acl acl2013 acl2013-386 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Claudiu Mihaila ; Sophia Ananiadou
Abstract: Current domain-specific information extraction systems represent an important resource for biomedical researchers, who need to process vaster amounts of knowledge in short times. Automatic discourse causality recognition can further improve their workload by suggesting possible causal connections and aiding in the curation of pathway models. We here describe an approach to the automatic identification of discourse causality triggers in the biomedical domain using machine learning. We create several baselines and experiment with various parameter settings for three algorithms, i.e., Conditional Random Fields (CRF), Support Vector Machines (SVM) and Random Forests (RF). Also, we evaluate the impact of lexical, syntactic and semantic features on each of the algorithms and look at er- rors. The best performance of 79.35% F-score is achieved by CRFs when using all three feature types.
Reference: text
sentIndex sentText sentNum sentScore
1 uk Abstract Current domain-specific information extraction systems represent an important resource for biomedical researchers, who need to process vaster amounts of knowledge in short times. [sent-8, score-0.309]
2 Automatic discourse causality recognition can further improve their workload by suggesting possible causal connections and aiding in the curation of pathway models. [sent-9, score-0.924]
3 We here describe an approach to the automatic identification of discourse causality triggers in the biomedical domain using machine learning. [sent-10, score-0.986]
4 Using biomedical text mining technology, text can now be enriched via the addition of semantic metadata and thus can support tasks such as analysing molecular pathways (Rzhetsky et al. [sent-26, score-0.401]
5 The notion of discourse can be defined as a coherent sequence of clauses and sentences. [sent-30, score-0.176]
6 These are connected in a logical manner by discourse relations, such as causal, temporal and conditional, which characterise how facts in text are related. [sent-31, score-0.176]
7 These relations can be either explicit or implicit, depending whether or not they are expressed in text using overt discourse connectives (also known as triggers). [sent-33, score-0.295]
8 (1) In the case of PmrB, a normal response to mild acid pH requires not only a periplasmic histidine 38 Sofia, BulPgraoricea,ed Ainugguss otf 4 t-h9e 2 A0C13L. [sent-35, score-0.185]
9 tc ud2e0n1t3 R Aess eoacricahti Wonor foksrh Coopm, ppaugteasti 3o8n–a4l5 L,inguistics but also several glutamic acid residues. [sent-37, score-0.111]
10 Therefore, regulation of PmrB activity may involve protonation of one or more of these amino acids. [sent-38, score-0.124]
11 Thus, by identifying this causal relation, search engines become able to discover relations between biomedical entities and events or between experimental evidence and associated conclusions. [sent-39, score-0.919]
12 However, phrases acting as causal triggers in certain contexts may not denote causality in all cases. [sent-40, score-1.03]
13 In this paper, we explore several supervised machinelearning approaches to the automatic identification of triggers that actually denote causality. [sent-42, score-0.349]
14 2 Related Work A large amount of work related to discourse pars- ing and discourse relation identification exists in the general domain, where researchers have not only identified discourse connectives, but also developed end-to-end discourse parsers (Pitler and Nenkova, 2009; Lin et al. [sent-43, score-0.735]
15 , 2008), a corpus of lexically-grounded annotations of discourse relations. [sent-46, score-0.176]
16 Until now, comparatively little work has been carried out on causal discourse relations in the biomedical domain, although causal associations between biological entities, events and processes are central to most claims of interest (Kleinberg and Hripcsak, 2011). [sent-47, score-1.682]
17 The equivalent of the PDTB for the biomedical domain is the BioDRB corpus (Prasad et al. [sent-48, score-0.309]
18 The number of purely causal relations annotated in this corpus is 542. [sent-52, score-0.61]
19 There are another 23 relations which are a mixture between causality and one ofeither background, temporal, conjunction or reinforcement relations. [sent-53, score-0.202]
20 , 2013), containing over 850 manually annotated causal discourse relations in 19 full-text open-access journal articles from the infectious diseases domain. [sent-55, score-0.918]
21 Using the BioDRB corpus as data, some researchers explored the identification of discourse connectives (Ramesh et al. [sent-56, score-0.276]
22 However, they do not distinguish between the types of discourse relations. [sent-58, score-0.215]
23 Also, they prove that there exist differences in discourse triggers between the biomedical and general domains by training a model on the BioDRB and evaluating it against PDTB and viceversa. [sent-63, score-0.803]
