The rapid deployment of artificial intelligence (AI) models demands a thorough investigation of biases and risks inherent in these models to understand their impact on individuals and society. This study extends the focus of bias evaluation in extant work by examining bias against social stigmas on a large scale. It focuses on 93 stigmatized groups in the United States, including a wide range of conditions related to disease, disability, drug use, mental illness, religion, sexuality, socioeconomic status, and other relevant factors. We investigate bias against these groups in English pre-trained Masked Language Models (MLMs) and their downstream sentiment classification tasks. To evaluate the presence of bias against 93 stigmatized conditions, we identify 29 non-stigmatized conditions to conduct a comparative analysis. Building upon a psychology scale of social rejection, the Social Distance Scale, we prompt six MLMs: RoBERTa-base, RoBERTa-large, XLNet-large, BERTweet-base, BERTweet-large, and DistilBERT. We use human annotations to analyze the predicted words from these models, with which we measure the extent of bias against stigmatized groups. When prompts include stigmatized conditions, the probability of MLMs predicting negative words is approximately 20 percent higher than when prompts have non-stigmatized conditions. In the sentiment classification tasks, when sentences include stigmatized conditions related to diseases, disability, education, and mental illness, they are more likely to be classified as negative. We also observe a strong correlation between bias in MLMs and their downstream sentiment classifiers (r =0.79). The evidence indicates that MLMs and their downstream sentiment classification tasks exhibit biases against socially stigmatized groups.
Although large language models (LLMs) has shown great performance on natural language processing (NLP) in the financial domain, there are no publicly available financial tailtored LLMs, instruction tuning datasets, and evaluation benchmarks, which is critical for continually pushing forward the open-source development of financial artificial intelligence (AI). This paper introduces PIXIU, a comprehensive framework including the first financial LLM based on fine-tuning LLaMA with instruction data, the first instruction data with 136K data samples to support the fine-tuning, and an evaluation benchmark with 5 tasks and 9 datasets. We first construct the large-scale multi-task instruction data considering a variety of financial tasks, financial document types, and financial data modalities. We then propose a financial LLM called FinMA by fine-tuning LLaMA with the constructed dataset to be able to follow instructions for various financial tasks. To support the evaluation of financial LLMs, we propose a standardized benchmark that covers a set of critical financial tasks, including five financial NLP tasks and one financial prediction task. With this benchmark, we conduct a detailed analysis of FinMA and several existing LLMs, uncovering their strengths and weaknesses in handling critical financial tasks. The model, datasets, benchmark, and experimental results are open-sourced to facilitate future research in financial AI.
NLP models often degrade in performance when real world data distributions differ markedly from training data. However, existing dataset drift metrics in NLP have generally not considered specific dimensions of linguistic drift that affect model performance, and they have not been validated in their ability to predict model performance at the individual example level, where such metrics are often used in practice. In this paper, we propose three dimensions of linguistic dataset drift: vocabulary, structural, and semantic drift. These dimensions correspond to content word frequency divergences, syntactic divergences, and meaning changes not captured by word frequencies (e.g. lexical semantic change). We propose interpretable metrics for all three drift dimensions, and we modify past performance prediction methods to predict model performance at both the example and dataset level for English sentiment classification and natural language inference. We find that our drift metrics are more effective than previous metrics at predicting out-of-domain model accuracies (mean 16.8% root mean square error decrease), particularly when compared to popular fine-tuned embedding distances (mean 47.7% error decrease). Fine-tuned embedding distances are much more effective at ranking individual examples by expected performance, but decomposing into vocabulary, structural, and semantic drift produces the best example rankings of all considered model-agnostic drift metrics (mean 6.7% ROC AUC increase).
