在USYD上NLP课程的时候对于w2v一直十分的困惑，只知道其输入输出的区别以及目的，没有深究其原理以及实现方式。今天对于其中两个模块进行详细的研究，算是补齐了之前的一小块拼图。其中的两个优化方法都是对于复杂度的降低，一个是对于词表的降低，另一个是对于损失函数计算方法的降低。是通过近似和模拟的方式去考虑以实现降低的计算。写篇文章记录一下，下次忘记了有地方找（bushi （：

Word2vec

If the encoding use the one hot encoding to represent the word. The dimension will be huge, and the matrix will be sparse. Even the PCA could be applied.

Also, this way could have another limitation, that is the semantic could not be represent well. For instance, the value of COS in similar wards are 0, they show the totally different.

Thus, the word2vec come up. Its not a model, but a tool.

It concludes CBOW and Skip-gram

First, define the size of window, if the window is 2, that will be [t-2,t-1,t,t+1,t+2]

For Skip-gram, the t is used to predict the context, i.e., 2 backward and forward.

For CBOW, the context is used to predict the word in the middle.

In terms of the input of CBOW, that is the 4 words, calculate the average values, and pass into the function of softmax to get the result.

Skip-gram is similar, remove the operation of average, due to it is itself, no matter need the average.

In the word2vec, there are 2 optimised method, negative sampling and hierarchical softmax.

negative sampling means “ABCDE” - middle word is C, context is AB and DE. postive sample is CA, CB, CD, CE, which is the concurrence of word in sentence. However, it need go through the whole dictionary, waste too much resource. So, create the negative samples CC,CF, etc. It also only choose K sample randomly for the simplicity.

Hierarchical softmax: first create the tree with order, 霍夫曼树, the final possibility is the value times among the path that searching in the tree. binary classification for each path, get the possibility, then multiply it. It means to reduce the computational resources, comparing to the go through whole data. The operation is the similar value, not exactly accurate.

Also, there is another concept called subsampling, the possibility decided by the frequent, aiming to remove the high frequent word, like ‘the’, which no include much meaning in the word.

Comparison: CBOW train faster than Skip-gram, but Skip-gram could get better vector representation.

CBOW only predict one word, but the skip-gram predict 2K times, K means the window size. So, skip-gram will get better vector representation for low-frequent, due to its training repeat more times.

XLNET

Original address of paper

https://arxiv.org/abs/1906.08237

Reason the idea come up

XLNET uses the AR method to implement the language model.

AR has the unidirectional limitation. And, AE has issue in MASK. In fine tuning phase, input of BERT without data with MASK, showing the discrepancy, and author thought that BERT assume the predicted token as independent word is oversimplified the natural language. Therefore, the XLNET is created to leverage best of AR and AE to avoid the limitation.

Main idea of the technology

Firstly, standing at the conventional AG mode

permutation

instead of forward or backward of sequence, XLNET have all possible permutations

this operation aims to get the bidirectional context like BERT.

In detail, XLNET create the all permutations and randomly choose some cases to predict some masked words in end of sequence. e.g., 1,2,3,4,5,6,7 → use 1,2,4,5,6,7 to predict 3. No matter the order is.

Meanwhile, how to decide the length of predict token? there is a hyper parameter K, which is equal the length of sequence dividing by length of predicted token.

The implementation of permutation is the mask in transformer’s attention.

Two-Stream Self-Attention

In the case of 1342 and 1324, when we want to predict the place of 3, i.e. (target 4 or 2), the word they could see are same, which is not right. Therefore, we introduce the Two-Stream Self-Attention.

1. h, the content representation
2. g, the query representation, access to contextual and position information, but without content

Thinking

The paper also incorporating the idea from Transformer-XL. (seems aims for long text task) (relative position encoding and fragment loop?)

