Rapid FrameNet annotation of spoken conversation transcripts

Rapid FrameNet annotation of spoken conversation transcripts
Rapid FrameNet annotation of spoken conversation transcripts
Jeremy Trione, Frederic Bechet, Benoit Favre, Alexis Nasr
Aix Marseille University, CNRS, LIF
Marseille, France
firstname.lastname@lif.univ-mrs.fr
Abstract
This paper presents the semantic annotation process of a corpus of spoken conversation transcriptions recorded in the Paris transport authority call-centre. The semantic model used is a FrameNet
model developed for the French language. The methodology proposed for the rapid annotation of this
corpus is a semi-supervised process where syntactic dependency annotations are used in conjunction
with a semantic lexicon in order to generate frame candidates for each turn of a conversation. This
first hypotheses generation is followed by a rule-based decision module in charge of filtering and
removing ambiguities in the frames generated. These rules are very specific, they don’t need to generalize to other examples as the final goal of this study is limited to the annotation of this given corpus,
on which a statistical frame parser will finally be trained. This paper describes this methodology and
give examples of annotations obtained. A first evaluation of the quality of the corpus obtained is also
given on a small gold corpus manually labeled.
1
Introduction
Parsing human-human conversations consists in enriching text transcription with structural and semantic information. Such information include sentence boundaries, syntactic and semantic parse of each
sentence, para-semantic traits related to several paralinguistic dimensions (emotion, polarity, behavioral
patterns) and finally discourse structure features in order to take into account the interactive nature of a
conversation.
The applicative context of this work is the automatic processing of human-human spoken conversations recorded in customer service telephone call centers. The goal of processing such data is to take
advantage of cues in order to automatically obtain relevant summaries and reports of such conversations
for speech mining applications. These processes are needed because coarse-grained analyses, such as
keyword search, are unable to capture relevant meaning and are therefore unable to understand human
dialogs.
Performing semantic parsing on spoken transcriptions is a challenging task Coppola et al. (2009).
Spoken conversation transcriptions have characteristics that make them very different to process from
written text Tur and De Mori (2011).
• non-canonical language: spontaneous speech represents a different level of language than the
canonical one used in written text such as newspaper articles;
• noisy messages: for spoken messages, automatic speech transcription systems make errors, especially when dealing with spontaneous speech;
• relevant and superfluous information: redundancy and digression make conversation messages
prone to contain superfluous information that need to be discarded;
• conversation transcripts are not self-sufficient: for spoken messages, even with a perfect transcription, non-lexical information (prosody, voice quality) has to be added to the transcription in order
to convey speakers’ intention (sentiment, behavior, polarity).
The general process of parsing conversations can be divided into three levels: conversational data
pre-processing; syntactic parsing; semantic parsing.
The pre-processing level involves the transcription (automatic or manual) of the spoken content and
the segmentation into speakers’ turns and sentence-like units.
The syntactic parsing level aims to uncover the word relationships (e.g. word order, constituents)
within a sentence and support the semantic layer of the language-processing pipeline. Shallow syntactic
processes, including part-of-speech and syntactic chunk tagging, are usually performed in a first stage.
One of the key activities described in this paper is the adaptation of a syntactic dependency parser to the
processing of spontaneous speech. The syntactic parses obtained are used in the next step for semantic
parsing.
The semantic parsing level is the process of producing semantic interpretations from words and other
linguistic events that are automatically detected in a text conversation or a speech signal. Many semantic
models have been proposed, ranging from formal models encoding deep semantic structures to shallow
ones considering only the main topic of a document and its main concepts or entities. We use in this
study a FrameNet-based approach to semantics that, without needing a full semantic parse of a message,
goes further than a simple flat translation of a message into basic concepts: FrameNet-based semantic
parsers detect in a sentence the expression of frames and their roles Gildea and Jurafsky (2002). Because
frames and roles abstract away from syntactic and lexical variation, FrameNet semantic analysis gives
enhanced access to the meaning of texts: of the kind who does what, and how where and when ?.
