Aspect Based Sentiment Analysis: A Python Demo

Summary

This video introduces aspect-based sentiment analysis (ABSA), a method to extract structured data on aspects, opinions, and sentiments from reviews. It explains ABSA's advantages over traditional sentiment analysis and demonstrates a paper's model for extracting aspect sentiment triplets. The video includes a Python demo showcasing how to train a model on a specific dataset and evaluate its performance using F1 score.

Takeaways

• 🍽 Aspect-Based Sentiment Analysis (ABSA) focuses on identifying specific aspects of a product or service mentioned in reviews and the sentiment towards those aspects.
• 📊 ABSA provides structured data that includes aspects, opinions, and sentiments, offering more detailed insights compared to traditional sentiment analysis.
• 🔍 The example given in the script illustrates how ABSA can distinguish between positive sentiment towards 'waiters' and negative sentiment towards 'pasta' in a restaurant review.
• 📚 Terminology in ABSA includes 'aspects' or 'targets' (like 'waiters' and 'pasta'), and 'subjective opinions' (like 'friendly' and 'average').
• 📊 ABSA can quantify sentiments towards aspects by counting mentions and categorizing them as positive, negative, or neutral.
• 🔍 The script mentions clustering and dimension reduction as methods to group similar aspects when there are too many to analyze individually.
• 📖 The paper 'Learning Span-level Interactions for Aspect Sentiment Triplet Extraction' is highlighted, which introduces a model that outputs 'aspect sentiment triplets'.
• 🛠️ The model architecture involves tokenization, span enumeration, pruning, and classification to identify aspects, opinions, and their sentiment relations.
• 💻 The paper's authors have made their code publicly available on GitHub, allowing others to replicate their work and use their model.
• 🐍 A Python demo using the pre-trained model is provided, showing how to predict aspect sentiment triplets for new sentences.
• 📈 The model's performance varies depending on the dataset and encoding method used, with BERT showing better performance than LSTM.

Q & A

What is Aspect-Based Sentiment Analysis (ABSA)?

Aspect-Based Sentiment Analysis is a method used to analyze specific aspects or targets of a product or service mentioned in a review or text, identifying the sentiment expressed towards these aspects.

How does ABSA differ from traditional sentiment analysis?

ABSA provides more detailed insights by extracting information about specific aspects people are discussing, whether they like or dislike each aspect, and the reasons behind their sentiments, whereas traditional sentiment analysis typically focuses on overall positive, negative, or neutral sentiments.

What is an example of structured data output for ABSA?

Structured data output for ABSA might include aspects like 'waiters' and 'pasta' from a restaurant review, with opinions such as 'friendly' and 'average', and sentiments like 'positive' and 'negative' respectively.

What terminology is used in ABSA for specific parts of a review?

In ABSA, 'aspects' or 'targets' refer to specific parts of a product or service being reviewed, while 'subjective opinions' are the adjectives or phrases used to describe those aspects.

What does the output of ABSA look like when analyzing a large number of reviews?

The output of ABSA on a large dataset might show the frequency of mentions for each aspect, along with the associated opinions and sentiments, such as 'food' being mentioned 523 times with various opinions like 'weird', 'authentic', 'great', and 'dry'.

How can clustering and dimension reduction be applied in ABSA?

When there are too many aspects identified in the reviews, clustering and dimension reduction techniques can be used to group similar aspects together, simplifying the analysis and making it more manageable.

What is an 'aspect sentiment triplet' as mentioned in the paper walkthrough?

An 'aspect sentiment triplet' refers to a combination of an aspect, the opinion expressed about that aspect, and the sentiment associated with that opinion, such as 'windows 8' being the aspect, 'did not enjoy' the opinion, and 'negative' the sentiment.

What is the process for breaking down a sentence into tokens in ABSA?

In the model described, sentences are broken down into tokens, which are then combined to form spans. These spans represent phrases that could be aspects or opinions, like 'windows 8' or 'did not enjoy'.

How does the model handle the computational complexity of enumerating spans?

