| { | |
| "paper_id": "Y12-1026", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T13:46:00.518671Z" | |
| }, | |
| "title": "Predicting Answer Location Using Shallow Semantic Analogical Reasoning in a Factoid Question Answering System", | |
| "authors": [ | |
| { | |
| "first": "Hapnes", | |
| "middle": [], | |
| "last": "Toba", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Universitas Indonesia", | |
| "location": { | |
| "postCode": "16424", | |
| "settlement": "Depok", | |
| "country": "Indonesia" | |
| } | |
| }, | |
| "email": "hapnes.toba@ui.ac.id" | |
| }, | |
| { | |
| "first": "Mirna", | |
| "middle": [], | |
| "last": "Adriani", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Universitas Indonesia", | |
| "location": { | |
| "postCode": "16424", | |
| "settlement": "Depok", | |
| "country": "Indonesia" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Ruli", | |
| "middle": [], | |
| "last": "Manurung", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Universitas Indonesia", | |
| "location": { | |
| "postCode": "16424", | |
| "settlement": "Depok", | |
| "country": "Indonesia" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "In this paper we report our work on a factoid question answering task that avoids namedentity recognition tool in the answer selection process. We use semantic analogical reasoning to find the location of the final answer from a textual passage.We demonstrate that without employing any linguistic tools during the answer selection process, our approach achieves a better accuracy than a typical factoid question answering architecture.", | |
| "pdf_parse": { | |
| "paper_id": "Y12-1026", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "In this paper we report our work on a factoid question answering task that avoids namedentity recognition tool in the answer selection process. We use semantic analogical reasoning to find the location of the final answer from a textual passage.We demonstrate that without employing any linguistic tools during the answer selection process, our approach achieves a better accuracy than a typical factoid question answering architecture.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "The task of a question answering system (QAS) is to provide a single answer for a given natural language question. In a factoid QAS, the system tries to give the best answer of an open-domain fact-based question. For example, the question \"Where was an Oviraptor fossil sitting on a nest discovered?\". A QAS should return 'Mongolia's Gobi Desert' as the final answer.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "A typical pipeline architecture in a fact-based QAS consists of four main processes, i.e.: question analysis, query formulation, information retrieval and answer selection. The main source of complexity in a QAS lies in the question analysis and answer selection process rather than in the information retrieval (IR) phase, which is usually achieved by utilizing third-party modules such as Lucene, Indri, or a web search engine.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The question analysis process seeks to determine the type of a given question, which in turn provides the expected answer type (EAT) of that question as a specific fact type, such as person, organization or location. The EAT will be used to select the best answer during the answer selection process, usually by utilizing a named-entity recognizer (NER) tool in a factoid QAS (Schlaefer et al., 2006) . Different approaches have been used in order to improve the performance of the answer selection component. Ko et al. (2010) employed probabilistic models for answer ranking of NER-based answer selection by utilizing external semantic resources such as WordNet. More advanced techniques utilizing linguistic tools have been proposed in Sun et al. (2005) , which uses syntactic relation analysis to extract the final answer, and Moreda et al. (2010) , which employs semantic roles to improve NER-based answer selection. Recent work by Moschitti and Quarteroni (2011) proposed classification of paired texts that learn to select answers by applying syntactic tree kernels to pairs of questions and answers.", | |
| "cite_spans": [ | |
| { | |
| "start": 376, | |
| "end": 400, | |
| "text": "(Schlaefer et al., 2006)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 510, | |
| "end": 526, | |
| "text": "Ko et al. (2010)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 738, | |
| "end": 755, | |
| "text": "Sun et al. (2005)", | |
| "ref_id": "BIBREF10" | |
| }, | |
| { | |
| "start": 830, | |
| "end": 850, | |
| "text": "Moreda et al. (2010)", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 936, | |
| "end": 967, | |
| "text": "Moschitti and Quarteroni (2011)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In our current work, we try to reduce the dependency of the answer selection process on linguistic tools such as NER systems. Our main concern is that in reality we do not always have a complete N-ER tool for every fact type. In our example mentioned above, the answer has a fact type which is neither an exact location, person nor an organization, i.e.: 'Mongolia's Gobi Desert'. In such case, a NER-based system might fail to extract the answer. Further, if we have a complete NER-tool, it is still a complex problem to predict the location of the exact answer in a retrieval result.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "We propose an approach which we call semantic analogical reasoning (SAR). Our approach tries to predict the location of the final answer in a textual passage by employing the analogical reasoning Silva et al. (2010) . We hypothesize that similar questions give similar answers. Based on the retrieved similar questions, our approach tries to provide the best example of question-answer pairs and use the influence level (weights) of the semantic features to predict the location of the final answer.", | |
