This paper presents a comparative study evaluating the performance of Large Language Models (LLMs) in generating timeline summaries from construction delay documents. We assessed seven open-source LLMs and two commercial chatbots (ChatGPT and Claude) on their ability to extract, organize, and summarize delay events from twenty-one carefully curated synthetic snippets of text. The evaluation framework combined automatic metrics (BERTScore and ROUGE scores) with expert human assessment across four dimensions: event description accuracy, date accuracy, event capture completeness, and language quality.Results demonstrate that while commercial solutions, particularly Claude, achieved superior performance, several open-source alternatives showed comparable capabilities. Notably, Llama-3.1-70B-Instruct showed robust performance in event capture and source tracking, while Llama-3.1-8B-Instruct offered efficient processing with balanced performance among smaller models. A critical finding was the widespread challenge in temporal information processing, with only Claude achieving complete accuracy in date extraction and event association. The study's findings suggest that open-source LLMs can serve as practical tools for construction document analysis, although model selection is a critical consideration based on specific accuracy and efficiency requirements, and resource constraints.

Text embeddings are numerical representations of text data, where words, phrases, or entire documents are converted into vectors of real numbers. These embeddings capture semantic meanings and relationships between text elements in a continuous vector space. The primary goal of text embeddings is to enable the processing of text data by machine learning models, which require numerical input. Numerous embedding models have been developed for various applications. This paper presents our work in evaluating different embeddings through a comprehensive comparative analysis of four distinct models, focusing on their text classification efficacy. We employ both K-Nearest Neighbors (KNN) and Logistic Regression (LR) to perform binary classification tasks, specifically determining whether a text snippet is associated with 'delay' or 'not delay' within a labeled dataset. Our research explores the use of text snippet embeddings for training supervised text classification models to identify delay-related statements during the document review process of construction delay disputes. The results of this study highlight the potential of embedding models to enhance the efficiency and accuracy of document analysis in legal contexts, paving the way for more informed decision-making in complex investigative scenarios.

Corporate litigation costs associated with manually reviewing documents in legal matters continue to rise, driven by the ever-expanding volumes of data generated by businesses. To address this challenge, legal teams have increasingly turned to text classification techniques, commonly referred to in the legal community as predictive modeling, predictive coding, or "Technology Assisted Review (TAR)", to identify relevant documents more efficiently during the discovery phase of legal matters. By streamlining the classification process, predictive coding has proven instrumental in minimizing both the time and cost associated with legal discovery. As more corporations expand their operations across global regions, legal matters increasingly involve multilingual datasets, introducing significant complexities for the application of predictive coding. Traditional predictive models exhibit bias towards the dominant language within the model’s training data. The prevalence of multiple languages within the same document further complicates the challenge. This can lead to inefficiencies, elevated costs, and risks of inaccurate classifications of nondominant language documents.This research empirically examines two methods for applying predictive coding to multilingual datasets: a single multilingual model approach and a language-specific modeling approach. This group of collaborators posits that language-specific models will outperform the single multilanguage model approach in evaluating document relevance. The research used a multilingual dataset from a recent confidential real-world legal matter containing predominantly English and Chinese document text Performance was assessed using precision and recall measures. Our results show that the ensemble approach – employing language-specific models for classifying documents in their respective languages and employing the mixed language model across all remaining documents—outperforms a single multilanguage model. The outcomes of this research aim to inform the workflow considerations for legal professionals undertaking multilingual discovery efforts. The findings will help legal practitioners make data-driven decisions on the most efficient and accurate modeling approach for multilingual document review, shaping the future of predictive coding in cross-linguistic e-discovery.

