3rd International Conference on Innovative Computing and Communication

Automatic Legal Judgment Summarization Using Graph

Vishnuprabha V a, Daleesha M Viswanathan b,*, R. Rajesh c

a
School of Engineering, Cochin University of Science and Technology, Kerala, India
b
School of Engineering, Cochin University of Science and Technology, Kerala, India
c
Scientist-F, NPOL, Kerala, India

Abstract:
Summarization is the process of generating summaries out of corpus. If the corpus is a legal document, the process is called legal summarization.
Headnotes are concise notes prepared by lawyers for quick reference of a past Judgment. Judgments are always lengthy, and it is essential to make
headnotes for legal practitioners for quick reference analysis. Legal experts generate headnotes, and it is a tedious task to prepare them by reading the
whole judgment. With the advancements in technologies, a lot of methods are available to generate a summary of a given document. But for legal
documents, the usual summarization technique have the impediment to precisely concise as judgments follow a legal parlance. The proposed method picks
relevant paragraphs from a judgment at first, and then it is used to create an extractive summary.

2. Related Works
1. Introduction
Text summarization is not a new technique in the field of Natural
Text processing is becoming an important research area of practical
Language Processing. It includes Extractive and Abstractive techniques.
applicability in this era of advanced computing. Now, most of the
In Extractive summarization, relevant sentences are picked from the
documents and tasks are getting digitalized for effortless processing and
document, and they are aligned to generate a summary. Various text
retrieval of information. Digitalization has made a significant impact on
summarization approaches are there, but the main idea behind every
the economy, especially in the legal domain. Legal advisors and Lawyers
approach is scoring the sentences based on some techniques. Whereas in
often need to access enormous texts for referring the past cases. Legal
abstractive summarization, the summary is similar to a human-generated
advisors use a previous case history to make valuable argument points.
summary as there will be logical connections. This approach is better
They make use of headnotes of similar past cases for understanding the
when compared to the extractive summarization technique because an
overall structure of a particular case, and if they feel it may help in
extractive summarization picks sentences from different parts of the
creating an argument, they read the entire judgment. Headnotes are
document, and it doesn't guarantee a relationship between the former and
nothing but a summary of the entire case. So, developing such a summary
the latter sentences.
from a 50-70 page judgment is a tedious task. Headnotes are generated by
knowledge experts in the domain and not by a language expert. A lawyer
The abstractive summarization approaches use a natural language
chooses a case/judgment as a reference by reading the headnote, so the
generator for creating an abstractive summary. But creating a natural
headnote should be generated carefully without losing the relevant
language generator is not an easy task. Modelling a natural language
information.
generator is done mostly using deep learning structures.

With the advancements in the field of Natural Language Processing and

There were attempts to create a summary of the legal text using the above-
Artificial Intelligence, most of the complex tasks are getting less
said methods. One of the most critical tasks applied for a legal judgment
complicated. Natural Language Processing can be used for summarizing
before feeding into the summary generator is the application of
legal judgments without human interference. Since the judgments are now
Conditional Random Field. Conditional Random Field is a segmentation
available in digital format, automatic legal text summarization is an
technique that helps to identify the segments in a document. A judgment
important area of research.
has a specific format, and they follow a similar sequence in every
judgment. Such partitions can be identified easily from the document
The extractive and abstractive summarization approaches can be used for
using a conditional random field approach. After segmentation,
summarizing the legal text. But the identification of relevant portions
summarization techniques are applied to relevant portions of the
from a lengthy judgment is not at all easy whereas recognizing those
summary. The significant attempts to legal text summarization include:
relevant paragraphs in a judgment can help make the summarization
process easier.
Farzindar et al. (2004) presented a thematic segmentation-based approach
for extractive summarization. They use keywords, position, title, etc. for
thematic segmentation of the document. It generates a table style

