|
--- |
|
dataset_info: |
|
- config_name: qa |
|
features: |
|
- name: q |
|
dtype: string |
|
- name: a |
|
dtype: string |
|
- name: rf |
|
dtype: string |
|
- name: song_title |
|
dtype: string |
|
- name: song_id |
|
dtype: string |
|
- name: id |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 67824 |
|
num_examples: 648 |
|
download_size: 134589 |
|
dataset_size: 67824 |
|
- config_name: song |
|
features: |
|
- name: id |
|
dtype: string |
|
- name: title |
|
dtype: string |
|
- name: name |
|
dtype: string |
|
- name: lyric |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 253605 |
|
num_examples: 181 |
|
download_size: 276024 |
|
dataset_size: 253605 |
|
- config_name: song_index |
|
features: |
|
- name: dic |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 2872 |
|
num_examples: 1 |
|
download_size: 4168 |
|
dataset_size: 2872 |
|
license: apache-2.0 |
|
task_categories: |
|
- question-answering |
|
tags: |
|
- music |
|
- art |
|
size_categories: |
|
- n<1K |
|
--- |
|
|
|
## How to Start |
|
|
|
``` |
|
from datasets import load_dataset |
|
import json |
|
|
|
qas = load_dataset("hobeter/JJQA","qa")["train"] |
|
songs = load_dataset("hobeter/JJQA","song")["train"] |
|
song_index=json.loads(load_dataset("hobeter/JJQA","song_index")["train"]["dic"][0])[0] |
|
``` |
|
|
|
# JJQA: a Chinese QA dataset on the lyrics of JJ Lin's songs |
|
|
|
**GitHub: https://github.com/bebetterest/JJQA** |
|
|
|
Large Language Models (LLMs) have shown powerful capability of text understanding, analysis and generation. It seems a good tool for text-style knowledge based question answering (QA) where semantically retrieving related texts, understanding them and generating correct answers are required. |
|
|
|
However, many feasible QA datasets are not challenging enough. First, given text-style knowledge might be easy to perceive and analyse. Second, the questions & answers follow commonsense. Thus, LLMs may benefit from training of language modeling and even take a shortcut. In this case, we want to build a new text-style knowledge based logical QA dataset where the text-style knowledge is tricky and LLMs are not likely to give correct answers without successfully retrieving and reasoning related texts. |
|
|
|
Chinese is a language where each single character could contain abundant meanings while just a few words, especially pieces of lyrics, are able to express complex conceptions, feelings and impressions. <font color=Purple size="">Besides, Junjie Lin, known as **[JJ Lin](https://www.jjlin.com/home/news/all)**, is a famous Singaporean Mandarin singer.</font> The lyrics of his songs are always imaginative, poetic and romantic. |
|
|
|
Hense, we propose JJQA, a Chinese text-style knowledge based question answering dataset on the lyrics of JJ Lin's songs, where related lyrics are provided as text-style knowledge for retrieval while the questions and answers are based on the lyrics. The Q&As are always abstract and follow anti-commonsense. For example, according to the related lyrics of a song called "爱情Yogurt", the question is "热量有什么作用？" ("What is the impact of heat?") and the answer is "降低爱情的过敏反应。" ("Ease the anaphylaxis of love"). It is indeed ridiculous and funny (you could find more in the dataset)🤪. Even human beings could not give the right answer without knowing the related lyrics. In addition, LLMs are not likely to naturally generate the right answer with the capability from training. Therefore, only if the related lyrics are retrieved and understood, are the right answers possibly generated By LLMs. |
|
|
|
|
|
## Dataset Details |
|
