Automatic Extraction of Task Statements

from Structured Meeting Content

Katashi Nagao, Kei Inoue, Naoya Morita and Shigeki Matsubara

Department of Media Science, Graduate School of Information Science, Nagoya University, Nagoya, Japan

Keywords: Discussion Mining, Discussion Structure, Task Statement, Automatic Extraction, Probability Model.

Abstract: We previously developed a discussion mining system that records face-to-face meetings in detail, analyzes

their content, and conducts knowledge discovery. Looking back on past discussion content by browsing

documents, such as minutes, is an effective means for conducting future activities. In meetings at which some

research topics are regularly discussed, such as seminars in laboratories, the presenters are required to discuss

future issues by checking urgent matters from the discussion records. We call statements including advice or

requests proposed at previous meetings “task statements” and propose a method for automatically extracting

them. With this method, based on certain semantic attributes and linguistic characteristics of statements, a

probabilistic model is created using the maximum entropy method. A statement is judged whether it is a task

statement according to its probability. A seminar-based experiment validated the effectiveness of the proposed

extraction method.

1 INTRODUCTION

Evidence-based research, such as research on life-

logging (Sellen and Whittaker, 2010) and big data

applications (Mayer-Schönberger and Cukier, 2013),

has been receiving much attention and has led to the

proposal of techniques for improving the quality of

life by storing and analyzing data on daily activities

in large quantities. These types of techniques have

been applied in the education sector, but a crucial

problem remains to be overcome: it is generally

difficult to record intellectual activities and

accumulate and analyze such data on a large scale.

Since this kind of data is not possible to compress in

a manner, such as taking the average, it is necessary

to maintain the original data as the instances of cases.

Such human intellectual-activity data should be

treated as big data in the near future.

The aim of the study was to develop an

environment in which the skills of students are

empowered by analysis of abundant discussion data.

We have developed a “discussion mining” system

(Nagao et al., 2005) that generates meeting minutes

linked to videos and audio data of the discussions. It

also creates metadata for use in clarifying the

semantic structure of the discussions. Statements

made in meetings are classified into two types: “start-

up,” which means the statement starts a discussion of

a new topic, and “follow-up,” which means the

statement continues the current topic of the

discussion. The discussions are then segmented into

discussion chunks corresponding to topics on the

basis of the statement type. A discussion chunk is a

set of statements that are semantically associated with

each other.

Looking back and reconsidering the content of

past discussions by browsing the recorded meeting

content is an effective means for efficiently

conducting future activities. In meetings at which

some research topics are regularly discussed, such as

seminars in laboratories, the presenters are required

to discuss future issues by checking urgent matters

from the structured discussion content.

In this paper, we call statements including advice

or requests proposed at previous meetings “task

statements” and propose a method for automatically

extracting them. With this method, based on certain

semantic attributes and linguistic characteristics of

statements, a probabilistic model is created using the

maximum entropy method (Wu, 1997).

We first discuss related work then describe our

mining system. We then explain the proposed

automatic extraction method of task statements from

structured meeting content and describe our

evaluation of this proposed method through a

statistical hypothesis test.

Nagao, K., Inoue, K., Morita, N. and Matsubara, S..

Automatic Extraction of Task Statements from Structured Meeting Content.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 1: KDIR, pages 307-315

ISBN: 978-989-758-158-8

307

2 RELATED WORK

There is not much research on the method of

extracting useful information from the minutes of

face-to-face meetings. The reason is that it is very

costly to record all statements in meetings as text and

to maintain them to analyze using a machine learning

technique. We have solved this problem by the

development and deployment of a variety of

specialized tools.

To extract useful information, such as advice and

requests, by using the records of the communication

in online forums has been widely studied. Since it is

common in terms of performing information

extraction from the text of communication records,

we describe the relevance of these studies to ours in

this section.

Extraction of request representations from public

comments embedded in online meetings organized by

government agencies has been conducted (Kanayama

and Nasukawa, 2008). Our proposed method covers

not only requests by participants but also the

presenter's responses and comments on future tasks.