24 3 Methodology In this section, we describe our data and the features of causal triggers. [sent-64, score-0.591]
25 BioCause is a collection of 19 openaccess full-text journal articles pertaining to the biomedical subdomain of infectious diseases, manually annotated with causal relationships. [sent-68, score-0.94]
26 Two types of spans of text are marked in the text, namely causal triggers and causal arguments. [sent-69, score-1.508]
27 Each causal relation is composed of three text-bound annotations: a trigger, a cause or evidence argument and an effect argument. [sent-70, score-0.56]
28 Some causal relations have implicit triggers, so these are excluded from the current research. [sent-71, score-0.61]
29 Figure 1 shows an example of discourse causality from BioCause, marking the causal trigger and the two arguments with their respective relation. [sent-72, score-1.101]
30 BioCause contains 381 unique explicit triggers in the corpus, each being used, on average, only 2. [sent-74, score-0.318]
31 The number decreases to 347 unique triggers when they are lemmatised, corresponding to an average usage of 2. [sent-76, score-0.318]
32 Both count settings show the diversity of causality-triggering phrases that are used in the biomedical domain. [sent-78, score-0.309]
33 2 Features Three types of features have been employed in the development ofthis causality trigger model, i. [sent-80, score-0.435]
34 1 Lexical features The lexical features are built from the actual tokens present in text. [sent-86, score-0.148]
35 We included the token immediately to the left and the one immediately to the right of the current token. [sent-92, score-0.096]
36 Firstly, in the case of tokens to the left, most triggers are found either at the beginning of the sentence (3 11instances) or are preceded by a comma (238 instances). [sent-94, score-0.365]
37 Secondly, for the tokens to the right, almost 45% of triggers are followed by a determiner, such as the, a or an, (281 instances) or a comma (71 instances). [sent-96, score-0.396]
38 Figure 2 depicts a partial lexical parse tree of a sentence which starts with a causal trigger, namely Our results suggest that. [sent-100, score-0.67]
39 From the lexical parse trees, several types of features have been generated. [sent-101, score-0.15]
40 For instance, the figure shows that the token that has the part-ofspeech IN. [sent-103, score-0.096]
41 These features are included due to the fact that either many triggers are lexicalised as an adverb or conjunction, or are part of a verb phrase. [sent-104, score-0.349]
42 For the same reason, the syntactical category path from the root of the lexical parse tree to the token is also included. [sent-105, score-0.235]
43 The path also encodes, for each parent constituent, the position of the token in its subtree, i. [sent-106, score-0.096]
44 , beginning (B), inside (I) or end (E); if the token is the only leaf node of the constituent, this is marked differently, using a C. [sent-108, score-0.127]
45 Secondly, for each token, we extracted the pred- Figure 2: Partial lexical parse tree of a sentence starting with a causal trigger. [sent-110, score-0.67]
46 icate argument structure and checked whether a relation exista between the token and the previous and following tokens. [sent-111, score-0.096]
47 Thirdly, the ancestors of each token to the third degree are instantiated as three different features. [sent-113, score-0.154]
48 , the root of the lexical parse tree is less than three nodes away), a ”none” value is given. [sent-116, score-0.11]
49 For instance, the token that in Figure 2 has as its first three ancestors the constituents marked with CX, CP and VP. [sent-117, score-0.185]
50 Finally, the lowest common ancestor in the lexical parse tree between the current token and its left neighbour has been included. [sent-118, score-0.339]
51 In the example, the lowest common ancestor for that and suggest is VP. [sent-119, score-0.133]
52 These last two feature types have been produced on the observation that the lowest common ancestor for all tokens in a causal trigger is S or VP in over 70% of instances. [sent-120, score-0.992]
53 Furthermore, the percentage of cases of triggers with V or ADV as lowest common ancestor is almost 9% in each case. [sent-121, score-0.482]
54 Also, the aver40 age distance to the lowest common ancestor is 3. [sent-122, score-0.133]
55 3 Semantic features We have exploited several semantic knowledge sources to identify causal triggers more accurately, as a mapping to concepts and named entities acts as a back-off smoothing, thus increasing performance. [sent-125, score-0.986]
56 All documents annotated for causality in BioCause had been previously manually annotated with biomedical named entity and event information. [sent-127, score-0.58]
57 Moreover, another advantage of having a gold standard annotation is the fact that it is now possible to separate the task of automatic causal trigger recognition from automatic named entity recognition and event extraction. [sent-131, score-0.892]