We describe in this paper the rapid semantic annotation of a corpus of human-human conversations
recorded in the Paris public authority call-centre, the RATP-DECODA corpus presented in Bechet et al.
(2012). This corpus is presented in section 2. The methodology followed is a semi-supervised process
where syntactic dependency annotations are used in conjunction with a semantic lexicon in order to
generate frame candidates for each turn of a conversation. This first hypotheses generation is followed
by a rule-based decision module in charge of filtering and removing ambiguities in the frames generated.
Section 3 describes the adaptation of syntactic parsing models to the processing of spontaneous speech.
Section 4 presents the FrameNet semantic model derived for annotating these call-centre conversations,
and finally section 5 reports some evaluation results on a small gold corpus manually annotated.
2
The RATP DECODA corpus
The RATP-DECODA1 corpus consists of 1514 conversations over the phone recorded at the Paris public
transport call center over a period of two days Bechet et al. (2012). The calls are recorded for the caller
and the agent, totaling over 74 hours of French-language speech.
The main problem with call-center data is that it often contains a large amount of personal data
information, belonging to the clients of the call-center. The conversations collected are very difficult
to anonymized, unless large amounts of signal are erased, and therefore the corpus collected can’t be
distributed toward the scientific community. In the DECODA project we are dealing with the call-center
of the Paris transport authority (RATP). This applicative framework is very interesting because it allows
us to easily collect large amount of data, from a large range of speakers, with very few personal data.
Indeed people hardly introduce themselves while phoning to obtain bus or subway directions, ask for a
lost luggage or for information about the traffic. Therefore this kind of data can be anonymized without
erasing a lot of signal.
Conversations last 3 minutes on average and usually involve only two speakers but there can be more
speakers when an agent calls another service while putting the customer on wait. Each conversation is
anonymized, segmented and transcribed. The call center dispenses information and customer services,
and the two-day recording period covers a large range of situations such as asking for schedules, directions, fares, lost objects or administrative inquiries.
Because speech that can be found in a call-centre context is highly spontaneous, many speech-specific
phenomenon such as disfluencies appear with a high frequency. In the RATP-DECODA corpus the
1
The RATP-DECODA corpus is available for research at the Ortolang SLDR data repository: http://sldr.org/sldr000847/fr
disfluencies considered correspond to repetitions (e.g. le le), discourse markers (e.g. euh, bien) and false
starts (e.g. bonj-).
Table 1 displays the amount of disfluencies found in the corpus, according to their types, as well as the
most frequent ones. As we can see, discourse markers are by far the most frequent type of disfluencies,
occurring in 28% of the speech segments.
disfluency type
discourse markers
# occ.
39125
% of turns
28.2%
repetitions
9647
8%
false starts
1913
1.1%
10 most frequent forms
[euh] [hein] [ah] [ben] [voila ]
[bon] [hm] [bah] [hm hm] [coutez]
[oui oui] [non non] [c’ est c’ est] [le
le] [de de]
[ouais ouais] [je je] [oui oui oui] [non
non non] [a a]
[s-] [p-] [l-] [m-] [d-]
[v-] [c-] [t-] [b-] [n-]
Table 1: Distribution of disfluencies in the RATP-DECODA corpus
Because of this high level of spontaneity, syntactic models such as Part-Of-Speech models or dependency models that were trained on written text have to be adapted. This semi-supervised annotation
method is presented in the next section.
3
Semi-supervised syntactic annotation
It has been shown in Bechet et al. (2014) that a great improvement in tagging and parsing performance
can be achieved by adapting models to the specificities of speech transcripts. Disfluencies can be integrated into the models without negative impact on the performance, if some annotated adaptation data is
available.