To manage computational complexity, the model performs pruning operations that remove invalid spans and keep only candidate aspects or targets and opinions.

What is the role of the sentiment relation classifier in the ABSA model?

The sentiment relation classifier determines whether each aspect-opinion pair conveys a positive, negative, or neutral sentiment, or if it is invalid due to being an objective fact without a subjective opinion.

How can one train a model from scratch using the provided ABSA framework?

The provided notebook demonstrates how to train a model from scratch using one's own labeled training data and evaluate its performance using metrics like the F1 score.

Outlines

00:00

📊 Aspect-Based Sentiment Analysis Explained

This paragraph introduces Aspect-Based Sentiment Analysis (ABSA), emphasizing its ability to provide structured data that includes aspects, opinions, and sentiments. The script uses a restaurant review as an example to illustrate how ABSA can distinguish between positive and negative sentiments towards different aspects of a product or service. It also explains the terminology involved, such as 'aspects' or 'targets' for the subjects being reviewed and 'subjective opinions' for the descriptors. The paragraph concludes by highlighting the advantages of ABSA over traditional sentiment analysis, which is its capacity to extract detailed feedback on specific aspects and reasons behind the sentiments.

05:00

🔍 Deep Dive into ABSA Model Architecture

The second paragraph delves into the model architecture for ABSA, focusing on a paper titled 'Learning Span-level Interactions for Aspect Sentiment Triplet Extraction.' It outlines the process of breaking sentences into tokens and then combining them into spans to identify aspects and opinions. The model performs pruning to filter out invalid spans, leaving only potential aspects and opinions. These are then paired and classified to determine the sentiment relationship. The paragraph also discusses the model's performance, mentioning that it has shown significant improvement over previous models. It notes the use of LSTM or BERT for sentence encoding and the different datasets used for training and testing, including restaurant and laptop reviews from 2014. The paragraph concludes with a mention of a Python demo on Google Colab that demonstrates how to use the model for prediction and evaluation.

Keywords

💡Aspect-Based Sentiment Analysis (ABSA)

Aspect-Based Sentiment Analysis is a subfield of sentiment analysis that focuses on identifying and categorizing opinions expressed about specific aspects or features of a product, service, or topic. Unlike general sentiment analysis, which determines the overall sentiment towards an entity, ABSA drills down to the granular level to understand sentiments towards particular attributes. In the video, ABSA is used to analyze an online restaurant review, distinguishing between sentiments towards 'waiters' and 'pasta'.

💡Structured Data

Structured data refers to information that is organized into a formatted repository, typically a database, where it can be easily accessed, managed, and analyzed. In the context of the video, structured data is the output format of ABSA, which shows the aspect, opinion, and sentiment in an organized manner. For instance, the review 'waiters are very friendly but the pasta is simply average' is translated into structured data indicating 'waiters' as an aspect with a positive sentiment, and 'pasta' with a negative sentiment.

💡Aspect

In ABSA, an 'aspect' is a specific feature or component of a product or service that is being evaluated. It is also known as a 'target'. The video script uses 'waiters' and 'pasta' from a restaurant review as examples of aspects where the sentiment analysis focuses on these particular elements to determine the sentiment expressed towards them.

💡Opinion

An 'opinion' in the context of ABSA refers to the subjective evaluation or view expressed about an aspect. It is a descriptive term that captures the sentiment-bearing phrase that is associated with an aspect. For example, in the transcript, 'friendly' and 'average' are opinions about the aspects 'waiters' and 'pasta', respectively.

💡Sentiment

Sentiment, in the video, refers to the emotional tone or attitude expressed in an opinion, which can be positive, negative, or neutral. It is a key output of ABSA, indicating the feeling or opinion towards an aspect. The script gives examples such as 'friendly' indicating a positive sentiment and 'average' indicating a negative sentiment.

💡Target-Based Sentiment Analysis

Target-Based Sentiment Analysis is synonymous with ABSA and focuses on analyzing sentiments directed at specific targets or aspects. The video emphasizes that ABSA, or target-based sentiment analysis, provides more detailed insights about individual aspects of a product or service compared to traditional sentiment analysis.