Extracting contexts and answers of questions

from the online travel forum “TripAdvisor” by using

a structural support vector machine (SVM) was

conducted (Yang, Cao and Lin, 2009). Since a target

of this research was to assign the labels Context,

Question, and Answer to each of the conversational

sentences with the proposed method, it seems to be

difficult to directly apply the method to task statement

extraction. Assuming that if there is a statement that

indicates the emergence of a task statement, the

proposed method may be applied to our task

extraction problem.

A rule-based approach to information extraction

from online discussion boards was studied

(Sarencheh et al., 2010). Some discussion boards are

created with software such as SMF, phpBB, and

vBulletin. The authors of that study developed a rule-

base that includes rules regarding the relationships

between the discussion structure and article content

formatted in HTML tags. Since these rules are

customized for each forum creation software and

several versions, the versatility of the proposed

method is not high with this approach. To increase the

accuracy of task statement extraction, it is

conceivable to combine machine learning and rule-

based approaches.

Qu and Liu (Qu and Liu, 2012) investigated

sentence dependency tagging of question and answer

(QA) threads in online forums. They defined the

thread tagging task as a two-step process. In the first

step, sentence types (they defined 13 types such as

Problem, Answer, and Confirmation) are labelled. In

the second step, dependencies between sentences are

tagged. With our approach, discussions are tagged

manually by speakers during a meeting in a very

convenient way and there is no need to consider all

statements and their relationships.

Wicaksono and Myaeng (Wicaksono and Myaeng,

2013) provided a methodology for extracting advice-

revealing sentences from online travel forums. They

identified three different types of features (i.e.,

syntactic, context, and sentence informativeness) and

proposed a hidden Markov model (HMM)-based

method for labelling sequential sentences. Their

features are similar to ours. Since the structure of the

discussion is determined in advance of information

extraction, our approach is easier to use than

extracting the advice sentences from general online

forums.

3 DISCUSSION MINING SYSTEM

Our discussion mining system promotes knowledge

discovery from the content of face-to-face meeting

discussions. Based on the meeting environment

shown in Figure 1, multimedia minutes are generated

for meetings in real time semi-automatically and

linked with audiovisual data. The discussions are

structured using a personal device called a

“discussion commander” that captures relevant

information. The content created from this

information is then viewed using a “discussion

browser,” which provides a search function that

enables users to browse the discussion details.

Figure 1: Overview of discussion mining system.

3.1 Recording and Structuring

Discussions

Meeting discussions are automatically recorded, and

the content is composed of structured multimedia data

including text and video. The recorded meeting

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

308

content is segmented on the basis of the discussion

chunks. The segments are connected to visual and

auditory data corresponding to the segmented

meeting scenes.

Previous studies on structuring discussions

include the issue-based information system (IBIS),

graphical IBIS (gIBIS) (Conklin and Begeman,

1988), and argument diagramming of meeting

conversations (Rienks, Heylen and van der Weijden,

2005), which take into account the structures of

semantic discussions. However, such a semantic

structure of discussions is still not at a practical level,

and most studies on technology for generating

discussion minutes have focused on devices, such as

meeting recorders (Lee et al., 2002), for automatically

recognizing auditory and visual data.

Our system uses natural language processing to

not only support comprehension of the arguments in

discussions but also form diversified perspectives

using auditory and visual information in slides as well

as other presentation media. It uses metadata to

clarify the semantic structures of discussion content.

Overall, our discussion mining system supports the

creation of minutes of face-to-face meetings, records

meeting scenes with cameras and microphones, and

generates meta-information that relates elements in

the meeting content.

In addition, the system graphically displays the

structure of the discussions to facilitate understanding

of the meeting content; therefore, improving the

effectiveness of statements made during the

discussions. The discussion commander has several

functions for facilitating discussions, including one

for pointing to and/or highlighting certain areas in

presentation slides and one for underlining text in the

slides displayed on the main screen.

Each statement is one of two types: “start-up” and

“follow-up.” The “start-up” type is assigned to a

statement that introduces a new topic, while the

“follow-up” type is assigned to a statement that is on

the same topic as the previous statement (i.e., it

inherits the predecessor’s topic). Each discussion

chunk begins with a start-up statement, as shown in

Figure 2. Speakers are required to manually associate

their statements with these attribute types with their

discussion commanders when they start speaking

during a meeting.