In order to adapt the tagger and parser to the specificities of oral French, we have parsed the RATPDECODA corpus with the MACAON tagger and dependency parser Nasr et al. (2011) and developed an
iterative process consisting in manually correcting errors found in the automatic annotations thanks to a
WEB-based interface Bazillon et al. (2012).
This interface allows writing regular expressions on the POS and dependency tags and the lexical
forms in order to correct the annotations on the whole RATP-DECODA corpus. Then the parser is
retrained with this corrected corpus. When the error rate computed on a development set is considered
acceptable, this correction process stops. The resulting corpus, although not perfect, constitutes our
training corpus, obtained at a reasonably low price compared to the whole manual annotation process of
the corpus. This process is described by figure 1.
The accuracy of the new parser is far above the accuracy of the parser trained on written text (French
TreeBank) : from 65.8% to 85.9% for Unlabeled Attachment Score (UAS) and from 58.3% to 83.8% for
Labeled Attachment Score (LAS). The performances of the parser can be compared to the performances
of a parser for written data despite the fact that the parser has been trained on a partially manually
corrected corpus.
Two reasons can explain this result. The first one is that the DECODA corpus has a quite restricted
and specific vocabulary and the parser used is quite good at learning lexical affinities. The second one
is that the DECODA corpus has a rather simple syntax with utterances generally restricted to simple
clauses and less common ambiguities, such as prepositional attachment and coordination, than written
texts.
One crucial issue is the amount of manual supervision needed to update the models. If a whole annotation of the corpus is needed, the process will be too costly whatever gain in performance is achieved.
We display in 2 the learning curve of the POS tagger, starting from a generic model trained on the French
DECODA
corpus
Manual
annotation
GOLD
corpus
gold
evaluation
TRAIN
corpus
MACAON
tools
Automatic
annotations
auto
Manual
correction
MACON model
retraining
Corrected
annotations
Figure 1: Semi-supervised adaptation process
TreeBank, and including some manual annotation on the target corpus. As we can see, even a very limited annotated subset of the corpus can boost performance: by adding as little as 20 dialogs, the POS
error rate drops by more than half (green curve) from 19% to 8%.
20
training=RATP-DECODA
training=FTB+RATP-DECODA
18
error rate
16
14
12
10
8
6
4
0
20
40
60
80
100
# of dialogs in training
Figure 2: Learning curve of the POS tagger with and without the FTB on the RATP-DECODA corpus
4
From syntactic to semantic annotation
Annotating manually with frame labels a corpus like the RATP-DECODA corpus is very costly. The
process we followed in this study is to take advantage of both syntactic annotations and external semantic
resources for performing this annotation at a very low cost.
We use in this study a FrameNet model adapted to French through the ASFALDA project. The
current model, under construction, is made of 106 frames from 9 domains. Each frame is associated to a
set of Lexical Units (LU) that can trigger the occurrence of a frame in a text. The first step, in annotating
a corpus with FrameNet, is to detect triggers and generate frame hypotheses for each detection. We did
this process on the RATP-DECODA corpus and found 188,231 potential triggers from 94 different frame
definitions.
The semi-supervised annotation process presented in this paper consists, for each LU, in searching
in the output of the parser for the dependencies (such as subject or object) of each trigger. This first
annotation is further refined thanks to semantic constraints on the possible dependent of a given LU,
considering the domain of the corpus.
The first step in our annotation process is to select a set of triggers on which frame selection will be
applied. In this study we limited our trigger set to the 200 most frequent verbs. By analyzing several
triggers on the corpus, we have defined the main domains and frames that we will use to annotate the
corpus.
Seven domains were considered:
• Motion.
Motion frames involve a theme which goes to an area. A vehicle can be use, and several other
parameter can be used like the source, the path used, the time, ...
Most used frames: Motion, Path shape, Ride vehicle, Arriving.
Examples:
Je
voudrais
Theme
I
would like to
Theme
aller
a Juvisy.
Motion
Area
go
to Juvisy.
Motion
Area
• Communication.