58 The named entity and event annotation in the BioCause corpus is used to extract information about whether a token is part of a named entity or event trigger. [sent-132, score-0.334]
59 Furthermore, the type of the named entity or event is included as a separate feature. [sent-133, score-0.119]
60 Using this resource, the hypernym of every token in the text has been included as a feature. [sent-135, score-0.096]
61 Only the first sense of every token has been considered, as no sense disambiguation technique has been employed. [sent-136, score-0.096]
62 Thus, we included a feature to say whether a token is part of a UMLS type and another for its semantic type if the previous is true. [sent-138, score-0.128]
63 3 Experimental setup We explored with various machine learning algorithms and various settings for the task ofidentifying causal triggers. [sent-140, score-0.56]
64 org/ s o ftware / crfsuit e On the other hand, we modelled trigger detection as a classification task, using Support Vector Machines and Random Forests. [sent-146, score-0.213]
65 A lexicon is populated with all annotated causal triggers and then this is used to tag all instances of its entries in the text as connectives. [sent-156, score-0.878]
66 This is mainly due to triggers which are rarely used as causal triggers, such as and, by and that. [sent-160, score-0.878]
67 Building on the previously mentioned observation about the lowest common ancestor for all tokens in a causal trigger, we built a baseline system that checks all constituent nodes in the lexical parse tree for the S, V, VP and ADV tags and marks them as causal 41 triggers. [sent-161, score-1.41]
68 68%, largely due to the high number of intermediate nodes in the lexical parse tree that have VP as their category. [sent-165, score-0.11]
69 The third baseline is a combination of Dict and Dep: we consider only constituents that have the necessary category (S, V, VP or ADV) and include a trigger from the dictionary. [sent-166, score-0.213]
70 Although the recall decreases slightly, the precision increases to almost twice that of both Dict and Dep. [sent-167, score-0.09]
71 2 Sequence labelling task As a sequence labelling task, we have modelled causal trigger detection as two separate tasks. [sent-171, score-0.839]
72 Firstly, each trigger is represented in the B-I-O format (further mentioned as the 3-way model). [sent-172, score-0.213]
73 Thus, the first word of every trigger is tagged as B (begin), whilst the following words in the trigger are tagged as I(inside). [sent-173, score-0.53]
74 Hence, a sequence of contiguous tokens marked as part of a trigger form one trigger. [sent-177, score-0.291]
75 sion is obtained by using the syntactic features only, reaching over 92%, almost 3% higher than when all three feature types are used. [sent-191, score-0.101]
76 In the three-way model, syntactic and semantic features produce the best recall (almost 65%), which is just under 1% higher than the recall when all features are used. [sent-192, score-0.154]
77 3 Classification task As a classification task, an algorithm has to decide whether a token is part of a trigger or not, similarly to the previous two-way subtask in the case of CRF. [sent-194, score-0.309]
78 Due to the fact that causal triggers do not have a semantic mapping to concepts in the named entity and UMLS annotations, the trees in the random forest classifier can easily produce rules that distinguish triggers from non-triggers. [sent-199, score-1.303]
79 As such, the use of semantic features alone produce a very good precision of 84. [sent-200, score-0.092]
80 Also, in all cases where semantic features are combined with other feature types, the precision increases by 0. [sent-202, score-0.092]
81 Secondly, we explored the performance of SVMs in detecting causal triggers. [sent-207, score-0.56]
82 As can be observed, the best performance is obtained when combining the lexical and semantic feature types (69. [sent-212, score-0.11]
83 The combination of all features produces the best precision, whilst the best recall is obtained by combining lexical and semantic features. [sent-214, score-0.18]
84 This applies to 107 trigger types, whose number offalse positives (FP) is higher than the number of true positives (TP). [sent-217, score-0.213]
85 In fact, 64 trigger types occur only once as a causal instance, whilst the average number of FPs for these types is 14. [sent-218, score-0.899]
86 One such example is and, for which the number of non-causal instances (2305) is much greater than that of causal instances (1). [sent-220, score-0.56]
87 Other examples of trigger types more commonly used as causal triggers, are suggesting (9 TP, 54 FP), indicating (8 TP, 41 FP) and resulting in (6 TP, 14 FP). [sent-221, score-0.841]
88 (2) Buffer treated control cells showed intense green staining with syto9 (indicating viability) and a lack of PI staining (indicating no dead/dying cells or DNA release). [sent-223, score-0.088]
