Extractive Question Answering on Queries in Hindi and Tamil Adhitya Thirumala Elisa Ferracane

2025-04-26 0 0 247.58KB 8 页 10玖币

侵权投诉

Extractive Question Answering on Queries

in Hindi and Tamil

Adhitya Thirumala, Elisa Ferracane

Abstract

Indic languages like Hindi and Tamil are underrepresented in the natural language

processing (NLP) ﬁeld compared to languages like English. Due to this under-

representation, performance on NLP tasks (such as search algorithms) in Indic

languages are inferior to their English counterparts. This diﬀerence dispropor-

tionately aﬀects those who come from lower socioeconomic statuses because they

consume the most Internet content in local languages. The goal of this project is

to build an NLP model that performs better than pre-existing models for the task

of extractive question-answering (QA) on a public dataset in Hindi and Tamil.

Extractive QA is an NLP task where answers to questions are extracted from

a corresponding body of text. To build the best solution, we used three diﬀerent

models. The ﬁrst model is an unmodiﬁed cross-lingual version of the NLP model

RoBERTa, known as XLM-RoBERTa, that is pretrained on 100 languages. The

second model is based on the pretrained RoBERTa model with an extra classiﬁca-

tion head for the question answering, but we used a custom Indic tokenizer, then

optimized hyperparameters and ﬁne tuned on the Indic dataset. The third model

is based on XLM-RoBERTa, but with extra ﬁnetuning and training on the Indic

dataset. We hypothesize the third model will perform best because of the variety

of languages the XLM-RoBERTa model has been pretrained on and the additional

ﬁnetuning on the Indic dataset. This hypothesis was proven wrong because the

paired RoBERTa models performed the best as the training data used was most

speciﬁc to the task performed as opposed to the XLM-RoBERTa models which had

much data that was not in either Hindi or Tamil.

I. Introduction

Indic languages are underrepresented in

the natural language processing (NLP)

ﬁeld compared to languages like English.

For example, in the corpora provided by

the Natural Language Toolkit (NLTK)

(Arora [2020]), over half of them are in

English. Indic languages include Hindi,

Tamil, Telugu, Malayalam, and others

that originate from the Indian subconti-

nent, and are spoken by the 1.7 billion

people that live on the Indian subconti-

nent. As seen in Table 1, around 57%

speak Hindi, while only 10.6% speak En-

glish. An English-centric view is bad for

the NLP community for a variety of rea-

sons. First, English’s written script is dif-

ferent from other languages, and a focus

on English or only English-like languages

(such as French, German, Spanish, etc.)

can make current systems work poorly on

languages with diﬀerent writing systems,

arXiv:2210.06356v1 [cs.CL] 27 Sep 2022

such as Chinese or Korean. Second, tok-

enization, which is the process of splitting

a document into smaller pieces, such as

paragraphs, sentences, or words, is more

straightforward in English and other lan-

guages that have an easier way to deﬁne

a “word.” However, in languages such

as Tamil where one does not tradition-

ally put spaces between words, the task

of tokenizing changes and becomes com-

pletely diﬀerent (Bender [2021]). Due

to this under-representation, NLP tasks

in Indic languages, such as search al-

gorithms and sentiment analysis, have

performance inferior to their English

counterparts (Arora [2020]). This diﬀer-

ence disproportionately aﬀects those who

come from lower socioeconomic statuses

because they consume the most Inter-

net content in local languages (S [2019]).

One way to allow people to access infor-

mation easier is to improve the perfor-

mance of extractive question-answering

algorithms on search engines in Indic lan-

guages. Extractive QA is an NLP task

that extracts an answer to questions from

a given body of text. This body of text

is known as the context and normally

contains the answer to the question pro-

vided. In the dataset we used, the con-

texts (and answers) came from Wikipedia

articles, and the questions were gener-

ated by native speakers of both Hindi

and Tamil. For this project, we will use

Hindi and Tamil as train data as they are

the ones provided by the Kaggle compe-

tition (Google [2021]) that inspired this

project.

Table 1:

Total speakers of languages in In-

dia (Source: 2011 Indian Census)

Language

Speakers(M) Speaker%

in Pop.

Hindi 692 57.1

English 129 10.6

Bengali 107 8.9

Marathi 99 8.2

Telugu 95 7.8

Tamil 77 6.3

Gujarati 60 5

Urdu 63 5.2

Kannada 59 4.9

Malayalam 36 2.9

Punjabi 36 3

Assamese 24 2

Maithili 14 1.2

Sanskrit 0.025 0.3

II. Related Work

Previously, the task of extractive QA

on English datasets was popularized in

SQuAD: 100,000+ Questions for Machine

Comprehension of Text (Rajpurkar et al.

[2016]). This paper outlined the creation

process for an English dataset for the

task of Extractive QA. The questions

were proposed by human annotators and

based on Wikipedia articles. Similarly,

the dataset that we used was also de-

rived from Wikipedia articles and had

questions produced by human annota-

tors. However, our dataset is diﬀerent

because it is one of the ﬁrst datasets pro-

duced in Hindi and Tamil. Currently, in

English, the best models on extractive

question answering use transformer mod-

els. Transformers are a computationally

eﬃcient deep learning model that intro-

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ExtractiveQuestionAnsweringonQueriesinHindiandTamilAdhityaThirumala,ElisaFerracaneAbstractIndiclanguageslikeHindiandTamilareunderrepresentedinthenaturallanguageprocessing(NLP)eldcomparedtolanguageslikeEnglish.Duetothisunder-representation,performanceonNLPtasks(suchassearchalgorithms)inIndiclanguage...

展开>> 收起<<

Extractive Question Answering on Queries in Hindi and Tamil Adhitya Thirumala Elisa Ferracane.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Extractive Question Answering on Queries in Hindi and Tamil Adhitya Thirumala Elisa Ferracane

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: