| { |
| "paper_id": "N07-1004", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T14:48:01.924312Z" |
| }, |
| "title": "What Decisions Have You Made: Automatic Decision Detection in Conversational Speech", |
| "authors": [ |
| { |
| "first": "Pei-Yun", |
| "middle": [], |
| "last": "Hsueh", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Edinburgh", |
| "location": { |
| "postCode": "EH9 8WL", |
| "settlement": "Edinburgh", |
| "country": "UK" |
| } |
| }, |
| "email": "p.hsueh@ed.ac.uk" |
| }, |
| { |
| "first": "Johanna", |
| "middle": [], |
| "last": "Moore", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Edinburgh", |
| "location": { |
| "postCode": "EH9 8WL", |
| "settlement": "Edinburgh", |
| "country": "UK" |
| } |
| }, |
| "email": "j.moore@ed.ac.uk" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "This study addresses the problem of automatically detecting decisions in conversational speech. We formulate the problem as classifying decision-making units at two levels of granularity: dialogue acts and topic segments. We conduct an empirical analysis to determine the characteristic features of decision-making dialogue acts, and train MaxEnt models using these features for the classification tasks. We find that models that combine lexical, prosodic, contextual and topical features yield the best results on both tasks, achieving 72% and 86% precision, respectively. The study also provides a quantitative analysis of the relative importance of the feature types.", |
| "pdf_parse": { |
| "paper_id": "N07-1004", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "This study addresses the problem of automatically detecting decisions in conversational speech. We formulate the problem as classifying decision-making units at two levels of granularity: dialogue acts and topic segments. We conduct an empirical analysis to determine the characteristic features of decision-making dialogue acts, and train MaxEnt models using these features for the classification tasks. We find that models that combine lexical, prosodic, contextual and topical features yield the best results on both tasks, achieving 72% and 86% precision, respectively. The study also provides a quantitative analysis of the relative importance of the feature types.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "Making decisions is an important aspect of conversations in collaborative work. In the context of meetings, the proposed argumentative models, e.g., in Pallotta et al. (2005) and Rienks et al. (2005) , have specified decisions as an essential outcome of meetings. Whittaker et al. (2005) have also described how reviewing decisions is critical to the re-use of meeting recordings. For example, a new engineer who just get assigned to a project will need to know what major decisions have been made in previous meetings. Unless all decisions are recorded in meeting minutes or annotated in the speech recordings, it is difficult to locate the decision points by the browsing and playback utilities alone. Banerjee and Rudnicky (2005) have shown that it is easier for users to retrieve the information they seek if the meeting record includes information about topic segmentation, speaker role, and meeting state (e.g., discussion, presentation, briefing). To assist users in identifying or revisiting decisions in meeting archives, our goal is to automatically identify the dialogue acts and segments where decisions are made. Because reviewing decisions is indispensable in collaborative work, automatic decision detection is expected to lend support to computerassisted meeting tracking and understanding (e.g., assisting in the fulfilment of the decisions made in the meetings) and the development of group information management applications (e.g., constructing group memory).", |
| "cite_spans": [ |
| { |
| "start": 152, |
| "end": 174, |
| "text": "Pallotta et al. (2005)", |
| "ref_id": "BIBREF15" |
| }, |
| { |
| "start": 179, |
| "end": 199, |
| "text": "Rienks et al. (2005)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 264, |
| "end": 287, |
| "text": "Whittaker et al. (2005)", |
| "ref_id": "BIBREF20" |
| }, |
| { |
| "start": 704, |
| "end": 732, |
| "text": "Banerjee and Rudnicky (2005)", |
| "ref_id": "BIBREF0" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Spontaneous face-to-face dialogues in meetings violate many assumptions made by techniques previously developed for broadcast news (e.g., TDT and TRECVID), telephone conversations (e.g., Switchboard), and human-computer dialogues (e.g., DARPA Communicator).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "In order to develop techniques for understanding multiparty dialogues, smart meeting rooms have been built at several institutes to record large corpora of meetings in natural contexts, including CMU (Waibel et al., 2001 ), LDC (Cieri et al., 2002) , NIST (Garofolo et al., 2004) , ICSI (Janin et al., 2003) , and in the context of the IM2/M4 project (Marchand-Mailet, 2003) . More recently, scenario-based meetings, in which partic-ipants are assigned to different roles and given specific tasks, have been recorded in the context of the CALO project (the Y2 Scenario Data) (CALO, 2003) and the AMI project (Carletta et al., 2005) .", |
| "cite_spans": [ |
| { |
| "start": 200, |
| "end": 220, |
| "text": "(Waibel et al., 2001", |
| "ref_id": "BIBREF19" |
| }, |
| { |
| "start": 228, |
| "end": 248, |
| "text": "(Cieri et al., 2002)", |
| "ref_id": "BIBREF4" |
| }, |
| { |
| "start": 256, |
| "end": 279, |
| "text": "(Garofolo et al., 2004)", |
| "ref_id": "BIBREF6" |
| }, |
| { |
| "start": 287, |
| "end": 307, |
| "text": "(Janin et al., 2003)", |
| "ref_id": "BIBREF11" |
| }, |
| { |
| "start": 351, |
| "end": 374, |
| "text": "(Marchand-Mailet, 2003)", |
| "ref_id": "BIBREF13" |
| }, |
| { |
| "start": 575, |
| "end": 587, |
| "text": "(CALO, 2003)", |
| "ref_id": null |
| }, |
| { |
| "start": 608, |
| "end": 631, |
| "text": "(Carletta et al., 2005)", |
| "ref_id": "BIBREF3" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "The availability of meeting corpora has enabled researchers to begin to develop descriptive models of meeting discussions. Some researchers are modelling the dynamics of the meeting, exploiting dialogue models previously proposed for dialogue management. For example, use the Issue-Based Information System (IBIS) model (Kunz and Ritte, 1970) to incorporate the history of dialogue moves into the Multi-Modal Discourse (MMD) ontology. Other researchers are modelling the content of the meeting using the type of structures proposed in work on argumentation. For example, Rienks et al. (2005) have developed an argument diagramming scheme to visualize the relations (e.g., positive, negative, uncertain) between utterances (e.g., statement, open issue), and Marchand et al. (2003) propose a schema to model different argumentation acts (e.g., accept, request, reject) and their organization and synchronization. Decisions are often seen as a by-product of these models.", |