Communication frames involve a communicator sending a message to an addressee. While our
corpus is about call center conversation these frame are really important to describe the structure
of the call.
Most used frames: Communication, Request, Communication response.
Examples:
je
vous
appelle
parce qu’on m’a redirige vers vous.
Communicator
Addressee
Request
Message
I
call
you
because I was redirected to you.
Communicator
Request
Addressee
Message
• Sentiment expression.
Sentiment expression frames involve a communicator and an addressee. In our case these sentimental detections can evaluate the behavior of the people in the conversation.
Most used frames: Judgment direct address, Desiring.
Examples:
je
vous
remercie
beaucoup.
Communicator
Addressee
Judgment direct address
Degree
I
thank
you
a lot.
Communicator
Addressee
Judgment direct address
Degree
• Commerce.
Commerce frames involve a buyer, some goods and sometimes a seller. These frames are pretty
frequent in every call about tariff or fine paying.
Most used frames: Commerce Buy, Commerce pay.
Examples:
Vous
devez
Buyer
You
have
acheter
un ticket.
Commerce buy
Goods
to
Buyer
buy
a ticket.
Commerce buy
Goods
• Action.
We call action frames every frames that involve an action linked to a person. These kind of frames
are frequent in conversations that deal with misfortune of the caller.
Most used frames: Losing, Giving, Intentionally affect.
Examples:
J’
ai
Owner
perdu
mon telephone
dans le bus 38.
Losing
Possession
Place
I
lost
my phone
in the bus 38.
Owner
Losing
Possession
Place
As mentioned above, all these frames are triggered by the 200 most frequent verbs in the corpus.
However, FrameNet was not specially designed for spoken conversations and we had to extend it with
two new frames specific to this kind of data:
• Greetings.
This frame is triggered to represent the opening and the closing of a call. We use the same frame
in both cases (”Hello!”, ”Goodbye!”).
Examples:
Bonjour
monsieur.
Hello
Addressee
Hello
sir.
Hello
Addressee
• Agreement.
Agreement is a crucial frame in a dialog context. Detecting positive or negative answers to direct
Boolean questions in the context of a call-centre dialog is very important. The Agreement frames
refer to every mark of agreement (”yes”, ”of course”, . . . ).
Examples:
[...]
d’accord
merci.
Agreement
[...]
alright
Agreement
thank you.
Once this Frame selection process has been done, we are able to produce Frame hypotheses directly
from our trigger list of verbs and derive Frame elements from syntactic annotations. There can be only
one frame candidate by trigger. If a trigger can correspond to several frames, we use a rule-based approach to choose one frame according to the context. Because the semantic domain of our corpus is rather
limited, there are not many ambiguities and most of the verbs only corresponds to one frame, therefore
the set of rules needed to remove ambiguities is very limited and restricted to the 5 most frequent verbs
(such as aller - to go).
To write these rules we selected examples of these ambiguous verbs on our corpus, and wrote rules
taking into account the lexical and syntactic context of these verbs. Only six rules were needed, five
of them focused on disambiguating motion frames which are the most ambiguous frame in our corpus.
Table 2 show an example of rule.
Trigger
Rule
Example 1
Example 2
Aller
Trigger + non verb = motion frame
Un conseiller va prendre votre appel.
Il faut aller directement en agence.
Not a motion frame
Motion frame
Table 2: Example of syntactic rule.
This example illustrates the ambiguity of the trigger verb ”aller” (to go). This verb is very frequent
in French, particularly in spontaneous conversations. Similarly to English, this is a polysemic verb that
can means ”motion” as well as an ongoing action (e.g. ”I’m going to do something”). A simple rule
checking if this verb is associated to another verb or to an object can remove this ambiguity (example 1
in 2).
For each rule proposed, we checked on a reference corpus (gold corpus presented in the next section)
how many ambiguities were correctly resolved, and we kept only the most efficient ones. This process
was quite fast as it was done on the Frame hypotheses already produced and checked automatically on a
small gold corpus. Just a few iterations allowed us to produce the small set of rules that removed most
of the ambiguities of the most frequent verbs.
