Content uploaded by Frank Schilder
Author content
All content in this area was uploaded by Frank Schilder on Dec 11, 2016
Content may be subject to copyright.
Temporal information extraction from legal
documents
Frank Schilder1and Andrew McCulloh1
R&D, Thomson Legal & Regulatory
610 Opperman Drive, Eagan 55123, U.S.A.
{Frank.Schilder|Andrew.McCulloh}@Thomson.com
Abstract. The aim of this paper is to analyze what kinds of temporal
information can be found in different types of legal documents. In partic-
ular, it provides a comparison of different legal document types (case law,
statute or transactional document) and how one can do further reasoning
with the extracted temporal information.
Keywords. extraction of temporal information, temporal reasoning, le-
gal documents
1 Introduction
In the recent past, only few research has been carried out in legal reasoning
looking at formalizing temporal information. This should come in particular as
a surprise since laws, regulations and legal documents in general are normally
filled with temporal information:
(1) Celltech owns a family of patents called the ”Adair” patents and sought to
claim royalties from Medimmune under a patent licence dated 19 January
1998.
Although temporal information is actually ubiquitous in legal text, systems for
legal reasoning deal normally only on an ’ad-hoc-basis’ with this important phe-
nomenon [1]. With the exception of the special issue of Information & Communi-
cations Technology Law in 1998 [1,2,3], there is hardly any research on temporal
information in legal text carried out. A couple of recent attempts focused on the
specification of legal text in XML including temporal information [4,5,6]. Apart
from these few research projects the extraction of temporal information has not
been looked at in the literature. Traditionally, legal reasoning has been the focus
of AI-related research, where the content of laws and regulations may, for exam-
ple, become formalized in the event calculus [7]. Time may play a role within
such a formalization, but it has not been the main focus of the formalization
apart from a few exceptions.
However, it is important to note that legal reasoning is not the main focus of
the current paper. Instead of looking at temporal information in legal reasoning,
we are interested in temporal information in legal text and doing reasoning
Dagstuhl Seminar Proceedings 05151
Annotating, Extracting and Reasoning about Time and Events
http://drops.dagstuhl.de/opus/volltexte/2005/313
2 F. Schilder, A. McCulloh
with the temporal information. We want to look at different types of legal text
and investigate what kind of temporal information they can contain and after
discussing how this information could be automatically extracted, how one could
do reasoning with the temporal information in order to add more value to the
document.
Section 2contains an overview of different kinds of legal documents and
provides a brief introduction on how temporal information and constraints can
be important for researching these legal documents. Section 3focuses on three
types of legal documents and discusses in more detail how temporal information
can be extracted from them. Section 4concludes and discusses possible avenues
of future research.
2 Legal documents and temporal information
Legal documents can be categorized in different ways. For this paper, we make
the following distinction for different U.S. legal documents:
–Statutes (issued by the federal government)
–Proclamations, code of Federal Regulations, administrative decisions (issued
by the President, Executive Departments and administrative departments
(e.g. National Labor Relations Board (NLRB))
–Case law (authorized by trial courts, appellate courts or the supreme courts)
–Transactional documents (written by lawyers)
–Documents used as evidence for a case
–News documents that mention parties or people relevant to a case
There are different ways of how to look at temporal information and legal
documents. For one thing we can look at the documents and their creation date
or the date when the law described by them takes effect. Legal documents can be
ordered along a time line according to these dates. This ordering of documents
could be called extrinsic temporal ordering.
Another ordering would be an intrinsic temporal ordering of the events de-
scribed within the document and placing them onto a time line. This type of
temporal extraction is clearly more sophisticated and requires deep NLP process-
ing techniques.
Another way of processing temporal information derived from legal docu-
ments is the mining of information about the participating parties mentioned in
the document. Based on the creation date, one can derive that a lawyer works
for a particular company at that time. A different case may show the same
lawyer working for a different company at a latter point in time. Other text
types such as news messages about companies, law firms or lawyers may also
give information about the current affiliation of the people mentioned in the
text. This information could be used to update databases on companies, law
firms or lawyers.