| "cite_spans": [ |
| { |
| "start": 320, |
| "end": 342, |
| "text": "(Kunz and Ritte, 1970)", |
| "ref_id": "BIBREF12" |
| }, |
| { |
| "start": 571, |
| "end": 591, |
| "text": "Rienks et al. (2005)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 757, |
| "end": 779, |
| "text": "Marchand et al. (2003)", |
| "ref_id": "BIBREF13" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "Automatically extracting these argument models is a challenging task. However, researchers have begun to make progress towards this goal. For example, Gatica et al. (2005) and Wrede and Shriberg (2003) automatically identify the level of emotion in meeting spurts (e.g., group level of interest, hot spots). Other researchers have developed models for detecting agreement and disagreement in meetings, using models that combine lexical features with prosodic features (e.g., pause, duration, F0, speech rate) (Hillard et al., 2003) and structural information (e.g., the previous and following speaker) (Galley et al., 2004) . More recently, Purver et al. (2006) have tackled the problem of detecting one type of decision, namely action items, which embody the transfer of group responsibility. However, no prior work has addressed the problem of automatically identifying decision-making units more generally in multiparty meetings. Moreover, no previous research has provided a quantitative account of the effects of different feature types on the task of automatic decision detection.", |
| "cite_spans": [ |
| { |
| "start": 151, |
| "end": 171, |
| "text": "Gatica et al. (2005)", |
| "ref_id": "BIBREF7" |
| }, |
| { |
| "start": 176, |
| "end": 201, |
| "text": "Wrede and Shriberg (2003)", |
| "ref_id": "BIBREF21" |
| }, |
| { |
| "start": 509, |
| "end": 531, |
| "text": "(Hillard et al., 2003)", |
| "ref_id": "BIBREF9" |
| }, |
| { |
| "start": 602, |
| "end": 623, |
| "text": "(Galley et al., 2004)", |
| "ref_id": "BIBREF5" |
| }, |
| { |
| "start": 641, |
| "end": 661, |
| "text": "Purver et al. (2006)", |
| "ref_id": "BIBREF16" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "Our aim is to develop models for automatically detecting segments of conversation that contain decisions directly from the audio recordings and transcripts of the meetings, and to identify the feature combinations that are most effective for this task.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Research Goal", |
| "sec_num": "3" |
| }, |
| { |
| "text": "Meetings can be viewed at different levels of granularity. In this study, we first consider how to detect the dialogue acts that contain decision-related information (DM DAs). Since it is often difficult to interpret a decision without knowing the current topic of discussion, we are also interested in detecting decision-making segments at a coarser level of granularity: topic segments. The task of automatic decision detection can therefore be divided into two subtasks: detecting DM DAs and detecting decisionmaking topic segments (DM Segments).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Research Goal", |
| "sec_num": "3" |
| }, |
| { |
| "text": "In this study we propose to first empirically identify the features that are most characteristic of decision-making dialogue acts and then computationally integrate the characteristic features to locate the DM DAs in meeting archives. For the latter task, previous research on automatic meeting understanding and tracking has commonly utilized a classification framework, in which variants of generative and conditional models are computed directly from data. In this study, we use a Maximum Entropy (MaxEnt) classifier to combine the decision characteristic features to predict DM DAs and DM Segments.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Research Goal", |
| "sec_num": "3" |
| }, |
| { |
| "text": "In this study, we use a set of 50 scenario-driven meetings (approximately 37,400 dialogue acts) that have been segmented into dialogue acts and annotated with decision information in the AMI meeting corpus. These meetings are driven by a scenario, wherein four participants play the role of Project Manager, Marketing Expert, Industrial Designer, and User Interface Designer in a design team in a series of four meetings. Each series of meeting recordings uses four distinctive speakers different from other series. The corpus includes manual transcripts for all meetings. It also comes with individual sound files recorded by close-talking headmounted microphones and cross-talking sound files recorded by desktop microphones.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Decision Annotation", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "In fact, it is difficult to determine whether a dialogue act contains information relevant to any decision point without knowing what decisions have been made in the meeting. Therefore, in this study DM DAs are annotated in a two-phase process: First, annotators are asked to browse through the meeting record and write an abstractive summary directed to the project manager about the decisions that have been made in the meeting. Next, another group of three annotators are asked to produce extractive summaries by selecting a subset (around 10%) of dialogue acts which form a summary of this meeting for the absent manager to understand what has transpired in the meeting.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Decision-Making Dialogue Acts", |
| "sec_num": "4.1.1" |
| }, |
| { |
| "text": "Finally, this group of annotators are asked to go through the extractive dialogue acts one by one and judge whether they support any of the sentences in the decision section of the abstractive summary; if a dialogue act is related to any sentence in the decision section, a \"decision link\" from the dialogue act to the decision sentence is added. For those extracted dialogue acts that do not have any closely related sentence, the annotators are not obligated to specify a link. We then label the dialogue acts that have one or more decision links as DM DAs.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Decision-Making Dialogue Acts", |