89 5 Conclusions and Future Work We have presented an approach to the automatic identification of triggers of causal discourse relations in biomedical scientific text. [sent-224, score-1.444]
90 Shallow approaches, such as dictionary matching and lexical parse tree matching, perform very poorly, due to the high ambiguity of causal triggers (with F-scores of approximately 15% each and 24% when combined). [sent-226, score-0.988]
91 We have explored various machine learning algorithms that automatically classify tokens into triggers or non-triggers and we have evaluated the impact of multiple lexical, syntactic and semantic features. [sent-227, score-0.397]
92 The performance of SVMs prove that the task of identifying causal triggers is indeed complex. [sent-228, score-0.878]
93 As future work, integrating the causal relations in the BioDRB corpus is necessary to check whether a data insufficiency problem exists and, if so, estimate the optimal amount of necessary data. [sent-231, score-0.61]
94 Furthermore, evaluations against the general domain need to be performed, in order to establish any differences in expressing causality in the biomedical domain. [sent-232, score-0.461]
95 Finally, our system needs to be extended in order to identify the two arguments of 43 causal relations, the cause and effect, thus allowing the creation of a complete discourse causality parser. [sent-235, score-0.888]
96 The unified medical language system (UMLS): integrating biomedical terminology. [sent-248, score-0.349]
97 BioCause: Annotating and analysing causality in the biomedical domain. [sent-296, score-0.491]
98 Using syntax to disambiguate explicit discourse connectives in text. [sent-321, score-0.245]
99 Static relations: a piece in the biomedical information extraction puzzle. [sent-334, score-0.309]
100 Overview of the infectious diseases (ID) task of BioNLP shared task 2011. [sent-339, score-0.132]
wordName wordTfidf (topN-words)
[('causal', 0.56), ('triggers', 0.318), ('biomedical', 0.309), ('trigger', 0.213), ('biocause', 0.196), ('discourse', 0.176), ('causality', 0.152), ('sophia', 0.114), ('miwa', 0.1), ('token', 0.096), ('biodrb', 0.087), ('mcnaught', 0.087), ('pmrb', 0.087), ('pyysalo', 0.084), ('tomoko', 0.084), ('ancestor', 0.083), ('ohta', 0.08), ('ananiadou', 0.076), ('bionlp', 0.076), ('infectious', 0.071), ('pdtb', 0.071), ('umls', 0.071), ('connectives', 0.069), ('bioinformatics', 0.069), ('acid', 0.067), ('prasad', 0.067), ('ichi', 0.062), ('diseases', 0.061), ('miyao', 0.058), ('informatics', 0.058), ('bmc', 0.058), ('ancestors', 0.058), ('mih', 0.053), ('sampo', 0.053), ('relations', 0.05), ('dict', 0.05), ('lowest', 0.05), ('fp', 0.049), ('whilst', 0.048), ('tokens', 0.047), ('jun', 0.047), ('rashmi', 0.046), ('named', 0.045), ('event', 0.044), ('tp', 0.044), ('tsuruoka', 0.044), ('adv', 0.044), ('acids', 0.044), ('fukuda', 0.044), ('glutamic', 0.044), ('histidine', 0.044), ('periplasmic', 0.044), ('protonation', 0.044), ('regulation', 0.044), ('staining', 0.044), ('crf', 0.042), ('parse', 0.041), ('vp', 0.041), ('medical', 0.04), ('lexical', 0.039), ('types', 0.039), ('zweigenbaum', 0.039), ('manche', 0.039), ('claudiu', 0.039), ('rzhetsky', 0.039), ('kernels', 0.038), ('makoto', 0.038), ('rf', 0.037), ('hong', 0.037), ('amino', 0.036), ('pathway', 0.036), ('savova', 0.033), ('ramesh', 0.033), ('labelling', 0.033), ('semantic', 0.032), ('dep', 0.032), ('almost', 0.031), ('marked', 0.031), ('identification', 0.031), ('svms', 0.031), ('features', 0.031), ('molecular', 0.03), ('mild', 0.03), ('tsujii', 0.03), ('analysing', 0.03), ('il', 0.03), ('entity', 0.03), ('recall', 0.03), ('tree', 0.03), ('indicating', 0.029), ('yoshimasa', 0.029), ('kleinberg', 0.029), ('pitler', 0.029), ('syntactical', 0.029), ('precision', 0.029), ('secondly', 0.028), ('tagged', 0.028), ('ph', 0.028), ('biological', 0.027), ('forests', 0.027)]
simIndex simValue paperId paperTitle
same-paper 1 0.99999911 386 acl-2013-What causes a causal relation? Detecting Causal Triggers in Biomedical Scientific Discourse
Author: Claudiu Mihaila ; Sophia Ananiadou
Abstract: Current domain-specific information extraction systems represent an important resource for biomedical researchers, who need to process vaster amounts of knowledge in short times. Automatic discourse causality recognition can further improve their workload by suggesting possible causal connections and aiding in the curation of pathway models. We here describe an approach to the automatic identification of discourse causality triggers in the biomedical domain using machine learning. We create several baselines and experiment with various parameter settings for three algorithms, i.e., Conditional Random Fields (CRF), Support Vector Machines (SVM) and Random Forests (RF). Also, we evaluate the impact of lexical, syntactic and semantic features on each of the algorithms and look at er- rors. The best performance of 79.35% F-score is achieved by CRFs when using all three feature types.