All these three dimensions of temporal extraction and reasoning can be found
if we look at the normal life circle of a case. Traditionally, the search for precedent
Short Headline Title for Dagstuhl Seminar Proceedings 3
cases is the centerpiece for the American legal system and most often the starting
point for the legal researcher. Hence, it is absolutely essential to find precedent
cases relevant to the current case that are also not superseded by decisions of
a higher court made at a later date. Services such as KeyciteTM offer a legal
researcher the tool to search the history and status of U.S. and state court cases
and statutes. In order to ensure accuracy this information is annotated by editors
a couple of hours after the decisions have become public.
Apart from this classic case of ordering legal cases according to a time line,
there are other applications where the automatic temporal ordering of docu-
ments can become crucial for a legal researcher. In the following, we will look
at three different kinds of legal text: legal narratives, statutes and transactional
documents. We will discuss the different kind of temporal expressions these doc-
uments can contain and how a standard off-the-shelf temporal tagger performs
on these different kinds of data.
3 Types of temporal information in legal documents
This section discusses three different types of legal documents in more detail.
First, we discuss fact-based narratives in case law which are most similar to
news messages, because they mention mainly actual events that are linked to
temporal expressions. Second, we investigate what kind of temporal expressions
can be found in statutes. They are concerned with normative legal concepts
rather than with concrete events. Consequently, event types are described that
are linked to a temporal expressions. We found a higher number of durations than
that is normally the case in news messages. Third, we looked at transactional
documents that are similar to the normative laws presented in statutes but also
contain more concrete dates and events (e.g. of a purchase event).
3.1 Legal narratives in case law
Narrative language describing the facts of the case most often contains temporal
expressions. At the beginning of a case the judge normally describes the facts
and the reasoning that follows should be based on the relevant laws, statutes or
regulations relevant to these facts.
(2) On November 12, 1998, Illinois State Police Trooper Daniel Gillette stopped
defendant on Interstate Route 80 in La Salle County for driving 71 miles
per hour in a zone with a posted speed limit of 65 miles per hour. Trooper
Gillette radioed the police dispatcher that he was making the traffic stop.
Such narratives are very similar to news messages and off-the-shelf temporal
tagger could extract temporal expressions reasonably well from this type of text.
In addition research focusing on temporal information derived from narratives
[8] could be leverages for deriving a formal representation of the chain of events.
Having derived the temporal constraints on the event described in the case,
4 F. Schilder, A. McCulloh
searches could be carried out that contain temporal constraints. A query such
as ”Banana /s slip /before fall” would return only cases where a slipping event
occurred before an falling event. Note that this is a (temporal) relation between
events and not sentences.
3.2 Temporal restrictions in statutes or regulations
Statutes and regulations contain several different types of temporal expressions.
In contrast to the fact-based narratives one finds in case law, they often contain
periods of time (e.g. 30 days ) or sets of times (e.g. every year). These two types
of temporal expressions are used to add time constraints to event types rather
than to an actual event, as this is the case in news messages or the facts sections
of a case.
(3) ATTORNEY GENERAL OPTION TO ELECT TO APPLY NEW
PROCEDURES.- In a case described in paragraph (1) in which an eviden-
tiary hearing under section 236 or 242 and 242B of the Immigration and
Nationality Act has not commenced as of the title III-A effective date, the
Attorney General may elect to proceed under chapter 4 of title II of such
Act (as amended by this subtitle). The Attorney General shall provide notice
of such election to the alien involved not later than 30 days before the
date any evidentiary hearing is commenced. If the Attorney General makes
such election, the notice of hearing provided to the alien under section 235
or 242(a) of such Act shall be valid as if provided under section 239 of such
Act (as amended by this subtitle) to confer jurisdiction on the immigration
judge.
The anchor for the duration in (3) is found in the date an evidentiary hearing
is commenced. It is important to note that the link between the temporal ex-
pression and this event is conditional. Only if such an evidentiary hearing exists
does the 30-days restriction apply.