AUTHOR 1

Electronic copy available at: https://ssrn.com/abstract=3563105

 3rd International Conference on Innovative Computing and Communication
summary. 1. Appellant/Petitioner
2. Respondents
Saravanan et al. (2008) proposed a machine learning method to identify 3. Appendix
the rhetorical role present in legal documents. The annotated roles in legal 3. 1 Appellant’s Annexures
documents with the help of legal knowledge experts and took it as training 3. 2 Respondent's Annexures
data. The conditional random field approach has been trained to perform 4. Overview of appeal
rhetorical role identification. 5. Appellant's arguments
6. Respondent's arguments
Hachey et al. (2006) identified rhetorical features manually and created an 7. Judicial analysis
annotated corpus. That annotated corpus is then used for predicting 8. Repeat 5,6,7
rhetorical roles automatically using machine learning techniques. Finally, 9. Ratio decidendi (Reason of deciding)
a sentence scoring technique is applied to create an extractive summary. 10. Final Judgment
Varun et al. (2019), in his paper, a hybrid method for automatic text
summarization of legal cases using k-means clustering technique and tf- 4. Proposed Method
idf(term frequency-inverse document frequency) word vectorizer is
proposed. In legal text summarization, conditional random field-based approaches
are the most successful ones according to the studies (Megala, 2018). The
Polsley et al. (2016) presented CaseSummarizer, a tool for automated text conditional random field approach segments the whole document into
summarization of legal documents that use standard summary methods segments according to the standard format of a judgment. The
based on word-frequency augmented with additional domain-specific segmentation approach needs training data for model creation. Segmented
knowledge. Summaries are then provided through an informative interface data is needed for accomplishing a CRF approach, and creating such
with abbreviations, significance heat maps, and other flexible controls. segmented data for a massive set of documents is a time-consuming
process. Even after segmentation, finding relevant portions from each
Galgani et al. (2010) proposed LEXA: Towards Automatic Legal Citation segment is a complicated task, if it is a lengthy judgment.
Classification. It is an approach towards legal citation classification by
employing an incremental knowledge acquisition model. The proposed approach identifies the relevant paragraph from the
judgment. Abstractive summarization approach for legal judgments is a
Another approach by Galgani et al. (2012) is a rule-based system using bit complicated because of the presence of special structures/references,
manual knowledge acquisition, where human intuition, supported by data, such as previous cases and sections. Also, a vast data set is needed for
this approach specifies not only attributes and algorithms, but the contexts training an efficient abstractive summarization model.
where these are used for summarization
The proposed model can identify relevant paragraphs from judgment, and
Schilder et al. (2008) proposed a work based on processing a set of legal it can be aligned to create a summary if needed. Also, those relevant
questions with a system consisting of a semi-automatic Web blog search paragraphs can be used to feed an extractive summarizer to create a
module and FastSum, a fully automatic extractive multi-document summary. The proposed model uses a graph-based approach for
sentiment summarization system. summarization. The proposed approach includes the following steps:

Most of the new systems use machine learning methods for 1. Pre-process the document.
summarization. Machine learning technique requires many training data Pre-processing the judgment is the most complicated task, even though
which are not available in digital format. Also, for conditional random there is a specific structure for every judgment. Making a standard pre-
field technique requires a segmented judgment for training; creating such processing system that can parse all judgments is not easy using regular
training data is possible only with the help of legal experts. expressions. Here a generalized pre-processing system that can pre-
process every judgment from the website https://indiankanoon.org/ is
proposed.
2. Structure of a Judgment
2. Model the whole document into a set of paragraphs.
Every official document follows a specific structure. For, e.g., a research Judgments are reproduced into a set of paragraphs after pre-processing the
paper has an Abstract, Introduction, Literature Survey, Methodology, document.
Analysis, and Conclusion. Researchers identify the relevant papers related
to their area of research by reading the abstract. Likewise, Lawyers make 3. Calculate the dice coefficient.
use of headnotes to identify whether to read the whole judgment or not. “Dice coefficient” measures the similarity of two strings. Here each
Every legal judgment follows a similar pattern. Judgment analysis can be paragraph is treated as a single string, depending on which the dice
done efficiently utilizing the document structure. Legal judgment is coefficient is calculated.
segmented as follows:
4. Create a graph.