| "cite_spans": [ | |
| { | |
| "start": 196, | |
| "end": 215, | |
| "text": "Silva et al. (2010)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In the remainder of this paper, our basic idea and related works of semantic analogical reasoning will be presented in Section 2. The system architecture, procedures, experiments, and performance evaluation will be presented in Sections 3 and 4. Finally, our conclusions and future work will be drawn in Section 5.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The basic idea of semantic analogical reasoning is to find a portion of text in a passage which is considered useful during the answer selection process. Consider the two pairs of question and answer in Figure 1 . Both questions need a fact type as the final answer, i.e. Mongolia's Gobi Desert (a) and Niagara Falls, N.Y. (b). We postulate that both questions have a high probability to share common answer features that will be useful to find the location of the final answer.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 203, | |
| "end": 211, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "If we investigate the structure of the answer passage of question (a), we can see that the final answer is a noun phrase (NP), which is surrounded by a preposition (PP) and a stop sign (O). In question (b), we also found that the final answer is located in a sequence of PP-NP-O. Thus, if we can learn these kinds of related structures between question answer pairs for any EAT, we will have useful information to predict the location of the final answer in a textual passage. In this sense, we focus our work in learning the relational feature similarity between question answer pairs. Silva et al. (2007; has investigated a statistical-based analogical reasoning (AR) framework. It is a method for ranking relations based on the Bayesian similarity criterion. The underlying idea of AR is to learn model parameters and priors from related objects (question and answer pairs in our case), and update the priors during the retrieval process of a query. The objective of the AR framework is to obtain a marginal probability that relates a new object pair (query) with a set of objects that have been learnt.", | |
| "cite_spans": [ | |
| { | |
| "start": 587, | |
| "end": 606, | |
| "text": "Silva et al. (2007;", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Most methods of classification or similarity measures focus on the similarity between the features of objects in a candidate pair and the features of objects in a query pair. AR focuses instead on the similarity between functions that map pairs to links. To some extent, this is the main reason that AR is appropriate for our idea. Wang et al. (2009) has shown that AR is effective in retrieving similar question-answer pairs in a community-based QAS. They use statistical features such as term frequency, common n-gram length, and question answer length ratio. In contrast to our approach which tries to validate the location of a final answer; their work is limited to the retrieval of similar question.", | |
| "cite_spans": [ | |
| { | |
| "start": 332, | |
| "end": 350, | |
| "text": "Wang et al. (2009)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "The SAR approach, that we develop in this research is an extension of our previous work (Toba et al., 2011) , which showed that AR can be used to construct EAT patterns. In our previous research, we used named-entity occurrences as features to relate the question and answer pairs. This time, instead of using named-entities as features, we use semantic information -which is based on syntactic featuresto predict the corresponding named-entities.", | |
| "cite_spans": [ | |
| { | |
| "start": 88, | |
| "end": 107, | |
| "text": "(Toba et al., 2011)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Moschitti and Quarteroni (2011) use predicate argument structures, syntactic and shallow semantic tree kernel features to train question and answer pairs on SVM rank. Two consequences of using complex linguistic features is high computational cost and the requirement to have access to adequate linguistic resources and tools. For these reasons, we propose to use a simpler feature set, i.e. the trigram sequences of syntactic chunk. Unlike the research in Moschitti and Quarteroni (2011) that uses the whole syntactic tree, in this research we only keep the order of the root of any partial tree segment in trigrams of part-of-speech (POS) sequences.", | |
| "cite_spans": [ | |
| { | |
| "start": 457, | |
| "end": 488, | |
| "text": "Moschitti and Quarteroni (2011)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "In short, we develop a set of procedures to determine the best similar question-answer pair and predict the final answer location of a given factoid ques- tion. Our SAR approach extends the above mentioned related works in the following aspects:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "1. We extend the AR framework from Silva, et al. (2007; to re-rank the AR retrieval process according to the most influential semantic features.", | |
| "cite_spans": [ | |
| { | |
| "start": 35, | |
| "end": 55, | |
| "text": "Silva, et al. (2007;", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "2. We extend the question-answer retrieval process of Wang, et al. (2009) and Moschitti and Quarteroni (2011) to find the most possible final answer location in a textual passage by utilizing POS sequences as semantic features.", | |
| "cite_spans": [ | |
| { | |
| "start": 54, | |
| "end": 73, | |
| "text": "Wang, et al. (2009)", | |
| "ref_id": "BIBREF11" | |
| }, | |
| { | |
| "start": 78, | |
| "end": 109, | |