Real-time visualization of the discussion structure

and visual referents (pointed texts and images)

facilitate the current discussion. Moreover, the

discussion structures can be modified by changing the

parent nodes of the follow-up statements and by

referring again to previous visual referents. A

participant can perform these modifications by

Figure 2: Discussion structure.

using his or her discussion commander. The

participant can also use the discussion commander for

marking the current statement by pressing the

marking button. When these buttons are pressed, the

system records who pressed the button and the target

statement. Presenters mark the statement that they

want to check later during the meeting and retrieve

the marked statements by using the discussion

browser mentioned in the next subsection.

3.2 Discussion Browser

The information accumulated with the discussion

mining system is presented synchronously in the

discussion browser with the timeline of the

corresponding meeting, as shown in Figure 3. It

consists of a video view, slide view, discussion view,

search menu, and layered seek bar. The discussion

browser provides a function for searching and

browsing details about the discussions. For example,

a participant can refer to a certain portion of a

preceding discussion by doing a search using

keywords or speaker names then browsing the details

of the statements in the search results.

Figure 3: Discussion browser interface.

Automatic Extraction of Task Statements from Structured Meeting Content

309

People who did not participate in the meeting can

search and browse the important meeting elements

displayed in the layered seek bar by searching for

statements in discussions that were marked using a

discussion commander or by surveying the frequency

distributions of keywords.

The discussion browser has the following five

components:

(1) Video view

The video view provides recorded videos of the

meeting, including the participants, presenter, and

screen.

(2) Slide view

Thumbnail images of presentation slides used in the

meeting are listed in this view. The images are placed

in the list in the order in which they are displayed on

the main screen.

(3) Discussion view

The discussion view consists of text forms in which

the content of the minutes primarily constitute

information input by a secretary and relationship

links, which visualize the structure of the discussions.

(4) Search menu

Three types of search queries are available in the

search menu: speaker name, search target (either the

content of a slide, a statement, or both), and

keywords. The search results are shown in the layered

seek bar (matched elements in the timeline are

highlighted) and in the discussion view (discussions

where the matched elements appear are highlighted).

(5) Layered seek bar

The elements comprising the meeting content are

displayed in the layered seek bar. Various bars are

generated depending on the element type.

4 AUTOMATIC EXTRACTION

OF TASK STATEMENTS

Remembering past discussion content helps us to

seamlessly carry out future activities. For example, in

laboratory seminars, presenters can remember

suggestions and requests about their research

activities from the discussion content recorded in

detail. The meeting content contains useful

information for the presenters, but it is burdensome to

read the information. Necessary information is

concealed in a large amount of statements, so it is not

easy to find. It is problematic if past discussions are

not being reviewed, even for other speakers not only

presenters. Therefore, it is necessary to extract the

information concerning unsolved issues from

previous discussions. We call statements including

future tasks “task statements.”

Our proposed method determines whether the

statements are about future tasks (i.e., task

statements). Some attributes including linguistic

characteristics, structures of discussions, and speaker

information are used to create a probabilistic model.

4.1 Model of Task Statements

A task statement can include any of the following

content:

1. Proposals, suggestions, or requests provided

during the meeting

The presenter has determined that they should be

considered.

2. Problems to be solved

The presenter has determined that they should be

solved.

3. Tasks not yet carried out before the meeting

Sometimes the presenter has already noticed them.

Candidates of task statements are fragments of a

discussion chunk, as mentioned earlier. A typical

discussion chunk is made from one or more questions

and comments of the meeting participants and the

presenter’s responses to them. A coherent piece of

discussion content related to tasks consists of

questions/comments and their responses. Thus,

“participants’ questions/comments + presenter’s

response” is a primary candidate and a target of

retrieval. “Participants’ questions/comments and no

response” is a secondary candidate.

Figure 4 shows example candidates of task

statements.

Figure 4: Candidates of task statements.

By using the correct data that were manually

created from past meeting content, the method

generates a probability model by using the maximum

entropy method. For each candidate, the method

calculates the probabilities of candidates of a task

statement using the generated probabilistic model. A

candidate whose probability value exceeds a certain

threshold (e.g., 0.5) is extracted as a task statement.

Figure 5 shows the overall process of extracting task

statements.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

310

Figure 5: Overall process of extraction.

