聚焦焦点关系的隐性语义关系识别

133 浏览量更新于2024-07-15 收藏 1.3MB PDF 举报

本文主要探讨了"Focus-Sensitive Relation Disambiguation for Implicit Discourse Relation Detection"这一主题，它属于计算机科学领域的研究论文。在自然语言处理领域，隐性语义关系检测是一项极具挑战性的任务，特别是在识别和限定那些难以察觉、模糊不清的隐性语义关联时。作者团队，由来自苏州大学计算机科学技术学院的Yu Hong、Siyuan Ding、Yang Xu、Jianmin Yao、Qiaoming Zhu和Guodong Zhou等人组成，专注于处理这类特别棘手的问题。论文的核心创新在于提出了一个名为"Implicit Discourse Relation Disambiguation (IDRD)"的新任务，旨在解决隐性语义关系的辨识难题。他们认识到，当隐性关系被焦点句子成分触发时，其真实性尤为关键。因此，他们设计了一个聚焦敏感的关联辨识模型，该模型能够准确确认由焦点句引发的真实关系。在实现这一目标的过程中，论文提出了一种基于话题驱动的焦点识别方法。这种方法利用文本中的主题信息来辅助确定焦点句，这对于理解隐性关联的上下文至关重要。此外，他们还开发了一套名为"Relation Search System (RSS)"的工具，作为支持关系辨识的重要组成部分。RSS可能包括了诸如语义分析、句法解析以及可能的数据挖掘技术，以便于在大量文本数据中定位和分析潜在的隐性关系。通过这些新颖的方法，论文试图改进现有的关系检测系统，提升对隐性语义关系的准确识别能力。这不仅有助于提高文本理解和分析的精确度，也为后续的研究工作提供了新的视角和工具，推动了隐性语义分析在自然语言处理领域的进一步发展。这篇论文对于深化对隐性语义的理解，以及提高文本处理系统的智能水平具有重要的理论价值和实践意义。

Front. Comput. Sci.

relation with the meaning of "after then" or Causal with that

of "because". It normally confuses the relation detection sys-

tems which favors the usage of connectives in relation de-

tection, e.g., [13] and [14] connective-to-relation mapping in

relation inference.

[4] gives a ﬁrst step at disambiguating the relation senses

of a small subset of connectives (e.g., the connectives "since",

"while" and "as") at the paragraph level. By means of syntac-

tic features derived from PDTB and a MaxEnt model, they

distinguish the relations of Temporal, causal Contingency,

contrastive Comparison for the connectives.

But, [5] shows connective disambiguation can only pro-

vide no more than 1% improvement for relation detection (on

nearly all sections of PDTB). Correspondingly, [14] proves

that the disambiguation doesn’t work when connectives ei-

ther occur as mood particles or are purely pragmatic failures.

In this paper, we specialize in implicit relation disam-

biguation. As will being seen, ambiguous implicit relations

are much harder to identify and eliminate than explicit ones,

more easily cheating relation detection systems because of

their imperceptibility.

3 Deﬁnition of IDRD

IDRD aims to identify and eliminate ambiguous implicit re-

lations, reducing confusion of multiple relations implied in

pairwise arguments. Ambiguous implicit relation has the

characteristics as:

• Implicit: relation between arguments isn’t directly sig-

naled by an overt connective or any other explicit mark.

• Ambiguous: a pair of arguments implies more than one

type of potential relation but only one is truly correct.

• Deceptive: some of the implied relations are pseudo-

correct, which easily cheats the relation detection sys-

tems.

Given a pair of arguments as input, IDRD is required to gen-

erate results in two forms. One is to output all the relations

implied in the arguments along with corresponding triggers,

where a trigger is made of sentence constituents that can be

linked to reﬂect a speciﬁc relation. The other is to give the

correct relation. See (4), which shows the standard output of

IDRD for the instance of (3).

(4)INPUT:

Tom is young (Arg1), even younger than old Billy (Arg2).

OUTPUT:

Relation 1→ Expansion.Instantiation

Trigger 1→ "young"(Arg1) + "younger than Billy"(Arg2)

Relation 2→ Comparison.Contrast

Trigger 2→ "young"(Arg1) + "old"(Arg2)

(Correct Relation→Relation1)

(Pseudo-Correct Relation→Relation2)

where, there are two relations probably considered as reason-

able, Contrast and Instantiation. In this case, however, the

truly correct relation is Instantiation. It is because that Arg2

actually presents an example (younger than Billy) to support

the opinion in Arg1 (Tom is young). The other relation,

Contrast, is incorrect because the main point is not to prove

that Tom is younger than Bill. It is easily-misjudged to be

correct, therefore named as a pseudo-correct relation.

A general discourse relation detection system is only re-

quired to output the type tag of the relation held between ar-

guments. Though the system may output more than one type

of relation no matter whether all of them are correct, such as

Instantiation and Contrast in (4). By contrast, an IDRD sys-

tem will provide richer information, including relation tags

and triggers. Besides, if there are more than one type of re-

lation detected, the system is required to make a ﬁnal detec-

tion on the correctness, marking out the sole genuine relation

type. For example, it adds a ”Correct Relation" to Relation1

in (4). Correspondingly the Relation2 is pseudo-correct.

As will be seen, we use a Relation Search System (RSS) to

obtain implied relations and triggers. On the basis, we detect

truly correct relations and eliminate pseudo-correct ones with

a focus-sensitive relation disambiguation model. But in the

writing of this paper, we begin with a discussion of the dis-

ambiguation model (Section 4), aimed at getting to the point

quickly, secondly we show RSS (Section 5), thirdly we intro-

duce eight relation detection systems in Section 6, including

two the-state-of-the-art systems and six modiﬁed ones by our

disambiguation model, and ﬁnally we show the experiments

in Section 7.

4 Focus-sensitive Relation Disambiguation

4.1 Disambiguation Model

We suppose that truly correct relations between arguments

are triggered by focal sentence constituents while pseudo-

correct ones are triggered by complementary or ancillary sen-

tence constituents. Accordingly, if a pair of arguments im-

plies multiple relations, the correct one can be determined by

the following model:

∗

= argmaxP(r| f ) (1)

剩余14页未读，继续阅读

weixin_38550459

粉丝: 4
资源: 956

聚焦焦点关系的隐性语义关系识别

Using-Android-Implicit-Intents:安卓

Deep Cost-Sensitive Kernel Machine for Binary Software Vulnerability Detection

A hypergraph based context-sensitive representation technique for VHR remote sensing image change detection.

ActiveRec: A Novel Context-Sensitive Ranking Method for Active Movie Recommendation

Deep Ranking Based Cost-sensitive Multi-label Learning for Distant Supervision Relation Extraction

Topic-sensitive PageRank - a context-sensitive ranking algorithm

Topic-sensitive_pagerank_A_context-sensitive_ranking_algorithm_for_web_search

Words-away:防止文本敏感词检测 - Prevent the text-sensitive-word detection of coolapk

Query-Aware Locality-Sensitive Hashing for Approximate Nearest Neighbor Search

Highly-sensitive and large-dynamic diffuse optical tomography system for breast tumor detection

最新资源