💡Aspect Sentiment Triplet

An 'Aspect Sentiment Triplet' is a model output that consists of an aspect, an opinion, and a sentiment. It is a structured way to represent the relationship between an aspect and the sentiment expressed towards it through an opinion. The video describes a model that outputs these triplets, such as 'Windows 8' as an aspect, 'did not enjoy' as the opinion, and 'negative' as the sentiment.

💡Tokenization

Tokenization is the process of breaking down text into individual elements or 'tokens', which are typically words or phrases. In the video, tokenization is the first step in the model's architecture, where a sentence is broken into tokens to later form spans that represent aspects or opinions.

💡Span

A 'span' in the context of the video refers to a sequence of tokens that together represent an aspect or an opinion. It is a multi-word phrase that is considered as a single unit for sentiment analysis. For example, 'Windows 8' is a span that represents an aspect in a review about a laptop.

💡Pruning

Pruning, in the context of the video, is a technique used to reduce computational complexity by eliminating irrelevant or unnecessary data. The model prunes spans that are classified as invalid, keeping only those that are potential aspects, targets, or opinions for further analysis.

💡Sentiment Relation Classifier

The 'Sentiment Relation Classifier' is a component of the model that determines the sentiment relationship between an aspect and an opinion. It classifies each aspect-opinion pair as positive, negative, neutral, or invalid, as not all aspects have a sentiment associated with them. The video uses the example of 'new Windows 8' being a fact without sentiment, thus classified as invalid.

Highlights

Introduction to Aspect-Based Sentiment Analysis (ABSA)

ABSA outputs structured data showing aspect, opinion, and sentiment

Terminology: aspects (targets), subjective opinions, and sentiment

ABSA provides more detailed sentiment analysis compared to traditional methods

Example of ABSA output for a restaurant review

ABSA can cluster and reduce dimensions for numerous aspects

Paper walkthrough on 'Learning Span-level Interactions for Aspect Sentiment Triplet Extraction'

Model outputs aspect sentiment triplets

Model architecture involves tokenization, span enumeration, and pruning

Aspect term extraction and opinion term extraction as part of the model

Sentiment relation classifier determines the sentiment between aspect and opinion