| "sec_num": "4.1.1" |
| }, |
| { |
| "text": "In the 50 meetings we used for the experiments, the annotators have on average found four decisions per meeting and specified around two decision links to each sentence in the decision summary section. Overall, 554 out of 37,400 dialogue acts have been annotated as DM DAs, accounting for 1.4% of all dialogue acts in the data set and 12.7% of the orginal extractive summary (which is consisted of the extracted dialogue acts). An earlier analysis has established the intercoder reliability of the two-phase process at the level of kappa ranging from 0.5 to 0.8. In this round of experiment, for each meeting in the 50-meeting dataset we randomly choose the DM DA annotation of one annotator as the sourec of its ground truth data.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Decision-Making Dialogue Acts", |
| "sec_num": "4.1.1" |
| }, |
| { |
| "text": "Topic segmentation has also been annotated for the AMI meeting corpus. Annotators had the freedom to mark a topic as subordinated (down to two levels) wherever appropriate. As the AMI meetings are scenario-driven, annotators are expected to find that most topics recur. Therefore, they are given a standard set of topic descriptions that can be used as labels for each identified topic segment. Annotators will only add a new label if they cannot find a match in the standard set. The AMI scenario meetings contain around 14 topic segments per meeting. Each segment lasts on average 44 dialogue acts long and contains two DM DAs.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Decision-Making Topic Segments", |
| "sec_num": "4.1.2" |
| }, |
| { |
| "text": "DM Segments are operationalized as topic segments that contain one or more DM DAs. Overall, 198 out of 623 (31.78%) topic segments in the 50-meeting dataset are DM Segments. As the meetings we use are driven by a predetermined agenda, we expect to find that interlocutors are more likely to reach decisions when certain topics are brought up. Analysis shows that some topics are indeed more likely to contain decisions than others. For example, 80% of the segments labelled as Costing and 58% of those labelled Budget are DM Segments, whereas only 7% of the Existing Product segments and none of the Trend-Watching segments are DM Segments. Functional segments, such as Chitchat, Opening and Closing, almost never include decisions.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Decision-Making Topic Segments", |
| "sec_num": "4.1.2" |
| }, |
| { |
| "text": "To provide a qualitative account of the effect of different feature types on the task of automatic decision detection, we have conducted empirical analysis on four major types of features: lexical, prosodic, contextual and topical features.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Features Used", |
| "sec_num": "4.2" |
| }, |
| { |
| "text": "Previous research has studied lexical differences (i.e., occurrence counts of N-grams) between various aspects of speech, such as topics (Hsueh and Moore, 2006) , speaker gender (Boulis and Ostendorf, 2005) , and story-telling conversation (Gordon and Ganesan, 2005 ). As we expect that lexical differences also exist in DM conversations, we generated language models from the DM Dialogue Acts in the corpus. The comparison of the language models generated from the DM dialogue Acts and the rest of the conversations shows that some differences exist between the two models: (1) decision making conversations are more likely to contain we than I and You;", |
| "cite_spans": [ |
| { |
| "start": 137, |
| "end": 160, |
| "text": "(Hsueh and Moore, 2006)", |
| "ref_id": "BIBREF10" |
| }, |
| { |
| "start": 178, |
| "end": 206, |
| "text": "(Boulis and Ostendorf, 2005)", |
| "ref_id": "BIBREF1" |
| }, |
| { |
| "start": 240, |
| "end": 265, |
| "text": "(Gordon and Ganesan, 2005", |
| "ref_id": "BIBREF8" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Lexical Features", |
| "sec_num": "4.2.1" |
| }, |
| { |
| "text": "(2) in decision-making conversations there are more explicit mentions of topical words, such as advanced chips and functional design; (3) in decision- In an exploratory study using unigrams, as well as bigrams and trigrams, we found that using bigrams and trigrams does not improve the accuracy of classifying DM DAs, and therefore we include only unigrams in the set of lexical features in the experiments reported in Section 6.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Lexical Features", |
| "sec_num": "4.2.1" |
| }, |
| { |
| "text": "Functionally, prosodic features, i.e., energy, and fundamental frequency (F0), are indicative of segmentation and saliency. In this study, we follow Shriberg and Stolcke's (2001) direct modelling approach to manifest prosodic features as duration, pause, speech rate, pitch contour, and energy level. We utilize the individual sound files provided in the AMI corpus. To extract prosodic features from the sound files, we use the Snack Sound Toolkit to compute a list of pitch and energy values delimited by frames of 10 ms, using the normalized cross correlation function. Then we apply a piecewise linearisation procedure to remove the outliers and average the linearised values of the units within the time frame of a word. Pitch contour of a dialogue act is approximated by measuring the pitch slope at multiple points within the dialogue act, e.g., the first and last 100 and 200 ms. The rate of speech is calculated as both the number of words spoken per second and the number of syllables per second. We use Festival's speech synthesis front-end to return phonemes and syllabification information. An exploratory study has shown the benefits of including immediate prosodic contexts, and thus we also include prosodic features of the immediately preceding and following dialogue acts. Table 1 contains a list of automatically generated prosodic features used in this study.", |
| "cite_spans": [ |
| { |
| "start": 149, |
| "end": 178, |
| "text": "Shriberg and Stolcke's (2001)", |
| "ref_id": "BIBREF18" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 1291, |
| "end": 1298, |
| "text": "Table 1", |
| "ref_id": "TABREF0" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Prosodic Features", |
| "sec_num": "4.2.2" |
| }, |
| { |
