classical rule-based pattern-matching methods and transformer-based language models that framed our RE task as a natural language inference (NLI) task. Using our hybrid approach for RE, we were able to extract related biodiversity entities from text even without a large training dataset. In this work, we implement an information extraction (IE) pipeline comprised of a named entity recognition (NER) tool and our hybrid relation extraction (RE) tool. The NER tool is a transformer-based language model that was pretrained on scientific text and then fine-tuned using COPIOUS (Conserving Philippine Biodiversity by Understanding big data; Nguyen et al. 2019), a gold standard corpus containing named entities relevant to species occurrence. We applied the NER tool to automatically annotate geographical location, temporal expression and habitat information contained within sentences. A dictionary-based approach is then used to identify mentions of reproductive conditions in text (e.g., phrases such as "fruited heavily" and "mass flowering"). We then use our hybrid RE tool to extract reproductive condition - temporal expression and habitat - geographical location entity pairs. We test our IE pipeline on the forestry compendium available in the CABI Digital Library (Centre for Agricultural and Biosciences International), and show that our work enables the enrichment of descriptive information on reproductive and habitat conditions of species. This work is a step towards enhancing a biodiversity database with the inclusion of habitat and reproductive condition information extracted from text.

Knowledge plays an essential role in inference, but is less explored by previous works in the Natural Language Inference (NLI) task. Although traditional neural models obtained impressive performance on standard benchmarks, they often encounter performance degradation when being applied to knowledge-intensive domains like medicine and science. To address this problem and further fill the knowledge gap, we present a simple Evidence-Based Inference Model (EBIM) to integrate clues collected from knowledge graphs as evidence for inference. To effectively incorporate the knowledge, we propose an efficient approach to retrieve paths in knowledge graphs as clues and then prune them to avoid involving too much irrelevant noise. In addition, we design a specialized CNN-based encoder according to the structure of clues to better model them. Experiments show that the proposed encoder outperforms strong baselines, and our EBIM model outperforms other knowledge-based approaches on the SciTail benchmark and establishes a new state-of-the-art performance on the MedNLI dataset.

Machine reading comprehension has been an interesting and challenging task in recent years, with the purpose of extracting useful information from texts. To attain the computer ability to understand the reading text and answer relevant information, we introduce ViMMRC 2.0 — an extension of the previous ViMMRC for the task of multiple-choice reading comprehension in Vietnamese Textbooks which contain the reading articles for students from Grade 1 to Grade 12. This dataset has 699 reading passages which are prose and poems, and 5273 questions. The questions in the new dataset are not fixed with four options as in the previous version. Moreover, the difficulty of questions is increased, which challenges the models to find the correct choice. The computer must understand the whole context of the reading passage, the question, and the content of each choice to extract the right answers. Hence, we propose a multi-stage approach that combines the multi-step attention network (MAN) with the natural language inference (NLI) task to enhance the performance of the reading comprehension model. Then, we compare the proposed methodology with the baseline BERTology models on the new dataset and the ViMMRC 1.0. From the results of the error analysis, we found that the challenge of the reading comprehension models is understanding the implicit context in texts and linking them together in order to find the correct answers. Finally, we hope our new dataset will motivate further research to enhance the ability of computers to understand the Vietnamese language.

Cross-lingual natural language processing relies on translation, either by humans or machines, at different levels, from translating training data to translating test sets. However, compared to original texts in the same language, translations possess distinct qualities referred to as translationese. Previous research has shown that these translation artifacts influence the performance of a variety of cross-lingual tasks. In this work, we propose a novel approach to reducing translationese by extending an established bias-removal technique. We use the Iterative Null-space Projection (INLP) algorithm, and show by measuring classification accuracy before and after debiasing, that translationese is reduced at both sentence and word level. We evaluate the utility of debiasing translationese on a natural language inference (NLI) task, and show that by reducing this bias, NLI accuracy improves. To the best of our knowledge, this is the first study to debias translationese as represented in latent embedding space.

In recent years, methods that selectively fine-tune or reduce the number of layers in large language models (LLMs) have garnered attention as an efficient alternative to traditional fine-tuning, where all layers are trained. In this study, we revisit the classical concept of layer freezing and propose a simple, effective strategy that selectively fine-tunes only a portion of transformer layers. We show that freezing the bottom 25% or 50% of layers in small-scale LLMs with sub-3 billion parameters yields significant improvements in memory efficiency and training speed while maintaining, or even surpassing, the performance of full fine-tuning and Low-Rank Adaptation (LoRA). Through experiments on Natural Language Inference (NLI) tasks using LLMs with fewer than 3 billion parameters, our approach achieves up to 50% memory savings and 30% faster training. Notably, our method does not require architectural modifications or additional parameters, making it particularly suitable for resource-constrained environments.

Sentence representation learning benefits from data augmentation strategies to improve model performance and generalization, yet existing approaches often encounter issues such as semantic inconsistencies and feature suppression.To address these limitations, we propose a method for generating Syntactically Aligned Negative (SAN) samples through a semantic importance-aware Masked Language Model (MLM) approach.Our method quantifies semantic contributions of individual words to produce negative samples that have substantial textual overlap with the original sentences while conveying different meanings.We further introduce Hierarchical-InfoNCE (HiNCE), a novel contrastive learning objective employing differential temperature weighting to optimize the utilization of both in-batch and syntactically aligned negative samples.Extensive evaluations across seven semantic textual similarity benchmarks demonstrate consistent improvements over state-of-the-art models 1 .