The details of Transformer XL is not clear for me, maybe after learning for it and get deeper understanding for XLNET

Terms appear in the paper

AR: Autoregressive Language model

using the sequential direction to predict the next word, that is the single direction, e.g. GPT2

AE: Autoencoding Language model

utilise the context. i.e., BERT. It includes the MLM (masked language model) and NSP (next sentence prediction). NSP is to deduce the inference relation, MLM replace some word with MASK, and to predict the position with MASK.

tBERT

Original address of paper

https://www.aclweb.org/anthology/2020.acl-main.630.pdf

Reason the idea come up

The semantic similarity could provide additional signal and information to model. Also, the integration of the domain could effectively help the model get better performance. That is the reason tBERT come up.

Main idea of the technology

First create the place for CLS in BERT, representing the domain of the input.

The input of sentences put into the topic model and get the average value.

At the meantime, the input the sentences into the BERT

The paper not mention the detail design of loss, from my understanding, the CLS should be included in the loss, and no other digest article for the tBERT, may it should be confirmed for the further research.

Thinking

Fine tuning could also helps model to deal with the domain shift, but the experiment showed that the result of tBERT is better than the fine tuning, especially for the specific domain, e.g., medical, finance, etc.

UniLM

The term stands for Unified Language Model Pre-training for Natural Language Understanding and Generation

Original address of paper

https://arxiv.org/pdf/1905.03197.pdf

Reason the idea come up

The reason of the BERT not good for NLG is the MLM(mask language model), which is not the same with the target of generation task. Except the masked position, other words are seen for model. Therefore, to improve the NLG performance of BERT, one way to handle is get better NLG ability.

Main idea of the technology

The author thought that the natural way to control the model is the mode of information input in the model. E.g., the access to the context of the word token to be predicted. The paper proposed the 3 ways: bidirectional LM, unidirectional LM, sequence-to-sequence LM.

For the special token, SOS, EOS stand for start and end of sentence. They helps to improve the ability of NLU and NLG, it marks the end of sequence, and help model to learn the when to terminate the NLG task.

In the unidirectional LM, comprised by left to right and right to left LM. Taking the example of left to right, the x1, x2, [mask], x4. The representation of mask only x1 and x2

In the bidirectional LM contains both directions context to represent the masked token.

In the sequence to sequence LM, the input has segment, that is the source and the target segments. SOS, t1 ,t2, EOS t3, t4, t5, EOS. t2 could access the first 4 tokens, t5, could only get the first 6 tokens. (including the SOS and EOS)

Also, the paper mentioned that the 1/3 time use the bidirectional LM objective, 1/3 time used the seq2seqLM objective, and the left to right and right to left are 1/6 of the time be set as the objective. By this strategy, the final objective of UniLM could be achieve, i.e., generative performance.

In more details, the initialisation of UniLM used BERT-large, and the strategy of mask is same with BERT, but 20% of chance will be bigram or trigram, to improve the predictive performance.

Terms appear in the paper

There are two new terms of NLP:

NLU: Natural language understanding

NLG: Natural language Generation

The concepts are not brand new, but the abbreviations are good for extension

RoBERTa

Original address of paper

https://arxiv.org/abs/1907.11692

https://zhuanlan.zhihu.com/p/149249619

根据知乎的文章，总结出以下几点特征：

1. 用更长的时间和更大的BS，训练更多的数据
2. 去掉BERT中的NSP（next sentence prediction)
3. 在更长的句子上训练
4. 根据训练数据动态的改变mask模式

Reason the idea come up

从BERT的角度来说，masking是在预处理的时候执行Masked Language Model. In random masking, BERT will select the 15% word in each sentence, and with 80% chance replaced by MASK, 10% chance replaced by other random word, 10% keep the same. Recall the COMP 5046 unit, assignment, there are same. Thus, the model target is to predict the place be marked.

Also, BERT could predict the adjacent position relationship between sentences, that is called Next Sentence Prediction(NSP)

Difference:

其中比较大的变化是对于mask的处理方式不同

BERT是在pre-precessing的时候对于masking进行处理，这样一来masking是静态的，也就是对于每一个epoch来说masking的结果是一致的。为了避免这样的情况，会对数据进行复制，作出不同的静态masking进行训练。但还是会出现重复的现象。

而RoBERTa使用动态masking，在序列进入模型之前才处理masking的任务。

RoBERT remove the task of NSP, it consider this task is not improve the performance in expected.