| "text": "From our qualitative analysis, we expect that contextual features specific to the AMI corpus, such as the speaker role (i.e., PM, ME, ID, UID) and meeting type (i.e., kick-off, conceptual design, functional design, detailed design) to be characteristic of the DM DAs. Analysis shows that (1) participants assigned to the role of PM produce 42.5% of the DM DAs, and (2) participants make relatively fewer decisions in the kick-off meetings. Analysis has also demonstrated a difference in the type, the reflexivity 1 and the number of addressees, between the DM DAs and the non-DM DAs. For example, dialogue acts of type inform, suggest, elicit assessment and elicit inform are more likely to be DM DAs.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Contextual Features", |
| "sec_num": "4.2.3" |
| }, |
| { |
| "text": "We have also found that immediately preceding and following dialogue acts are important for identifying DM DAs. For example, stalls and fragments preceding and fragments following a DM DA are more likely than for non-DM DAs. 2 In 1 According to the annotation guideline, the reflexivity reflects on how the group is carrying on the task. In this case, the interlocutors pause to evaluate the group performance less often when it comes to decision making.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Contextual Features", |
| "sec_num": "4.2.3" |
| }, |
| { |
| "text": "2 STALL is where people start talking before they are ready, or keep speaking when they haven't figured out what to say; FRAGMENT is the segment which is not really speech or is unclear enough to be transcribed, or where the speaker did not contrast, there is a lower chance of seeing suggest and elicit-type DAs (i.e., elicit-inform, elicitsuggestion, elicit-assessment) in the preceding and following DM DAs.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Contextual Features", |
| "sec_num": "4.2.3" |
| }, |
| { |
| "text": "As reported in Section 4.1.2, we find that interlocutors are more likely to reach decisions when certain topics are brought up. Also, we expect decisionmaking conversations to take place towards the end of a topic segment. Therefore, in this study we include the following features: the label of the current topic segment, the position of the DA in a topic segment (measured in words, in seconds, and in %), the distance to the previous topic shift (both at the toplevel and sub-topic level)(measured in seconds), the duration of the current topic segment (both at the top-level and sub-topic level)(measured in seconds).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Topical Features", |
| "sec_num": "4.2.4" |
| }, |
| { |
| "text": "Detecting DM DAs is the first step of automatic decision detection. For this purpose, we trained Max-Ent models to classify each unseen sample as either DM DA (POS) or non-DM DA (NEG). We performed a 5-fold cross validation on the set of 50 meetings. In each fold, we trained MaxEnt models from the feature combinations in the training set, wherein each of the extracted dialogue acts has been labelled as either POS or NEG. Then, the models were used to classify unseen instances in the test set as either POS or NEG. In Section 4.2, we described the four major types of features used in this study: unigrams (LX1), prosodic (PROS), contextual (CONT), and topical (TOPIC) features. For comparison, we report the naive baseline obtained by training the models on the prosodic features alone, since the prosodic features can be generated fully automatically. The different combinations of features we used for training models can be divided into the following four groups: (A) using prosodic features alone (BASELINE), (B) using lexical, contextual and topical features alone (LX1, CONT, TOPIC); (C) using all available features except one of the four types of features (ALL-LX1, ALL-PROS, ALL-CONT, ALL-TOPIC); and get far enough to express the intention.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Classifying DM DAs", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "(D) using all available features (ALL).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Classifying DM DAs", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "Detecting DM segments is necessary for interpreting decisions, as it provides information about the current topic of discussion. Here we combine the predictions of the DM DAs to classify each unseen topic segment in the test set as either DM Segment (POS) or non-DM Segment (NEG). Recall that we defined a DM Segment as a segment that contains one or more hypothesized DM DAs. The task of detecting DM Segments can thus be viewed as that of detecting DM Dialogue Acts in a wider window. Table 2 reports the performance on the test set. The results show that models trained with all features (ALL), including lexical, prosodic, contextual and topical features, yield substantially better performance than the baseline on the task of detecting DM DAs. We carried out a one-way ANOVA to examine the effect of different feature combinations on overall accuracy (F1). The ANOVA suggests a reliable effect of feature type (F (9, 286) = 3.44; p < 0.001). Rows 2-4 in Table 2 report the performance of models in Group B that are trained with a single type of feature. Lexical features are the most predictive features when used alone. We performed sign tests to determine whether there are statistical differences among these models and the baseline. We find that when used alone, only lexical features (LX1) can train a better model than the baseline (p < 0.001). However, none of these models yields a comparable performance to the ALL model.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 487, |
| "end": 494, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| }, |
| { |
| "start": 960, |
| "end": 967, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Classifying DM Segments", |
| "sec_num": "6.1" |
| }, |
| { |
| "text": "To study the relative effect of the different feature types, Rows 5-8 in the table report the performance of models in Group C, which are trained with all available features except LX1, PROS, CONT and TOPIC features respectively. The amount of degradation in the overall accuracy (F1) of each of the models in relation to that of the ALL model indicates the contribution of the feature type that has been left out of the model. We performed sign tests to examine the differences among these models and the ALL model. We find that the ALL model outperforms all of these models (p < 0.001) except the model trained by leaving out contextual features (ALL-CONT). A closer investigation of the precision and recall of the ALL-CONT model shows that the contextual features are detrimental to recall but beneficial for precision. The mixed result is due to the fact that models trained with contextual features are tailored to recognize particular types of DM dialogue acts. Therefore, using these contextual features improves the precision for these types of DM DAs but reduces the overall recognition accuracy. The last three columns of Table 2 are the results obtained using a lenient match measure, allowing a window of 10 seconds preceding and following a hypothesized DM DA for recognition. The better results show that there is room for ambiguity in the assessment of the exact timing of DM DAs.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 1133, |
