School of Computing

Information Retrieval

NUS SoC, 2016/2017, Semester II Video Conferencing Room (COM1 02 VCRm) / Fridays 11:00-13:00

Last updated: Sunday, April 9, 2017 11:37:03 PM SGT - Updated the link to the competition framework, more details on the query files and evaluation. All signaled by updated or new and italicized text.
Monday, March 27, 2017 01:11:58 AM SGT - Updated the description of the corpus, put more details on the competition framework.
Sunday, March 26, 2017 02:09:51 AM SGT - Updated much of the assignment and added many new hints.

Homework #4 » Legal Case Retrieval Mini Project

Intelllex Logo

In our final Homework 4, we will hold an information retrieval contest with real-world documents and queries: the problem of legal case retrieval. As described in lecture, legal retrieval is a case where structured documents are prevalent, so serve as a good testbed for a variety of different retrieval approaches.

updated (now accepting submissions!) Jump to the competition framework, the current leaderboard.

Commonalities with Homeworks #2 and #3

The indexing and query commands will use an (almost) identical input format to Homeworks #2 and #3, so that you need not modify any of your code to deal with command line processing. To recap:

Indexing: $ python index.py -i directory-of-documents -d dictionary-file -p postings-file

Searching: $ python search.py -d dictionary-file -p postings-file -q query-file -o output-file-of-results

The difference from Homeworks #2 and #3 is that the query-file specifies a single query, and not a list of queries.

However, significantly different from the previous homework, we will be using a legal corpus, provided by Intelllex, a company with origins partially from NUS (Disclaimer: My research group may be helping with their work outside of our class homework assignment on better legal search technologies).

Problem Statement: Given 1) a legal corpus (to be posted in chunks to IVLE) as the candidate document collection to retrieve from, and 2) a set of queries, return the list of the IDs of the relevant documents for each query, in sorted order or relevance. Your search engine should return the entire set of relevant documents (don't threshold to the top K relevant documents).
Your system should return the results for the query query-file on a single line. Separate the IDs of different documents by a single space ' '. Return an empty line if no cases are relevant.

For this assignment, no holds are barred. You may use any type of preprocessing, post-processing, indexing and querying process you wish. You may wish to incorporate or use other python libraries or external resources; however, for python libraries, you'll have to include them with your submission properly -- I will not install new libraries to grade your submissions.

Intelllex, the company we are working with for this contest, is particularly interested in good IR systems for this problem and thus is cooperating with us for this homework assignments. They have provided the corpus (the documents are in the public domain, as is most government information, but the additional structuring the Intelllex team has done is their own work) and relevance judgments for a small number of queries. Teams that do well may be approached by Intelllex to see whether you'd like to work further on your project to help them for pay. Note: Your README will be read by both Min and may be read by the Intelllex team, but your code will not be given to their team to use; if they are interested in what you have done, you may opt to license your work to them.

More detail on the inputs: Queries and Cases

The legal cases and the information needs have a particular structure in this task. Let's start with the information needs.

Queries:

In Intelllex's own system, searchers (lawyers or paralegals) use the familiar search bar to issue a Boolean phrasal query, such as in the training query q1.txt: "intentional tort" AND "remoteness of damage". You'll notice that the training and test queries are simple boolean AND queries with keywords/keyphrases in quotes -- even single words are in quotes -- e.g., q2.txt, which has the contents "murder" AND "provocation" AND "loss of self-control". Phrasal queries are 2 or 3 words long, max; so you if you are able to deal with phrasal queries, you can support them using n-word indices or with positional indices. There are no ORs, NOTs or parentheses in the queries issued by us so you can simplify your query parsing code if you choose.

Query Relevance Assessments:

The query is the first line of the query file. updated The sample training query files also come with (very few) relevance judgments, as subsequent lines, which are only available to you in the sample queries (the test queries won't supply you with these answers for obvious reasons).

Each line marks a positive (relevant) or negative (irrelevant) legal case identified within the corpus. You should ideally have your system rank documents from the positive list before any other documents. As relevance judgments are expensive (lawyers assigned the judgments made available to you), the bulk of the Intelllex corpus was not assessed for their relevance. That is, there may be additional documents that are relevant in the corpus that are not listed. However, you will be assessed only on those that show up in the positive. We show the example for the above q1.txt.