Recent studies have demonstrated that natural-language prompts can help to leverage the knowledge learned by pre-trained language models for the binary sentence-level sentiment classification task. Specifically, these methods utilize few-shot learning settings to fine-tune the sentiment classification model using manual or automatically generated prompts. However, the performance of these methods is sensitive to the perturbations of the utilized prompts. Furthermore, these methods depend on a few labeled instances for automatic prompt generation and prompt ranking. This study aims to find high-quality prompts for the given task in a zero-shot setting. Given a base prompt, our proposed approach automatically generates multiple prompts similar to the base prompt employing positional, reasoning, and paraphrasing techniques and then ranks the prompts using a novel metric. We empirically demonstrate that the top-ranked prompts are high-quality and significantly outperform the base prompt and the prompts generated using few-shot learning for the binary sentence-level sentiment classification task.
Sentiment analysis (SA) has been a long-standing research area in natural language processing. It can offer rich insights into human sentiments and opinions and has thus seen considerable interest from both academia and industry. With the advent of large language models (LLMs) such as ChatGPT, there is a great potential for their employment on SA problems. However, the extent to which existing LLMs can be leveraged for different sentiment analysis tasks remains unclear. This paper aims to provide a comprehensive investigation into the capabilities of LLMs in performing various sentiment analysis tasks, from conventional sentiment classification to aspect-based sentiment analysis and multifaceted analysis of subjective texts. We evaluate performance across 13 tasks on 26 datasets and compare the results against small language models (SLMs) trained on domain-specific datasets. Our study reveals that while LLMs demonstrate satisfactory performance in simpler tasks, they lag behind in more complex tasks requiring deeper understanding or structured sentiment information. However, LLMs significantly outperform SLMs in few-shot learning settings, suggesting their potential when annotation resources are limited. We also highlight the limitations of current evaluation practices in assessing LLMs' SA abilities and propose a novel benchmark, \textsc{SentiEval}, for a more comprehensive and realistic evaluation. Data and code during our investigations are available at \url{https://github.com/DAMO-NLP-SG/LLM-Sentiment}.
Weakly-supervised text classification trains a classifier using the label name of each target class as the only supervision, which largely reduces human annotation efforts. Most existing methods first use the label names as static keyword-based features to generate pseudo labels, which are then used for final classifier training. While reasonable, such a commonly adopted framework suffers from two limitations: (1) keywords can have different meanings in different contexts and some text may not have any keyword, so keyword matching can induce noisy and inadequate pseudo labels; (2) the errors made in the pseudo label generation stage will directly propagate to the classifier training stage without a chance of being corrected. In this paper, we propose a new method, PIEClass, consisting of two modules: (1) a pseudo label acquisition module that uses zero-shot prompting of pre-trained language models (PLM) to get pseudo labels based on contextualized text understanding beyond static keyword matching, and (2) a noise-robust iterative ensemble training module that iteratively trains classifiers and updates pseudo labels by utilizing two PLM fine-tuning methods that regularize each other. Extensive experiments show that PIEClass achieves overall better performance than existing strong baselines on seven benchmark datasets and even achieves similar performance to fully-supervised classifiers on sentiment classification tasks.
Africa has over 2000 indigenous languages but they are under-represented in NLP research due to lack of datasets. In recent years, there have been progress in developing labeled corpora for African languages. However, they are often available in a single domain and may not generalize to other domains. In this paper, we focus on the task of sentiment classification for cross domain adaptation. We create a new dataset, NollySenti - based on the Nollywood movie reviews for five languages widely spoken in Nigeria (English, Hausa, Igbo, Nigerian-Pidgin, and Yoruba. We provide an extensive empirical evaluation using classical machine learning methods and pre-trained language models. Leveraging transfer learning, we compare the performance of cross-domain adaptation from Twitter domain, and cross-lingual adaptation from English language. Our evaluation shows that transfer from English in the same target domain leads to more than 5% improvement in accuracy compared to transfer from Twitter in the same language. To further mitigate the domain difference, we leverage machine translation (MT) from English to other Nigerian languages, which leads to a further improvement of 7% over cross-lingual evaluation. While MT to low-resource languages are often of low quality, through human evaluation, we show that most of the translated sentences preserve the sentiment of the original English reviews.
