Homework 2

Description

• Probabilistic Language Identi cation

Your detectives arrived at company X only to nd an important piece of ev-idence, a business letter, has been shredded. After painstakingly recovering the pieces, the detectives reported to you that all they could do was to count how many times each alphabet character appeared in the letter. Your mission, should you choose to accept it, is to estimate whether the letter was written in English or Spanish.

1.1 Build your own digital shredder

We gave you a python program skeleton hw2.py. You should complete the program. The rst thing it should do is to read in a plain text letter and digitally shred it: print out the character counts.

• Your program should read from the input le named \letter.txt” in the same directory as hw2.py. We will supply this text le when we grade your program. For developing your code, copy any one of the samples=letter[1 4]:txt les as letter.txt. Do not hardcode your input lenames as let-ter0.txt, letter1.txt, and so on.

• The input le may contain any printable ASCII characters and can be short (e.g. a single word) or long (e.g. an article). In particular, there can be a mix of upper and lower case characters; numbers and other printable ASCII characters are possible too.

• The letter is known to be either entirely written in English or entirely written in Spanish. For the sake of this homework, we assume Spanish is written in printable ASCII characters, namely without accent marks etc.

• Your program should ignore case, i.e. merge \A” and \a” counts together, and so on. This is known as case-folding.

• Your program should only count characters A to Z (after case-folding), and ignore all other characters such as space, punctuations, etc.

• For this part, your program should output the string Q1 on its own line, followed by exactly 26 lines. Each line contains the upper case letter, a

space, the count of that letter (after case-folding). You can see the sample output in samples=letter[1 4] out:txt.

For example, assume letter.txt is the following:

Hi! I’ll go 🙂

This part of your program should output:

Q1

• 0 B 0 C 0 D 0 E 0 F 0 G 1 H 1 I 2 J 0 K 0 L 2 M 0 N 0 O 1 P 0 Q 0 R 0 S 0 T 0 U 0 V 0 W 0 X 0 Y 0 Z 0

1.2 Language identi cation via Bayes rule

We arrange the 26 counts into a 26-dimensional count vector

X = (X₁;:::;X₂₆);

where X₁ is the count of A, X₂₆ is the count of Z, and so on. X is the observed evidence. We are interested in the source language y 2 fEnglish; Spanishg. Our goal is to compute the conditional probability

P (Y = y j X):

where we removed C(X) and P (X). Equivalently, we can compute (recall p depends on y)

26

Y

f(y) = P (Y = y) p^X_iⁱ

i=1

and then normalize:

f(y)

^{P (Y = y j X) =} f(English) + f(Spanish) ^:

We also do not like multiplying many small numbers for fear of under ow, so we want to compute things in the log domain:

26

X

F (y) = log f(y) = log P (Y = y) + X_i log p_i:

i=1

We will use the default base in python which is natural log. From F (y) we can compute the desired conditional probability by

_eF (y)

^{P (Y = y j X) =} _eF (English) _{+ e}F (Spanish) ^:

However, F () can be very negative if the letter is long. This leads to e^{F ()} under ow, and we may have a divide-by-zero error. To make the computation more robust, let’s x y = English and note that

P (Y = English j			eF (English)	1
		X) =		=		:
			eF (English) + eF (Spanish)		1 + eF (Spanish) F (English)
Now, F (Spanish)		F (English) can be very positively large and eF (Spanish) F (English)
over ows toward in nity. But we know the ratio should approximate 0 in
that case.	Conversely, F (Spanish) F (English) can be very negative and
eF (Spanish)	F (English) under ows toward zero. But we know the ratio should

approximate 1 in that case. Thus we suggest in your python program you do an if-else check as follows:

• If F (Spanish) F (English) 100 then simply set P (Y = English j X)=0;

• If F (Spanish) F (English) 100 then simply set P (Y = English j X)=1;

• Otherwise compute

1

^{P (Y = English j X) =} _{1 + e}F (Spanish) F (English) ^:

1.4 Program output speci cation

We should be able to run your program with the following command:

python3 hw2.py

Your program should produce output that follow the speci cation in the follow-ing 4 questions.

Q1 ([25] points): Print \Q1″ followed by the 26 character counts for let-ter.txt. Please see sample output les for the exact format. We will test it on di erent letter.txt to check your count output.

Q2 ([25] points): Compute X₁ log e₁ and X₁ log s₁. Print \Q2″ then these values up to 4 decimal places on two separate lines. Sample output for let-ter.txt is as follows:

-0.0000

-0.0000

Output format explanation: The rst line is the value of X₁ log e₁ and the second line is the value of X₁ log s₁. There is no empty line in between or additional blank spaces. Notice that the values are 0 since ’A’ (or ’a’) does not appear in the text! The negative sign is due to python implementation, we also accept 0.0000 as an answer

Q3 ([25] points): Compute F (English) and F (Spanish). Similarly, print \Q3″ followed by their values up to 4 decimal places on two separate lines.

Q4 ([25] points): Compute P (Y = English j X). Print \Q4″ then this value up to 4 decimal places.

You are expected to print out the answers to the above questions to stdout. In other words, your program should print the values to your screen. You can use the python print() function. Do not add additional blank lines between your outputs. Carefully follow the format described above as your submission will be automatically graded using an autograder. We provided four sample input and output les for your reference.

1.5 Files in the assignment

The homework is released as a single hw2.zip which contains the following les and directories:

• English character probability le e.txt

• Spanish character probability le s.txt

• directory samples contains sample letters and the desired output. Your outputs should be identical to the ones provided. HINT: One way to check whether your output matches the expected output is to use vimdi on Linux Terminals (use CSL Machines) for a line by line comparison between two les. Usage:

>> vimdiff file1 file2

You can save your output to a le and then compare using vimdi as follows

• • • python3 hw2.py | tee your_output

• • • vimdiff your_output letter_out.txt

• Skeleton code hw2.py to help you get started. It contains the following two functions:

{ shred(): Implements digital shredder functionality described in sec-tion 1.1. You are provided the de nition and return value. You will need to complete the function to answer Q1.

{ get parameter vectors(): Reads “e.txt” and “s.txt” and returns the e and s vectors described in section 1.2. We have already im-plemented this function for you. You can directly use the returned values in your code.

You are free to implement the rest of the assignment as you wish using your python coding skills.

1.6 Deliverables

Please submit your les in a zip le named hw2 <netid>.zip, where you replace <netid> with your netID (your wisc.edu login). Inside your zip le, there should be only one le named: hw2.py. Do not submit a Jupyter notebook .ipynb le. Make sure you use python3!

ALL THE BEST!

6