| "end": 1140, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "EXP1: Classifying DM DAs", |
| "sec_num": "6.2" |
| }, |
| { |
| "text": "As expected, the results in Table 3 are better than those reported in Table 2 , achieving at best 83% overall accuracy.The model that combines all features (ALL) yields significantly better results than the baseline. The ANOVA shows a reliable effect of different feature types on the task of detecting DM Segments (F (11, 284) = 2.33; p <= 0.01). Rows 2-4 suggest that lexical features are the most predictive in terms of overall accuracy. Sign tests confirm the advantage of using lexical features (LX1) over the baseline (PROS) (p < 0.05). Interestingly, the model that is trained with topical features alone (TOPIC) yields substantially better precision (p < 0.001). The increase from 49% precision for the task of detecting DM DAs (in Table 2 ) to 91% for that of detecting DM Segments stems from the fact that decisions are more likely to occur in certain types of topic segments. In turn, training models with topical features helps eliminate incorrect predictions of DM DAs in these types of topic segments. However, the accuracy gain of the TOPIC model on detecting certain types of DM Segments does not extend to all types of DM Segments. This is shown by the significantly lower recall of the TOPIC model over the baseline (p < 0.001).", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 28, |
| "end": 35, |
| "text": "Table 3", |
| "ref_id": "TABREF4" |
| }, |
| { |
| "start": 70, |
| "end": 77, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| }, |
| { |
| "start": 740, |
| "end": 747, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "EXP2: Classifying DM Segments", |
| "sec_num": "6.3" |
| }, |
| { |
| "text": "Finally, Rows 5-8 report the performance of the models in Group (C) on the task of detecting DM Segments. Sign tests again show that the model that is trained with all available features (ALL) outperforms the models that leave out lexical, prosodic, or topical features (p < 0.05). However, the ALL model does not outperform the model that leaves out contextual features. In addition, the contextual features degrade the recall but improve the precision on the task of detecting DM Segments. Calculating how much the overall accuracy of the models in Group C degrades from the ALL model shows that the most predictive features are the lexical features, followed by the topical and prosodic features.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "EXP2: Classifying DM Segments", |
| "sec_num": "6.3" |
| }, |
| { |
| "text": "As suggested by the mixed results obtained by the model that is trained without the contextual features, the two-phase decision annotation procedure (as described in Section 4.1) may have caused annotators to select dialogue acts that serve different functional roles in a decision-making process in the set of DM DAs. For example, in the dialogue shown in Figure 1 , the annotators have marked dialogue act (1), (5), (8), and (11) as the DM DAs related to this decision: \"There will be no feature to help find the remote when it is misplaced\". Among the four DM DAs, (1) describes the topic of what this decision is about; (5) and (8) describe the arguments that support the decision-making process; (11) indicates the level of agreement or disagreement for this decision. Yet these DM DAs which play different functional roles in the DM process may each have their own characteristic features. Training one model to recognize DM DAs of all functional roles may have degraded the performance on the classification tasks. Developing models for detecting DM DAs that play different functional roles requires a larger scale study to discover the anatomy of general decision-making discussions.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 357, |
| "end": 365, |
| "text": "Figure 1", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Discussion", |
| "sec_num": "7" |
| }, |
| { |
| "text": "This is the first study that aimed to detect segments of the conversation that contain decisions. We have (1) empirically analyzed the characteristic features of DM dialogue acts, and (2) computational developed models to detect DM dialogue acts and DM topic segments, given the set of characteristic features. Empirical analysis has provided a qualitative account of the DM-characteristic features, whereas training the computational models on different feature combinations has provided a quantitative account of the effect of different feature types on the task of automatic decision detection. Empirical analysis has exhibited demonstrable differences", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions and Future Work", |
| "sec_num": "8" |
| }, |
| { |
| "text": "(1) A: but um the feature that we considered for it not getting lost.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions and Future Work", |
| "sec_num": "8" |
| }, |
| { |
| "text": "(2) B: Right. Well (3) B: were talking about that a little bit (4) B: when we got that email (5) B: and we think that each of these are so distinctive, that it it's not just like another piece of technology around your house. (6) B: It's gonna be somewhere that it can be seen. (7) A: Mm-hmm. (8) B: So we're we're not thinking that it's gonna be as critical to have the loss (9) D: But if it's like under covers or like in a couch you still can't see it.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions and Future Work", |
| "sec_num": "8" |
| }, |
| { |