The Frame selection process consists now, for each trigger in a conversation, to check first if this
trigger is ambiguous or not. If it is, a rule should be applied to disambiguate it. If the trigger is not
ambiguous, we simply annotate the sentence with the corresponding frame from the dictionary. Due to
our very specific corpus, we have a low number of ambiguities and therefore a low number of rules.
5
Evaluation of Frame selection
A small gold corpus was manually defined and annotated. The automatic rule-based Frame selection
process is evaluated on this corpus, as presented in figure 1. Our gold corpus is a set on 21 conversations
from the RATP-DECODA corpus. These conversations were fully manually annotated by one annotator.
The tables below give a representation of the distribution of the frames on this subcorpus, comparing
manual annotation and automatic annotation.
Table 3 show us that on average there is at least one trigger per speaker turn. Moreover, we can
already tell that the automatic annotation predicts more triggers than the human annotator, and get more
variability in the frame chosen. In Table 4 we find our main domain on the RATP-DECODA corpus
through the frames. In fact ”Hello” and ”Judgment direct address” represent the structure of the call
(opening and closing), while ”Request”, ”Losing”, ”Motion” and ”Commerce buy” can easily represent
the reason of the call.
Number of Frames per Conversation
Number of Frames per speaker turn
Number of different frames
Manual annotation
23.67
0.97
26
Automatic annotation
31.33
1.24
37
Table 3: Frames distribution on the gold corpus.
Manual Annotation
Frame name
Occurrences
Agreement
161
Hello
95
Judgment direct address
59
Motion
33
Request
21
Waiting
20
Awareness
18
Communication
15
Losing
14
Commerce buy
9
Automatic annotation
Frame name
Occurrences
Agreement
216
Hello
95
Motion
45
Communication
34
Judgment direct address
27
Desiring
20
Awareness
19
Intentionally affect
16
Possibility
12
Waiting
11
Table 4: Top 10 used frames on the gold corpus.
The quality of the automatic prediction, with respect to the gold corpus, is presented in Table 5.
There are different levels of evaluation (trigger selection, frame level, frame element level, span, . . . ).
We chose to evaluate our annotation at the frame level. In other words, we evaluate if a trigger produced
the correct frame.
Automatic annotation
Recall
83.33
Precision
94.54
f-measure
88.58
Table 5: Evaluation on the automatic annotation on the gold corpus.
These first results are satisfying at the precision level is 94.5% of Frame predictions are correct. The
recall measure is lower but satisfactory considering that we limited the frame selection process to only
the most frequent verbs. A bigger gold corpus is now needed in order to assess the final quality of this
corpus.
6
Conclusion
We have presented in this paper a methodology for the rapid annotation of spoken conversation corpus
recorded in a French call-centre. This semi-supervised process uses syntactic dependency annotations
in conjunction with a FrameNet semantic lexicon. The rule-based decision module in charge of filtering
and removing ambiguities in the frames generated is evaluated at each learning cycle on a small manually
labelled gold corpus. The first evaluation described in this paper validate this approach by showing good
precision scores with an acceptable recall. This corpus will now be used to train a statistical frame parser
such as Das et al. (2014) that will be evaluated on other call-centre conversation transcriptions.
References
Bazillon, T., M. Deplano, F. Bechet, A. Nasr, and B. Favre (2012). Syntactic annotation of spontaneous
speech: application to call-center conversation data. In Proceedings of LREC, Istambul.
Bechet, F., B. Maza, N. Bigouroux, T. Bazillon, M. El-Beze, R. De Mori, and E. Arbillot (2012). Decoda:
a call-centre human-human spoken conversation corpus. In LREC, pp. 1343–1347.
