Talking the Talk Does Not Entail Walking the Walk: On the Limits of Large Language Models

本文是LLM系列文章,针对《Talking the Talk Does Not Entail Walking the Walk: On the Limits of Large Language Models in Lexical Entailment Recognition》的翻译。

说到不一定做到:论大语言模型在词汇蕴涵识别中的局限性

摘要

动词构成了语言的支柱,为句子提供了结构和意义。然而,它们错综复杂的语义细微差别构成了一个长期的挑战。通过词汇蕴涵的概念理解动词关系对于理解句子含义和把握动词动态至关重要。这项工作研究了八种大型语言模型通过不同设计的提示策略和对来自两个词汇数据库(即 WordNet 和 HyperLex)的动词对的零/小样本设置来识别动词之间的词汇蕴涵关系的能力。我们的研究结果揭示了这些模型可以以中等好的性能处理词汇蕴涵识别任务,尽管有效性程度不同,条件不同。此外,使用小样本提示可以提高模型的性能。然而,完美解决这项任务对所有接受研究的 LLM 来说都是一个未解决的挑战,这为该主题的进一步研究发展提出了前景。

1 引言

2 相关工作

3 本研究中使用的资源

4 方法

5 结果

6 结论

我们介绍了 LLM 代表机构在处理词

Sure, I can help you develop a reasoner for EL or its extensions and apply the normalization procedure to the given TBOX. Here's an implementation in Python: ``` from rdflib import Graph, Namespace, OWL, RDFS from rdflib.plugins.sparql import prepareQuery # Define the TBOX axioms T = """ @prefix : <http://example.org/> . @prefix owl: <http://www.w3.org/2002/07/owl#> . @prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> . :A owl:equivalentClass [ a owl:Restriction ; owl:onProperty :r ; owl:someValuesFrom :C ] . :C rdfs:subClassOf [ a owl:Restriction ; owl:onProperty :s ; owl:someValuesFrom :D ] . :B rdfs:subClassOf [ a owl:Class ; owl:intersectionOf ( [ a owl:Restriction ; owl:onProperty :r ; owl:someValuesFrom [ a owl:Restriction ; owl:onProperty :s ; owl:someValuesFrom :T ] ] :D ) ] . """ # Create a graph and parse the TBOX axioms g = Graph() g.parse(data=T, format="turtle") # Define the query to check for entailment QUERY = prepareQuery(""" PREFIX : <http://example.org/> PREFIX owl: <http://www.w3.org/2002/07/owl#> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> SELECT ?x WHERE { :A rdfs:subClassOf ?x . ?x rdfs:subClassOf :D . } """, initNs={"owl": OWL, "rdfs": RDFS}) # Apply normalization procedure by adding the inverse role hierarchy g.add((OWL.InverseRoleOf, RDF.type, OWL.Axiom)) g.add((OWL.InverseRoleOf, OWL.inverseOf, OWL.InverseRoleOf)) # Run the query to check for entailment results = g.query(QUERY) if len(results) > 0: print("A entails D") else: print("A does not entail D") ``` In this implementation, we use the `rdflib` library to represent and manipulate the TBOX axioms as an RDF graph. We then define a SPARQL query to check whether `A ⊑ D` is entailed by the TBOX. Finally, we apply the normalization procedure by adding the inverse role hierarchy to the graph, and run the query to check for entailment. Note that this implementation assumes the EL profile with inverse roles. If you need to support other extensions of EL, you may need to modify the normalization procedure accordingly. Also, this implementation does not handle the addition of new axioms to the TBOX or the removal of existing axioms, which would require additional reasoning and consistency checks.
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

UnknownBody

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值