"intentional tort" AND "remoteness of damage"
+ 3110812
+ 3110916
+ 2750632
- 2154455
updatedThe above indicates that there are three documents, with document_ids of 3110812, 3110916 and 2750632, that are relevant to the query; and one document, 2154455, that is known to be irrelevant.

Cases:

As per the guest lecture, legal cases are structured documents. For the purposes of our assignment, we are going to use an XML representation of a relevant case. Below are snippets of a document, ID 3110812, a case from the Singapore jurisdiction, relevant to the above example query:

    <str name="document_id">3110812</str>
    <str name="title">Ngiam Kong Seng and another v Lim Chiew Hock[2008] 3 SLR(R) 674; [2008] SGCA 23</str>
    <str name="url">http://www.singaporelaw.sg/sglaw/laws-of-singapore/case-law/cases-in-articles/negligence/1608-ngiam-kong-seng-and-another-v-lim-chiew-hock-2008-3-slr-r-674-2008-sgca-23</str>
    <str name="content">Ngiam Kong Seng and another v Lim Chiew Hock


                                            Ngiam Kong Seng and another v Lim Chiew Hock[2008] 3 SLR(R) 674; [2008] SGCA 23


                                            Suit No
                                            :
                                            Civil Appeal No 38 of 2007


                                            Decision Date
                                            :
                                            29 May 2008


                                            Court
                                            :
                                            Court of Appeal


                                            Coram
                                            :
                                            Chan Sek Keong CJ, Andrew Phang Boon Leong JA and V K Rajah JA


                                            Counsel
                                            :
                                            Cecilia Hendrick and Wee Ai Tin Jayne (Kelvin Chia Partnership) for the appellants; Quentin Loh SC (Rajah & Tann) and Anthony Wee (United Legal Alliance LLC) for the respondent.

                                Tort¬†‚¬†Negligence¬†‚¬†Duty of care¬†‚¬†Psychiatric harm¬†‚¬†Applicable test to determine existence of duty of care¬†‚¬†Application of two-stage test set out in Spandeck Engineering (S) Pte Ltd v Defence Science & Technology Agency¬†‚¬†First stage involving consideration of whether there was sufficient legal proximity with three factors set out by Lord Wilberforce in McLoughlin v O'Brian playing important role¬†‚¬†Second stage involving consideration of whether there are any public policy factors militating against the court imposing duty of care¬†‚¬†Threshold considerations of recognisable psychiatric illness and factual foreseeability                               Tort¬†‚¬†Negligence¬†‚¬†Duty of care¬†‚¬†Psychiatric harm¬†‚¬†Applicable test to determine existence of duty of care¬†‚¬†Whether type of damage claimed should result in different test from two-stage test set out in Spandeck Engineering (S) Pte Ltd v Defence Science & Technology Agency¬†‚¬†Application of two-stage test irrespective of type of damage claimed
                                Tort¬†‚¬†Negligence¬†‚¬†Duty of care¬†‚¬†Whether tortfeasor owing duty of care not to cause psychiatric harm¬†‚¬†Whether communication of matters relating to accident sufficient to found duty of care

                                Facts
                                The first appellant, who was riding a motorcycle, was involved in an accident which was allegedly caused by the respondent, who was driving a taxi. As a result of the accident, the first appellant sustained severe injuries which rendered him a tetraplegic. Both immediately after and during the period following the accident, the respondent represented himself to be a helpful bystander who had rendered assistance to the first appellant. The second appellant was, accordingly, led to believe that the respondent was a good Samaritan and developed feelings of gratitude towards him. The inquiries by the appellants' solicitors eventually led to the second appellant being told that the respondent had been involved in the accident. She subsequently suffered from major depression and suicidal tendencies resulting from, she claimed, having been "betrayed". The appellants eventually started an action in negligence.

...

                                Held, dismissing the appeal:

                                (1) The appeal, as far as the first appellant was concerned, should be dismissed. There was no basis for finding that the trial judge's decision was plainly wrong as, inter alia, there were inconsistencies in the first appellant's evidence; the evidence from the only independent eyewitness (ie, the respondent's passenger at the material time) did not support the first appellant's evidence; there was a lack of objective evidence in support of the first appellant's case: at [14] to [18], [148].

...