2 0.49449569 387 acl-2013-Why-Question Answering using Intra- and Inter-Sentential Causal Relations
Author: Jong-Hoon Oh ; Kentaro Torisawa ; Chikara Hashimoto ; Motoki Sano ; Stijn De Saeger ; Kiyonori Ohtake
Abstract: In this paper, we explore the utility of intra- and inter-sentential causal relations between terms or clauses as evidence for answering why-questions. To the best of our knowledge, this is the first work that uses both intra- and inter-sentential causal relations for why-QA. We also propose a method for assessing the appropriateness of causal relations as answers to a given question using the semantic orientation of excitation proposed by Hashimoto et al. (2012). By applying these ideas to Japanese why-QA, we improved precision by 4.4% against all the questions in our test set over the current state-of-theart system for Japanese why-QA. In addi- tion, unlike the state-of-the-art system, our system could achieve very high precision (83.2%) for 25% of all the questions in the test set by restricting its output to the confident answers only.
3 0.21602362 206 acl-2013-Joint Event Extraction via Structured Prediction with Global Features
Author: Qi Li ; Heng Ji ; Liang Huang
Abstract: Traditional approaches to the task of ACE event extraction usually rely on sequential pipelines with multiple stages, which suffer from error propagation since event triggers and arguments are predicted in isolation by independent local classifiers. By contrast, we propose a joint framework based on structured prediction which extracts triggers and arguments together so that the local predictions can be mutually improved. In addition, we propose to incorporate global features which explicitly capture the dependencies of multiple triggers and arguments. Experimental results show that our joint approach with local features outperforms the pipelined baseline, and adding global features further improves the performance significantly. Our approach advances state-ofthe-art sentence-level event extraction, and even outperforms previous argument labeling methods which use external knowledge from other sentences and documents.
4 0.20345193 56 acl-2013-Argument Inference from Relevant Event Mentions in Chinese Argument Extraction
Author: Peifeng Li ; Qiaoming Zhu ; Guodong Zhou
Abstract: As a paratactic language, sentence-level argument extraction in Chinese suffers much from the frequent occurrence of ellipsis with regard to inter-sentence arguments. To resolve such problem, this paper proposes a novel global argument inference model to explore specific relationships, such as Coreference, Sequence and Parallel, among relevant event mentions to recover those intersentence arguments in the sentence, discourse and document layers which represent the cohesion of an event or a topic. Evaluation on the ACE 2005 Chinese corpus justifies the effectiveness of our global argument inference model over a state-of-the-art baseline. 1
5 0.15724391 229 acl-2013-Leveraging Synthetic Discourse Data via Multi-task Learning for Implicit Discourse Relation Recognition
Author: Man Lan ; Yu Xu ; Zhengyu Niu
Abstract: To overcome the shortage of labeled data for implicit discourse relation recognition, previous works attempted to automatically generate training data by removing explicit discourse connectives from sentences and then built models on these synthetic implicit examples. However, a previous study (Sporleder and Lascarides, 2008) showed that models trained on these synthetic data do not generalize very well to natural (i.e. genuine) implicit discourse data. In this work we revisit this issue and present a multi-task learning based system which can effectively use synthetic data for implicit discourse relation recognition. Results on PDTB data show that under the multi-task learning framework our models with the use of the prediction of explicit discourse connectives as auxiliary learning tasks, can achieve an averaged F1 improvement of 5.86% over baseline models.