AUTHOR 2

Electronic copy available at: https://ssrn.com/abstract=3563105

 3rd International Conference on Innovative Computing and Communication
Treat each paragraph as a node of a graph and add a connection/ edge would depend on its own peculiar facts. The Hon'ble Supreme
between two nodes if the value of the dice coefficient is higher than the Court in CIT vs. M/s. Vinay Cement Ltd17. had dismissed the SLP,
calculated threshold value. as it was not a fit case for grant of a SLP therefore cannot be
said to be the law decided on the subject and it was not a
5. Identify the relevant nodes binding precedent as per Article 141 of the Constitution of
Relevant paragraphs/nodes are estimated from the graph by considering India. Further, in the case of Hari Singh Etc v State Of
the degree of the nodes. When subgraphs are there in a graph, the node Haryana, the Supreme Court said:18, 15Kunhayammed and Ors v.
with the highest degree should be selected. In the case of a subgraph with State of Kerala and Anr [(2000) 6 SCC 359],
equal value for the degree, all nodes should be selected. https://indiankanoon.org/doc/1940266/ Dhakeswar1 Cotton Mills
Ltd vs Commissioner Of Income Tax,West Bengal, 1955 AIR 65,
6. Apply an Extractive summarization method for the chosen 1955 SCR (1) 941, Bench: Mahajan, Mehar Chand (Cj), Das, Sudhi
nodes/paragraphs. Ranjan, Hasan, Ghulam, Bhagwati, Natwarlal H., Aiyyar, T.L.
A simple extractive summarization technique that scores the sentence with Venkatarama, JJ. https://indiankanoon.org/doc/1837602/ 17 CIT
the help of term frequency is applied to the chosen paragraphs. vs. M/s. Vinay Cement Ltd. SLP, [(2007) 213 CTR 268] 18 Hari
Singh Etc v State Of Haryana, 1993 SCR (3) 61, 1993 SCC (3) 114
The doctrine of precedent is not applicable to an order passed
by this Court rejecting a Special Leave Petition. Any such order
cannot be held to be stare decisis so that it is a binding on
the Court. The mere rejection of the Special Leave Petition of
co- accused persons cannot seal the fate of the appeals of the
appellants which have been entertained after leave having been
granted by this Court. The appellants to whom leave has been
granted can urge all questions within the framework of Article
136 of the Constitution for consideration. By this Court and a
relief to which such appellants may be entitled cannot be denied
to them merely on the ground that a Special Leave Petition in
respect of co-accused persons with more or less similar charges,
evidence and convictions has already been rejected.

75
Fig. 1 – Graph created using the proposed system for the Judgment
between AM Kalra vs Ministry of Human Resource on 14 February, The principle laid down in Girish Ramachandra Despande cannot be
2017 used by the public authority to deny the information sought. In
fact, if that principle is universally quoted in every case
Relevant paragraphs chose by the proposed system for Judgment claiming the protection of sec 8(1)(j) then information relating
between AM Kalra vs Ministry of Human Resource on 14 February, to any public servant in respect of his conduct or illegal
2017 (Indian Kanoon) activities, especially corruption could be withheld as secret.
Such an application will defeat very purpose of the Right to

30 Information Act. Is it proper to say official conduct of a

public servant in a public authority with reference to the

The Supreme Court may reverse, modify or affirm the judgment- disciplinary action taken on charges of a bogus TA claim is not

decree or order appealed against while exercising its appellate related to official or public activity of the authority? It is

jurisdiction and not while exercising the discretionary in public interest to reveal why the punishment was imposed on a

jurisdiction under Article 136 disposing petition for special public servant and then why it was revoked?

leave to appeal. The doctrine of merger can therefore be applied

to the former and not to the latter. The Hon'ble Supreme Court 11

also observed in Dhakeswari Cotton Mills16 that it is not

possible to define the limitations on the exercise of the The petitioner in the instant case has not made a bona fide

discretionary jurisdiction vested in this Court by Article 136 public interest in seeking information, the disclosure of such

of Constitution of India. It being an exceptional and overriding information would cause unwarranted invasion of privacy of the

power, naturally, it has to be exercised sparingly and with individual under Section 8(1)(j) of the RTI Act.

