Large Language Models (LLMs) are prone to generating content that exhibits gender biases, raising significant ethical concerns. Alignment, the process of fine-tuning LLMs to better align with desired behaviors, is recognized as an effective approach to mitigate gender biases. Although proprietary LLMs have made significant strides in mitigating gender bias, their alignment datasets are not publicly available. The commonly used and publicly available alignment dataset, HH-RLHF, still exhibits gender bias to some extent. There is a lack of publicly available alignment datasets specifically designed to address gender bias. Hence, we developed a new dataset named GenderAlign, aiming at mitigating a comprehensive set of gender biases in LLMs. This dataset comprises 8k single-turn dialogues, each paired with a “chosen” and a “rejected” response. Compared to the “rejected” responses, the “chosen” responses demonstrate lower levels of gender bias and higher quality. Furthermore, we categorized the gender biases in the “rejected” responses of GenderAlign into 4 principal categories. The experimental results show the effectiveness of GenderAlign in reducing gender bias in LLMs.

Psycholinguistic knowledge resources have been widely used in constructing features for text-based human trait and behavior analysis. Recently, deep neural network (NN)-based text analysis methods have gained dominance due to their high prediction performance. However, NN-based methods may not perform well in low resource scenarios where the ground truth data is limited (e.g., only a few hundred labeled training instances are available). In this research, we investigate diverse methods to incorporate Linguistic Inquiry and Word Count (LIWC), a widely-used psycholinguistic lexicon, in NN models to improve human trait and behavior analysis in low resource scenarios. We evaluate the proposed methods in two tasks: predicting delay discounting and predicting drug use based on social media posts. The results demonstrate that our methods perform significantly better than baselines that use only LIWC or only NN-based feature learning methods. They also performed significantly better than published results on the same dataset.

In recent years pre-trained language models (PLM) such as BERT have proven to be very effective in diverse NLP tasks such as Information Extraction, Sentiment Analysis and Question Answering. Trained with massive general-domain text, these pre-trained language models capture rich syntactic, semantic and discourse information in the text. However, due to the differences between general and specific domain text (e.g., Wikipedia versus clinic notes), these models may not be ideal for domain-specific tasks (e.g., extracting clinical relations). Furthermore, it may require additional medical knowledge to understand clinical text properly. To solve these issues, in this research, we conduct a comprehensive examination of different techniques to add medical knowledge into a pre-trained BERT model for clinical relation extraction. Our best model outperforms the state-of-the-art systems on the benchmark i2b2/VA 2010 clinical relation extraction dataset.

We propose a novel supervised open information extraction (Open IE) framework that leverages an ensemble of unsupervised Open IE systems and a small amount of labeled data to improve system performance. It uses the outputs of multiple unsupervised Open IE systems plus a diverse set of lexical and syntactic information such as word embedding, part-of-speech embedding, syntactic role embedding and dependency structure as its input features and produces a sequence of word labels indicating whether the word belongs to a relation, the arguments of the relation or irrelevant. Comparing with existing supervised Open IE systems, our approach leverages the knowledge in existing unsupervised Open IE systems to overcome the problem of insufficient training data. By employing multiple unsupervised Open IE systems, our system learns to combine the strength and avoid the weakness in each individual Open IE system. We have conducted experiments on multiple labeled benchmark data sets. Our evaluation results have demonstrated the superiority of the proposed method over existing supervised and unsupervised models by a significant margin.

Text analytics is a useful tool for studying malware behavior and tracking emerging threats. The task of automated malware attribute identification based on cybersecurity texts is very challenging due to a large number of malware attribute labels and a small number of training instances. In this paper, we propose a novel feature learning method to leverage diverse knowledge sources such as small amount of human annotations, unlabeled text and specifications about malware attribute labels. Our evaluation has demonstrated the effectiveness of our method over the state-of-the-art malware attribute prediction systems.

We describe the systems developed by the UMBC team for 2018 SemEval Task 8, SecureNLP (Semantic Extraction from CybersecUrity REports using Natural Language Processing). We participated in three of the sub-tasks: (1) classifying sentences as being relevant or irrelevant to malware, (2) predicting token labels for sentences, and (4) predicting attribute labels from the Malware Attribute Enumeration and Characterization vocabulary for defining malware characteristics. We achieve F1 score of 50.34/18.0 (dev/test), 22.23 (test-data), and 31.98 (test-data) for Task1, Task2 and Task2 respectively. We also make our cybersecurity embeddings publicly available at http://bit.ly/cyber2vec.

In this paper, we demonstrate how the state-of-the-art machine learning and text mining techniques can be used to build effective social media-based substance use detection systems. Since a substance use ground truth is difficult to obtain on a large scale, to maximize system performance, we explore different unsupervised feature learning methods to take advantage of a large amount of unsupervised social media data. We also demonstrate the benefit of using multi-view unsupervised feature learning to combine heterogeneous user information such as Facebook “likes” and “status updates” to enhance system performance. Based on our evaluation, our best models achieved 86% AUC for predicting tobacco use, 81% for alcohol use and 84% for illicit drug use, all of which significantly outperformed existing methods. Our investigation has also uncovered interesting relations between a user’s social media behavior (e.g., word usage) and substance use.