6 0.14306951 41 acl-2013-Aggregated Word Pair Features for Implicit Discourse Relation Disambiguation
7 0.12844074 2 acl-2013-A Bayesian Model for Joint Unsupervised Induction of Sentiment, Aspect and Discourse Representations
8 0.11404721 85 acl-2013-Combining Intra- and Multi-sentential Rhetorical Parsing for Document-level Discourse Analysis
9 0.091983534 69 acl-2013-Bilingual Lexical Cohesion Trigger Model for Document-Level Machine Translation
10 0.080607757 130 acl-2013-Domain-Specific Coreference Resolution with Lexicalized Features
11 0.068613857 52 acl-2013-Annotating named entities in clinical text by combining pre-annotation and active learning
12 0.063145511 224 acl-2013-Learning to Extract International Relations from Political Context
13 0.060282022 296 acl-2013-Recognizing Identical Events with Graph Kernels
14 0.058009923 144 acl-2013-Explicit and Implicit Syntactic Features for Text Classification
15 0.056748945 134 acl-2013-Embedding Semantic Similarity in Tree Kernels for Domain Adaptation of Relation Extraction
16 0.055773057 45 acl-2013-An Empirical Study on Uncertainty Identification in Social Media Context
17 0.054404289 339 acl-2013-Temporal Signals Help Label Temporal Relations
18 0.053260572 306 acl-2013-SPred: Large-scale Harvesting of Semantic Predicates
19 0.052808229 118 acl-2013-Development and Analysis of NLP Pipelines in Argo
20 0.05272764 169 acl-2013-Generating Synthetic Comparable Questions for News Articles
topicId topicWeight
[(0, 0.159), (1, 0.059), (2, -0.065), (3, -0.078), (4, -0.021), (5, 0.197), (6, -0.019), (7, 0.081), (8, 0.035), (9, 0.129), (10, 0.119), (11, -0.016), (12, -0.087), (13, 0.089), (14, 0.028), (15, -0.087), (16, 0.019), (17, -0.162), (18, -0.106), (19, -0.105), (20, 0.071), (21, -0.04), (22, 0.05), (23, -0.079), (24, -0.034), (25, -0.012), (26, 0.005), (27, -0.016), (28, -0.012), (29, -0.074), (30, 0.008), (31, 0.078), (32, -0.193), (33, 0.147), (34, -0.134), (35, 0.122), (36, 0.105), (37, 0.205), (38, 0.061), (39, -0.013), (40, 0.075), (41, 0.117), (42, 0.075), (43, 0.03), (44, -0.155), (45, 0.033), (46, -0.102), (47, 0.094), (48, 0.002), (49, -0.09)]
simIndex simValue paperId paperTitle
same-paper 1 0.9223401 386 acl-2013-What causes a causal relation? Detecting Causal Triggers in Biomedical Scientific Discourse
Author: Claudiu Mihaila ; Sophia Ananiadou
Abstract: Current domain-specific information extraction systems represent an important resource for biomedical researchers, who need to process vaster amounts of knowledge in short times. Automatic discourse causality recognition can further improve their workload by suggesting possible causal connections and aiding in the curation of pathway models. We here describe an approach to the automatic identification of discourse causality triggers in the biomedical domain using machine learning. We create several baselines and experiment with various parameter settings for three algorithms, i.e., Conditional Random Fields (CRF), Support Vector Machines (SVM) and Random Forests (RF). Also, we evaluate the impact of lexical, syntactic and semantic features on each of the algorithms and look at er- rors. The best performance of 79.35% F-score is achieved by CRFs when using all three feature types.
2 0.78194529 387 acl-2013-Why-Question Answering using Intra- and Inter-Sentential Causal Relations
Author: Jong-Hoon Oh ; Kentaro Torisawa ; Chikara Hashimoto ; Motoki Sano ; Stijn De Saeger ; Kiyonori Ohtake
Abstract: In this paper, we explore the utility of intra- and inter-sentential causal relations between terms or clauses as evidence for answering why-questions. To the best of our knowledge, this is the first work that uses both intra- and inter-sentential causal relations for why-QA. We also propose a method for assessing the appropriateness of causal relations as answers to a given question using the semantic orientation of excitation proposed by Hashimoto et al. (2012). By applying these ideas to Japanese why-QA, we improved precision by 4.4% against all the questions in our test set over the current state-of-theart system for Japanese why-QA. In addi- tion, unlike the state-of-the-art system, our system could achieve very high precision (83.2%) for 25% of all the questions in the test set by restricting its output to the confident answers only.