Bechet, F., A. Nasr, and B. Favre (2014). Adapting dependency parsing to spontaneous speech for open
domain spoken language understanding. In Fifteenth Annual Conference of the International Speech
Communication Association.
Coppola, B., A. Moschitti, and G. Riccardi (2009). Shallow semantic parsing for spoken language
understanding. In Proceedings of Human Language Technologies: The 2009 Annual Conference of
the North American Chapter of the Association for Computational Linguistics, Companion Volume:
Short Papers, pp. 85–88. Association for Computational Linguistics.
Das, D., D. Chen, A. F. Martins, N. Schneider, and N. A. Smith (2014). Frame-semantic parsing. Computational Linguistics 40(1), 9–56.
Gildea, D. and D. Jurafsky (2002, September). Automatic labeling of semantic roles. Comput. Linguist. 28(3), 245–288.
Nasr, A., F. Béchet, J.-F. Rey, B. Favre, and J. Le Roux (2011). Macaon: An nlp tool suite for processing word lattices. In Proceedings of the 49th Annual Meeting of the Association for Computational
Linguistics: Human Language Technologies: Systems Demonstrations, pp. 86–91. Association for
Computational Linguistics.
Tur, G. and R. De Mori (2011). Spoken language understanding: Systems for extracting semantic information from speech. John Wiley & Sons.
7
Annex
Abandonment
Activity prepare
Agreement
Arriving
Attaching
Awareness
Becoming aware
Borrowing
Breathing
Bungling
Causation
Cause harm
Cause to perceive
Change event time
Chatting
Coming to be
Commerce pay
Communication
Compliance
Containing
Control
Defending
Desirable event
Duration description
Accuracy
Activity resume
Agree or refuse to act
Assessing
Attempt
Becoming
Being in effect
Breaking apart
Bringing
Canceling
Cause change
Cause motion
Certainty
Change operational state
Choosing
Commerce buy
Commerce sell
Communication response
Conferring benefit
Contingency
Cotheme
Departing
Desiring
Duration relation
Activity pause
Adjusting
Amalgamation
Assistance
Avoiding
Becoming a member
Being in operation
Breaking off
Building
Categorization
Cause change of strength
Cause to experience
Change accessibility
Change position on a scale
Closure
Commerce collect
Commitment
Complaining
Contacting
Contrition
Deciding
Deserving
Difficulty
Emitting
Emphasizing
Estimating
Existence
Experiencer focus
Feeling
Forming relationships
Giving
Halt
Hiding objects
Ingestion
Judgment
Labeling
Locale closure
Losing
Motion
Operating a system
Path shape
Placing
Posture
Preference
Process end
Questioning
Reliance
Repayment
Request
Reshaping
Respond to proposal
Scrutiny
Sending
Simultaneity
Storing
Success or failure
Taking sides
Theft
Trap
Using resource
Waiting
Working
Emptying
Event
Expend resource
Experiencer obj
Filling
Getting
Givinig
Having or lacking access
Hiring
Intentionally affect
Judgment direct address
Leadership
Locating
Making arrangements
Name conferral
Opinion
Perception active
Possession
Practice
Prevarication
Processing materials
Receiving
Removing
Replacing
Required event
Residence
Ride vehicle
Self motion
Sign
Spelling and pronouncing
Studying
Sufficiency
Telling
Topic
Triggering
Verification
Warning
Negation
Erasing
Evidence
Expensiveness
Explaining the facts
Forgiveness
Give impression
Grasp
Hello
Impact
Intentionally create
Justifying
Lending
Location in time
Memory
Offering
Participation
Performers and roles
Possibility
Predicting
Process continue
Process start
Redirecting
rentraire
Reporting
Reserving
Resolve problem
Run risk
Self otion
Similarity
Statement
Subscribing
Surpassing
Text creation
Transfer
Using
Wagering
Work
Table 6: Semantic Frames chosen to annotate the RATP-DECODA
corpus
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Related manuals

Download PDF

advertising