Benchmarking Knowledge-Enhanced Commonsense Question Answering via Knowledge-to-Text Transformation

Ning Bian, Xianpei Han,* Bo Chen, Le Sun

---

작성중...

Abstract
1. how far can we get by exploiting external knowledge for CQA?
2. how much potential knowledge has been exploited in current CQA model?
3. which are the most promising directions for future CQA?

we use a simple(not specialized to specific models) and effective knowledge-to-text transformation framework!

Result: ktt SOP on CommonsenseQA

Direction:

• potential of knowledge is still far from being fully exploited with golden knowledge(?),
• context-sensitive knowledge selection
• heterogeneous knowledge exploitation, commonsense-rich language models are promising

---

Intro

knowledged-enhanced cqa --> external knowledge should be incorporated in a simple way that is not specialized to specific models/components

challenging bc:
1. struct knowl and unstruct text knowl heterogeneity
2. context sensitivity: kb contain many facts but only several relevant

ktt framework 3 stages:
1. retrieve facts from KG
2. these facts to textual desc(tkd) via 3 transformation algo: template-based, paraphrasing-based, retrieval based -> hetero solve
3. machine reading comprehension models to predict answers by using q and tkd

tkd of ktt:
1. template
2. paraphrasing: template -> top-M paraphrases using beam-search decoding; encoder-decoder paraphrasing model trained on PPDB and WikiAnswers
3. retrieval:

• above two only gen pseudo textual desc -> retrieve from real world corpus (wikipedia), distant supervision assumption; "if a sentence contains the entities on a knowledge path, it will express the meaning of the knowledge path"
• split all wiki docs into seperate sents and build wiki sent retrieval sys using ElasticSearch(?)
• template as query to retrieve wiki sent containing concepts on knowledge paths via BM25(?) algo
• rank 1 sent -> description

mrc answ pred:
mrc bc:
1. automatically learn to identify relevant information in a document; here select q-rel knowl
2. mrc well studied tech
mrc == lm(?)