Change Log

When What
March 31, 2005 Donated by Jane Huffman Hayes

Notes from the Author


Open source MODIS dataset[NASA. Dr.Hayes and Dr.Dekhtyar modified the original dataset and created an answerset with the help of analysts.

Past Usage

  • Improving Requirements Tracing via Information Retrieval, Jane Huffman Hayes, Alex Dekhtyar, and James Osborne, in Proceedings, 11th International Requirements Engineering Conference (RE 2003), pp. 151-161, September 2003, Monterey Bay, CA.
  • Helping Analysts Trace Requirements: An Objective Look (2004) Jane Huffman Hayes, Alex Dekhtyar, Senthil Karthikeyan Sundaram, and Sarah Howard, in Proceedings, 12th International Requirements Engineering Conference (RE 2004), pp. 249-261, September 2004, Kyoto, Japan.
  • A Framework for Comparing Requirements Tracing Experiments, Jane Huffman Hayes, Alex Dekhtyar, accepted, pending revisions, International Journal on Software Engineering and Knowledge Engineering (IJSEKE), special issue, 2004(5).
  • Jane Huffman Hayes, Alex Dekhtyar, and James M. Carigan, “Recommending a Framework for Comparison of Requirements Tracing Experiments”, in on-line proceedings of the Workshop on Empirical Studies of Software Maintenance (WESS 2004), Chicago, IL, September 2004.
  • Text Mining for Software Engineering: How Analyst Feedback Impacts Final Results, Jane Huffman Hayes, Alex Dekhtyar and Senthil Karthikeyan Sundaram, accepted, MSR’2005: Second International Workshop on Mining Software Repositories, St. Louis, MO, May 2005.

Relevant Information

Requirements tracing is defined as “the ability to describe and follow the life of a requirement, in both a forward and backward direction, through the whole systems lifecycle1.” It helps us in assuring that all requirements have been implimented.

We have implemented and evaluated a variety of IR methods including tf-idf vector retrieval, tf-idf retrieval with simple thesaurus, and Latent Semantic Indexing (LSI).

TF-IDF model: “vector model (also known as tf-idf model) for information retrieval is defined as follows. Let V = {k1,…,kN} be the vocabulary of a given document collection. Then, a vecto r model of a document d is a vector (w1,…,wN) of keywords weights, where wi is computed as wi = tfi(d)*idf , where tfi(d) is the so-called term frequency: the frequency of keyword ki in the document d, and idfi, called inverse document frequency is computed as idfi = log(n/dfi), where n is the number of documents in the document collection and dfi is the number of documents in which keyword ki occurs. Given a document vector d=(w1,…,wN) and a similarly computed query vector q=(q1,…,qN) the similarity between d and q is defined as the cosine of the angle between the vectors.[+ Simple Thesaurus: “This method extends tf-idf model with a simple thes aurus of terms and key phrases. A simple thesaurus T is a set of triples <t1,t2,a>, where t1 and t2 are matching thesaurus terms and a is the similarity coefficient between them. Thesaurus terms can be either single keywords or key phrases(sequences of two or more keywords). The vector model is augmented to account for thesaurus matches as follows. First, all thesaurus terms that are not keywords (i.e., thesaurus terms that consist of more than one keyword) are added as separate keywords to the document collection vocabulary. Given a thesaurus T={<ki,kj,aij>}, and document/query vectors d=(w1,…,wN), q=(q1,…,qN), the similarity between d and q is compute d[2](2]”


Latent Semantic Indexing (LSI) : “LSI is a dimension reduction technique based on Singular Value Decomposition (SVD) of the term-by-document matrix that can be constructed by putting tf-idf vectors of all documents in a single matrix. SVD transforms the original matrix into a product of two orthogonal matrices and a diagonal matrix of eigenvalues. By considering only the top k eigenvalues, we can obtain an appro ximation of the original matrix by a smaller matrix. Rows of the matrix can be compared to each other using the cosine similarity described above.[Feedback : To incorporate interactive work with analyst into RETRO, we have implemented relevance feedback for the IR methods studied. Relevance feedback works as follows: the analyst conveys to RETRO both positive (true link found) and negative (false positive found) information. The relevance feedback processor recomputes the vector qnew for the query q by adding to it positive information and subtracting negative information as specified in [2](3]”


Dataset details : This dataset is a modified dataset from [based on open source NASA Moderate Resolution Imaging Spectroradiometer (MODIS) documents. The dataset contains 19 high level and 49 low-level requirements. The trace for the dataset was manually verified. The “theoretical true trace” (answerset) built for this dataset consisted of 41 correct links. Each of the high and low-leve files contain the text of one requirement element. Take, for example, SDP3.2-2. The file with the name SDP3.2-2 is a text file that contains:

Each MODIS standard data product shall be produced within the data volume and processing load allocation shown in Table B-1 in plain text.

