PhD Seminar - Masud Rahman, University of Saskatchewan

Editor's Notes

• #2: Hello everyone! Good afternoon! Thanks for attending this meeting. My name is Masud Rahman. I am a PhD Candidate from Software Research Lab. I work with Dr. Chanchal K. Roy. Today, I will be talking about automated query reformulations for code search.
• #3: So, Who am I? I came to Canada back in 2012 as a young graduate student. This the first winter. I got these two lovely young ladies by my side 24/7. I am a member of Software Research Lab, University of Saskatchewan. I work with Dr. Chanchal Roy. We got a pretty big group there, and do a lot of stuffs besides research.
• #4: A little bit of background about Me: Currently, I am a PhD Candidate at USASK. I completed my MSc in Software Engineering from the same university in 2014. Before that, I completed my BSc in Computer Science & Engineering from Khulna University, back in 2009.
• #5: Today, my talk will be divided into four sections. In the first section, I will discuss the research problem I am trying to solve in my PhD. In the second section, I will discuss about my PhD Thesis proposals to solve that research problem. In the third section, I will summarize my PhD contributions. Finally, we will have a Q&A session and interesting discussions.
• #6: Part 1: Research Problem
• #13: Now, lets say, a Boeing customer has submitted a bug report. Now, a Boeing developer is responsible to locate and repair the faulty code triggering that bug. As a frequent practice, developer chooses a few important keywords and attempts to locate the buggy code within the Boeing codebase. But the study shows that 88% of the keywords chosen by the developer could be incorrect. That is, they do not return the buggy code. So, the obvious next step is to reformulate the query through automated tool supports, so that the buggy code could be located. There are also tools that take a bug report and suggest appropriate search queries in the first place. So, we are interested into these part of the process, and my PhD focuses on this. As you can also see that, Google does not have any jurisdiction in this case.
• #14: So far, we deal with query reformulation for a single codebase. But you know, besides this local codebase, developers also look for source code in Internet-scale cross domain codebases. Now, that is a whole new game. Study shows that developers might fail 88% of the times to retrieve the correct code segments. So, we did two other studies, and we extensively used Stack Overflow.
• #16: Now, we are done with Background concepts, Part 1. Now, we are going into Part 2 -- PhD Thesis
• #17: So what we did? We did a systematic literature survey using 56 primary studies on query reformulation for code search. During this study, we found 3 major issues in the literature.
• #20: Let us see an example. This is a bug report, this is title and this is the description. Now, developer JOE would use this bug report to localize the bug from source code. Now he chose some ad hoc queries. Which one is the best do you think, here? PAUSE! Well, lets see. This one returns the correct result at this position. That means, the developer needs to check 1300+ results b4 reaching to the correct result he tries this query. … oh… this one is the best. So, selecting appropriate keywords from the bug report is not that simple.
• #22: Similarly, we can see the phrases and dependencies among the terms in the bug report texts as well. Our job is to identify the keywords from these texts, right? So, did we do? We consider the co-occurrences among the terms. That is, how terms occur with other terms within a certain context. We encode such co-occurrences as edges, and transform the texts into a graph like this.
• #23: Besides term co-occurrences, we consider another aspect called syntactic dependencies. For this, we used Jespersen Rank Theory, a theory developed back in 1925. According this theory, parts of speech of sentence can be divided into three ranks – nouns (first), verbs + adjectives in the second rank and the rest are the third ranks According to Jespersen, verb and adjective modifies noun. That is there are some syntactic dependencies for between element and reported and element and plain to covey the overall meaning of the sentence. Now, we capture such syntactic dependencies as well, and transform the report texts into a POS graph as well.