caution and only in special and extraordinary situations. No

rules or principles as to when such leave ought to be granted 76

and when it ought to be refused can be laid down, as each case

AUTHOR 3

Electronic copy available at: https://ssrn.com/abstract=3563105

 3rd International Conference on Innovative Computing and Communication
The Commission observes that the conduct of the public authority of People's Act, and passing of Lokpal Act by the Parliament.
in the case of Dr. R.R. Singh was in relation to the false claim The disciplinary action or complaints against a public servant,
of TA, enquiry thereafter, disciplinary action taken and its cannot fall under the categories of private domain such as
straightaway revocation which is directly connected to the family, marriage, procreation, motherhood, child bearing and
public activity of public authority and it also deals with education. Compared to an order of division bench of Supreme
misuse of public money. In view of the Supreme Court judgments Court in dismissal of SLP in Girish under Article 136, the
cited, various other legal aspects and the facts produced, the decision of the division bench in Writ Appeal in R Rajgopal
Commission concludes that the information sought pertains to under Article 141 has binding value as declaration of law. The
public interest/public activity hence it does not attract the information sought under Girish cannot primarily be considered
exemptions under sec 8(1)(j). as personal. Even if assumed as personal, the public relation,
invasion of privacy, public interest factors make that
51 information to be shared because of Section 8(1)(j) of RTI Act.

What is Personal? Conflict with Section 8 of RTI Act: The

Supreme Court categorically stated that Point 3 was not just an
exception to Point 1 but an independent rule by itself. This 5. Evaluation
rule is similar to Section 8(1)(j) of RTI Act, which says the
information which is related to public activity is not private Evaluation is the most tedious task, especially in the case of legal text.
information. The order in Girish is contrary to both express The existing method for evaluating the quality of a summary is the
provision of law and well laid down ratio by the apex court in ROUGE score. The ROUGE-1 method compares the human-generated
division bench. The order in Girish does not even mention R summary and system-generated summary by counting the matching
Rajgopal's proposition and there was neither any analysis nor tokens in both the summary and length of the summary. If more than one
its overruling. When compared and analysed, the ratio in headnote is available for the summary, then it can match more tokens, and
Rajgopal was in full-fledged writ appeal which was heard on the quality of the summary can be evaluated more precisely.
merits was not over ruled by the order of dismissal of SLP in
Girish. The ratio decidendi in Rajgopal is in tune with Section In this paper, evaluation of the system is done manually by comparing the
8(1)(j) while dismissal of SLP in Girish is in conflict with paragraphs identified by the proposed system with the already available
that express provision of the law. The Girish order does not headnotes manually as well as using ROUGE-1 scores. The proposed
explain why the provision of RTI Act is not the law and by what system can pick relevant paragraphs from the judgment using a paragraph
rule of interpretation the new principle that employee related scoring technique. The chosen paragraphs were able to convey the content
information is personal information has been evolved? Without of the judgment. The relevance of the paragraph is evaluated manually by
over ruling the ratio in Rajagopal, it is not jurisprudentially comparing the human-generated summary.
possible to establish a different principle of law, that too in
conflict with express provision of law. As it was not writ The final summary created out of relevant paragraphs using extractive
appeal there was no opportunity to bring the ratio in Rajagopal summarization (Sentence scoring) is compared with the help of ROUGE
to the notice of the Division Bench of Supreme Court. Another scores. ROUGE-1, ROUGE-2, and ROUGE-SU4 are used for the
vital point is that the memos, complaints or disciplinary action evaluation of summaries. In this paper, Income Tax (IT) judgments are
related information is not unconnected with public activity, used for evaluation as its headnotes are readily available on the website
which was also not discussed in the order. It might have http://itatonline.org/. On evaluating around 10 Income Tax judgments, an
happened because it was not a full writ appeal but a hearing on average ROUGE-1 score of 0.4 is obtained. The Table 2 shows the
SLP only. ROUGE-1, ROUGE-2, and ROUGE-SU4 scores of 10 judgments.