                                [Observation: Whether or not reform of the tort of negligence vis-a-vis psychiatric harm was to be effected was one that was best left to t\
he Legislature. Many issues which had to be grappled with laid wholly outside the expertise of the court and related to policy matters which required the Legislature's con\
sideration: at [120].]
                                Case(s) referred to
                                Alcock v Chief Constable of South Yorkshire Police [1992] 1 AC 310 (refd)
                                Anns v Merton London Borough Council [1978] AC 728 (refd)
                                Barnard v Santam Bpk 1999 (1) SA 202 (refd)
                                Bolitho v City and Hackney Health Authority [1998] AC 232 (refd)
                                Bourhill v Young [1943] AC 92 (refd)
                                Brown v The Mount Barker Soldiers' Hospital Incorporated [1934] SASR 128 (refd)
                                Caparo Industries Plc v Dickman [1990] 2 AC 605 (refd)
                                Childs v Desormeaux (2006) 266 DLR (4th) 257 (refd)
                                Cooper v Hobart (2001) 206 DLR (4th) 193 (refd)
                                Corr v IBC Vehicles Ltd [2008] 2 WLR 499 (refd)
                                Council of the Shire of Sutherland, The v Heyman (1984-1985) 157 CLR 424 (refd)
                                Customs and Excise Commissioners v Barclays Bank plc [2007] 1 AC 181 (refd)

...

                                [Editorial note: This was an appeal from the decision of the High Court in [2007] SGHC 38  .]

                                            29 May 2008

                                Andrew Phang Boon Leong JA (delivering the grounds of decision of the court):

                                Introduction

                                1       This was an appeal against the decision of the trial judge ("the Judge"), who dismissed the appellants' claims for damages (see Ngi\
am Kong Seng v CityCab Pte Ltd [2007] SGHC 38  ) ("the GD")). We dismissed the appeal and now give the reasons for our decision.

...

                                The facts

...

Conclusion
                                148 Our consideration of the evidence and the parties' arguments demonstrated that the Judge was not plainly wrong in arriving at her decision that the respondent was not responsible for the first appellant's injuries. In so far as the second appellant was concerned, we found that there was no duty of care owed by the respondent to her. In the circumstances, we dismissed the appeal with costs and the usual consequential orders.
                                149 That the Accident was a tragic one is undeniable, and we have the utmost sympathy for the very real plight of both the appellants. In this regard, we must commend Mr Quentin Loh SC, who was lead counsel for the respondent. At the conclusion of this appeal, he not only expressed his sympathy for the appellants, but also assured the court that, although he had no instructions as to costs, he would consult with his client and recommend that costs not be enforced against the appellants. He added that, if necessary, his own costs would be reduced.
                                Reported by Nathaniel Khng.</str>
    <str name="source_type">primary</str>
    <str name="content_type">case</str>
    <str name="court">SGCA</str>
    <date name="date_posted">2008-05-29T00:00:00Z</date>
    <arr name="jurisdiction">
      <str>SG</str>
    </arr>
    <arr name="tag">
      <str>Negligence</str>
    </arr>
    <str name="domain">singaporelaw.sg</str>
    <bool name="show">true</bool>
    <bool name="hide_url">false</bool>
    <bool name="hide_blurb">false</bool>
    <bool name="modified">true</bool>
    <date name="date_modified">2017-01-26T00:00:00Z</date>
    <long name="_version_">1560853972779008000</long>

You will notice that there are a lot of fields in the case. However, not all fields things are relevant to assessing a case's relevance to the query (and thus you may not want to index them), but are included for the sake of completeness. Importantly the content has much structure itself, consisting of subheaders, numbered statements and substatements. You may decide to try to treat such work using preprocessing in your indexing if you think you can capitalize on it. Note that different jurisdictions may have differences in formatting, or even a different court's format compared to others.

new Please note that the corpus distributed consists of three parts (intelllex-sg.zip, intellex-au.zip and intellex-etc.zip), due to different release times. Please dump all of these files into a single directory (e.g., intellex) to use with the -i command line switch. You should have a total of 29,869 documents after consolidation (ignoring duplicate documents).

Zones and Fields

As introduced in Week 8, Zones are free text areas usually within a document that holds some special significance. Fields are more akin to database columns (in a database, we would actually make them columns), in that they take on a specific value from some (possibly infinite) enumerated set of values.

