Semantic roles and semantic relations
Semantic role is the role of a dependent daughter with respect to its head constituent. On the other hand, semantic relation means relations between any two related constituents. Therefore, semantic relations have a broader coverage than head-dependent relations, since other than head-dependent relations, they also include coordinate relations and discourse relations.
Conventional text annotation, such as treebanks, might annotate syntactic dependent structures and semantic roles of constituents, such as Sinica Treebank (Chen et al. 2003). How is the semantic role of a dependent daughter determined? Usually, it is a result of considering the parameters of head constituent, dependent daughter, role marker, and phrasal/sentential pattern. By those factors, we decide a best role to describe this dependent daughter. The premise of above role assignment scheme is that there is only one semantic role for each dependent daughter and there is a limited set of predetermined semantic roles which you may choose from. Such a simple schema annotates only the semantic role of dependent daughter of head-dependent relations without considering coordinate relations and relations across sentences, such as discourse relations. Furthermore, a naïve role labeling system may cause the following problems.
First of all, each constituent may have dependent relations with multiple heads. For instance, in (3), an object of a verb-result compound in Chinese has two dependent relations. One is relation with respect to verb part and the second relation is with result part. Each bears different semantic relation. However, only single role is assigned to each dependent daughter. As a matter of fact, multiple relations commonly exist in a discourse structure, for example topic-comment chains.
-
(3)
張三打破花瓶 (The vase is the patient of 打 dǎ ‘hit’ or theme of 破 pò ‘broken’.)
zhāngsān__dǎpò__huāpíng
Zhangsan__hit__broken__vase
Zhangsan breaks the vase.
-
(4)
花瓶打破了 (The vase is the theme of 破 pò ‘broken’.)
huāpíng__dǎpò__le
vase__hit__broken__LE
The vase was broken.
The second problem is none-exclusiveness of semantic roles/relations. Semantic roles/relations may categorize into different feature dimensions. The relations between two constituents may be described by different categories of relations. For example in (3), the role of 花瓶 huāpíng ‘vase’ may be described by either dynamic relation of patient or static relation of theme. Similarly temporal relations, cause-result relations are falling into two different feature dimensions and the two relations may co-exist in CPCs. Conventionally, semantic roles may each other share the same characteristics and have idiosyncrasies. Each semantic role can be characterized by a few semantic features. The best role describing a dependent daughter is the role which matches most semantic features of the target constituent in its contextual environment (Dowty 1991). It causes competition among several possible candidate roles. Then, to determine the best role, in addition to feature matching, should each feature assign different weights in different contextual environments?
It results in a third practical problem: if a semantic role system is too complicated, it is very hard to annotate the best role/relation. Since a semantic role of a constituent is determined not only by the semantic relation with its dependent head but also by the entire contexture environment of the constituent. As we had mentioned, the major parameters for semantic role determination are head, dependent daughter, role marker, and phrasal/sentential pattern. Such criteria cause the complication of determining the best semantic role as exemplified in (5–7).
-
(5)
-
a.
張三被屠殺。(Zhangsan is the patient of “slaughter”; it focused on the dynamic feature dimension.)
zhāngsān__bèi__túshā
Zhangsan__BEI__slaughtered
Zhangsan was slaughtered.
-
b.
張三遭屠。 (Zhangsan is the theme of “being slaughtered”; it focused on the static feature dimension.)
zhāngsān__zāotú
Zhangsan__ZAO__slaughtered
Zhangsan was slaughtered.
-
(6)
-
a.
張三把牛肉軟化了 (Role marker bǎ delimits “beef” as a patient, i.e. instead of the main verb, bǎ becomes the head.)
zhāngsān__bǎ__niúròu__ruǎnhuà__le
Zhangsan__BA__beef__soften
Zhangsan softened the beef.
-
b.
牛肉軟化了 (“Beef” is the theme of “soften”.)
niúròu__ruǎnhuà__le
beef__soften
The beef is softened.
-
(7)
他買票(V1)上車(V2)。 (Do temporal relation, condition-consequence, cause-result, or elaboration all exist between V1 and V2?)
tā__mǎipiào__shàngchē
he__buy__ticket__get-on-the-bus
He bought a ticket (V1) and got on the bus (V2).
A better relational system for human annotators and also for future automation should meet the criteria of uniqueness and completeness; Uniqueness means two CPCs may assign a semantic relation (mostly are formed by a pair of semantic roles) which best describes their semantic relation. Completeness means two CPCs may be assigned some semantic role(s) to describe their semantic relations. That is, semantic relations between two CPCs are best described by one of the relation labels in the system without ambiguity.
Therefore, in designing a semantic relation system for CPCs, we encounter the following problems: How many relations are needed to meet the completeness criterion, that is, are all semantic relations between CPCs covered by the proposed classification system? How to meet the uniqueness criterion, that is, achieve a unique and consistent best labeling for CPCs? In the meanwhile, should we allow multiple relations while different interpretations occur? In the following subsections, by studying the related work and summarizing previous systems and offering our viewpoints, we attempt to answer the above questions and come up with a workable design methodology and a practical semantic relation system.