In natural language processing (NLP), deep neural networks (DNNs) could model complex interactions between context and have achieved impressive results on a range of NLP tasks. Prior works on feature interaction attribution mainly focus on studying symmetric interaction that only explains the additional influence of a set of words in combination, which fails to capture asymmetric influence that contributes to model prediction. In this work, we propose an asymmetric feature interaction attribution explanation model that aims to explore asymmetric higher-order feature interactions in the inference of deep neural NLP models. By representing our explanation with an directed interaction graph, we experimentally demonstrate interpretability of the graph to discover asymmetric feature interactions. Experimental results on two sentiment classification datasets show the superiority of our model against the state-of-the-art feature interaction attribution methods in identifying influential features for model predictions. Our code is available at https://github.com/StillLu/ASIV.
NLP datasets are richer than just input-output pairs; rather, they carry causal relations between the input and output variables. In this work, we take sentiment classification as an example and look into the causal relations between the review (X) and sentiment (Y). As psychology studies show that language can affect emotion, different psychological processes are evoked when a person first makes a rating and then self-rationalizes their feeling in a review (where the sentiment causes the review, i.e., Y -> X), versus first describes their experience, and weighs the pros and cons to give a final rating (where the review causes the sentiment, i.e., X -> Y ). Furthermore, it is also a completely different psychological process if an annotator infers the original rating of the user by theory of mind (ToM) (where the review causes the rating, i.e., X -ToM-> Y ). In this paper, we verbalize these three causal mechanisms of human psychological processes of sentiment classification into three different causal prompts, and study (1) how differently they perform, and (2) what nature of sentiment classification data leads to agreement or diversity in the model responses elicited by the prompts. We suggest future work raise awareness of different causal structures in NLP tasks. Our code and data are at https://github.com/cogito233/psych-causal-prompt
Our contribution to the 2023 AfriSenti-SemEval shared task 12: Sentiment Analysis for African Languages, provides insight into how a multilingual large language model can be a resource for sentiment analysis in languages not seen during pretraining. The shared task provides datasets of a variety of African languages from different language families. The languages are to various degrees related to languages used during pretraining, and the language data contain various degrees of code-switching. We experiment with both monolingual and multilingual datasets for the final fine-tuning, and find that with the provided datasets that contain samples in the thousands, monolingual fine-tuning yields the best results.
We present the findings of SemEval-2023 Task 12, a shared task on sentiment analysis for low-resource African languages using Twitter dataset. The task featured three subtasks; subtask A is monolingual sentiment classification with 12 tracks which are all monolingual languages, subtask B is multilingual sentiment classification using the tracks in subtask A and subtask C is a zero-shot sentiment classification. We present the results and findings of subtask A, subtask B and subtask C. We also release the code on github. Our goal is to leverage low-resource tweet data using pre-trained Afro-xlmr-large, AfriBERTa-Large, Bert-base-arabic-camelbert-da-sentiment (Arabic-camelbert), Multilingual-BERT (mBERT) and BERT models for sentiment analysis of 14 African languages. The datasets for these subtasks consists of a gold standard multi-class labeled Twitter datasets from these languages. Our results demonstrate that Afro-xlmr-large model performed better compared to the other models in most of the languages datasets. Similarly, Nigerian languages: Hausa, Igbo, and Yoruba achieved better performance compared to other languages and this can be attributed to the higher volume of data present in the languages.
This paper presents a performance study of transformer language models under different hardware configurations and accuracy requirements and derives empirical observations about these resource/accuracy trade-offs. In particular, we study how the most commonly used BERT-based language models (viz., BERT, RoBERTa, DistilBERT, and TinyBERT) perform on embedded systems. We tested them on four off-the-shelf embedded platforms (Raspberry Pi, Jetson, UP2, and UDOO) with 2 GB and 4 GB memory (i.e., a total of eight hardware configurations) and four datasets (i.e., HuRIC, GoEmotion, CoNLL, WNUT17) running various NLP tasks. Our study finds that executing complex NLP tasks (such as "sentiment" classification) on embedded systems is feasible even without any GPUs (e.g., Raspberry Pi with 2 GB of RAM). Our findings can help designers understand the deployability and performance of transformer language models, especially those based on BERT architectures.