3 0.55617082 56 acl-2013-Argument Inference from Relevant Event Mentions in Chinese Argument Extraction
Author: Peifeng Li ; Qiaoming Zhu ; Guodong Zhou
Abstract: As a paratactic language, sentence-level argument extraction in Chinese suffers much from the frequent occurrence of ellipsis with regard to inter-sentence arguments. To resolve such problem, this paper proposes a novel global argument inference model to explore specific relationships, such as Coreference, Sequence and Parallel, among relevant event mentions to recover those intersentence arguments in the sentence, discourse and document layers which represent the cohesion of an event or a topic. Evaluation on the ACE 2005 Chinese corpus justifies the effectiveness of our global argument inference model over a state-of-the-art baseline. 1
4 0.55053002 206 acl-2013-Joint Event Extraction via Structured Prediction with Global Features
Author: Qi Li ; Heng Ji ; Liang Huang
Abstract: Traditional approaches to the task of ACE event extraction usually rely on sequential pipelines with multiple stages, which suffer from error propagation since event triggers and arguments are predicted in isolation by independent local classifiers. By contrast, we propose a joint framework based on structured prediction which extracts triggers and arguments together so that the local predictions can be mutually improved. In addition, we propose to incorporate global features which explicitly capture the dependencies of multiple triggers and arguments. Experimental results show that our joint approach with local features outperforms the pipelined baseline, and adding global features further improves the performance significantly. Our approach advances state-ofthe-art sentence-level event extraction, and even outperforms previous argument labeling methods which use external knowledge from other sentences and documents.
5 0.49984002 69 acl-2013-Bilingual Lexical Cohesion Trigger Model for Document-Level Machine Translation
Author: Guosheng Ben ; Deyi Xiong ; Zhiyang Teng ; Yajuan Lu ; Qun Liu
Abstract: In this paper, we propose a bilingual lexical cohesion trigger model to capture lexical cohesion for document-level machine translation. We integrate the model into hierarchical phrase-based machine translation and achieve an absolute improvement of 0.85 BLEU points on average over the baseline on NIST Chinese-English test sets.
6 0.45694965 229 acl-2013-Leveraging Synthetic Discourse Data via Multi-task Learning for Implicit Discourse Relation Recognition
7 0.43835643 41 acl-2013-Aggregated Word Pair Features for Implicit Discourse Relation Disambiguation
8 0.42124158 85 acl-2013-Combining Intra- and Multi-sentential Rhetorical Parsing for Document-level Discourse Analysis
9 0.39832103 42 acl-2013-Aid is Out There: Looking for Help from Tweets during a Large Scale Disaster
10 0.36942953 296 acl-2013-Recognizing Identical Events with Graph Kernels
11 0.3499625 2 acl-2013-A Bayesian Model for Joint Unsupervised Induction of Sentiment, Aspect and Discourse Representations
12 0.33957782 247 acl-2013-Modeling of term-distance and term-occurrence information for improving n-gram language model performance
13 0.31228897 159 acl-2013-Filling Knowledge Base Gaps for Distant Supervision of Relation Extraction
14 0.29355112 298 acl-2013-Recognizing Rare Social Phenomena in Conversation: Empowerment Detection in Support Group Chatrooms
15 0.28431866 224 acl-2013-Learning to Extract International Relations from Political Context
16 0.27614069 16 acl-2013-A Novel Translation Framework Based on Rhetorical Structure Theory
17 0.27394709 134 acl-2013-Embedding Semantic Similarity in Tree Kernels for Domain Adaptation of Relation Extraction
18 0.27243915 321 acl-2013-Sign Language Lexical Recognition With Propositional Dynamic Logic
19 0.26820382 144 acl-2013-Explicit and Implicit Syntactic Features for Text Classification
20 0.26816934 22 acl-2013-A Structured Distributional Semantic Model for Event Co-reference
topicId topicWeight
[(0, 0.053), (6, 0.03), (11, 0.051), (13, 0.264), (14, 0.017), (24, 0.057), (26, 0.051), (35, 0.054), (42, 0.095), (48, 0.043), (64, 0.018), (70, 0.041), (88, 0.052), (90, 0.025), (95, 0.062)]
simIndex simValue paperId paperTitle
same-paper 1 0.79060227 386 acl-2013-What causes a causal relation? Detecting Causal Triggers in Biomedical Scientific Discourse
Author: Claudiu Mihaila ; Sophia Ananiadou
Abstract: Current domain-specific information extraction systems represent an important resource for biomedical researchers, who need to process vaster amounts of knowledge in short times. Automatic discourse causality recognition can further improve their workload by suggesting possible causal connections and aiding in the curation of pathway models. We here describe an approach to the automatic identification of discourse causality triggers in the biomedical domain using machine learning. We create several baselines and experiment with various parameter settings for three algorithms, i.e., Conditional Random Fields (CRF), Support Vector Machines (SVM) and Random Forests (RF). Also, we evaluate the impact of lexical, syntactic and semantic features on each of the algorithms and look at er- rors. The best performance of 79.35% F-score is achieved by CRFs when using all three feature types.