Along with the standard notion of a document as a ordered set of words, handling either / both zones and fields is important for certain aspects of case retrieval.

Query Expansion

You might notice that many of the terms used in the text of the legal cases themselves do not overlap with the query terms used. This is known as the anomalous state of knowledge (ASK) problem or vocabulary mismatch, in which the searcher may use terminology that doesn't fit the documents' expression of the same semantics. A simple way that you can deal with the problem is to utilize query expansion.

In this technique, we use a first round of retrieval on the query terms used by a searcher to find some sample documents. Assuming that these documents are relevant, we can extract sometimes found these documents or use the entire documents themselves as queries, used in a second round of retrieval. The idea is that the sample documents have terminology that match the document corpus, overcoming the problem of vocabulary mismatch.

What to turn in?

You are required to submit README.txt, index.py, search.py, dictionary.txt, and postings.txt. Please do not include the legal case corpus.

Submission Formatting

You are allowed to do this assignment individually or as a team of up to 4 students. There will be no difference in grading criteria if you do the assignment as a large team or individually. For the submission information below, simply replace any mention of a matric number with the matric numbers concatenated with a separating dash (e.g., A000000X-A000001Y-A000002Z). Please ensure you use the same identifier (matric numbers in the same order) in all places that require a matric number.

For us to grade this assignment in a timely manner, we need you to adhere strictly to the following submission guidelines. They will help me grade the assignment in an appropriate manner. You will be penalized if you do not follow these instructions. Your matric number in all of the following statements should not have any spaces and any letters should be in CAPITALS. You are to turn in the following files:

  • A plain text documentation file README.txt: this is a text only file that describes any information you want me to know about your submission. You should not include any identifiable information about your assignment (your name, phone number, etc.) except your matric number and email (we need the email to contact you about your grade, please use your A*******@nus.edu.sg address, not your email alias). This is to help you get an objective grade in your assignment, as we won't associate matric numbers with student names. You should use the README.txt template given to you in Homework #1 as a start. In particular, you need to assert whether you followed class policy for the assignment or not.
  • All source code. We will be reading your code, so please do us a favor and format it nicely. Again, if you're using external libraries, make sure to include some nicely so they play well with our default ir_env environment (and acknowledge your external libraries as a source of help in your submission).

These files will need to be suitably zipped in a single file called <matric number>.zip. Please use a zip archive and not tar.gz, bzip, rar or cab files. Make sure when the archive unzips that all of the necessary files are found in a directory called <matric number>. Upload the resulting zip file to the IVLE workbin by the due date: 14 April 2017, 11:59:59 pm SGT. There will absolutely be no extensions to the deadline of this assignment. Read the late policy if you're not sure about grade penalties for lateness.

Grading Guidelines

The grading criteria for the assignment is below. You should note that there are no essay questions for this assignment.

  • 30% Documentation. This component is graded with a higher weightage in this assignment than in previous ones. This component breaks down into the following two subcomponents:
    • 5% For following the submission instructions and formatting your documentation accordingly.
    • 5% For code level documentation.
    • 10% For the originality of your ideas. Submissions attempting to do something unusual or interesting will be given higher marks. Sometimes attempting an interesting algorithm may have negative consequences on the performance of your system. In such cases, you can comment out the code that does not work as well. You should still document this code, so I can give you an appropriate documentation score. However, I will then assess your code's performance based on what actually runs from your submission.
    • 10% For your high level documentation, in your README document. This component comprises of an overview of your system, your system architecture and how your system deals with each of the optional components (query expansion, utilizing external resources, field/zone treatment, run-time optimizations, and the allocation of work to each of the individual members of the project team. Note that this component absorbs the weightage of the essay questions pounding previous assignments, so you should be especially thorough in your documentation.
  • 70% Performance of your code. updated Your systems will be evaluated for Mean Reciprocal Rank (MRR), at retrieving the documents marked as relevant in the query relevance portion of the query files. Note that files that do not match the Boolean phrasal search queries are not judged by the assessor (and thus may best be left out of the returned results) but could help you better rank documents that do match. This component breaks down into several subcomponents:
    • 30% I will compare it against a baseline TF×IDF ranked retrieval implementation, in which the entire document is treated without zones (i.e., all zone/field information is removed). If your system works at least as good as the standard baseline, you will receive all 30% for this component.
    • 35% We will use a competition framework to assign credit to teams and to show the leaderboard. Submissions will begin to be auto-run with a 24 hour delay (to prevent phishing for relevant documents). The framework will attempt to auto-run your code, with the training queries and a few other development queries (hidden from you, but run by the framework, please do not try to log these developmental queries. If in a team, please do not permute your matric numbers to allow yourselves additional runs; this is unfair to smaller teams. Please note the competition framework only uses either Python 2.7 or updated 3.6 (3.6 is backward compatible to 3.x), please check the submission instructions on how to declare which version your system requires. NLTK is installed on the system.
    • 5% will be due to the time efficiency of your system to answer queries (not testable by the competition framework, since you provide the answers to it). Your system should be able to answer a query within one minute. This requirement is mostly so that I can grade assignments in a timely manner.