| "text": "Moschitti and Quarteroni (2011)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Semantic Analogical Reasoning", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Our architecture is depicted in Figure 2 . There are two main process flows in the architecture. The first one is the training process (noted by the dashed lines), and the second one is the question answering process (noted by the solid lines).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 32, | |
| "end": 40, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "System Architecture", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "During the training process, the semantic features as described in Table 1 will be extracted and used in the AR training module. The training process will produce an AR model. Another important step in the training process is the evaluation of the features importance level. We need to know which semantic feature has the most influence in the model. This information will be important to select the best question answer pair later in the re-ranking process.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 67, | |
| "end": 74, | |
| "text": "Table 1", | |
| "ref_id": "TABREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Question Answering Framework", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "In the question answering step, the shallow semantic features of the question and the related answer passages -which have been retrieved during Figure 2 : System Architecture the IR process -are extracted. In this step, we will have a collection of ranked similar question answer pairs from the learnt AR model. Each similar pair needs to be evaluated (re-ranked), to make sure that we will have the best similar pair. In the final step, based on the best similar question answer pair, we search for the location of the answer from the textual passage by matching the sequence of the answer chunk to produce the final answer.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 144, | |
| "end": 152, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Question Answering Framework", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "In this part we summarize first the AR framework as introduced by Silva et al. (2007; . The framework consists of two phases, i.e. the training and retrieval process. Consider a collection of related objects with some unseen labels L ij 's, where", | |
| "cite_spans": [ | |
| { | |
| "start": 66, | |
| "end": 85, | |
| "text": "Silva et al. (2007;", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "L ij \u2208 {0, 1}", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "is an expected indicator of the existence of a relation between two related objects i and j. Consider then that we also have K-dimensional vectors, each consisting of features which relates the objects i and j :", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "\u0398 = [\u0398 1 . . . \u0398 k ] T .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "In general, this vector will represent the presence or absence of relation between two particular objects. Given the vectors of features \u0398 , the strength of the relation between two objects i and j is computed by performing logistic regression estimation as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P (L ij |x ij , \u0398) = logistic(\u0398 T X ij )", | |
| "eq_num": "(1)" | |
| } | |
| ], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "where logistic(x) is defined as:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "1 1 + e \u2212x", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "(2)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "During AR training phase, the framework learns the weight (prior) for each feature by performing the following equation:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P (\u0398) = N (\u0398, (cT ) \u22121 )", | |
| "eq_num": "(3)" | |
| } | |
| ], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "where\u0398 is the logistic estimator of \u0398, and", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "N (m, v)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "is a normal of mean m and variance v. Matrix T is the empirical second moment's matrix of the link object features, and c is a smoothing parameter which is set by the user.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "During the AR retrieval phase, a final score that indicates the rank of predicted relations between two new objects i and j (query) and the related objects that have been learnt in a given set S is compute as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "score(Q i , A j ) = log P (L ij |X ij , S, L S = 1) P (L ij = 1|X ij )", | |
| "eq_num": "(4)" | |
| } | |
| ], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "Silva et al. 2010use the variational logistic regression approach to compute the scoring function in equation 4. This score gives the rank of similarity of how \"analogous\" a new query is to other related objects in a given learnt set S.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "A drawback of AR as mentioned in Silva et al. (2010) is that by conditioning on the link indicators, the similarity score (eq. 4) between two objects i:j, and other objects x:y, is always a function of pairs (i,j) and (x,y) that is not in general decomposable as similarities between i and x, and j and y. Due to this limitation, we propose to evaluate the importance level (weight) of each feature which is used to relate the objects, and use the weights to re-rank the similarity score.", | |
| "cite_spans": [ | |
| { | |
| "start": 33, | |
| "end": 52, | |