Statutes may also contain date expression. These can be linked to an actual
event, as for an effective date (or termination date) in (e.g. (4)). But mostly,
even these date expressions are linked to an event type as a temporal constraint,
as in (5).
(4) Amendment by Pub. L. 99177 effective Dec. 12, 1985, and applicable with
respect to fiscal years beginning after Sept. 30, 1985, but with subsec. (c)
to expire Sept. 30, 2002, see section 275(a)(1), (b) of Pub. L. 99177, as
amended, set out as an Effective and Termination Dates note under section
900 of Title 2, The Congress.
(5) ) (. . . ) is an alien who entered the United States on or before December
31, 1990, who filed an application for asylum on or before December 31,
1991, and who, at the time of filing such application, was a national of
the Soviet Union, Russia, any republic of the former Soviet Union, Latvia,
Estonia, Lithuania, Poland, Czechoslovakia, Romania, Hungary, Bulgaria,
Albania, East Germany, Yugoslavia, or any state of the former Yugoslavia;
Short Headline Title for Dagstuhl Seminar Proceedings 5
In a preliminary study of the United State Code we investigated the perfor-
mance of an off-the-shelf temporal tagger (i.e. TempEx by [9] ) on a small test
set drawn from the United States Code by hand-annotating this test set with
respect to the links between temporal expressions and events or event types
(i.e. TLINK).
First we ran the TempEx tagger and computed precision and recall for a
randomly selected set of 26 statute sections extracted from the 8th United States
Code on Aliens and Nationality. Of the 64 temporal expressions in the sampled
sections, the temporal tagger identified 24. Of these four contained incorrect date
attributions. Results on this test are shown in table 1(a). Take into consideration
that the Tempex tagger was written for news messages and that such a test can
only be seen as a baseline for temporal taggers that are more fine-tuned for legal
language in statutes or regulations.
correct occurrences percent
Precision 20 24 83.33%
Recall 20 64 31.25%
count DT PT FT AE
raw 22 26 11 5
59 5
total 64
Table 1. (a) Tagging accuracy (b) Distribution of temporal expression
types
Then we hand-annotated all temporal expressions in these 26 sections accord-
ing to the subordinated link and temporal link between the temporal expression
and the event (type). We defined the following categories:1
PT Period linked to event type
FT Set of times linked to event type
DT Date links to event type
AE Date linked to actual event
The results of our preliminary study of the distribution of different types of
links between temporal expressions and event (types) can be found in table 1(b).
From the distribution of these different link types one can conclude that temporal
expressions in statutes serve a different function than in news messages or in the
facts sections of cases. Statutes define event types that can be restricted by
temporal constraints. A set of people may be defined by their actions within a
certain time frame in addition to other conditions that have to hold (e.g. (5)).
Such conditional definitions do not occur that often in factive text.
Nevertheless, the TimeML specification allows for such a link via an SLINK
[10]:
1We did not find any periods or sets of times linked to actual events (e.g. John wrote
the note within 2 minutes.)
6 F. Schilder, A. McCulloh
(6) On Dec. 2 Marcos promised to return to the negotiating table if the conflict
zone was demilitarized.
<SLINK eventInstanceID="ei1" subordinatedEventInstance="ei2"
signalID="s1" relType="CONDITIONAL"/>
Important signals for conditional SLINKs are conjunctions when or if, as
described in the TimeML annotation guide. Those signals, however, are not
found in statutes. Instead these temporal expressions are often used within a
modal context (cf. The Attorney General shall provide notice of such election to
the alien involved not later than 30 days).
Extracting these links can be useful for the shallow processing of statutes
where conditions including temporal ones are extracted and a matching algo-
rithm could filter those statutes or regulations relevant to a given case (e.g. for-
mer citizen of East Germany entered the United States om November 11th, 1990
and filled an application for asylum 20 days after he entered the country fulfills
all conditions stated in (5) ).