| "text": ". . . (10) A: Okay , that's a fair evaluation. (11) A: Um we so we do we've decided not to worry about that for now. Figure 1 : Example decision-making discussion in the words (e.g., we), the contextual features (e.g., meeting type, speaker role, dialogue act type), and the topical features. The experimental results have suggested that (1) the model combining all the available features performs substantially better, achieving 62% and 82% overall accuracy on the task of detecting DM DAs and that of detecting DM Segments, respectively, (2) lexical features are the best indicators for both the task of detecting DM DAs and that of detecting DM Segments, and (3) combining topical features is important for improving the precision for the task of detecting DM Segments. Many of the features used in this study require human intervention, such as manual transcriptions, annotated dialogue act segmentations and labels, annotated topic segmentations and labels, and other types of meeting-specific features. Our ultimate goal is to identify decisions using automatically induced features. Therefore, studying the performance degradation when using the automatically generated versions of these features (e.g., ASR words) is essential for developing a fully automated component on detecting decisions immediately after a meeting or even for when a meeting is still in progress. Another problem that has been pointed out in Section 6 and in Section 7 is the different functional roles of DM dialogue acts in current annotations. Purver et al. (2006) have suggested a hierarchical annotation scheme to accommodate the different aspects of action items. The same technique may be applicable", |
| "cite_spans": [ |
| { |
| "start": 1528, |
| "end": 1548, |
| "text": "Purver et al. (2006)", |
| "ref_id": "BIBREF16" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 117, |
| "end": 125, |
| "text": "Figure 1", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Conclusions and Future Work", |
| "sec_num": "8" |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "in a more general decision detection task.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "annex", |
| "sec_num": null |
| }, |
| { |
| "text": "This work was supported by the European Union Integrated Project AMI (Augmented Multi-party Interaction, FP6-506811, publication AMI-204).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Acknowledgement", |
| "sec_num": "9" |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "The necessity of a meeting recording and playback system, and the benefit of topic-level annotations to meeting browsing", |
| "authors": [ |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Banerjee", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [ |
| "I" |
| ], |
| "last": "Rose", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Rudnicky", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of the Tenth International Conference on Human-Computer Interaction", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Banerjee, C. Rose, and A. I. Rudnicky. 2005. The necessity of a meeting recording and playback system, and the benefit of topic-level annotations to meeting browsing. In Proceedings of the Tenth International Conference on Human-Computer Interaction.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "A quantitative analysis of lexical differences between genders in telephone conversation", |
| "authors": [ |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Boulis", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Ostendorf", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "C. Boulis and M. Ostendorf. 2005. A quantitative anal- ysis of lexical differences between genders in tele- phone conversation. In Proceedings of the 42nd An- nual Meeting of the Association for Computational Linguistics. ACM Press.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "The ami meeting corpus: A pre-announcement", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Carletta", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Ashby", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Bourban", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Flynn", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Guillemot", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Hain", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Kadlec", |
| "suffix": "" |
| }, |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Karaiskos", |
| "suffix": "" |
| }, |
| { |
| "first": "W", |
| "middle": [], |
| "last": "Kraaij", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Kronenthal", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Lathoud", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Lincoln", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Lisowska", |
| "suffix": "" |
| }, |
| { |
| "first": "I", |
| "middle": [], |
| "last": "Mccowan", |
| "suffix": "" |
| }, |
| { |
| "first": "W", |
| "middle": [], |
| "last": "Post", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Reidsma", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Wellner", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of 2nd Joint Workshop on Multimodal Interaction and Related Machine Learning Algorithms", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. Carletta, S. Ashby, S. Bourban, M. Flynn, M. Guille- mot, T. Hain, J. Kadlec, V. Karaiskos, W. Kraaij, M. Kronenthal, G. Lathoud, M. Lincoln, A. Lisowska, I. McCowan, W. Post, D. Reidsma, and P. Wellner. 2005. The ami meeting corpus: A pre-announcement. In Proceedings of 2nd Joint Workshop on Multi- modal Interaction and Related Machine Learning Al- gorithms.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Research methodologies, observations and outcomes in conversational speech data collection", |
| "authors": [ |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Cieri", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Miller", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Walker", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Proceedings of the Human Language Technologies Conference (HLT)", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "C. Cieri, D. Miller, and K. Walker. 2002. Research methodologies, observations and outcomes in conver- sational speech data collection. In Proceedings of the Human Language Technologies Conference (HLT).", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Identifying agreement and disagreement in conversational speech: Use of bayesian networks to model pragmatic dependencies", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Galley", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Mckeown", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Hirschberg", |
| "suffix": "" |