|
|
According to [QQMusicSpider](https://github.com/yangjianxin1/QQMusicSpider), we crawled lyrics of all songs of JJ Lin from [QQMusic](https://y.qq.com/). After data cleaning and label annotation, 648 Q&As with 181 related song lyrics are included. |
|
|
|
Three fields ("qa", "song", "song_index") are included in JJQA. |
|
|
|
"qa" contains Q&As with 6 features. "q" and "a" are a question and the corresponding answer. "song_title" and "song_id" are the title and the corresponding id of the related song. "id" is the id for the Q&A. "rf" locates the lines of lyrics for reference, splited by a space " ". |
|
|
|
"song" contains information of songs with 4 features. "title" and "name" are the title and the corresponding name of the song. "id" is the id of the song. "lyric" is the lyrics of the song, where each line is splited by "\n". |
|
|
|
"song_index" contains one dictionary, whose keys are the ids of songs and values are indexes of the corresponding song in "song" field, to align QAs with the corresponding songs. |
|
|
|
## Baselines |
|
|
|
We evaluate three baseline methods on JJQA. The first one (*wo_info*) is to "ask" the question directly without any additional lyric, which is to show the performance of uninformed LLMs; the second one (*w_song*) is to include whole lyrics of the related song as in-contexts; the third one (*w_rf*) is to just include related lyrics. *w_song* and *w_rf* are two reference lines for retrieval-based method. |
|
|
|
Six feasible LLMs (*ernie-turbo*, *chatglm2_6b_32k*, *qwen-turbo*, *baichuan2-7b-chat-v1*, *gpt-4*, *gpt-3.5-turbo*) are included. We apply *ernie-turbo* and *chatglm2_6b_32k* in [qianfan platform](https://cloud.baidu.com/product/wenxinworkshop); *qwen-turbo* and *baichuan2-7b-chat-v1* in [dashscope platform](https://dashscope.aliyun.com/); *gpt-4* and *gpt-3.5-turbo* in [openai platform](https://platform.openai.com/). |
|
|
|
We consider [BERTScore](https://github.com/Tiiiger/bert_score) with *rescale_with_baseline=True* as the metric. |
|
|
|
The results are as follows. |
|
|
|
|LLM|Method|Precision|Recall|F1|Date| |
|
|:---:|:---:|:---:|:---:|:---:|:---:| |
|
|*ernie-turbo*|*wo_info*|-0.0350|0.1568|0.0511|2023/11/06| |
|
|*ernie-turbo*|*w_song*|0.2472|0.5765|0.3895|2023/11/06| |
|
|*ernie-turbo*|*w_rf*|0.3600|0.6528|0.4864|2023/11/06|2023/11/05| |
|
|*chatglm2_6b_32k*|*wo_info*|0.0466|0.1787|0.1066|2023/11/05| |
|
|*chatglm2_6b_32k*|*w_song*|0.2361|0.4606|0.3335|2023/11/05| |
|
|*chatglm2_6b_32k*|*w_rf*|0.4650|0.6477|0.5436|2023/11/05| |
|
|*qwen-turbo*|*wo_info*|0.2331|0.2150|0.2208|2023/11/05| |
|
|*qwen-turbo*|*w_song*|0.7673|0.8041|0.7804|2023/11/05| |
|
|*qwen-turbo*|*w_rf*|0.8600|0.8251|0.8386|2023/11/05| |
|
|*baichuan2-7b-chat-v1*|*wo_info*|0.1755|0.2012|0.1857|2023/11/05| |
|
|*baichuan2-7b-chat-v1*|*w_song*|0.4635|0.6324|0.5371|2023/11/05| |
|
|*baichuan2-7b-chat-v1*|*w_rf*|0.6567|0.7272|0.6851|2023/11/05| |
|
|*gpt-3.5-turbo*|*wo_info*|0.2201|0.1983|0.2061|2023/11/06| |
|
|*gpt-3.5-turbo*|*w_song*|0.8031|0.7812|0.7884|2023/11/06| |
|
|*gpt-3.5-turbo*|*w_rf*|0.8110|0.7484|0.7758|2023/11/06| |
|
|*gpt-4*|*wo_info*|0.2426|0.2377|0.2376|2023/11/06| |
|
|*gpt-4*|*w_song*|0.8405|0.8587|0.8464|2023/11/06| |
|
|*gpt-4*|*w_rf*|0.8865|0.8643|0.8732|2023/11/06| |
|
|
|
It is worth noting that *Date* stands for the time (UTC+8) for evaluation. In addition, a small number of samples are not feasible in the dashscope platform because of its safety system. We just skip these Q&As. (1 sample for *qwen-turbo* *wo_info*; 3 samples for *qwen-turbo* *w_song*; 3 samples for *baichuan2-7b-chat-v1* *w_song*) |
|
|
