Skip to content
Snippets Groups Projects
Select Git revision
  • master default protected
1 result
Compare
Name Last commit Last update
.vscode
AndroidApp
CheckerCorrector
Crawler
.gitignore
INSTALL.md
README.md
README.md

Group 6 - Language Correction

Documentation

Summary

Our group aims to develop a Language Correction software that efficiently analyzes text for correct language usage by crawling through pages of the target language(s) on the Internet. The evaluator will compile common usage patterns of words and phrases from crawled texts and use them as a reference to identify words or phrases that are consistent or inconsistent with the gathered data. This project will assist users in improving their language skills and ensuring the accuracy of their written content.

Credits

Alex Melnick

Michael Harkess

Reza Sajjadi

Description of Implementation

How it Works

Crawler

Our crawler works as follows: URLs to crawl, URLs that have been crawl, and dissallowed domains are saved in HashSets. We used HashSets because they enforce uniqueness (we do not want to crawl the same URL multiple times) and operate in O(1) time for adding, removing, and contains operations. Seed URLs are added either by command line arguments or from a specified file. The seeds are loaded into the HashMap. From there, until we run out of URLs to crawl or hit our page crawl limit, we crawl each page one at a time. The crawl algorithm works as follows: the next URL is removed from the HashSet. We then check if its domain is known to allow crawling. If the domain has not been crawled before, we check the website's robots.txt file to see what pages are allowed to be crawled and save those preferences so we can respect them later. If the has been crawled before, then then we simply check if the URL is in the allowed Domains HashSet. If it is not allowed, the URL is skipped. If it is allowed, we then update the wait delay and write to a file as much data as allowed per website (default 1KB). After a page is finished being crawled (and we are below the crawled page limit), the crawler fetches the next page and the process repeats.

Checker

Our checker is using two different methods to assign confidence points. A confidence point is 0 if the checker didn't find any problem and 100 if it was completley skeptical. The first method uses a State Machine and the second is using an n-Grams inspired implementation. The final score will be a weighted sum of these scores provided by two different methods.

State Machine Checker

In order to have a good working State machine we need to first update the grammar and roles of each word manually. This step will be automatized to some extent in the next milestones. Tokens are provided in CheckerCorrector/SQLite/mydatabase.db, and also the basic graph provided by CheckerCorrector/DirectedGraph/BasicGraph.java. A sentence will first go through a typo checker and get updated if needed (it will also affect the confidence score). Then the sentence will be tokenized, and using the provided graph it will check whether the sentence is following the correct format or not, for each miss on any edge of the graph a penalty will be added to the confidence score. The typo corrector is calculating the distance of a word to dictionary(supposly provided by crawler). Here we used the small dictionary of homewrok2, and replace the misspelled word with closest distant word with some condition.

n-Grams checker

This checker used the crawled data and gave a score by summing up all the n_grams probabilities of phrases in a sentence. In order to store the data of the crawled data we used SHA-256 hash and store the result in CheckerCorrector/SQLite/hash_database.db.

Proof of Effictiveness

In order to show the effictiveness of our tool we used ChatGBT to write a script that is using a third-party Python library to make the exact same score for each sentence. These can be found in CheckerCorrector/samples/ directory. Both Json generated by our tool and the third party are available.

Corrector

The current corrector uses the typo corrector which was used in the checker and also the state machine to suggest possible corrections. Corrections of the typo corrector are based on the most similar path through the state machine. In order to find a similar path we are first doing a DFS on the graph starting from the first token and storing all the possible paths. Then, based on the most similar path we decide whether we should change/delete/add a token and suggest a token with a similar role. The correction at this point is limited to the state machine complexity, but it will be improved for the next milestone.

Learning Proceess and Database Updating

In order to have a checker/corrector that is working properly it needs to process the crawled data. This processing cosists of different stages. The process for n-grams is to break compute the hash of each new phrase and add up the number of its occurecs. The other updates is the update for tokens. The tool needs so somehow learn what is the role of each word in a sentence. It can be given a to it directly which is happen when a crawled data from a dictionary is given to it, or it can happen just by giving the raw data. Giving the raw data can cause some mis proccessing since the tool will learn the token by itslef. That is why we have an option where the user can decide whether the role predicted for the token by the tool is valid or not through a GUI.

