Multi-abstraction Concern Localization

Multi-Abstraction Concern Localization Tien-Duy B. Le, Shaowei Wang, and David Lo {btdle.2012, shaoweiwang.2010,davidlo}@smu.edu.sg

Multi-Abstraction Retrieval

Motivation • Concern Localization is the process of locating code units that match a particular textual description (bug reports or feature requests)

•We propose multi-abstraction Vector Space Model (VSMMA) by combining VSM with our abstraction hierarchy.

• Recent concern localization techniques compare documents at one level of abstraction (i.e. words/topics)

•In multi-abstraction VSM, document vectors are extended by adding elements corresponding to topics in the hierarchy.

•A word can be abstracted at multiple levels of abstraction. For example, Eindhoven can be abstracted to North Brabant, Netherlands, Western Europe, European Continent, Earth etc.

•Given a query q and a document d in corpus D, the similarity between q and d is calculated in VSMMA as follows:

•In multi-abstraction concern localization, we represent documents at multiple abstraction levels by leveraging multiple topic models.

V

L

 t f  idf (wi ,q,D) t f  idf (wi ,d,D)  i 1

k 1

d 

V

Preprocessing Hierarchy Creation

MultiAbstraction Retrieval

Level 1 Level 2 …. Level N Abstraction Hierarchy

+

Standard Retrieval Technique

Ranked Methods Per Concern Text Preprocessing • We remove Java keywords, punctuation marks, special symbols, and break identifiers into tokens based on Camel casing convention •Finally, we apply Porter Stemming algorithm to reduce English words into their root forms. Hierarchy Creation Step •We apply Latent Dirichlet Allocation (LDA), with different number of topics, a number of times, to construct an abstraction hierarchy •Each application of LDA creates a topic model, which corresponds to an abstraction level. •We refer to the number of topic models contained in a hierarchy as the height of the hierarchy

k

i

i

L K (Hk )

2 tf  idf (w , q , D )   i V

k 1

Hk 2 (   q ,t i ) i 1

L K (Hk )

2 tf  idf (w , d , D )   i i 1

Concerns

i 1

k

Where i 1

Method Corpus

H H     q ,t d ,t

q  d

q 

Overall Framework

K (Hk )

k 1

Hk 2 (   d ,t i ) i 1

•V is the size of the original document vector • wi is the ith word in d • L is the height of abstraction hierarchy H •Hi is the ith abstraction level in the hierarchy Hk •  d ,ti is the probability of topic ti to appear in d as assigned by the kth topic model in abstraction hierarchy H •tf-idf (w,d,D) is the term frequency-inverse document frequency of word w in document d given corpus D

Experiments Effectiveness of Multi-Abstraction VSM Number of Topics MAP Baseline (VSM) 0.0669 H1 50 0.0715 H2 50, 100 0.0777 H3 50, 100, 150 0.0787 H4 50, 100, 150, 200 0.0799

Improvement N/A 6.82% 16.11% 17.65% 19.36%

• The MAP improvement of H4 (over baseline) is 19.36% • The MAP is improved when the height of the abstraction hierarchy is increased

Future Work •Extend the experiments with combinations of Different numbers of topics in each level of the hierarchy Different hierarchy heights Different topic models (Pachinko Allocation Model, Syntactic Topic Model, Hierarchical LDA) Experiment with Panichella et al. ‘s method [1] to infer good LDA configurations for our approach [1]A. Panichella, B. Dit, R.Oliveto, M.D. Penta, D. Poshyvanyk, and A.D Lucia. How to effectively use topic models for software engineering tasks? an approach based on genetic algorithms. (ICSE 2013)