To discover the characteristics as clues to

extracting task statements, some past meeting content

was manually analyzed. The survey data included 11

types of meeting content and 598 groups of

statements (candidates). Each presenter of the

meeting manually selected task statements from each

type of content.

As a result of manually extracting task statements

of the survey data, 246 task statements were found

corresponding to 41.1% of all candidates. By

comparing the percentages of task statements, we

analyzed the characteristics of the task statements.

For example, the attribute of speakers of

statements and percentages of task statements to

which the attribute contributed are listed in Table 1.

Statements of teachers had a higher percentage of task

statements overall. Therefore, the speaker attribute is

helpful for calculating probabilities of task

statements.

Table 1: Speaker attributes and percentages of task

statements.

As mentioned earlier, presenters use their

discussion commanders to mark statements that they

want to check later during the meetings. We

investigated the effect of marking for discrimination

of task statements by calculating the percentage of

marked task statements of all task statements. The

percentage of marked task statements was 73.4%,

which was higher than that of the task statements for

all candidates.

To examine whether there is a characteristic

tendency in the number of characters in task

statements, we obtained a distribution of the

respective characters of a presenter’s and

participants’ statements. We divided the number of

characters into five groups and calculated the

percentages of task statements in each group. In the

participants’ statements, the percentage of task

statements increased when the number of characters

increased. This is because when the participants were

giving concrete requests and advice, the number of

characters of their statements increased. On the other

hand, in the presenter’s statement, the number of

characters of a higher percentage of task statements

was 20 or less. The more characters there are the

smaller percentage of task statements. It is believed

that if the presenter accepts the requests or advice

participants presented, his or her response would tend

to be brief.

We also investigated certain types of sentences

included in the task statements. In the participants’

statements, the percentage of task statements was

higher when sentences were in the present tense and

in the declarative form (56.1%). This was due to the

fact that a large amount of advice or requests were in

the pattern of “should be ...” or “I want to …” In the

presenter’s statement, the percentage of task

statements in the past tense and in the declarative

form was low (29.2%). This is because when the

presenter talked about future tasks, he or she did not

tend to use sentences in the past tense. In addition, the

percentage of task statements of the presenter in the

past tense and in the interrogative form was 0%.

The details of the statistics of sentence types are

presented in Table 2.

Table 2: Sentence types and percentages of task statements.

The start time and duration of statements were

also considered as characteristics to discriminate task

statements. To determine the distribution of the start

time of the participants’ statements, we divided the

entire meeting time into five intervals, each

consisting of 20% of the meeting. We then

determined the percentage of the task statements in

each interval. At the 0-20% interval, the percentage

of task statements was smaller. We assumed that this

was because there were more questions than advice

and requests in the early stages of the meetings. At

Automatic Extraction of Task Statements from Structured Meeting Content

311

the 20-40% and 80-100% intervals, the percentage of

task statements was higher. That is, at the middle

interval of the meeting, suggestions and advice about

the purpose and approach were given, and at the final

interval, future issues were presented as a summary

of the entire meeting.

Morphemes and collocations of morphemes in

statements are also important features. We generated

a morpheme bigram of nouns, verbs, adjectives, and

auxiliary verbs in the survey data by calculating the

number of occurrences of the morphemes. We then

determined a feature of morphemes and their bigrams

of the statements if their occurrences exceeded certain

thresholds. Specifically, the selected nouns had an

occurrence percentage that was greater than or equal

to 0.5% of all nouns. The selected verbs also had a

percentage greater than or equal to 0.5% for all verbs.

Morpheme bigrams were selected if their percentages

were greater than 0.05% for the total morpheme

bigrams. These selected morphemes and bigrams

were used as features for discrimination of task

statements.

Based on the above survey results, the following

features were selected for creating a prediction

model:

1 Attribute of presenter

2 Feature of participant's statement

2.1 Start time and duration of statement

2.2 Speaker type (teacher or student)

2.3 Statement type (start-up or follow-up)

2.4 Marking (0 or 1)

2.5 Length (number of characters)

2.6 Sentence types

2.7 Morphemes and morpheme bigrams

2.8 Response by presenter (0 or 1)

3 Feature of presenter's response

3.1 Marking (0 or 1)

3.2 Length (number of characters)

3.3 Sentence types

3.4 Morphemes and morpheme bigrams

For values of sentence type features, we used

answers (0 or 1) to the following questions:

1. Does the statement include a sentence in the

past tense and in the declarative form?