52 Also, the system is tested for Non-Income Tax Judgements manually. The
Judgment between A M Kalra vs Ministry Of Human Resource ... on 14
Privacy of Assets Information of public servants: The assets and February, 2017 is of that category.
properties owned by a person, spouse and children could be
private information of the public servant because this is in his Table 1 – Graphs generated for different IT judgments using the
personal domain as per Point 1, but when contrasted with Point 3 proposed method
of Rajagopal ratio assets related information is not personal.
The public servant is not expected to misappropriate the public
money, with which he will be dealing in his day-to-day
operations. In such background, the property he and his family
members earned could be revealed. This became the law because of
judicial pronouncements and also amendment to the Representation

AUTHOR 4

Electronic copy available at: https://ssrn.com/abstract=3563105

 3rd International Conference on Innovative Computing and Communication
training data. The Conditional Random Field approach is an efficient
technique for Legal Judgment segmentation, but it requires a massive
amount of segmented judgments for training. The proposed method in this
paper picks relevant paragraphs from judgment without any training data.
The paragraphs in the judgment are modeled as graph nodes, and nodes
with the highest degree are selected as relevant ones. Experimental results
shows that our method is able to achieve promising results.

REFERENCES

Farzindar, A., & Lapalme, G. (2004). Letsum, an automatic legal text

summarizing system. Legal knowledge and information systems,
JURIX, 11-18.
Saravanan, M., Ravindran, B., & Raman, S. (2008). Automatic
identification of rhetorical roles using conditional random fields for
legal document summarization. In Proceedings of the Third
International Joint Conference on Natural Language Processing:
Volume-I.
Hachey, B., & Grover, C. (2006). Extractive summarisation of legal texts.
Artificial Intelligence and Law, 14(4), 305-345.
Pandya, V. (2019). Automatic Text Summarization of Legal Cases: A
Hybrid Approach. arXiv preprint arXiv:1908.09119.
Polsley, S., Jhunjhunwala, P., & Huang, R. (2016, December).
Casesummarizer: a system for automated summarization of legal
texts. In Proceedings of COLING 2016, the 26th International
Conference on Computational Linguistics: System Demonstrations (pp.
258-262).
Galgani, F., & Hoffmann, A. (2010, December). Lexa: Towards
Table 2 - ROUGE Scores of summaries generated using IT automatic legal citation classification. In Australasian Joint
Judgments
Conference on Artificial Intelligence (pp. 445-454). Springer, Berlin,
JUDGMENT ROUGE -1 ROUGE- ROUGE- Heidelberg.
2 SU4 Galgani, F., Compton, P., & Hoffmann, A. (2012, April). Combining
IT JUDG 1 0.4 0.24 0.25 different summarization techniques for legal text. In Proceedings of
IT JUDG 2 0.21 0.06 0.07 the Workshop on Innovative Hybrid Approaches to the Processing of
Textual Data (pp. 115-123). Association for Computational
IT JUDG 3 0.6 0.36 0.34
Linguistics.
IT JUDG 4 0.32 0.15 0.06
Schilder, F., & Kondadadi, R. (2008, June). FastSum: fast and accurate
IT JUDG 5 0.52 0.23 0.20 query-based multi-document summarization. In Proceedings of the
IT JUDG 6 0.41 0.19 0.16 46th annual meeting of the association for computational linguistics
IT JUDG 7 0.31 0.0 0.04
on human language technologies: Short papers (pp. 205-208).
Association for Computational Linguistics.
IT JUDG 8 0.50 0.22 0.25
Megala, S. S. (2018). Classification of Legal Judgement Summary using
IT JUDG 9 0.56 0.19 0.28 Conditional Random Field Algorithm.
IT JUDG 10 0.31 0.13 0.11 https://indiankanoon.org/
http://itatonline.org/

6. Conclusion

Legal judgment summarization is a challenging problem in the field of

Natural Language Processing (NLP). The application of newly developed
techniques such as Machine Learning & Deep Learning can achieve better
results for Legal judgment summarization but only with a large amount of

AUTHOR 5

Electronic copy available at: https://ssrn.com/abstract=3563105