Coreference and anaphora语用标注一例
- 主题发起人 xujiajin
- 时间
SPAACy - A semi-automated tool for annotating dialogue acts
这里还有一个语用标注的工具。
. Speech act level
The most important assumption on the speech act level is that by first identifying
the grammatical type of an utterance, the choices for its associated speech
act can be greatly limited in order to simplify the analysis. Thus, in service dialogues,
if an utterance has been identified as a question, it is most likely that
its pragmatic force is that of a request, either for an action to be performed on
the part of the hearer or for the hearer to provide some information. For utterances
with declarative form, on the other hand, the choices are more varied
as it could represent C depending on the context C an answer, a statement or,
given an appropriate intonation, also a kind of request for information, etc.
However, identifying the grammatical category of an utterance on its own,
even within its immediate context, will often also not be sufficient, but possibly
needs to be augmented by additional information contained in the appropriate
attributes, such as topic, mode, polarity, etc. For example, the declarative
“if you can (yes3) just bear with me please”, through its initial if-clause,
seems to state a condition. The annotation tool would include such information
into the mode attribute, but it does also pick up on the sub-phrase “bear
withme”, which is classified as a key-phrase indicating dialogue management,
and could therefore reinterpret the speech-act of the utterance as being a holddirective,
rather than an answer or inform. In a similar way, the following previously
unclassified example is reinterpreted, on the basis of the information
contained in the topic and mode attributes, as containing two informs, which
are functionally types of declaratives.
<frag sp-act=“inform” polarity=“positive” topic=“time-location-journey”
mode=“reason”>
because of the time that you’re travelling out {#} of Birmingham New
Street
</frag>
</utt>
<utt id=“ ”>
<decl sp-act=“inform” polarity=“negative” topic=“fare-journey”
mode=“exists”> there’s not any advance purchase tickets left {#}
</decl>
At the current stage, and with only a few rules for speech-act identification implemented
so far, we have already achieved a recall rate of up to 70.1% on some
dialogues, with an average of 59.9% over the first fifteen dialogues of Trainline
data. From a cursory analysis, e.g. by manually post-editing some of these dialogues,
we estimate that precision will be close to 95%, depending on the complexity
of the dialogue.Without adding additional rules, we have also achieved
up to 62.5% recall on a different type of service dialogue ofmore diverse nature,
although the average over the first ten dialogues here drops to 45.4%.
这里还有一个语用标注的工具。
. Speech act level
The most important assumption on the speech act level is that by first identifying
the grammatical type of an utterance, the choices for its associated speech
act can be greatly limited in order to simplify the analysis. Thus, in service dialogues,
if an utterance has been identified as a question, it is most likely that
its pragmatic force is that of a request, either for an action to be performed on
the part of the hearer or for the hearer to provide some information. For utterances
with declarative form, on the other hand, the choices are more varied
as it could represent C depending on the context C an answer, a statement or,
given an appropriate intonation, also a kind of request for information, etc.
However, identifying the grammatical category of an utterance on its own,
even within its immediate context, will often also not be sufficient, but possibly
needs to be augmented by additional information contained in the appropriate
attributes, such as topic, mode, polarity, etc. For example, the declarative
“if you can (yes3) just bear with me please”, through its initial if-clause,
seems to state a condition. The annotation tool would include such information
into the mode attribute, but it does also pick up on the sub-phrase “bear
withme”, which is classified as a key-phrase indicating dialogue management,
and could therefore reinterpret the speech-act of the utterance as being a holddirective,
rather than an answer or inform. In a similar way, the following previously
unclassified example is reinterpreted, on the basis of the information
contained in the topic and mode attributes, as containing two informs, which
are functionally types of declaratives.
<frag sp-act=“inform” polarity=“positive” topic=“time-location-journey”
mode=“reason”>
because of the time that you’re travelling out {#} of Birmingham New
Street
</frag>
</utt>
<utt id=“ ”>
<decl sp-act=“inform” polarity=“negative” topic=“fare-journey”
mode=“exists”> there’s not any advance purchase tickets left {#}
</decl>
At the current stage, and with only a few rules for speech-act identification implemented
so far, we have already achieved a recall rate of up to 70.1% on some
dialogues, with an average of 59.9% over the first fifteen dialogues of Trainline
data. From a cursory analysis, e.g. by manually post-editing some of these dialogues,
we estimate that precision will be close to 95%, depending on the complexity
of the dialogue.Without adding additional rules, we have also achieved
up to 62.5% recall on a different type of service dialogue ofmore diverse nature,
although the average over the first ten dialogues here drops to 45.4%.
[本贴已被 作者 于 2005年06月28日 16时43分50秒 编辑过]
[本贴已被 作者 于 2005年06月28日 16时44分26秒 编辑过]