2. Does the statement include a sentence in the

present tense and in the declarative form?

3. Does the statement include a sentence in the

past tense and in the interrogative form?

4. Does the statement include a sentence in the

present tense and in the interrogative form?

5. Does the statement include a sentence of the

other type?

As mentioned earlier, a probabilistic model for

extracting task statements is created using the

maximum entropy method based on the above

features. We used the Apache OpenNLP library

(https://opennlp.apache.org/) for implementing this

method.

Among the features used, morpheme, morpheme

bigram, and sentence type are dependent on the

language (in this case, Japanese). However, other

languages, such as English, seem to have almost the

same properties; therefore, it is necessary to analyze

in detail.

4.2 Results of Task Statement

Extraction

We give examples of task statements that were

correctly extracted with the proposed method.

Example 1

Participant (regarding listeners’ comments on a

presentation rehearsal): Is it possible for presenters to

ask their deep intensions of the comments?

Presenter: I think our system should deal with such

situations.

The presenter expressed the intention to handle

the requests from the participants. This task statement

was not marked, so it was very difficult to find in the

browsing of the meeting content.

Example 2

Participant (regarding self-driving cars): although a

goal is to make vehicles run precisely according to

their routes, I think it is difficult. So it is better to

decide the acceptable range of the target route.

Presenter: We calculated an acceptable margin for

each route, but there is a need to ascertain how far the

vehicles deviated from the route.

The proposed method correctly extracted the

description of the work the presenter should do and

also the advice from the participant.

Example 3

Participant (regarding gamification): Because it is

good that there is a sense of tension, I think it is better

to reduce the goals and to achieve them repeatedly.

This statement is a proposal by the participant, but

the presenter did not reply to it. Statements without a

response from the presenter can also be extracted with

the proposed method.

We also give examples of extraction failure.

Example 4

Participant (regarding document retrieval and

summarization): Do you have any idea of

summarization?

Presenter: No specific idea has been considered yet.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

312

This is an example of statements that were not

extracted despite a task statement. It can be

considered to have failed in extracting because the

concrete content of the task has not been stated.

Example 5

Participant (regarding user adaptation of an

authoring tool): If you wanted to take the data as part

of a rigorous evaluation, I think that you should have

to do it exactly from start to end.

Presenter: I think that it is useless to say this now, I

should do so.

This is an example of statements extracted as a

task statement by mistake. Because it contains the

auxiliary verb “should” indicating the meaning of

duty and suitable, it can be considered to have been

misclassified as a task statement.

By learning from these failures, we consider

additional features if phrases such as “not considered”

and “I think it should be …” are included in the

candidates of task statements.

5 VERIFICATION OF

EXTRACTION RESULTS

5.1 Experimental Results

To confirm the effectiveness of the proposed method,

10-fold cross-validation was applied to the extraction

results. The data used for verification included 42

types of meeting content and 1,637 groups of

statements (candidates). Each presenter created

correct data of task statements in each type of meeting

content as well as the survey data mentioned earlier.

The data used for verification were totally different

from the survey data.

We also compared the results of the proposed

method to the extraction results of alternative

methods that just select a set of statements that

included any of the following features: (1) statements

from teachers, (2) statements marked by presenters,

(3) statements that have features (1) and (2).

We confirmed the effectiveness of the proposed

method based on high precision (index for extraction

accuracy), recall (index for extraction leakage), and

F-measure (harmonic mean of precision and recall),

as shown in Table 3.

Table 3: Experimental results.

The extraction results of the task statements with

the proposed method are as follows: precision was

75.8%, recall was 64.2%, and F-measure was 69.5%.

On the other hand, the results of the three alternative

extraction methods were as follows: selecting the

statements that were marked by the presenter had the

highest precision (68.9%), selecting the statements

from teachers or statements that were marked by the

presenter had the highest recall (44.1%) and F-

measure (48.7%). The proposed method obtained the

highest values compared to these other extraction

methods.