| }, |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Shriberg", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proceedings of the 42nd Annual Meeting of the ACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "M. Galley, J. McKeown, J. Hirschberg, and E. Shriberg. 2004. Identifying agreement and disagreement in con- versational speech: Use of bayesian networks to model pragmatic dependencies. In Proceedings of the 42nd Annual Meeting of the ACL.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "The nist meeting room pilot corpus", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [ |
| "S" |
| ], |
| "last": "Garofolo", |
| "suffix": "" |
| }, |
| { |
| "first": "C", |
| "middle": [ |
| "D" |
| ], |
| "last": "Laprun", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Michel", |
| "suffix": "" |
| }, |
| { |
| "first": "V", |
| "middle": [ |
| "M" |
| ], |
| "last": "Stanford", |
| "suffix": "" |
| }, |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Tabassi", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proceedings of LREC04", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. S. Garofolo, C. D. Laprun, M. Michel, V.M. Stanford, and E. Tabassi. 2004. The nist meeting room pilot corpus. In Proceedings of LREC04.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Detecting group interest level in meetings", |
| "authors": [ |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Gatica-Perez", |
| "suffix": "" |
| }, |
| { |
| "first": "I", |
| "middle": [], |
| "last": "Mccowan", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Zhang", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Bengio", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "IEEE Int. Conf. on Acoustics, Speech, and Signal Processing (ICASSP)", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "D. Gatica-Perez, I. McCowan, D. Zhang, and S. Bengio. 2005. Detecting group interest level in meetings. In IEEE Int. Conf. on Acoustics, Speech, and Signal Pro- cessing (ICASSP).", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Automated story extraction from conversational speech", |
| "authors": [ |
| { |
| "first": "Andrew", |
| "middle": [ |
| "S" |
| ], |
| "last": "Gordon", |
| "suffix": "" |
| }, |
| { |
| "first": "Kavita", |
| "middle": [], |
| "last": "Ganesan", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of the Third International Conference on Knowledge Capture (K-CAP 05", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Andrew S. Gordon and Kavita Ganesan. 2005. Auto- mated story extraction from conversational speech. In Proceedings of the Third International Conference on Knowledge Capture (K-CAP 05).", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Detection of agreement vs. disagreement in meetings: Training with unlabeled data", |
| "authors": [ |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Hillard", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Ostendorf", |
| "suffix": "" |
| }, |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Shriberg", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proc. HLT-NAACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "D. Hillard, M. Ostendorf, and E. Shriberg. 2003. Detec- tion of agreement vs. disagreement in meetings: Train- ing with unlabeled data. In Proc. HLT-NAACL.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Automatic topic segmentation and lablelling in multiparty dialogue", |
| "authors": [ |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Hsueh", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Moore", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "IEEE/ACM workshop on Spoken Language Technology", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "P. Hsueh and J. Moore. 2006. Automatic topic segmen- tation and lablelling in multiparty dialogue. In the first IEEE/ACM workshop on Spoken Language Technol- ogy (SLT). IEEE/ACM.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "The icsi meeting corpus", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Janin", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Baron", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Edwards", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Ellis", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Gelbart", |
| "suffix": "" |
| }, |
| { |
| "first": "N", |
| "middle": [], |
| "last": "Morgan", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Peskin", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Pfau", |
| "suffix": "" |
| }, |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Shriberg", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Stolcke", |
| "suffix": "" |
| }, |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Wooters", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proceedings of ICASSP-2003", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "A. Janin, D. Baron, J. Edwards, D. Ellis, D. Gelbart, N. Morgan, B. Peskin, T. Pfau, E. Shriberg, A. Stol- cke, and C. Wooters. 2003. The icsi meeting corpus. In Proceedings of ICASSP-2003, Hong Kong.", |
| "links": null |
| }, |
| "BIBREF12": { |
| "ref_id": "b12", |
| "title": "Issue as elements of information system", |
| "authors": [ |
| { |
| "first": "W", |
| "middle": [], |
| "last": "Kunz", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [ |
| "W J" |
| ], |
| "last": "Ritte", |
| "suffix": "" |
| } |
| ], |
| "year": 1970, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "W. Kunz and H. W. J. Ritte. 1970. Issue as elements of information system. Technical Report Working Pa- per 131, Institute of Urban and Regional Development Research, University of California, Berkeley.", |
| "links": null |
| }, |
| "BIBREF13": { |
| "ref_id": "b13", |
| "title": "Meeting record modeling for enhanced browsing", |
| "authors": [ |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Marchand-Mailet", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "S. Marchand-Mailet. 2003. Meeting record modeling for enhanced browsing. Technical report, Computer Vi- sion and Multimedia Lab, Computer Centre, Univer- sity of Geneva, Switzerland.", |
| "links": null |
| }, |
| "BIBREF14": { |
| "ref_id": "b14", |
| "title": "Ontology-based discourse understanding for a persistent meeting assistant", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Niekrasz", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Purver", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Dowding", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Peters", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proc. of the AAAI Spring Symposium", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. Niekrasz, M. Purver, J. Dowding, and S. Peters. 2005. Ontology-based discourse understanding for a persis- tent meeting assistant. In Proc. of the AAAI Spring Symposium.", |
