The legal counsellors will always look up on the prior cases and statutes to ensure justice. Referring all prior cases is a time consuming process since they are vast. Artificial intelligence can be made use to select the most relevant among all the documents. The existing systems have made use of different word embeddings, machine learning and deep learning methods to select the relevant ones. In our approach, different vectorization methods such as Word2Vec, Glove, Tf-Idf and count vectorizer are used and then similarity is calculated using Jaccard similarity and cosine similarity for ranking the prior cases and statutes. The work is evaluated on the AILA@FIRE-2019 dataset, in which it provides two tasks of finding the prior documents namely the relevant cases and statutes.
The population which needs to attain any legal assistance have to search for
the prior cases and statutes for the current case. The search goes innumerable
since there are many. Aiding the help of artificial intelligence for the search and
retrieval seems to be a effective idea when compared to the manual search
retrieval. Different machine learning and deep learning methods include Doc2Vec,
Tf-Idf, LSTM , etc. are made use for retrieving the prior cases [
For retrieving the relevant cases and statutes, initially the documents are vectorized using Word2Vec, Tf-Idf, count vectorizer and a ensembling method of Glove and Word2Vec. After vectorizing, the documents are all ranked by finding the similarity using cosine similarity and jaccard similarity. 2.2
The query document and the case document are initially vectorized using
Word2Vec model of dimension ’300’. After vectorizing the whole document, the
max-min vector of the documents are considered to represent each document as
a vector. Max-min vector of query is considered as ’a’ and ’b’ be the max-min
vector of case document. The case documents are ranked by finding the cosine
similarity [
In this method, the query and documents are all vectorized using both Word2Vec
and Glove and then they are concatenated. Considering a single query and case
document, vectorize the case document using Glove which forms the vector ’a1’
and vectorize the case document using Word2Vec model which forms the vector
’a2’. Concatenate the two vectors a1 and a2 which becomes the vector ’a’. The
same process of vectorization is done for the query document which forms
vector ’b’. Then, the case documents are ranked by finding the cosine similarity [
In this method, the documents and queries are all vectorized using Tf-Idf
vectorizer. The entire corpus (queries and cases) is made use to form vocabulary which
is used to fit the documents. The vector of query forms ’a’ and that of document
forms ’b’. Rank the documents by finding the cosine similarity [
In this method, the documents and queries are all vectorized using Tf-Idf
vectorizer. The entire corpus (queries and statues) is made use to form vocabulary
which is used to fit the documents. The vector of query forms ’a’ and that of
document forms ’b’. Rank the statute documents by finding the cosine similarity
[
This method is done by finding out the Jaccard similarity. The vocabulary of
query document forms ’a’ and vocabulary of statute document which forms ’b’.
The documents are ranked by finding the jaccard similarity [
In this method, the documents and queries are all vectorized using count
vectorizer (i.e) the count of each word in the documents (query and statute). Frequent
words of the query document forms the dictionary of the file whose vector forms
’a’. Frequent words of the statute document forms the dictionary of the file
whose vector forms ’b’. The statutes are ranked by finding the cosine similarity
[
vectorization methods for identifying the relevant cases and Tf-Idf vectorizing method seems to be better for identifying the relevant statutes. 4
Artificial Intelligence helps us in many ways for identifying the relevant documents among the entire prior documents in legal domain. Different word embedding methods of finding out similarity are experimented on ALIA@FIRE-2019 dataset. Various word embeddings like Word2Vec, Glove, ensembling Word2Vec and Glove, Tf-Idf vectorizer and count vectorizer are used for vectorization. After vectorization, cosine similarity is calculated to rank the documents. Among these Word2Vec seems to perform better than the other vectorization process for Task 1 of identifying the relevant cases and Tf-Idf seems to perform better than the other vectorizing methods for task 2 of identifying the relevant prior
Teams and Runs SSN_NLP Run 1 SSN_NLP Run 2 SSN_NLP Run 3
Yunqiu Shao
UBLTM
Sara Renjit Soumil Mandal - JU_SRM
HLJIT2019
Kavya S Ganesh Baban Gain - IITP
MAP score 0.077 0.061 0.051 0.156 0.102 0.096 0.083 0.081 0.068 0.036 statutes. The performance can be further improved by other machine learning and deep learning methods.
We would like to thank the Science and Engineering Research Board (SERB), Department of Science and Technology for funding the GPU system (EEQ/2018/000262) where this work has been carried out.