Table 4 lists the results without certain features of

the probabilistic model. The F-measure significantly

decreased to 56.9% when the features of morphemes

and morpheme bigrams were not used. Since the F-

measures of the methods lacking any features were

reduced, the validity of the features used with the

proposed method was confirmed.

Table 4: Experimental results without features.

As mentioned earlier, the proposed method

calculates the probabilities of candidates of a task

statement using the generated probabilistic model. A

candidate whose probability value exceeds a certain

threshold is extracted as a task statement.

We first set the threshold to 0.5. It is not

guaranteed that this threshold value is optimal.

Therefore, we re-evaluated the outputs of the system

by lowering the threshold by 0.1 from 0.5. The results

are listed in Table 5.

Table 5: Experimental results with different thresholds.

Automatic Extraction of Task Statements from Structured Meeting Content

313

We found that the F-measure at a threshold of 0.4

was highest (71.4%). In the future, it should be

conducted to extract task statements by setting a

threshold to 0.4.

Since we used the maximum entropy method as a

classifier, we also confirmed that this method works

better than other classifiers. Table 6 shows the results

of the SVM and naive Bayes classifiers. We used the

“kernlab” package for the SVM and the “e1071”

package for the naive Bayes of R language.

Table 6: Experimental results with alternative classifiers.

The difference column in this table shows the

differences between the F-measures of the subject

classifier and the maximum entropy method in the

case of 0.4 (best performance) and 0.5 (initial setting)

thresholds. We found that our method is slightly

better that other traditional classifiers. While the

performance of the maximum entropy method did not

have a very significant advantage, the results obtained

as probability values can contribute to flexible control

of the presentation of results by using techniques such

as sorting and filtering.

5.2 Permutation Test

Since comparison with simple baselines (i.e., teachers’

statements and marked statements) is not sufficient

for proving the reliability of the proposed method, we

require another technique for this proof.

As well as evaluating the performance of the

proposed method, we also determined if the results

are statistically reliable. Therefore, we conducted a

statistical hypothesis test regarding the

misclassification rate calculated from cross-

validation. The statistical hypothesis test is a kind of

contradiction to prove the significance by rejecting a

hypothesis in which a complementary event of the

hypotheses is to be clarified. Since the correctness of

some results is generally difficult to prove directly,

the concept based on this contradiction is used in the

statistical hypothesis test.

In cross-validation, the misclassification rate is

calculated by

numberofmisclassifications

thecandidatesaremisclassifiedintotaskstatements

numberofdiscriminations

Even though this statistical hypothesis test is

based on the misclassification rate and unknown null

distribution, it is possible to estimate the null

distribution in a nonparametric manner by using a

permutation test (Good, 1994). In a permutation test,

a sample of a label is repeated many times to be sorted

randomly (here a label corresponds to whether it is a

task statement), and a null distribution is virtually

constituted. A ratio of statistics in this manner is

produced for each permutation that becomes equal to

or less than the value of the original test statistics. The

ratio is called a p-value, which is a measure of the

probability of events observed under the null

hypothesis. When the p-value is less than the

significance level that was set in advance, the

observed events under the null hypothesis do not

occur by chance, that is, the null hypothesis is

rejected. Then, we use an alternative hypothesis in

which the prediction model is statistically significant.

In this experiment, we set the significance level to

0.05 and conducted a permutation test from 1,000

iterations. The results are listed in Table 7. The

misclassification rate was 0.2260, and the p-value for

this was less than 0.001. It was confirmed that our

probabilistic model of task statements is statistically

significant below the level of significance. Since a p-

value is calculated from 1000 iterations, its accuracy

will rise in 0.001 increments. In other words, the

actual p-value is also considered much less likely than

0.001.

Table 7: Results of permutation test.

6 FUTURE WORK

Future work includes improvement in the accuracy of

our proposed method and in the usability of our

application system. We are considering the use of the

sentence end representation of statements and

planning to enhance the application system to

automatically generate a summary statement

indicating the content of the task from a set of

statements and to send feedback of users’ quotation

data of the task statements to the extraction module

for modification of the probability model.

Future work also includes creating a more

semantic structuring of discussions. In particular, we

aim to develop a system that can automatically

determine to what extent a discussion proceeds

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

314

depending on the topic. For example, if the topic in a

discussion chunk changes, the system should sub-

divide the chunk accordingly and determine whether

the previous topic is convergent.