• #24: So, we have created two graphs, right? Now, we have two graphs developed from the bug report based on two different dimensions --Word co-occurrence and syntactic dependence. Once we have graphs, we apply this famous algorithm called PageRank algorithm. This is the backbone of Google search. Now, the algorithmic details are a bit complex, but I will try to provide an overview here. Why do you think, this guy is laughing? Because, it is getting the maximum votes. Similarly, in the graph, the node that is connected to most of the nodes is the winner. That is, a term’s importance will be determined by its connectivity with other nodes. More importantly, since this is a recursive algorithm, the importance depends on the weights of the connected node as well. Once the computation is done, we get a reformulation candidate from each graph. What is the reformulation candidate? – a ranked list of keywords like this. So, we collect two candidates from two graph, apply machine learning and suggest the best one as our suggested query from the bug report.
• #25: Now, for the experiments, we chose 8 subject systems from Apache and Eclipse. We collect about 3000 bug reports, and try to map them with the version control history at GitHub. Through such mapping we extract the ground truth for the bug reports. This is a standard process followed by the existing literature.
• #26: Now let me explain the metrics a bit since we will be using these a lot. Hit@K is the percentage of the queries for which at least one ground truth is found within the top K results. MAP is the standard precision + result position. The detailed is much complex which I can discuss later. MRR is the inverse of the rank of first ground truth within the result. QE also stands for query effectiveness is just the opposite of MRR
• #30: We did an empirical study with 5K+ bug reports in our ICSE poster. And we discovered that bug reports could be very different in terms of quality. There could be different types of bug reports. It could be noisy with stack traces which is 16% It could be really poor that does not contain any structured entities, which is 30% Or it could be rich bug reports that include source code, test case and other stuffs, which is 54%
• #31: So, clearly there are different quality levels for bug reports. Now, what do the existing approaches do? They do not look at the bug reports or its quality. Rather they apply the same treatment to all. But one size does not fit all, as you know. So what we do? We propose a much more balanced way to deal with this. We prefer equity rather than equality in our treatment to bug reports. So, here comes our work!
• #32: So, first we take a bug report as input. Then we apply regular expressions to identify the structured components. We then classify whether this is a a noisy report containing stack traces. a poor bug report containing only regular texts a rich bug report containing source code and texts. Once the quality level is identified, what’s the next step? Well, we do query reformulation unlike the earlier studies. We separate signals from noise from noisy report, feed the poor bug report with appropriate keywords. We mostly keep the rich bug report as is. So, that is the equity approach.
• #33: Now, lets go a bit deeper with this. From a noisy bug report, we first extract the stack traces. It might contain hundreds of traces like this. We choose any two consecutive traces. Lets call them I and J
• #34: Now, each trace will contain three piece of information – a package name, a class name and a method name. We know that, in a single line, the method and class are statically connected for sure. However, classes and methods are also hierarchically dependent across trace lines due to caller-callee relationships. We capture such static and hierarchical relationships from consecutive trace lines, and develop a trace graph like this.
• #36: So, from a noisy report, we extract The report title The encountered exception The most important keywords from the stack traces. Then we do the search with this newly constructed query. For example, the baseline noisy query returns the result at 53rd position. Whereas our query returns the correct result at the topmost position.
• #37: In the case of poor bug report, we also apply a similar PageRank approach. But we collect the keywords from the source code using pseudo-relevance feedback. The details can be found in the paper. So, the bug report texts are merged with the keywords from relevant source code. While the bug report texts return the result at 30th position. After feeding poor report with appropriate keywords, the correct result is returned at the top most position.
• #38: The rich bug report is inherently good for IR-based localization. But we found that selecting appropriate keywords can make it better. For example, if we reduce the bug report is reduced to these keywords, the correct result comes to the top of the list.
• #39: Now, for the experiments, we chose 8 subject systems from Apache and Eclipse. We collect about 3000 bug reports, and try to map them with the version control history at GitHub. Through such mapping we extract the ground truth for the bug reports. This is a standard process followed by the existing literature.
• #45: First, we construct a semantic hyperspace using Stack Overflow corpus. What is hyperspace? Now, if we have more than 3 dimensions, then we call that space as hyperspace. How do we do it? First we Stack Overflow data dump that contain software specific texts. Our corpus contains about 2.1 million questions and answers. We do pre-processing and feed the contents to FastText. Now FastText generates a three-layer neural network model. This model essentially represents the whole vocabulary like this in a hyperspace. Now how does it help?