思考：BERT输出信息的继续挖掘以达到更好的语义理解

On the sentence embeddings from pre-trained language models

this is the address for original paper

https://arxiv.org/pdf/2011.05864.pdf

The problems proposed in the paper mainly discuss the:

1. why low performance of text matching in BERT for unsupervised task. Is that lacking of the sentiment information or wrong way to digging these information?
2. If lacking of digging, how to utilize them.

论文给出的结论是我们对于BERT的输出信息利用不够强

论文中提出了一个方式，去映射BERT sentence embedding, called BERT-flow.

其启发是来自于1. 高频的偏差，对于高频词语的space来说，可能会导致向量的偏移，这会导致对于向量之间关系不按照我们所想的方式进行发展。2.sparse word会在BERT空间中比较稀疏（低频词），以及产生一些洞。这些洞会导致语义理解的难度

基于这两个问题， paper提出了一个转换的方式：

BERT embedding space to standard Gaussian latent space

从而实现了基于流的标定来帮助建立更加正确的BERT词向量

FastBERT

https://aclanthology.org/2020.acl-main.537/

首先给出模型的提出原文地址

https://zhuanlan.zhihu.com/p/127869267

对于FastBERT的定义和解读，知乎这篇文章讲的十分优秀，可以仔细阅读

简单来理解，单纯使用BERT对样本进行预测工作需要经过完整的网络，FastBERT的主要目的是为了让降低计算。

1. 从蒸馏的角度来说是创建一个全新的学生网络去拟合本地的大BERT模型从而实现对模型的简化。
2. 但是FastBERT的思想是对于模型进行改造。聚焦于一些简单的样本，也就是不需要通过整个网络就可以进行分类的样本进行优化。解决方案为：给每一层后接上分类器。这样主要的结构就出现了。

作者称之BERT模型为主干，Backbone

每一个分类器为分支，Branch

这一过程作者称之为自蒸馏，在pre-train & fine-tuning 的时候只更新主干，tuning结束之后freeze主干用branch获得蒸馏结果

总结，fastBERT主要思想是提前输出简单样本，来减少模型的计算压力，从而提高推理速度。优点：不需要改变主要的BERT结构（相比于Distill-BERT）

思考：如何去决定当前批数据经过的层数，如何去定义？

在原文中提到，不确定性高于speed的将会被传入下一层，低于speed的就停止传递

其中，对于不确定性的定义是当前批的数据的煽值（正则化），其中包括一些假设：低不确定性对应高精度。这样子理解，speed在原文中的意义就是threshold，去定义以及划分不确定性的高低。越高的speed，越高的层收到的样本就越少，也就意味着门槛变低，更多的样本被直接输出，从而提高了模型的计算速度。

至于模型的loss计算，GPT给出的解释为分类loss和层选择的loss（CE, MSE[选择的层和全部的层]）但是觉得与原文有些出入，还需要进一步研究。（对于损失函数的观点未经证实）

本文记录了阅读FastBERT paper的笔记以及对于前辈的知识总结进行归纳。大致了解了FastBERT的思路和实现原理，但对于原文中的KL散度分布还有些困惑，需要进一步研究。

Knowledge Distillation

Original address of paper

https://zhuanlan.zhihu.com/p/75031938

The main idea of the knowledge distillation is “forget the models in the ensemble and the way they are parameterised and focus on the function”

Main idea of the technology

https://arxiv.org/abs/1503.02531

在Hinton原文中指出训练时就如同毛毛虫吃树叶来积攒能量，在使用神经网络的时候既做了吃树叶又繁殖的任务，导致效率低下。因此对于这样的问题，希望将一个复杂的模型转变到简单的模型，也就是知识蒸馏做的事情。

蒸馏的过程就是将训练的大模型的内容总结到小模型中，已达到模型复杂度和精度的平衡。在这里会有两个模型，较大的为老师模型，小的模型为学生模型，老师模型通过对于hard targert 进行训练，而学生模型通过老师模型的输出进行收敛。

To achieve training of the student network, the paper mentioned that adding a temperature to scale down the target before input into the softmax. (i.e., exp(zi/T) / sum(exp(zi/T))) The more larger T in formula, the more smooth distribution the result will be.