Our previous study revealed that some follow-up

statements were about a topic different from that of

the start-up statement (Tsuchida, Ohira and Nagao,

2008). The discussion may thus become unsettled and

be abandoned because the participants do not know

whether the discussion on the previous topic reached

a conclusion. We may be able to develop a

mechanism that can automatically identify such

unsolved topics and suggest that participants discuss

them again.

7 CONCLUSIONS

We proposed an automatic extraction method of task

statements from meeting content. With 10-fold cross-

validation and permutation test, we evaluated the

effectiveness and reliability of the proposed method.

We also compared the results with those from

alternative methods without certain features and

confirmed the validity of the features used with the

proposed method.

Although our discussion mining system is able to

record face-to-face meetings in detail, analyze their

content, and conduct knowledge discovery, it is

unable to structure the discussions so that the topic of

each discussion is classified. To overcome this

problem, we aim to achieve more semantic

structuring of discussions by deeply analyzing

linguistic characteristics of statements and by

applying certain machine learning techniques.

REFERENCES

Sellen, A. J. and Whittaker, S., 2010, Beyond Total

Capture: A Constructive Critique of Lifelogging.

Commun. ACM, vol. 53, no. 5, pp. 70–77.

Mayer-Schönberger, V., Cukier, K., 2013. Big Data: A

Revolution that Will Transform How We Live, Work,

and Think, Houghton Mifflin Harcourt.

Nagao, K., Kaji, K., Yamamoto D. and Tomobe, H., 2005.

Discussion Mining: Annotation-Based Knowledge

Discovery from Real World Activities, Advances in

Multimedia Information Processing – PCM 2004,

LNCS, vol. 3331, pp. 522–531. Springer.

Wu, N., 1997. The Maximum Entropy Method, Springer

Series in Information Sciences, 32, Springer.

Kanayama, H. and Nasukawa, T., 2008. Textual Demand

Analysis: Detection of User’s Wants and Needs from

Opinions, In Proceedings of the 22nd International

Conference on Computational Linguistics (COLING-

2008), pp. 409–416.

Yang, W.-Y., Cao, Y. and Lin, C.-Y., 2009, A Structural

Support Vector Method for Extracting Contexts and

Answers of Questions from Online Forums, In

Proceedings of the Conference on Empirical Methods

in Natural Language Processing, pp. 514–523.

Sarencheh, S., Potdar, V., Yeganeh, E. A. and Firoozeh, N.,

2010, Semi-automatic Information Extraction from

Discussion Boards with Applications for Anti-Spam

Technology, In Proceedings of ICCSA 2010, LNCS,

vol. 6017, pp. 370–382. Springer.

Qu, Z. and Liu, Y., 2012, Sentence Dependency Tagging in

Online Question Answering Forums, In Proceedings of

the 50th Annual Meeting of the Association for

Computational Linguistics, pp. 554–562.

Wicaksono, A. F. and Myaeng, S.-H., 2013, Automatic

Extraction of Advice-revealing Sentences for Advice

Mining from Online Forums, In Proceedings of the 7th

International Conference on Knowledge Capture (K-

CAP 2013).

Conklin, J. and Begeman, M. L., 1988. gIBIS: A Hypertext

Tool for Exploratory Policy Discussion, ACM

Transactions on Information Systems (TOIS), vol. 6, no.

4, pp. 140–152.

Rienks, R., Heylen, D., and van der Weijden, 2005.

Argument Diagramming of Meeting Conversations, In

Proceedings of the Multimodal Multiparty Meeting

Processing Workshop at the 7th International

Conference on Multimodal Interfaces (ICMI 2005).

Lee, D., Erol, B., Graham, J., Hull, J. and Murata, N., 2002.

Portable Meeting Recorder, In Proceedings of ACM

Multimedia 2002, pp. 493–502.

Good, P., 1994. Permutation Tests: A Practical Guide to

Resampling Methods for Testing Hypothesis, Springer.

Tsuchida, T., Ohira, S. and Nagao, K., 2008. Knowledge

Activity Support System Based on Discussion Content,

In Proceedings of the Fourth International Conference

on Collaboration Technologies.

Automatic Extraction of Task Statements from Structured Meeting Content

315