Related works
In the late 1980s, Mann and Thompson (1988) proposed a new interpretation of rhetorical structure theory (abbreviated RST) to describe natural texts, characterizing their structure primarily in terms of relations that hold between parts of the text. By examining real data, they provided a list which enumerates those which have proven the most useful 23 relations, including circumstance, solutionhood, elaboration, background, enablement, and motivation. Continuing the RST framework, Carlson and Marcu (2001; RST-DT) in their tagging reference manual claimed 16 classes covering a total of 78 relations, including attribution, background, cause, comparison, and condition.
Hovy and Maier (1992) summarized a survey of the conclusions of approximately 30 researchers who proposed more than 400 intersegment relations in different classification systems. Hovy then suggested using just as many relations as are required for determining the major aspects of English discourse structure, that is, approximately 70 relations, organized into a hierarchy of increasing specificity. The top-level classifications are divided into three parts: Ideational relation is defined between adjacent segments of material as those relations that express some experience of the world about us and within our imagination, for example, circumstance, cause/result, and general condition. Interpersonal relation is defined as holding between adjacent segments of textual material by which the author attempts to affect the address’s beliefs, attitudes, desire, and so on, by means of language, for example, interpretation and enablement. Textual relation is defined as holding between adjacent segments of text which exists solely due to the juxtaposition imposed by the nature of the presentation medium, for example conjunction and pre-sequence.
Miltsakaki et al. (2008) proposed four classes covering a total of 38 fine-grained relations which are annotated to the Penn Discourse Treebank (abbreviated PDTB), the largest-scale annotated corpus at the discourse level. Their relation structure is therefore most prevalent. The top-level classifications are divided into four parts: temporal is used when the situations described in the arguments are related temporally; contingency is used when the situations described in the arguments are causally influenced; comparison applies when a discourse relation is established between arguments in order to highlight prominent differences between the two situations; and expansion groups all the relations which expand the discourse and move forward its narrative or exposition.
Based on the lexically grounded approach of PDTB, Zhou and Xue (2012, 2015) presented a Chinese discourse annotation scheme and focused on the key characteristics of Chinese text which differs from English that we have mentioned in previous section. They claimed promising results on identifying a discourse relation; classifying the semantic type of explicit, implicit, or altLex; and determining the argument span. The agreements of the former are both over 95%, and the latter is over 80%.
Also based on PDTB classification, Huang et al. (2014) adopted discourse connectives to reveal explicit discourse relation in Chinese. They found there are 808 Chinese connectives which are eight times more than English connectives. Moreover, the Chinese discourse connectives have a variety of parts of speech that further deepen the difficulties of Chinese discourse relation labeling. By using semi-supervised learning method, the labeling result shows an F-score of 73.22%.
Zhou et al. (2014) followed the annotation scheme of PDTB and presented the first open discourse treebank for Chinese (abbreviated DTBC). They modified the PDTB sense hierarchy; three type level senses (i.e., CONTINGENCY. Inference, CONTINGENCY. Purpose and EXPANSION. Background) and two subtype level senses (i.e., EXPANSION. Conjunction. parallel and EXPANSION. Conjunction. progressive) were added to meet the needs of Chinese textual characteristics. They reported an over 90% inter-annotator agreement on discourse connective identification and an over 85% on sense annotation.
Li et al. (2014), based on the annotation of discourse connective in the Chinese discourse treebank (abbreviated CDTB), propose a three-level classification with a total of 17 fine-grained relations. The top-level classifications are divided into causality, transition, coordination, and explanation, and the third level merely lists the corresponding connective markers belonging to the middle-level types. The CDTB annotation is done by five long-term training annotators. The result showed a 94% inter-annotator agreement on explicit or implicit identification, an 82.3% agreement on explicit connective identification, and a 74.6% on implicit connective insertion.
Regarding practical relation identification problems, Wang et al. (2010) interpreted the implicit semantic relation between N-N compounds by adopting a dynamic approach using paraphrasing verbs. Not giving a set of relation candidates, they recognized each relation of N-N compounds by collocate verbs, for example, through Chinese word sketch engine finding people tell/seek a 愛情故事 àiqíng gùshì ‘love story’; and 民間故事 mínjiān gùshì ‘folk tales’ are stories that come from/spread in villages, then further recognize the fine-grained differences of meaning between the related compounds. On the contrary, Hong and Huang (2015) revealed the semantic relation between V-V compounds using an ontology-based conceptual classification, where three types of eventive relations, i.e., coordinate, modificational, and resultative, are predicted automatically. Xu and Huang (2014) discriminated sentences into general events, speech act, and modality types which is namely a task of event type classification.