Commit
•
6eaf2a5
1
Parent(s):
97e4014
Delete README.md
Browse files
README.md
DELETED
|
@@ -1,109 +0,0 @@
|
|
| 1 |
-
# Problem Description
|
| 2 |
-
|
| 3 |
-
This project aims to develop a system capable of automatically **summarizing short dialogue text**. This addresses the challenge of extracting concise yet informative summaries from conversational exchanges, enabling users to **quickly grasp the information of the dialogues**.
|
| 4 |
-
|
| 5 |
-
Summarizing these conversations can be valuable for various applications, such as:
|
| 6 |
-
- Streamlining information retrieval in customer service interactions
|
| 7 |
-
- Condensing meeting discussions for efficient review
|
| 8 |
-
- Providing concise overviews of chat conversations on social media platforms
|
| 9 |
-
|
| 10 |
-
This project tackles the task of automatically generating concise summaries, saving users time and effort while improving comprehension.
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
<p align="center"><i>Source: Google Research</i></p>
|
| 15 |
-
|
| 16 |
-
**Input:** Dialogue text
|
| 17 |
-
|
| 18 |
-
Example:
|
| 19 |
-
```
|
| 20 |
-
Matt: Do you want to go for date?
|
| 21 |
-
Agnes: Wow! You caught me out with this question Matt.
|
| 22 |
-
...
|
| 23 |
-
Agnes: See you on saturday.
|
| 24 |
-
Matt: Yes, looking forward to it.
|
| 25 |
-
Agnes: Me too.
|
| 26 |
-
```
|
| 27 |
-
|
| 28 |
-
**Output:** Summarized dialogue
|
| 29 |
-
|
| 30 |
-
Example:
|
| 31 |
-
```
|
| 32 |
-
Matt invites Agnes for a date to get to know each other better. They'll go to the Georgian restaurant in Kazimierz on Saturday at 6 pm, and he'll pick her up on the way to the place.
|
| 33 |
-
```
|
| 34 |
-
|
| 35 |
-
# Dataset
|
| 36 |
-
|
| 37 |
-
We'll utilize the `DialogSum` dataset accessible from 🤗**Hugging Face** (https://huggingface.co/datasets/knkarthick/dialogsum) and **Paper** (https://arxiv.org/pdf/2105.06762.pdf). This dataset comprises real-life dialogue scenarios paired with corresponding manually crafted summaries and dialogue topics.
|
| 38 |
-
|
| 39 |
-
`DialogSum` is a large-scale dialogue summarization dataset, consisting of **13,460** (Plus 100 holdout data for topic generation) dialogues with corresponding manually labeled summaries and topics.
|
| 40 |
-
|
| 41 |
-
Here's a sample of the `DialogSum` dataset structure:
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|id|dialogue|summary|topic|
|
| 45 |
-
|-|-|-|-|
|
| 46 |
-
|train_3|#Person1#: Why didn't you tell me you had a girlfriend? #Person2#: Sorry, I thought you knew. ... #Person1#: Oh, you men! You are all the same.|#Person1#'s angry because #Person2# didn't tell #Person1# that #Person2# had a girlfriend and would marry her.|have a girl friend|
|
| 47 |
-
|train_16|#Person1#: Tell me something about your Valentine's Day. ...#Person2#: Yeah, that is what the holiday is for, isn't it?|#Person2# tells #Person1# their Valentine's Day. #Person1# feels it's romantic.|Valentine's Day|
|
| 48 |
-
|...|...|...|...|
|
| 49 |
-
|
| 50 |
-
**Distribution of dataset**
|
| 51 |
-
|
| 52 |
-
|Dialogue|Summary|Dialogue + Summary|
|
| 53 |
-
|:-:|:-:|:-:|
|
| 54 |
-
||||
|
| 55 |
-
|
| 56 |
-
# Method
|
| 57 |
-
|
| 58 |
-
### Pre-trained Language Models:
|
| 59 |
-
|
| 60 |
-
This project explores two powerful LLMs well-suited for dialogue summarization:
|
| 61 |
-
|
| 62 |
-
- **FLAN-T5:** This model excels at understanding complex relationships within text, making it effective in summarizing the nuances of conversations.
|
| 63 |
-
- **BART:** This model boasts strong capabilities in text generation tasks, making it adept at generating informative and well-structured summaries.
|
| 64 |
-
|
| 65 |
-
### Fine-tuning Techniques:
|
| 66 |
-
|
| 67 |
-
To tailor these LLMs specifically for dialogue summarization, we will investigate several fine-tuning approaches:
|
| 68 |
-
|
| 69 |
-
- Instruction Fine-tuning
|
| 70 |
-
- Parameter Efficient Fine Tuning (PEFT)
|
| 71 |
-
+ Low-Rank Adaptation **(LoRA)**
|
| 72 |
-
+ Quantized Low-Rank Adaptation **(QLoRA)**
|
| 73 |
-
|
| 74 |
-
# Installation
|
| 75 |
-
|
| 76 |
-
```
|
| 77 |
-
!git clone "https://github.com/dtruong46me/dialogue-text-summarization.git"
|
| 78 |
-
```
|
| 79 |
-
|
| 80 |
-
# Contributions
|
| 81 |
-
|
| 82 |
-
**Supervisor:** Prof. Le Thanh Huong
|
| 83 |
-
|
| 84 |
-
**Student Group:**
|
| 85 |
-
|
| 86 |
-
|No.|Name|Student ID|Email|
|
| 87 |
-
|:-:|-|:-:|-|
|
| 88 |
-
|1|Phan Dinh Truong (Leader)|20214937|truong.pd214937@sis.hust.edu.vn|
|
| 89 |
-
|2|Nguyen Tung Luong|20214913|luong.nt214913@sis.hust.edu.vn|
|
| 90 |
-
|3|Vu Tuan Minh|20210597|minh.vt210597@sis.hust.edu.vn|
|
| 91 |
-
|4|Hoang Tu Quyen|20214929|quyen.ht214929@sis.hust.edu.vn|
|
| 92 |
-
|
| 93 |
-
# [Bonus] How to run Streamlit on Kaggle
|
| 94 |
-
|
| 95 |
-
```
|
| 96 |
-
!pip install -q streamlit
|
| 97 |
-
```
|
| 98 |
-
|
| 99 |
-
```
|
| 100 |
-
!wget -q -O - ipv4.icanhazip.com
|
| 101 |
-
```
|
| 102 |
-
|
| 103 |
-
```
|
| 104 |
-
!npm install -g localtunnel -q
|
| 105 |
-
```
|
| 106 |
-
|
| 107 |
-
```
|
| 108 |
-
!streamlit run "/kaggle/working/dialogue-text-summarization/streamlit_app.py" & npx localtunnel --port 8501
|
| 109 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|