The files in the high and the low directory are in the same format.

The handtrace.txt file is the answerset. It maps high-level requirements to their low level children.

The handtrace.txt file has this format:

% SDP3.2-2 L1APR01-I-3 % SDP3.3-1 L1APR01-I-1 % SDP3.3-2 L1APR03-I-2 L1APR01-I-2 % SDP3.3-4 L1APR03-I-2 L1APR01-I-2 L1APR01-I-1 .. .. .. ..

where SDP3.2-2 is the identifier of a high level requirement, and L1APR01-I-3 is the identifier of the only low level requirement that traces to it. The items are separated by tabs.

So, for example, high-level requirement SDP3.3-4 has three children requirements: L1APR03-I-2, L1APR01-I-2, L1APR01-I-1.


a) RECALL is the percentage of the actual matches that are found.

    recall = number of links found/ total number of links

b) PRECISION is the ratio of the true links found and the number of candidate links generated.

    precision = number of links found/ total number of candidate links

Please refer to 2 for the definition of other primary and secondary metrics.

Feedback strategy:

“For each method, four different feedback strategies or behaviors, called Top 1, Top 2, Top 3 and Top 4 were tested. The Top i behavior meant that at each iteration, we simulated correct analyst feedback for the top i unmarked candidate links from the list for each high-level requirement. For example, for each high level requirement, Top 1 behavior examined the top candidate link suggested by the IR procedure that had not yet been marked as true. If the link was found in the verified trace, it was marked as true, otherwise as false. After repeating the Top i relevance feedback procedure for each high level requirement, the answers were submitted to the feedback processing module. At this point, the Standard Rochio procedure was used to update query (high-level requirement) keyword weig hts, and to submit the new queries to the IR method. The process continued for a maximum of eight iterations or until the results had converged.[technique:

A filtering technique is a simple decision procedure that examines each candida te link produced by the IR method and decides whether to show it to the analyst. In our study, in addition to the test run involving no filtering, we used filters with different threshold values. For example, filter with threshold value 0.1 will throw out all the candidate links that have relevance less than 0.1.

Expreimental Results:


Method Prec. Recall Sel. tf-idf 7.9% 75.6% 41.9% tf-idf+TH 10.1% 100.0% 43.1% LSI (10/100) 6.3% 92.6% 64.1% LSI+TH (10/100) 6.5% 95.1% 63.7% LSI (19/200) 4.2% 63.4% 65.2% LSI+TH (29/200) 5.4% 80.4% 65.8%

The table shows the results of running different methods on Modis dataset. Please refer to [2](3]”

filtering) and for more results.


# J. Matthias. Requirements tracing. Communications of the ACM, 41(12), 1998. # Jane Huffman Hayes, Alexander Dekhtyar, Senthil Sundaram, Sarah Howard, “Helping Analysts Trace Requirements: An Objective Look”, in Proceedings of IEEE Requirements Engineering Congerence (RE) 2004, Kyoto, Japan, September 2004, pp. 249-261. # Baselines in Requirements Tracing, Senthil Karthikeyan Sundaram, Jane Huffman Hayes and Alex Dekhtyar, accepted, PROMISE’2005: International Workshop on Predictor Models in Software Engineering , St. Louis, MO, May 2005. # Level 1A (L1A) and Geolocation Processing Software Requirements Specification, SDST-0591, GSFC SBRS, September 11, 1997. # MODIS Science Data Processing Software Requirements Specification Version 2, SDST-089, GSFC SBRS, November 10, 1997.

Number of Instances

  • 49 low level requirements
  • 19 high level requirements

Directory Content

The ModisDataset folder contains answerset (handtrace.txt), the high directory and the low directory. The high directory contains the high-level requirements, the low directory contains the low level requirements.


See above section “Relevant Information” for further reference.