Model performance compared to previous papers

Use of LSTM or BERT for sentence encoding

Training and testing sets from the Semantic Evaluation Workshop

Code availability on GitHub and Python demo on Google Colab

Process of downloading and installing packages for the demo

Example of training data for the 2014 laptop dataset

Pre-trained model prediction example

Model sensitivity to the dataset it was trained on

Instructions on training a model from scratch with custom data

Evaluation of model using F1 score

Call to action for likes and subscriptions

Transcripts

hello everyone today i will introduce

what is aspect-based sentiment analysis

we will take a look at what absa outputs

look like followed by a paper

walkthrough and we will end with a

python demo

let's start with an example

let's say we have an online review for a

restaurant waiters are very friendly but

the pasta is simply average

ebsa should be able to output structured

data

showing the aspect the opinion and the

sentiment

in this case waiters are being described

as friendly and it's a positive

sentiment in contrast pasta is being

described as average is a negative

sentiment

as we saw on the previous page here we

have some terminology and jargons

waiters and pasta are called aspects

also called targets

friendly and average are subjective

opinions

to summarize

absa is also called target-based

sentiment analysis and opinion mining

compared to traditional sentiment

analysis ebsa provides more details

for a given product or service ebsa can

extract what aspects are people talking

about

do people like or dislike each aspect

and for what reason

next let's look at what absa outputs

should look like

let's say we scraped all the online

reviews for a restaurant this is a

google review page for a place i really

like in florida

the outputs should look something like

this for instance people may have

mentioned the word food 523 times

seven people may have said the food

tasted weird

six people may have said the food tasted

authentic

43 said the food was great 17 said the

food was dry and weir correspond to a

negative sentiment

authentic positive great positive

and so on and on the other hand people

may have mentioned the word service 326

times 34 people may have said the

service was slow

12 may have said the service was fast

six may have said the servers were rude

and 15

friendly they also correspond to

different sentiments

if you ended up having too many aspects

you can do clustering and dimension

reduction to group similar aspects

together

and now that we have a general

expectation of what outputs we're

looking for i'm going to do a paper walk

through so that we can understand the

model design and architecture

this paper is called learning span level

interactions for aspect sentiment

triplet extraction it has two first

authors who contributed to the work

equally

and their model will output what they

call

aspect sentiment triplets

for example we have a consumer review

here for a laptop did not enjoy the new

windows 8 and touch screen functions

here windows 8 is the aspect and so is

touch screen functions

they were both described by the opinion

not enjoy and the sentiment is negative

here we have the model architecture

first they break a sentence into tokens

so did not enjoy the new windows 8 they

were broken down into many tokens

and these tokens are then combined to

form a span

for example did

did not enjoy not enjoy enjoy the new

and we need to combine these words to

form a span because an aspect or an

opinion can be a phrase and a phrase can

have multiple words windows 8 should be

read together as one phrase it wouldn't

make sense to break it apart

and the enumeration of span is pretty

computational intensive because the

number of possible spans grows

exponentially as the sentence gets

longer

to control the computational complexity

next they did some pruning

here is a task called aspect term

extraction

and opinion term extraction

each span is then classified as either

invalid

an opinion

or a target also called an aspect

and a pruning operation here removes

everything that's invalid and only keeps

candidate aspects or targets as well as

candidate opinions

and finally each aspect is paired with

each opinion these target candidate

aspects and opinions are coupled

together so that we can determine the

sentiment relationship between them

here the sentiment relation classifier

determines whether each aspect opinion

pair

conveys a positive

or a negative or a neutral sentiment it

can also be invalid because the new

windows 8 is a fact it's not a

subjective opinion but rather an

objective fact so there is no sentiment

relationship there

and this model has a decent performance

which is significantly improved compared

to previous papers

the performance varies depending on the

model in the beginning they can either

do sentence encoding using by lstm or

bird

bird is a more complex model so it leads

to better performance

they also had four different training

and testing sets

here rest 14 means that it's a data set

about restaurant from 2014

and lab 14 means that it's a data set

about laptop from 2014.

these data sets were originally released

by the semantic evaluation workshop

the authors of this paper have made

their code publicly available on github

and we can see a python demo on google

colab

in the beginning we need to download and

install a number of packages as well as

loading the pre-trained model

there are a lot of packages that needed

to be downloaded it took me about four

minutes

the next cell shows what training data

looks like and here specifically they're

showing an example

for the 2014 laptop data

and this is what the data looks like for

example i charge it at night and skip

taking the chord with me because of the

good battery life

the target word is battery life and its

index goes from 16 to 17

and the opinion word is good its index

goes from 15 to 15

and also this is a

positive sentiment

and below we can use a pre-trained model

for prediction this is the training

example we saw from the paper did not

enjoy the new windows 8 and touch screen

functions

at the very end the output is going to

show you a triplet

so here the target word is windows 8 and

the opinion is not enjoy the sentiment

is negative this sentence has two

aspects so we also get an output for

touch screen

so the touch screen functions

the opinions also not enjoy and the

sentiment is negative

now let's try providing a new example

easy to use and set up

no problems after six months

and let's run the cell and

see what happens

and after about 48 seconds we have the

output

here the target is used the opinion is

that it's easy to use and the sentiment

is positive

so it looks pretty good

and let's try something else

so so this model is trained on laptop

data let's try

something not related to laptop

let's say the treatment is effective

and i already run this cell it ended up

not giving us anything

meaning that it wasn't able to detect

any sentiment aspect triplets

so the model is pretty sensitive to the

data set it was trained on because it

was trained on data about laptop it

wasn't able to detect

it wasn't able to detect anything

in the medical domain

but the nice thing is that the notebook

also show you how to train a model from

scratch using your own label training

data

they also show you how to evaluate

your model f1 score

this is the end of the video if you

learned anything useful today feel free

to like and subscribe

thank you for your time and take care