#dissertation progress

It’s been a year since I finished uni and since my dissertation has been presented. Life caught up with me and I’ve moved to a new city, got my first job and basically finally started adulthood. 

I want with this blog post to round up my dissertation progress and what I have learned by quoting bits and pieces from my evaluation.

As a computer science student I believe my final project rounded up my whole university experience and knowledge:

Logical thinking/problem solving

starting with pseudo-code that turned into real code and then into refactored code;

working with MySQL within Java and continuously trying to not repeat statements and methods that differ only by one table name or variable;

Working with external libraries, new software and understanding new APIs

JDBC, MySQL, XML

Natural Language Processing, Stanford Libraries

Machine Learning, Weka

Presenting my work

in technical terms - talking to my colleagues and teachers helped me consolidate my communication skills and consolidate my knowledge (through repetition)

in layman terms - how to explain things to people that might not understand programming rules and tasks made me see the project from outside, it’s uses and how it can be improved and adapted to a more general user.

Organisation

source control (and as I keep finding out even now GIT is an extraordinary tool)

Working on my own

the project was my own from start to finish, my plans, my ideas, my code;

there was no other students who to confer with or that could help;

the entire experience proved to me that when I have my next idea I don’t have to be afraid to approach it on my own.

Where everything can be found: My gitHub

I spent these last days reading article after article. It's been over 1000 for sure but ended up with 652 records in my database from which unique articles: 589.

Why the difference? Mostly due to the multiple domain classification. There were many articles that could have been classed as theft and fraud or as murder and assault.

What was tricky? The attempted murders, the articles about crime statistics, the ones with so many strange characters that you wouldn't recognize the words or understand the sentence. 

What I noticed?

- Conspiracy was added as a new class! Also property damage and misconduct (less found than initially thought)

- It get's more and more difficult to label when you find only murders and assaults and not enough fraud or theft;

- Quite a lot of duplicate articles with different IDs especially on conspiracy(I guess and attempt on the Queens life is interesting for all issues)

- I started to look at the articles words: with theft, assault and murder I kept noticing verbs: stealing, robbed, cut, beating, stabbed; with conspiracy was more about: dynamite, bomb, plot(nouns);  with fraud: false, false pretences, falsified, forgery, forged(a mix)

It's time for machine learning.. not much time left..

I have 15911 articles in my database! Words that I queried for were: crime, investigation, arrested and police.

Started the next step: Gathering my training data and test data by reading and labelling 500 articles with the help of my little Helper. This is a new java project that get's a random number between 1 and 15911 shows me the article, abstract, id, text and title and lets me input numbers from 1 to 5 to label the article. Everything then goes into the junction table that connects the Article table with the Domain table. Making things work faster. Already done 100 inputs.

  Now to talk a bit about the findings and the issues.

1. Might need a conspiracy class

"PLOT TO SEIZE THE ARSENAL AT RANGOON. KANGOOX, Thursday. A plot fcr seizing the Areenal and setting fire to the town has been dis- covered by the authorities. Many Burmese have betn arrested in consequence."

2. People assault animals :( .. I guess it goes under assault

"THE GREAT WYRLEY OUTRAGES. The cd tor of a Wolverhampton paper has received another anonymous letter, signed on behalf of the members of the Wyrley Gang, in which the writer asserts that young Edalji is as guilty of the crime of horse maiming, of which he was convicted, as any other member of the gang. The writer adds that the police arc watching the letterboxes, and the members of the gang are watching the police. He wai ns Inspector Campbell that his time is short, as he is \\"a r< gue and liar,\\" and says that the gang is determined to kill him."

3. Arrested because of misconduct and being drunk? ummm do I make a category for that?

"HER 81st APPEARANCE. Florence Jones made her 81st appearance at Cardiff Police-courr, this morning, wnen she wa1 cnarged with being drunk in Louiinunaqoare, on Thursday.Police-constable Wheeler (129) gave evidence in support of thu charge, and prisoner was sent to gaol for fourteen days, the magistrates on the bench being Dr. H. Jr Plline and Alderman Lewis."

4. Not sure how to label this: 

"CARMARTHEN. CRIME IN CARMARTHENSHIRE.The Welshman, in commenting upon the recent murder in Carmarthenshire, makes some startling assertions which, if trueand we cannot believe them without very strong evidence go far to prove that that part of Wales resembles Ireland, at least in one particularthe protection afforded to crime. According to our contemporary, the failure \\"istice in the case in question may most ibly be traced to the fact that witnes ses J their lives will not be safe if they spe e truth. The country is in a state of teri every man, woman, and child seems afraid to speak on the fearful topic;\\" and one man told a police officer that \\"one life has gone, and it was a pity to waste another.\\" Cannot the Welshman push its investigations a little further, and let us know why the murderer is protected? Are the people so much opposed to capital punishment that they would rather commit another murder than allow justice to be carried out, or is there some other reason ? And how comes it that in Wales generally crime is so rare, and in Carmarthenshire murder can be committed with impunity ?"