We present the first Africentric SemEval Shared task, Sentiment Analysis for African Languages (AfriSenti-SemEval) - The dataset is available at https://github.com/afrisenti-semeval/afrisent-semeval-2023. AfriSenti-SemEval is a sentiment classification challenge in 14 African languages: Amharic, Algerian Arabic, Hausa, Igbo, Kinyarwanda, Moroccan Arabic, Mozambican Portuguese, Nigerian Pidgin, Oromo, Swahili, Tigrinya, Twi, Xitsonga, and Yor\`ub\'a (Muhammad et al., 2023), using data labeled with 3 sentiment classes. We present three subtasks: (1) Task A: monolingual classification, which received 44 submissions; (2) Task B: multilingual classification, which received 32 submissions; and (3) Task C: zero-shot classification, which received 34 submissions. The best performance for tasks A and B was achieved by NLNDE team with 71.31 and 75.06 weighted F1, respectively. UCAS-IIE-NLP achieved the best average score for task C with 58.15 weighted F1. We describe the various approaches adopted by the top 10 systems and their approaches.
Recent decades have seen the rise of large-scale Deep Neural Networks (DNNs) to achieve human-competitive performance in a variety of artificial intelligence tasks. Often consisting of hundreds of millions, if not hundreds of billion parameters, these DNNs are too large to be deployed to, or efficiently run on resource-constrained devices such as mobile phones or IoT microcontrollers. Systems relying on large-scale DNNs thus have to call the corresponding model over the network, leading to substantial costs for hosting and running the large-scale remote model, costs which are often charged on a per-use basis. In this paper, we propose BiSupervised, a novel architecture, where, before relying on a large remote DNN, a system attempts to make a prediction on a small-scale local model. A DNN supervisor monitors said prediction process and identifies easy inputs for which the local prediction can be trusted. For these inputs, the remote model does not have to be invoked, thus saving costs, while only marginally impacting the overall system accuracy. Our architecture furthermore foresees a second supervisor to monitor the remote predictions and identify inputs for which not even these can be trusted, allowing to raise an exception or run a fallback strategy instead. We evaluate the cost savings, and the ability to detect incorrectly predicted inputs on four diverse case studies: IMDB movie review sentiment classification, Github issue triaging, Imagenet image classification, and SQuADv2 free-text question answering
Sentiment analysis, the automated process of determining emotions or opinions expressed in text, has seen extensive exploration in the field of natural language processing. However, one aspect that has remained underrepresented is the sentiment analysis of the Moroccan dialect, which boasts a unique linguistic landscape and the coexistence of multiple scripts. Previous works in sentiment analysis primarily targeted dialects employing Arabic script. While these efforts provided valuable insights, they may not fully capture the complexity of Moroccan web content, which features a blend of Arabic and Latin script. As a result, our study emphasizes the importance of extending sentiment analysis to encompass the entire spectrum of Moroccan linguistic diversity. Central to our research is the creation of the largest public dataset for Moroccan dialect sentiment analysis that incorporates not only Moroccan dialect written in Arabic script but also in Latin letters. By assembling a diverse range of textual data, we were able to construct a dataset with a range of 20 000 manually labeled text in Moroccan dialect and also publicly available lists of stop words in Moroccan dialect. To dive into sentiment analysis, we conducted a comparative study on multiple Machine learning models to assess their compatibility with our dataset. Experiments were performed using both raw and preprocessed data to show the importance of the preprocessing step. We were able to achieve 92% accuracy in our model and to further prove its liability we tested our model on smaller publicly available datasets of Moroccan dialect and the results were favorable.
Tsetlin Machines (TsMs) are a promising and interpretable machine learning method which can be applied for various classification tasks. We present an exact encoding of TsMs into propositional logic and formally verify properties of TsMs using a SAT solver. In particular, we introduce in this work a notion of similarity of machine learning models and apply our notion to check for similarity of TsMs. We also consider notions of robustness and equivalence from the literature and adapt them for TsMs. Then, we show the correctness of our encoding and provide results for the properties: adversarial robustness, equivalence, and similarity of TsMs. In our experiments, we employ the MNIST and IMDB datasets for (respectively) image and sentiment classification. We discuss the results for verifying robustness obtained with TsMs with those in the literature obtained with Binarized Neural Networks on MNIST.