2 0.75115699 322 acl-2013-Simple, readable sub-sentences
Author: Sigrid Klerke ; Anders Sgaard
Abstract: We present experiments using a new unsupervised approach to automatic text simplification, which builds on sampling and ranking via a loss function informed by readability research. The main idea is that a loss function can distinguish good simplification candidates among randomly sampled sub-sentences of the input sentence. Our approach is rated as equally grammatical and beginner reader appropriate as a supervised SMT-based baseline system by native speakers, but our setup performs more radical changes that better resembles the variation observed in human generated simplifications.
3 0.74884516 186 acl-2013-Identifying English and Hungarian Light Verb Constructions: A Contrastive Approach
Author: Veronika Vincze ; Istvan Nagy T. ; Richard Farkas
Abstract: Here, we introduce a machine learningbased approach that allows us to identify light verb constructions (LVCs) in Hungarian and English free texts. We also present the results of our experiments on the SzegedParalellFX English–Hungarian parallel corpus where LVCs were manually annotated in both languages. With our approach, we were able to contrast the performance of our method and define language-specific features for these typologically different languages. Our presented method proved to be sufficiently robust as it achieved approximately the same scores on the two typologically different languages.
4 0.6507498 68 acl-2013-Bilingual Data Cleaning for SMT using Graph-based Random Walk
Author: Lei Cui ; Dongdong Zhang ; Shujie Liu ; Mu Li ; Ming Zhou
Abstract: The quality of bilingual data is a key factor in Statistical Machine Translation (SMT). Low-quality bilingual data tends to produce incorrect translation knowledge and also degrades translation modeling performance. Previous work often used supervised learning methods to filter lowquality data, but a fair amount of human labeled examples are needed which are not easy to obtain. To reduce the reliance on labeled examples, we propose an unsupervised method to clean bilingual data. The method leverages the mutual reinforcement between the sentence pairs and the extracted phrase pairs, based on the observation that better sentence pairs often lead to better phrase extraction and vice versa. End-to-end experiments show that the proposed method substantially improves the performance in largescale Chinese-to-English translation tasks.
5 0.55014807 56 acl-2013-Argument Inference from Relevant Event Mentions in Chinese Argument Extraction
Author: Peifeng Li ; Qiaoming Zhu ; Guodong Zhou
Abstract: As a paratactic language, sentence-level argument extraction in Chinese suffers much from the frequent occurrence of ellipsis with regard to inter-sentence arguments. To resolve such problem, this paper proposes a novel global argument inference model to explore specific relationships, such as Coreference, Sequence and Parallel, among relevant event mentions to recover those intersentence arguments in the sentence, discourse and document layers which represent the cohesion of an event or a topic. Evaluation on the ACE 2005 Chinese corpus justifies the effectiveness of our global argument inference model over a state-of-the-art baseline. 1
6 0.54280245 225 acl-2013-Learning to Order Natural Language Texts
7 0.54232961 70 acl-2013-Bilingually-Guided Monolingual Dependency Grammar Induction
8 0.54166704 98 acl-2013-Cross-lingual Transfer of Semantic Role Labeling Models
9 0.53890014 80 acl-2013-Chinese Parsing Exploiting Characters
10 0.53876513 223 acl-2013-Learning a Phrase-based Translation Model from Monolingual Data with Application to Domain Adaptation
11 0.53812879 123 acl-2013-Discriminative Learning with Natural Annotations: Word Segmentation as a Case Study
12 0.53758883 38 acl-2013-Additive Neural Networks for Statistical Machine Translation
13 0.53698909 82 acl-2013-Co-regularizing character-based and word-based models for semi-supervised Chinese word segmentation
14 0.53655195 343 acl-2013-The Effect of Higher-Order Dependency Features in Discriminative Phrase-Structure Parsing
15 0.53586137 155 acl-2013-Fast and Accurate Shift-Reduce Constituent Parsing
16 0.53569293 132 acl-2013-Easy-First POS Tagging and Dependency Parsing with Beam Search
17 0.53539169 183 acl-2013-ICARUS - An Extensible Graphical Search Tool for Dependency Treebanks
18 0.53470516 164 acl-2013-FudanNLP: A Toolkit for Chinese Natural Language Processing
19 0.53347939 18 acl-2013-A Sentence Compression Based Framework to Query-Focused Multi-Document Summarization
20 0.53244245 127 acl-2013-Docent: A Document-Level Decoder for Phrase-Based Statistical Machine Translation