5. Treason, Bombing attempt, Property damage

"A BARCELONA MYSTERY. Barcelona, Sunday.At 7 o`clock this evening the police here were informed that a suspicious looking objeet had beon foond on tbe Ramble. On investigation the object proved to be an iron cylinder filled with liquid, It was taken to the police station and thance to the Law Courts, where it was placed on the`ground. Immediately a terrific explosion occurred, injuring an official  and doing much damage. Renter,"

6.  Crime against religion?

"A DABWIWIAN PaOTBSSOB INDICTED.  An Austrian Ultramontane paper reports that Professor Dr. Brnehl, the wellknown lecturer on aootomical science in Vienna University, has been indicted for a crime against Church and religion. The Minister of Public Works, Dr. Gautsck von Frankenthurn, asked the rector of the university to call upon Dr. Bruehl to explain some of his recsnt upon the Darwinian theories, in which he i» an ardent i( believer. Dr. Bruehl Is 70 years of age and highly respected."

7. Not enough fraud found and a lot of no crime labelled.

8. What to do with the articles that contain trials and time in court?

  If it's this difficult for me ..how would it be for the computer?

I'll be a SQL ninja by the end of this project. Fought with SQL statements within Java all this time but it's getting there. I like the feeling of accomplishment when after a few struggles I feel like I actually know how the code actually works in the end.

Had a basic understanding of SQL when I started this project and was very comfortable with Java so imagine my surprise when I got stuck in this area. But then again JDBC seems like a different type of programming. It felt clumsy and uncomfortable:  all those Statements and all those executeQueries and result sets and while loops and if statements. I feel like my code is a bit of a mess now (the Database class) and it needs a great deal of refactoring.

But I have done it in the end I have my database now(over 11000 articles and counting stored locally), my tables, even my little Labelling helper that takes a random article and offers me the option to add one or two or three .. of the domain_ids.

Machine learning is the ability of a computer to learn from patterns and improve prediction and behaviour through data .

Developing a machine learning model means having a good amount of labelled training and testing data. But the most important part is choosing the best attributes for the ML algorithm to find patterns that can differentiate between classes. For the project I have chosen to use nouns and verbs and their occurrence in each article as attributes. This mean using one of the Part Of Speech taggers libraries available online.

After a bit of research I found two in Java that are better suited for what I'm looking for: OpenNLP Apache and Stanford NLP.

Choosing one of them wasn't that easy as their results weren't too different but I ended up selecting Stanford due to the options available (multiple features under one umbrella) and its accuracy around 97% for the POS tagger.

The way the library was built was using a machine learning  model that tags the words in a sentence using morphology analysis techniques and an English vocabulary. The result is a string similar to:

The_DT above_JJ sketch_NN shows_VBZ the_DT position_NN of_IN the_DT boy_NN

  Because I am tagging both verbs and nouns I am using an if statement and an option variable.

The method will request two variables then: the text that will be analysed and the result wanted: verb or noun: taggIT(theText, "verb").

To end up with the actual required part of speech I have used just a few simple replace functions

.replace("_VBN", "").replace("_VBZ", "")

.replace("_NN ", " ").replace("_NNS", "")

and then split the sentence into words and only select the ones that don't contain an underscore "_"

for (String word : sentence.split(" ")) {

                     if (!word.contains("_")) {

                           attributeString.append(" " + word);

                     }

The method will return a String that will contain the verbs (or nouns) separated by spaces.

I have chosen to use a String as the result of this method due to the next stage of my attribute processing which is lemmatization. To find a words lemma is to find its general base form. This is particularly useful when using verbs as they can have different forms and tenses that can increase the margin of ML error when having such a variety of attributes in the training set that in the end mean the same thing.

e.g.: be, is, been, was, were / escape escaped escaping

The lemmatizing method uses the Stanford coreNLP library, it splits the string into sentences and tokens(words) and reduces them to the base form using the LemmaAnnotation.class

Not many weeks left, the plan is (as discussed in last meeting with supervisor):

 1. Get all training data in database:

Make list of words the initial search will be done on

Don't add duplicate articles just because they have multiple words from the above list

Test if article is split on two pages and connect the article text

Extract verbs and remove duplicates

2. Database design and set-up

Choose database and design structure

Add all articles to database

Label the training data manually (>500 articles)

3. Machine learning stage 1: Developing the model

Research and play with Weka  (look at Apache Mahout, Java-ML )

Develop the model that classifies the articles.

4. Machine learning stage 2: Testing of Model

Getting new articles to test the model with

Tweaking the model

5. Machine learning stage 3: Applying the model on the rest of the data.

Verify article is not part of training model

Get final result into final database.

All these should be accompanied by constant documentation writing and regular blog posts. (Even if the coding is really the fun/exciting part for me! )