The loss function is a * soft + (1-a) * hard, i.e. the trade-off between them.

in general, the larger soft distribution will get better performance in test.

1. 训练大模型，也就是使用hard taget， 也就是不对标签做处理
2. 通过老师模型计算soft target
3. 训练小模型
   1. 设置相同T计算结果与soft target 计算loss （from 2）
   2. 设置T为1，也就是hard target计算loss
4. 学生模型T设置为1作为最后的预测

Summary

The reason knowledge distillation worked is the too sharp distribution of result makes loss function not effect the model, i.e., except the true label, outputs are too close to 0, making the model could not learning anymore. Thus, the introducing of the T to scaling down the result helps to get smooth distribution. In this way, get simper model and better convergence of model.

SpaCy

Quick Start

1. It contains the pip installation in python
2. created a english based nlp object
3. input some text to create the object of document
4. operations demo that can get the text from token
5. span of the token in text

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

# 1. installation
!pip install spacy

# 2. Create a blank English nlp object
nlp = spacy.blank("en")

# 3. Created by processing a string of text with the nlp object
doc = nlp("Hello world!")

# 4. Index into the Doc to get a single Token
token = doc[1]

# Get the token text via the .text attribute
print(token.text)

# Iterate over tokens in a Doc
for token in doc:
    print(token.text)

# 5. Create another document with text
doc = nlp("Hello NLP class!")

# A slice from the Doc is a Span object
span = doc[1:3]

# Get the span text via the .text attribute
print(span.text)

demonstrations of is_alpha, is_punct, like_num

1
2
3
4
5
6
7
8
9
10
11
12
13
14

doc = nlp("It costs $5.")
print("Index:   ", [token.i for token in doc])
print("Text:    ", [token.text for token in doc])

print("is_alpha:", [token.is_alpha for token in doc])
print("is_punct:", [token.is_punct for token in doc])
print("like_num:", [token.like_num for token in doc])

# output:
# Index:    [0, 1, 2, 3, 4]
# Text:     ['It', 'costs', '$', '5', '.']
# is_alpha: [True, True, False, False, False]
# is_punct: [False, False, False, False, True]
# like_num: [False, False, False, True, False]

Part of Speech

Load the small english pipeline to achieve the pos

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

!python -m spacy download en_core_web_sm

# Load the small English pipeline
nlp = spacy.load("en_core_web_sm")

# Process a text
doc = nlp("She ate the pizza")

# Iterate over the tokens
for token in doc:
    # Print the text and the predicted part-of-speech tag
    print(token.text, token.pos_, token.pos)

# result: 
# She PRON 95
# ate VERB 100
# the DET 90
# pizza NOUN 92

# also the token.head returns the syntactic head token
for token in doc:
    print(token.text, token.pos_, token.dep_, token.head)
    
# She PRON nsubj ate
# ate VERB ROOT ate
# the DET det pizza
# pizza NOUN dobj ate

Name Entity Recognition

1
2
3
4
5
6
7

# Process a text
doc = nlp("Apple is looking at buying U.K. startup for $1 billion")

# Iterate over the predicted entities
for ent in doc.ents:
    # Print the entity text and its label
    print(ent.text, ent.label_)

Matching

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

# Import the Matcher
from spacy.matcher import Matcher

# Load a pipeline and create the nlp object
nlp = spacy.load("en_core_web_sm")

# Initialize the matcher with the shared vocab
matcher = Matcher(nlp.vocab)

# Add the pattern to the matcher
pattern = [{"TEXT": "iPhone"}, {"TEXT": "X"}]
matcher.add("IPHONE_PATTERN", [pattern])

# Process some text
doc = nlp("iPhone X news! Upcoming iPhone X release date leaked")

# Call the matcher on the doc
matches = matcher(doc)

# Iterate over the matches
for pattern_id, start, end in matches:
    # Get the matched span
    matched_span = doc[start:end]
    print('"{}" - match for pattern {} in span ({}, {})'.format(matched_span.text, pattern_id, start, end))
    