Implemented Features

  • Provide real-time status and statistics ( e.g., what URL is being processed, rate of processing, size of storage, etc.) feedback for the crawler. [10%]
    • Implemented by calculating statistics as the crawler crawls each page. We report the length of the page processed in bytes, the total amount of data crawled thus far, the number of links extracted from the page, the number of pages crawled, the number of URLs left available to crawl, the current crawling rate in pages per second, the current crawling rate in links extracted per second, and the current crawling rate in bytes per second. Usage: use "--stats" argument for the crawler.
  • Provide a list of reasonable corrections to a suspicious text, ranked in order of how different they are from the original text. [15%]
    • There are four possible corrections. First correction to typos will come, then correction that include remove/insertion or changing a word will come. Since corrections are fine-tuned each individual is improtant. Also, since a typo correction will lead to a big change it comes first.
    • The implementation is just by dumping the changes into correction_details.txt.
  • Graphical User Interface that highlights suspicious and non-suspicious textual elements in a given text. [15%]
    • This interface is interactive and choose each phrases with 3 words and check it the confidence for it and color it based on it on the specterum.
    • The implementation of this part was with JTextArea of java.awt.
  • Extend your crawler to crawling social media posts of some large network ( e.g., Twitter, Facebook, LinkedIn, Truth, MeWe…). [15%]
    • Implemented by adding a social media platform as a seed to our crawler. We are using Tumblr as our social media platform. It is considered a major platform with more than 500 million monthly users. Choosing a social media platform was difficult since most platforms no longer allow crawling as a safeguard against their data being used to train LLMs. After careful examination of possible sites, we decided that Tumblr was the best choice that still allowed crawlers. Usage: use "--social" argument for the crawler.
  • Provide a graphical human feedback system for deciding among possible phrase corrections, with feedback into the suspicion levels reported by the system. [15%]
    • This feature will run a GUI that represent the possible correction as a button where user can accept or reject them. By appling the changes a new set of suggestions will show. User can also go to the next sentence by another button.
    • For the implementation we use a GUI interface and implement it in the corrector so that we can interact with GUI and get and pass the information.
  • Develop and Android client for your checker. [15%]
    • For the most part, the Android client of the checker is interfacing with the Checker CLI tool to produce its output in a pop-up window on the app. To do this, the Checker tool was modified so that it did not write the checked result to a JSON file (and instead kept it as a JSON string) as well as not require CLI arguments for the script to function. In addition, helper class functions were used to display the output to the user in the app.
  • Extend your system to a language in which none of the team members have fluency. [15% per language, up to 3] - 1x
    • We added the ability to crawl Dutch websites by adding a Dutch webpage seed to our crawler. Usage: use "--dutchSeed" argument for the crawler. Additional Dutch websites to crawl can be specified by adding their URLs to a file and using the "--file [filename.txt]" argument for the crawler. Also, for the checker and corrector things are similar to english but you need to usee --dutch option.
  • Provide a reasonable translation from English to another language based on common language structures. [30%]
    • We added the ability to translation from Dutch to English and from English to Dutch. We accomplished this by crawling a Dutch to English online dictionary and translating the words literally. Usage: use ""--dutchDict" argument with crawler to re-crawl the dictionary. For translation we used crawled dictionary and translate the data on both side."--translateToEnglish" and "--translateToDutch" in corrector will do this.

Changed Features from Initial Defense Report

Added
  • Feature 1
Removed
  • Feature 1

References

https://docs.oracle.com/javase/tutorial/networking/urls/index.html

Code

  • Most Updated (Stable) Branch: Master
  • Folder with All data needed to run repo:
  • Folder with All testing Code:

Work Breakdown

Webcrawler, Data Parser, Webcrawler stats, Webcrawler social media, Extend system to other language (partial), Provide a reasonable translation (partial) by Alex Melnick

Android App by Michael Harkess

All of the package and module in CheckerCorrector including: Language checker and corrector with all of the GUI features for them, including feature 3 and 5 and other features like giving user feadback on tokens, and the list of possible corrections, The extention to dutch and translation for dutch on checker/corrector the by Reza Sajjadi

This work breakdown is signed by Everyone: Seyed Reza Sajjadinasab Alexander Ross Melnick Michael Harkess

Known Issues/Bugs/Limitation

Checker/Corrector

  • The tool is unable to detect the language and it is assumed that the user is giving the desired language by specifying the target.
  • The corrector is mostly effective for typo correction. There are some hyper paramters withing the code that can be tuned to help the corrector decide how aggresive it should rewrite the code, but still it has known problem with suggesting a replacement.
  • Both corrector and checker are limited by how large the crawled data was and how much the user fix them.
  • The corrector is unable to fix the relation of groups of word more than 2 because it's not using the n-grams.

Android App

  • If a dialog box pops up and the device/screen is rotated, then the app will unexpectedly crash
  • The confidence score of every word in a given sentence/phrase is 0