Social media is daily creating massive multimedia content with paired image and text, presenting the pressing need to automate the vision and language understanding for various multimodal classification tasks. Compared to the commonly researched visual-lingual data, social media posts tend to exhibit more implicit image-text relations. To better glue the cross-modal semantics therein, we capture hinting features from user comments, which are retrieved via jointly leveraging visual and lingual similarity. Afterwards, the classification tasks are explored via self-training in a teacher-student framework, motivated by the usually limited labeled data scales in existing benchmarks. Substantial experiments are conducted on four multimodal social media benchmarks for image text relation classification, sarcasm detection, sentiment classification, and hate speech detection. The results show that our method further advances the performance of previous state-of-the-art models, which do not employ comment modeling or self-training.
Transfer learning plays an essential role in Deep Learning, which can remarkably improve the performance of the target domain, whose training data is not sufficient. Our work explores beyond the common practice of transfer learning with a single pre-trained model. We focus on the task of Vietnamese sentiment classification and propose LIFA, a framework to learn a unified embedding from several pre-trained models. We further propose two more LIFA variants that encourage the pre-trained models to either cooperate or compete with one another. Studying these variants sheds light on the success of LIFA by showing that sharing knowledge among the models is more beneficial for transfer learning. Moreover, we construct the AISIA-VN-Review-F dataset, the first large-scale Vietnamese sentiment classification database. We conduct extensive experiments on the AISIA-VN-Review-F and existing benchmarks to demonstrate the efficacy of LIFA compared to other techniques. To contribute to the Vietnamese NLP research, we publish our source code and datasets to the research community upon acceptance.
Africa is home to over 2,000 languages from more than six language families and has the highest linguistic diversity among all continents. These include 75 languages with at least one million speakers each. Yet, there is little NLP research conducted on African languages. Crucial to enabling such research is the availability of high-quality annotated datasets. In this paper, we introduce AfriSenti, a sentiment analysis benchmark that contains a total of >110,000 tweets in 14 African languages (Amharic, Algerian Arabic, Hausa, Igbo, Kinyarwanda, Moroccan Arabic, Mozambican Portuguese, Nigerian Pidgin, Oromo, Swahili, Tigrinya, Twi, Xitsonga, and Yor\`ub\'a) from four language families. The tweets were annotated by native speakers and used in the AfriSenti-SemEval shared task (The AfriSenti Shared Task had over 200 participants. See website at https://afrisenti-semeval.github.io). We describe the data collection methodology, annotation process, and the challenges we dealt with when curating each dataset. We further report baseline experiments conducted on the different datasets and discuss their usefulness.
We introduce InstructABSA, an instruction learning paradigm for Aspect-Based Sentiment Analysis (ABSA) subtasks. Our method introduces positive, negative, and neutral examples to each training sample, and instruction tune the model (Tk-Instruct) for ABSA subtasks, yielding significant performance improvements. Experimental results on the Sem Eval 2014, 15, and 16 datasets demonstrate that InstructABSA outperforms the previous state-of-the-art (SOTA) approaches on Term Extraction (ATE), Sentiment Classification(ATSC) and Sentiment Pair Extraction (ASPE) subtasks. In particular, InstructABSA outperforms the previous state-of-the-art (SOTA) on the Rest14 ATE subtask by 5.69% points, the Rest15 ATSC subtask by 9.59% points, and the Lapt14 AOPE subtask by 3.37% points, surpassing 7x larger models. We also get competitive results on AOOE, AOPE, and AOSTE subtasks indicating strong generalization ability to all subtasks. Exploring sample efficiency reveals that just 50% train data is required to get competitive results with other instruction tuning approaches. Lastly, we assess the quality of instructions and observe that InstructABSA's performance experiences a decline of ~10% when adding misleading examples.