• #46: Here we see that burger is close sandwich. Why? They are eaten together. I do that all the time. Well, that is not the case. They are mentioned in the similar contexts by the people across the whole corpus. The model recognizes such occurrences and thus put burger and sandwich close together. Similarly, dumpling and ramen are close to each other. Now, we propose this. This is original query, and this is reformulated query. Now, a good reformulated query will cluster together the original query. A bad reformulated query will NOT be able to cluster with the original query. So, clustering tendency within the hyperspace is our weapon here. We calculated Hopkins statistic and Polygon Area for calculating the clustering tendency.
• #48: Now, here is how we do it. We extract three reformulation candidates from the source code based on IR and data analytics. Then we apply the data analytics as a proxy to query quality, do data resampling, machine learning and then identify the best reformulated query.
• #49: Now this is a poor bug report, does not contain any useful hints for bug localization. Now this is the baseline initial query. This is from the state-of-the-art And this is our performance.
• #52: Now, I am not going to discuss those studies in details. But here is the glimpse. Developers generally look for relevant code on the web using natural language query. Please note that we are not talking about simply web search, rather talking about source code repository such as GitHub. Now, GitHub provides this result. Now, you see it tries to match the query keywords with comment and identifiers. But what we are dealing with source code right? So, we need source code friendly query for a better result. So, we identify relevant API classes against this natural language query through extensive data mining and data analytics. And once again, Stack Overflow is our friend in this grand challenge.
• #58: For the API suggestion, we natural language queries from four tutorial sites such as KodeJava and others. We collect 300+ queries, we also collect the ground truth API classes from them. Then we try to determine our approach can suggest appropriate API classes for those queries by mining crowd knowledge from Stack Overflow. For the query reformulation part, we collect 4K code examples from GitHub, combine with our ground truth code segments from tutorial site. Then we determine whether our reformulated query actually works or not.
• #62: Now, I care about the replication and reproducibility of my works in order to grow as a research community. So, all of works could be replicated and potentially reproduced (hopefully). I was busy during PhD and my works are uploaded at GitHub. Many of them are open source. I am a member of replication workshop and a PC member of replication track, ICPC.
• #63: Now, how far I would to go in the next 5-10 years?
• #64: Now, if you remember this diagram. So far, I focused on Bug Localization in my PhD mostly. In the next 5 years, I would focus on Bug Understanding and fixing, the next logical steps of software debugging. Right now, I am working on bug reproducibility. That is, how to make a reported bug reproducible. Without that, a bug can not be fixed most probably.
• #65: So, I compared between TF-IDF and PageRank. But so far, I did only for query reformulation. But there 20+ software engineering tasks that used TF-IDF in one way or another. Now, a cool thing would be to investigate how PageRank can influence that tasks given PageRank outperformed TF-IDF in query reformulation. It would be really interesting to see.
• #67: Another plain I have to develop a programmer bot. It will accept a natural language query from developer JOE and return a high quality relevant code. Now, in the background, this will happen. First query will be translated into search engine friendly query with relevant API classes and thousands of examples will be collected. Then we will use some to determine the best quality code example. Now, I did figured out this part. sophisticated machine learning The rest part is still pending and hope to get it done by some brilliant grad students.
• #68: Over the years, I worked with experts from my domain, and learned from the bests. I worked with academia as well as industry. For example, Vendasta is a leading company based in Saskatoon, and we got our NSERC Industry Engage grant in collaboration with them. So my since thanks and gratitude to all the collaborators.
• #69: Now, these are some achievements of my work. Got several competitive awards over the last few years. Got funding from NSERC. Thanks U of S and my professor for all the support. Also got accepted in the flagship conferences/journals such ICSE, FSE, ASE and EMSE
• #70: Thanks for your time and attention. Now, I am ready to take your questions.