| "links": null |
| }, |
| "BIBREF15": { |
| "ref_id": "b15", |
| "title": "Collaborative and argumentative models of meeting discussions", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Pallotta", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Niekrasz", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Purver", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceeding of CMNA-05 workshop on Computational Models of Natural Arguments in IJCAI 05", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "V. Pallotta, J. Niekrasz, and M. Purver. 2005. Collab- orative and argumentative models of meeting discus- sions. In Proceeding of CMNA-05 workshop on Com- putational Models of Natural Arguments in IJCAI 05.", |
| "links": null |
| }, |
| "BIBREF16": { |
| "ref_id": "b16", |
| "title": "Shallow discourse structure for action item detection", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Purver", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Ehlen", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Niekrasz", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "the Workshop of HLT-NAACL: Analyzing Conversations in Text and Speech", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "M. Purver, P. Ehlen, and J. Niekrasz. 2006. Shallow discourse structure for action item detection. In the Workshop of HLT-NAACL: Analyzing Conversations in Text and Speech. ACM Press.", |
| "links": null |
| }, |
| "BIBREF17": { |
| "ref_id": "b17", |
| "title": "Argument diagramming of meeting conversations", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [ |
| "J" |
| ], |
| "last": "Rienks", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Heylen", |
| "suffix": "" |
| }, |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Van Der Weijden", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Multimodal Multiparty Meeting Processing Workshop at the ICMI", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "R. J. Rienks, D. Heylen, and E. van der Weijden. 2005. Argument diagramming of meeting conversations. In Multimodal Multiparty Meeting Processing Workshop at the ICMI.", |
| "links": null |
| }, |
| "BIBREF18": { |
| "ref_id": "b18", |
| "title": "Direct modeling of prosody: An overview of applications in automatic speech processing", |
| "authors": [ |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Shriberg", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Stolcke", |
| "suffix": "" |
| } |
| ], |
| "year": 2001, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "E. Shriberg and A. Stolcke. 2001. Direct modeling of prosody: An overview of applications in automatic speech processing.", |
| "links": null |
| }, |
| "BIBREF19": { |
| "ref_id": "b19", |
| "title": "Advances in automatic meeting record creation and access", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Waibel", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Bett", |
| "suffix": "" |
| }, |
| { |
| "first": "F", |
| "middle": [], |
| "last": "Metze", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Ries", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Schaaf Amd", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Schultz", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Soltau", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Zechner", |
| "suffix": "" |
| } |
| ], |
| "year": 2001, |
| "venue": "Proceedings of ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "A. Waibel, M. Bett, F. Metze, K. Ries, T. Schaaf amd T. Schultz, H. Soltau, H. Yu, and K. Zechner. 2001. Advances in automatic meeting record creation and ac- cess. In Proceedings of ICASSP.", |
| "links": null |
| }, |
| "BIBREF20": { |
| "ref_id": "b20", |
| "title": "Analysing meeting records: An ethnographic study and technological implications", |
| "authors": [ |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Whittaker", |
| "suffix": "" |
| }, |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Laban", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Tucker", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of MLMI 2005", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "S. Whittaker, R. Laban, and S. Tucker. 2005. Analysing meeting records: An ethnographic study and techno- logical implications. In Proceedings of MLMI 2005.", |
| "links": null |
| }, |
| "BIBREF21": { |
| "ref_id": "b21", |
| "title": "Spotting hot spots in meetings: Human judgements and prosodic cues", |
| "authors": [ |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Wrede", |
| "suffix": "" |
| }, |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Shriberg", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proceedings of EUROSPEECH", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "B. Wrede and E. Shriberg. 2003. Spotting hot spots in meetings: Human judgements and prosodic cues. In Proceedings of EUROSPEECH 2003.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "text": "spoken in current, previous and next subdialogue Duration (in seconds) of current, previous and next subdialogue Pause Amount of silence (in seconds) preceding a subdialogue Amount of silence (in seconds) following a subdialogue Speech rate Number of words spoken per second in current, previous and next subdialogue Number of syllables per second in current, previous and next subdialogue Energy Overall energy level Average energy level in the first, second, third, and fourth quarter of a subdialogue Pitch Maximum and minimum F0, overall slope and variance Slope and variance at the first 100 and 200 ms and last 100 and 200 ms, at the first and second half, and at each quarter of the subdialogue", |
| "type_str": "figure", |
| "num": null, |
| "uris": null |
| }, |
| "TABREF0": { |
| "type_str": "table", |
| "content": "<table/>", |
| "html": null, |
| "text": "Prosodic features used in this study.", |
| "num": null |
| }, |
| "TABREF2": { |
| "type_str": "table", |
| "content": "<table/>", |
| "html": null, |
| "text": "Effects of different combinations of features on detecting DM DAs.", |
| "num": null |
| }, |
| "TABREF4": { |
| "type_str": "table", |
| "content": "<table/>", |
| "html": null, |
| "text": "Effects of different combinations of features on detecting DM Segments.", |
| "num": null |
| } |
| } |
| } |
| } |
