In this work, we introduce a pioneering Neural Semantic Search Framework designed for legal information retrieval, known as Semantic Legal Searcher (SLS). The framework utilizes state-of-the-art neural information retrieval techniques to deliver precise legal information responses for user queries, irrespective of their legal background. This work encompasses a distinctive architecture, incorporating the split-merge technique for enhanced document embeddings and a multi-interactions paradigm to elevate semantic similarity measures. Additionally, the SLS framework integrates two pre-trained Legal Language Models(LLMs), KRLawBERT and KRLawGPT, specifically tailored for legal texts, surpassing the capabilities of conventional NLP models like BERT.

The SLS is presented as a groundbreaking conceptual search model based on neural information retrieval, aiming to revolutionize legal information accessibility. Unlike common methods in Figure 1 (a), we extend our search model by introducing two information retrieval techniques shown in (b). The SLS employs an optimized neural Information Retrieval (IR) approach for legal datasets, incorporating the split-merge technique to maximize information retention in document embeddings. The multi-interactions technique enhances semantic similarity measures, establishing SLS as a comprehensive search framework that combines semantic search and topic modeling for relevance identification and keyword extraction.

BERT(Bidirectional Encoder Representations from Transformers) and GPT(Generative Pre-trained Transformer) are both NLP models capable of understanding the semantic meaning of the text and have been applied to a variety of tasks including text classification, entity recognition, and more.

KRLawBERT, developed through benchmarking Masked Language Modeling (MLM) and Transformer-based Sequential Denoising Auto-Encoder (TSDAE), outperforms conventional BERT-based models in capturing sentence text similarity for legal information retrieval.

☞ https://github.com/Navy10021/KRLawBERT

Additionally, KRLawGPT, a decoder-only transformer, is introduced for generating expressive Korean legal text, broadening the language model capabilities.

☞ https://github.com/Navy10021/KRLawGPT

The process of the SLS is divided into four steps as shown in Figure 2 : document encoding and split-merge modelization, parallel clustering-based topic modeling for keyword extraction, multi-interactions for relevance estimation, and the provision of user search results based on relevance scores. Each step is intricately designed to optimize the efficiency and effectiveness of legal information retrieval.

Users have the flexibility to utilize existing models such as BERT or opt for the specialized KRLawBERT, pre-trained on a large-scale legal text dataset, enhancing performance for legal information retrieval.

import pandas as pd
# Load dataset(Cornell University., 2022)
df = pd.read_csv('./data/arxiv_meta.csv')

# Load pre-trained language model on English dataset
my_plms = "all-mpnet-base-v2"

KRLawBERT is a pre-trained Transformer's encoder-based model to generate high-quality embeddings and better understand texts in legal domains.

import pandas as pd
# Load Korean Legal Dataset(Korean Judicial precedent data)
df = pd.read_csv('./data/law_cases(20221020).csv')

# Load pre-trained & fine-tuned models (KRLawBERT)
my_plms = './output/tsdae-krlawbert'

A novel parallel clustering-based topic modeling technique is introduced, ensuring swift and efficient extraction of latent keywords from legal documents.

1. Parallel Clustering class

• dataframe : Dataframe based table
• tgt_col : Documents columns (str)
• model_name : PLMs (str)
• use_sentence_bert : Whether to generate sentence embeddings or not (bool)
• threshold : threshold (float)
• page_size : Number of initial centroids (int)
• iterations : Max iterations (int)

2. Keywords Extraction class

• dataframe : DataFrame including clustered documents
• n : Number of keywords to extract (int)
• en : Whether documents is English or not (bool)

# Import parallel_clustering-based topic modeling model
from models.parallel_clustering_TM import *

# Obtain Embeddings
target_text = 'abstract'
cluster = ParallelCluster(
    dataframe = df,
    tgt_col = target_text,
    model_name = my_plms,
    use_sentence_bert = True
    )

# Parallel Clustering
clusters, unclusters = cluster.parallel_cluster(
    clusters = None,
    threshold = 0.56,
    page_size = 2000,
    iterations = 30
    )
    
# Stack : Stack the clustered results in order of cluster size
col_list = ['title', 'abstract', 'year']
new_df = cluster.cluster_stack(
    col_list = col_list,
    clusters = clusters,
    unclusters = unclusters
    )
    
# Extract Keywords from each documents
top_n_words = cluster.extract_top_n_words_per_topic(
    dataframe = new_df,
    n = 20,
    en = True
    )
new_df['keywords'] = [', '.join(top_n_words[i]) for i in new_df['Topic'].values]

The research highlights the significance of the split-merge embedding modelization technique and the multi-interactions scoring mechanism, showcasing a substantial improvement of 14-20% in semantic search accuracy. The SLS class facilitates these processes for both English and Korean legal datasets.

SLS class

• dataframe : Dataframe based table
• doc_col : Documents columns (str)
• key_col : Keywords columns (str)
• model_name : PLMs name (str)
• use_sentence_bert : Whether to generate sentence embeddings or not (bool)
• split_and_merge : Whether to use split-merge embeddings modelization technique (bool)
• multi_inter : Whether to use multi-interactions scoring technique (bool)

# If you use the English dataset, import this semantic searcher model
from models.semantic_searcher_eng import *
# If you use the Korean case law dataset, import this semantic searcher model
from models.semantic_legal_searcher import *

# Load SLS framework
sls = SLS(
    dataframe = new_df,
    doc_col = 'abstract',
    key_col = 'keywords',
    model_name = my_plms,
    use_sentence_bert = True,
    split_and_merge = True,
    multi_inter = True,
    )

# Build the Index
# (Strategy 1) : All Distance Metric
all_index = sls.all_distance_metric()
# (Strategy 2) : Restricted Distance Metric
#restricted_index = sls.restricted_distance_metric(nlist = 200, nprobe = 6)

Users can seamlessly initiate semantic searches by inputting queries, leveraging the comprehensive SLS framework to retrieve relevant legal documents.

semantic search function

• user_query : Your input query (str)
• top_k : Number of documents related to your query (int)
• index : Index (variable)
• print_results : Whether to print search results or not (bool)

# Semantic documents(arXiv) search
your_query = "Research about the Transformer network architecture, based solely on attention mechanisms."

outputs, _ = sls.semantic_search(
    user_query = your_query,
    top_k = 10,
    index = all_index,
    print_results = True,
    )

• Currently, our paper is under review. It will be revealed in the future.

• Seoul National University NLP Labs
• Under the guidance of Navy Lee