• #73: For the API suggestion, we natural language queries from four tutorial sites such as KodeJava and others. We collect 300+ queries, we also collect the ground truth API classes from them. Then we try to determine our approach can suggest appropriate API classes for those queries by mining crowd knowledge from Stack Overflow. For the query reformulation part, we collect 4K code examples from GitHub, combine with our ground truth code segments from tutorial site. Then we determine whether our reformulated query actually works or not.
• #74: Now let me explain the metrics a bit since we will be using these a lot. Hit@K is the percentage of the queries for which at least one ground truth is found within the top K results. MAP is the standard precision + result position. The detailed is much complex which I can discuss later. MRR is the inverse of the rank of first ground truth within the result. QE also stands for query effectiveness is just the opposite of MRR
• #75: Here we see that burger is close sandwich. Why? They are eaten together. I do that all the time. Well, that is not the case. They are mentioned in the similar contexts by the people across the whole corpus. The model recognizes such occurrences and thus put burger and sandwich close together. Similarly, dumpling and ramen are close to each other. Now, we propose this. This is original query, and this is reformulated query. Now, a good reformulated query will cluster together the original query. A bad reformulated query will NOT be able to cluster with the original query. So, clustering tendency within the hyperspace is our weapon here. We calculated Hopkins statistic and Polygon Area for calculating the clustering tendency.
• #76: Now lets expand and generalize the problem a bit. So far, we discuss the code search within a local codebase. It could also be in the large-scale open source repository such as GitHub. Now, based on these contexts, there are different challenges in query reformulation. The local codebase is small, domain specific and organized. On the contrary, GitHub is huge, cross-domain and very noisy. So, yes, they need different strategies to suggest queries for them.
• #77: Now once such items are extracted, we split them. Now as we see, these single terms share some kind of semantics to convey a broader semantic. That is, they complement each other in this context. Now, we capture such semantic dependencies in the source code, and develop a term graph like this.
• #79: That is, each of three people, customer, past developer and JOE have their own vocabulary to describe a certain problem/concept. In fact, any people will discuss the same problem with the same vocabulary, this probability is only 15%-20% So, naturally, developer JOE finds it a great challenge to make a connection between bug report and the buggy code. This costs development time, money and valuable efforts.
• #80: Here we see that burger is close sandwich. Why? They are eaten together. I do that all the time. Well, that is not the case. They are mentioned in the similar contexts by the people across the whole corpus. The model recognizes such occurrences and thus put burger and sandwich close together. Similarly, dumpling and ramen are close to each other. Now, we propose this. This is original query, and this is reformulated query. Now, a good reformulated query will cluster together the original query. A bad reformulated query will NOT be able to cluster with the original query. So, clustering tendency within the hyperspace is our weapon here. We calculated Hopkins statistic and Polygon Area for calculating the clustering tendency.
• #82: So, from a noisy report, we extract The report title The encountered exception The most important keywords from the stack traces. Then we do the search with this newly constructed query. For example, the baseline noisy query returns the result at 53rd position. Whereas our query returns the correct result at the topmost position.
• #84: Now the question is, why is this so challenging? The answer is vocabulary mismatch problem. In fact, this is a common problem for any type of document search. Here we see both guys are looking at the same object, but they are explaining it differently. That is, they are both correct from their perspective, but wrong from other guy’s perspective. This also actually happens with bug reports as well. Both customer and developer will explain the same problem using the same terminologies, that probability is only 15% That is why selecting appropriate keywords from the bug report is very challenging.
• #85: Let us see an example. This is a bug report, this is title and this is the description. Now, developer JOE would use this bug report to localize the bug from source code. Now he chose some ad hoc queries. Which one is the best do you think, here? PAUSE! Well, lets see. This one returns the correct result at this position. That means, the developer needs to check 1300+ results b4 reaching to the correct result he tries this query. … oh… this one is the best. So, selecting appropriate keywords from the bug report is not that simple.