Another important temporal dimension one encounters with this type of
document is the history of the statute. Arnold-Moore describes a system that
keeps track of the amendments that were added to a statutes of regulation. This
system is currently being used for legislations in Tasmania.2
3.3 Dates in transactional documents
Some of the most common documents handled by lawyers in their daily work in-
clude transactional documents. These include contracts, purchase or sales agree-
ments, and others which represent some kind of legal transaction. These docu-
ments almost always contain time expressions important for the legal stature of
the document. The most important of these is the execution date, or the date
when the transaction takes effect. In addition transactional documents may also
contain duration clauses. These, for example, may establish a time frame for one
party to establish or meet some condition necessary to satisfy the contact. In
practice, an attorney may want to search their document management system
to find all contracts signed after a particular date. We developed a system to
recognize these dates in transactional documents.
Dates in legal documents are typically expressed in a form containing the
day, month, and year. Contractual documents are overly specific, often using
complicated language to rule out any possible future cause a party may have
to contest the document. Dates need to be fully defined in this sense and so a
reader is never required to infer the year or month based on other evidence in the
document. In addition to the many, often wordy, date expressions, transactional
documents typically have a particular date format not found in most document
collections. Because these agreements need to take effect the same day they are
signed, the author often leaves the actual day of the month blank, to be filled in
at the time of signing.
2http://www.thelaw.tas.gov.au/index.w3p
Short Headline Title for Dagstuhl Seminar Proceedings 7
Common off the shelf date recognition systems tended to over-match areas
of the document that might otherwise appear to be date information. Simple
rule or regular expression based systems often misconstrued other information
as dates. Common errors included plot numbers and acreage sizes in real-estate
transactions or citations to civil code that are found in many of these documents.
In addition, these systems were unable to cope with the cases where the day was
left to be filled in at the time of signing. These are especially important as they
usually represent the execution date and the paragraph containing them may
also have other information pertaining to the timeliness of the contract.
We undertook a study to see if it was possible to build a system which could
recognize dates in transactional documents. We received approximately 1000
documents of various types from a local law firm. We manually identified several
date types that we wished to recognize. In addition to the standard American
date form of MM/DD/YY there were many more verbose examples. Many in-
volving ordinal values, which were often spelled out. Some of the examples can
be seen here
January 1, 2001
15<sup>TH</sup> DAY OF JANUARY, A.D. 2002
January 15, 2002
15th day of January, 2002
January 31, 2000
the 24<sup>th</sup> day of January 2002
January ____, 2002
this _______ day of January, 2002
first (1st) day of June, 2002
this 25th day of August, 2002
Given the small number of date types and the relatively few variations, we
decided to write a recursive descent parser to identify dates. We used the Antlr
compiler toolkit [11] for the implementation. We first constructed a tokenizer,
which only recognized a limited number of token types. The most important
being numbers both cardinal and ordinal, months, and underlines. The grammar
for the parser could then be very specific. It was written to recognize either fully
specified dates (containing day, month, and year) or a partially specified date,
which contained a blank to be filled in at the time of signing. In practice, the
program would tokenize the document, and then scan through the token lists
until it located a token that could begin a date production. At this point the
recursive descent parsing mechanism would take over and attempt to recognize a
date. If successful, a date object would be created and stored with the document
as searchable meta-data in the law firms document management system.
We compared the output of our parser to the output of the same off-the-shelf
temporal tagger as before [9]. The test collection consisted of 6 documents and
contained 20 date references. The time tagging system was able to identify 13
dates spread across the bodies of the documents but could not correctly identify
the 6 partial dates (i.e. those with a blank for the date to be filled in when
8 F. Schilder, A. McCulloh
the document was signed.) In addition there were three false positives where
addresses were considered as years by the off-the-shelf tagger. The complete
results are in table 2. Because our parser only looks for fully specified dates it
does not confuse other numbers as parts of dates. In addition the off-the-shelf
tagger requires a separate pass to tag parts of speech before processing input.