Predictive modeling - known in the legal industry as ‘predictive coding’ or ‘Technology Assisted Review (TAR)’ - is a popular tool used by legal professionals to augment a historically manual document review and classification process in responding to data requests for various legal proceedings. It has become more commonplace over the past decade due to its proven ability to minimize manual document classification, thus reducing the time and cost associated with this aspect of legal proceedings. There is significant research supporting the effectiveness of this technology that includes topics, such as identifying the most performant machine learning algorithms (e.g., logistic regression) or establishing the best methodologies for selecting representative training and testing data. Primarily, this research has been performed without a focus on multilanguage data sets because many legal proceedings typically involve one primary, dominant language. As acceptance of predictive modeling technology grows, legal practitioners have developed independent preferences for handling multiple languages in data sets. These preferences have been primarily based on anecdotal experience rather than empirical assessments and have resulted in two modeling approaches. The first approach uses a single model, which is less complex, and more cost effective than the next approach. The second develops multiple language-specific models, which can create workflow complexity that may impact the overall cost savings that predictive coding seeks to achieve. Proponents of the second approach believe that creating models per language results in better performance over a single model approach. In this study, we empirically explore the performance differences between the two approaches - single model and multiple language-specific models. We hypothesize that a single model approach performs similarly to a language-specific modeling approach when classifying documents for relevance. We evaluate both approaches by comparing their resulting precision-recall curves across real-world test data from legal document review projects. Excitingly, our results demonstrate that in most scenarios, a single model (trained with multiple languages) approach will perform as well, or sometimes better than the approach that uses a group of language-specific models. The results of our research provide a roadmap for legal practitioners and other parties to thoughtfully engage in dialogue around the most effective way to deploy predictive coding within a multilanguage data set. This research will change how predictive models are deployed for multilanguage data sets across the legal industry and enable significant cost savings for clients and legal practitioners by utilizing the most efficient modeling method.

The increased integration of Large Language Models (LLMs) across industry sectors is enabling domain experts with new text classification optimization methods. These LLMs are pretrained on exceedingly large amounts of data; however, practitioners can perform additional training, or “fine-tuning,” to improve their text classifier’s results for their own use cases. This paper presents a series of experiments comparing a standard, pretrained DistilBERT model and a fine-tuned DistilBERT model, both leveraged for the downstream NLP task of text classification. Tuning the model using domain-specific data from real-world legal matters suggests fine-tuning improves the performance of LLM text classifiers.To evaluate the performance of text classification models, using these two Large Language Models, we employed two distinct approaches that 1) score a whole document’s text for prediction and 2) score snippets (sentence-level components of a document) of text for prediction. When comparing the two approaches we found that one prediction method outperforms the other, depending on the use case.

In the domain of natural language processing and text analysis, the Bidirectional Encoder Representations from Transformers (BERT) has emerged as a powerful tool for discerning the intricate nuances of textual data. Nonetheless, BERT’s inherent token limit of 512 tokens presents a noteworthy challenge when confronted with exceedingly long documents. Lengthy documents are commonly encountered during legal document review and often exceed the 512-token constraint. In response to this challenge, this study empirically compares two distinct applications of BERT, leveraging real-world, construction industry legal data. The approach compares applications of BERT to the entire document and on segmented text portions from each document. In the latter approach, the highest-scoring text segment from each document represents the document’s score. This research offers practical insights for effectively utilizing BERT in scenarios where document length exceeds the token limit. Our results allow practitioners and researchers to make informed choices when confronted with documents of significant length, thus contributing to a more effective and insightful application of BERT for text analysis.

In recent years, transformer-based, large-scale language models have greatly advanced deep learning in NLP tasks. These models allow transfer learning to be performed on NLP in similar methods to what was previously done for computer vision. Two of the latest models are DistilBERT and Longformer -- the former is a distilled version of BERT, which enables faster training and inferencing on off-cloud servers, and the latter offers capability of working with long texts. In this paper, we study empirical comparisons of the effectiveness of the two deep learning methods along with a logistical regression method for text classification, using three real-world datasets from legal document reviews. The study shows that Longformer performs better (up to 10%) than or at par with the other two methods. Cross-dataset evaluation is leveraged as well to validate the performance of Longformer as a viable method when labeled data is not available.