Hints

  • If there are certain online services or APIs (e.g., web services), you wish to invoke in running your system, you can use these through function calls to external services. You may also wish to pre-compile certain (case-specific) resources for use by your system at run-time. For using online resources, it maybe helpful to use python's utilities to retrieve web pages (URLs), and save it as a temporary page for more analysis. You may use temporary file names of the form: temp-*.* in the current directory for query processing.
  • Working in a group can be helpful but can be counter-productive too. I have observed that group work tends to make homework submissions grades slightly higher than single submissions but also more average (it's harder to have an outstanding grade with a group). If you think you work better in a larger group, please do. Make sure to clearly partition the work and demarcate the API between your modules.
  • While you don't need to print scores out for the official runs that your system will be graded on, you may find it useful to include such information in your debugging output when developing your solution.
  • Similar to homework assignments #2 and #3, we will only be giving you a few queries to work with and (incomplete) query relevance judgments. We can only give you a few query relevance judgments, as the legal case relevance process also takes time to do for our human expert at Intelllex to assemble. However, we suggest you use your peers to pose some queries yourself and assess whether they are relevant (this may be hard -- Intelllex tells us that legal background is required to understand many documents). Documentation and participation marks will be given to student teams who do this. updated You might also try to search the Web for legal texts and landmark cases that are tagged (usually those tagged by Intelllex are considered such landmark cases anyways) to construct other queries.
  • Bulletproof your input and output. Make sure directories (e.g., arguments to -i) are correctly interpreted (add trailing slash if needed). Check that your output is in the correct format (docIDs separated by single spaces, no quotations, no tabs).
  • If you're fishing for ideas about how to do legal case retrieval, you might find past iterations of the Legal Retrieval TREC task interesting. This was a yearly contest, that featured legal case retrieval from 2006-2012. You are encouraged, but not obliged, to use ideas from this research community. Note that the "e-discovery" task that are the retrieval tasks in some years deals with evidence for legal cases, and not the legal cases themselves; so you may want to disregard some of the strategies involved there, but overall, we believe some of the general methods in these research papers may be helpful.
  • Intelllex's Ari has stated that some of the final XML fields: <show>, <show>, <hide_url>, <hide_blurb>, <modified>, <date_modified>, and <_version_> are not relevant to the search task. You may want to omit this information in your indexing. He had also indicated that the <areaoflaw> tags (when available) can be quite helpful. Also as the queries were made and assessed by a lawyer working in Singapore, you may find documents from the SG jurisdiction more relevant.
  • You may find it helpful to try the searches on Intelllex's own platform for comparison. You can sign up for their platform and use the same searches to generate what their engine pulls up as relevant. Reproducing their ranking in part is a good sign you are heading the right way; but they know there's significant room for improvement in their own system too.
  • Here's a recommended strategy to tackle this assignment: 1) Try getting the <content> string indexed as a normal cosine, tf×idf field; 2) Support phrasal search queries, by adding a positional index or a n-word index; 3) Find a method to cut the document into different zones, which may further help you weight different parts of the documents in matching your search. There are plenty of other things you can try, so spend some time with the data and the relevant and irrelevant documents to decide what seems to work well.
  • Finally, please note that the documents are in XML format,updated as are the queries. You will want to choose a suitable Python library (some of which are included directly in the Python system libraries) to process the data. Please plan on spending a bit of time to familiarize yourself with XML processing.