To its credit the system did recognize date ranges and other indicators which
could be useful in the analysis of transactional documents. Our tagger could not
do this.
correct occurrences percent
Precision 13 16 81.25%
Recall 13 20 65.00%
Table 2. Off the shelf tagger on transactional data
4 Conclusions
This paper reports on work-in-progress on temporal information extraction tech-
niques to legal documents. More specifically, we focused on three types of legal
documents and discussed the applicability of temporal taggers to these different
types of documents.
–Legal narratives in case law are similar to news messages and off-the-shelf
temporal taggers should provide a good coverages with respect to extract-
ing temporal expressions. In addition, the narrative structure should give
additional clues for ordering the events of the current case. Applications
that could benefit from a temporal extraction techniques are more detailed
searches with temporal connectors or temporal reasoning of witness accounts
in order to detect inconsistencies among the witnesses’ statements.
–Statutes or regulations have a different languages and differ in many respect
from other legal texts by providing legal rules that should match the facts of
the current case. This is also reflected in the temporal information encoded
into these rules. In a preliminary study, we found a large amount of tem-
poral expressions that are linked to event types rather than actual event. A
temporal and event tagger has to take this into account when applied to this
kind of data. Consequently, an off-the-shelf temporal tagger we used had a
very low recall. Future applications could use the temporal constraints men-
tioned in the statutes and match them against the actual case and suggest
relevant passages.
–Transactional documents describe legal rules as well as actual dates. In addi-
tion, many numbers mentioned in the document could be confused by dates.
We also found underspecified temporal expressions with the day information
Short Headline Title for Dagstuhl Seminar Proceedings 9
left open in these documents. A temporal tagger tuned to this kind of data
was able to deal with these special requirements sufficiently.
References
1. Vila, L., Yoshino, H.: Time in automated legal reasoning. Information and Com-
munications Technology Law 7(1998) 173–197
2. Brian Knight, J.M., Nissan, E.: Representing temporal knowledge in legal dis-
course. Law, Computers, and Artificial Intelligence / Information and Communi-
cations Technology Law 7(1998) 199–211
3. Farook, D.Y., Nissan, E.: Temporal structure and enablement representation for
mutual wills:. Law, Computers, and Artificial Intelligence / Information and Com-
munications Technology Law 7(1998) 243–268
4. Arnold-Moore, T.: About time: legislation’s forgotten dimension. In: Proceedings
of the 3rd AustLII Law via the Internet Conference 2001, Sydney, Australia (2001)
5. Arnold-Moore, T.: Point in time publication for legislation (xml and legislation).
In: Proceedings ot the 6th Conference on Computerisation of Law via the Internet,
Paris, France (2004)
6. Grandi, F., Mandreoli, F., Tiberio, P., Bergonzini, M.: A temporal data model and
system architecture for the management of normative texts (extended abstract). In:
Proceedings of SEBD 2003 - Natl’ Conf. on Advanced Database Systems, Cetraro,
Italy (2003) 169–178
7. Kowalski, R., Sergot, M.: A logic-based calculus of events. New Gen. Comput. 4
(1986) 67–95
8. Mani, I., Pustejovsky, J.: Temporal discourse models for narrative structure. In
Webber, B., Byron, D.K., eds.: Proceedings of the ACL 2004 Workshop on Dis-
course Annotation, Barcelona, Spain, Association for Computational Linguistics
(2004) 57–64
9. Mani, I., Wilson, G.: Robust temporal processing of news. In: Proceedings of the
38th Annual Meeting of the Association for Computational Linguistics (ACL’2000),
Hong Kong (2000) 69–76
10. Pustejovsky, J., Ingria, B., Sauri, R., Castano, J., Littman, J., Gaizauskas, R.,
Setzer, A., Katz, G., Mani, I.: The specification language TimeML. In Mani, I.,
Pustejovsky, J., Gaizauskas, R., eds.: The Language of Time: A Reader. Oxford
University Press, Oxford (2005)
11. Parr, T.J., Quong, R.W.: Antlr: A predicated- ll(k) parser generator. Softw.,
Pract. Exper. 25 (1995) 789–810