# "iPhone X" - match for pattern 9528407286733565721 in span (0, 2)
# "iPhone X" - match for pattern 9528407286733565721 in span (5, 7)

in match pattern, LEMMA contains the multiple variations of the word

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

pattern = [
    {"LEMMA": "love", "POS": "VERB"},
    {"POS": "NOUN"}
]
matcher = Matcher(nlp.vocab)
matcher.add("LOVE_PATTERN", [pattern])

doc = nlp("I loved vanilla but now I love chocolate more.")

matches = matcher(doc)
for pattern_id, start, end in matches:
    matched_span = doc[start:end]
    print('"{}" - match for pattern {} in span ({}, {})'.format(matched_span.text, pattern_id, start, end))

# "loved vanilla" - match for pattern 4358456325055851256 in span (1, 3)
# "love chocolate" - match for pattern 4358456325055851256 in span (6, 8)

1
2
3
4
5
6

Operators and quantifiers let you define how often a token should be matched. They can be added using the "OP" key. "OP" can have one of four values:

- An "!" negates the token, so it's matched 0 times.
- A "?" makes the token optional, and matches it 0 or 1 times.
- A "+" matches a token 1 or more times.
- And finally, an "*" matches 0 or more times.

Word Vector

SpaCy contains the word vector in medium pipeline

It could output the word vectors and calculate the similarity of tokens

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

# Load a larger pipeline with vectors
nlp = spacy.load("en_core_web_md")

# Look at one word vector
doc = nlp("I love chocolate")
print(doc[2].vector)

# Compare two documents
doc1 = nlp("I like fast food")
doc2 = nlp("I like pizza")
print("Comparing sentences:", doc1.similarity(doc2))

# Compare two tokens
doc = nlp("I like pizza and pasta")
token1 = doc[2]
token2 = doc[4]
print("Comparing 'pizza' and 'paste':", token1.similarity(token2))

Reference

1. https://course.spacy.io/en/
2. https://spacy.io/

Lecture 1

N-Gram LMs

The 4 ways to deal with the unseen word sequence

smoothing:

discounting:

Interpolation: using the multiple N-Gram probability

Kneser-Ney smoothing:

Lecture2

Lecture3

Lecture4

Note the formula macro and micro F1 score — ass2

Not the find_best_socre initial number should be negative — ass2

directly using average will make some features vanish (+ and -), also the order is meaningless

in slide, NER page 66

labels number * length * length means:

each token in sequence is possible for start or end, that is length * length

then, each span can be predicted as each label

so, the output should be (labels * length * length)

Lecture 5

beam search

in nature, greedy search is to find the local optimal solution in each step

Lecture 6

the label in current word from LSTM purely, B I O possibility(sum as 1)

score = current label p * best (previous p * transfer p)

it from

score = current label p * best (i.e. DP * transfer model)

DP refers to the optimal route in choice

transfer model refers to the Markov model

co reference

dependency parsing

Lecture 7

Lecture 8

disadvantages of self-attention

problem:

non-linear lacking

order lacking

resolve:

feedforward layer

additional position vector

issue: the length of the position vector is fixed, not flexible enough
the mask is for ignoring the future word in the training process, it lets the model know what is known in this step, and what is the prediction in this step. In math, put the infinite negative value for future words and dot the product in the current word
the multiple layers of self-attention are used to build better representation of the word

query: curve line

key: embedding word

value: passing into the weighted value

Lecture 10

data can from

fineweb

common craw

Lecture 11

Review

The content below is the missing slide week.

parsing, 句法分析, identity the grammar structure by sub-phrases

The span in the sentence called the phrase

treebanks

Method: Dependency Parsing / Grammar

compute every possible dependency to find the best score

Or, choose a way to get the best overall score

co-reference