| "text": "Silva et al. (2010)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "We empirically calculate the weighting factors for each feature set in Table 1 , with respect to the expected answer-type, by performing chi-square (x 2 ) evaluation ( (Manning et al., 2008) , pp. 255-256) of overlapped features. The chi-square evaluation of the weighting factors are computed from the ARretrieval results of the training data. To calculate the importance of each feature, we performed a top-10 retrieval for each question during the training phase on several parameter settings. As suggested in Silva et al. (2010) , the value of the smoothing parameter is set as the 'number of positive links' in the training data. We took variations of this smoothing parameter by multiplying the 'number of positive links' by a factor of: 0.1, 0.5, 2, 4, 8, 10 and 16 during the chi-square evaluation. Finally, we compute an average value to form the final weighting factor of each feature, as can be seen in Table 2 .", | |
| "cite_spans": [ | |
| { | |
| "start": 168, | |
| "end": 190, | |
| "text": "(Manning et al., 2008)", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 513, | |
| "end": 532, | |
| "text": "Silva et al. (2010)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 71, | |
| "end": 78, | |
| "text": "Table 1", | |
| "ref_id": "TABREF0" | |
| }, | |
| { | |
| "start": 914, | |
| "end": 921, | |
| "text": "Table 2", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The final answer selection strategy is started by selecting the best question-answer analogous pair. To select the best pair we performed first a top-10 AR retrieval, re-ranked them by using the feature weighting factors, and finally took the best score pair. This pair is considered as the best pair which has the most overlapped features to the new question. To select the final answer in a passage, we performed a feature matching process of the answer features, i.e.: the overlap of trigram POS chunks sequences '[left chunk -answer chunk -right chunk]'.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analogical Reasoning, Re-rank Process and Final Answer Selection", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The main objectives of our experiments are twofold: on one hand, we try to find the importance level of the feature set that we use. On the other hand, we evaluate the potential of our approach to locate factoid answers in snippets and document retrieval scenarios without using any NER-tool. For the second objective we run two kinds of experiments. The first one is by using the gold standard snippets and the second one is by performing a document retrieval process.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments and Evaluation", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In our experiments we use the question answer pairs from CLEF 1 English monolingual of the year 2006, 2007 and 2008. For the training data we use the 2007 and 2008 collections. In total it consists of 321 factoid question answer pairs. For the testing data we use the 2006 collection, consisting of 75 factoid questions (Magnini et al., 2006) .", | |
| "cite_spans": [ | |
| { | |
| "start": 96, | |
| "end": 116, | |
| "text": "2006, 2007 and 2008.", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 320, | |
| "end": 342, | |
| "text": "(Magnini et al., 2006)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments and Evaluation", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In our empirical experiments, by performing chisquare statistics, we find that the answer chunk is the most important feature. The right-chunk of an answer is the least significant feature. The complete weighting factors of the feature set can be seen in Table 2 . In our experiments, we also add the EAT parameter as one of the factors which will be important in the re-ranking process.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 255, | |
| "end": 262, | |
| "text": "Table 2", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experiments and Evaluation", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We use the accuracy metric during evaluation (Schlaefer et al., 2006) (Pe\u00f1as et al., 2010) , which covers the proportion the number of questions correctly answered in the test set. We choose this kind of evaluation because we are interested in the potential of our approach to predict the location of an answer in a given snippet / document.", | |
| "cite_spans": [ | |
| { | |
| "start": 45, | |
| "end": 69, | |
| "text": "(Schlaefer et al., 2006)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 70, | |
| "end": 90, | |
| "text": "(Pe\u00f1as et al., 2010)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments and Evaluation", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In this first experiment we assume that the IR process performed perfectly and returns the best snippet which covers the final answer. We choose Open Ephyra (Schlaefer et al., 2006) as our competing pipeline. This decision is based on the fact that Open Ephyra employs two types of NER integrated in it. The first type is the model-based NER which consists of OpenNLP 2 and Stanford NER 3 . The second type is a dictionary-based NER that was specially design for TREC-QA competition.To maintain the fairness of the evaluation, we decided to only use the first type (model-based NER) and build a special trained answer-type classifier for CLEF datasets as described in Toba et al. (2010) . In short, we hold the QA components of our approach and those of Open Ephyra all the same, except for the final answer selection.", | |
| "cite_spans": [ | |
| { | |
| "start": 157, | |
| "end": 181, | |
| "text": "(Schlaefer et al., 2006)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 668, | |
| "end": 686, | |
| "text": "Toba et al. (2010)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Gold Standard Snippets", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "The result of this experiment can be seen in Table 3. Our approach outperforms the overall result of Open Ephyra. The best accuracy of our approach is achieved in the OTHER and LOCATION answertype, both 0.83, whereas the worst accuracy is for the TIME-typed questions, 0.45. In particular, our approach performs exceptionally well in the 'OTH-ER' type. We believe this is due to the fact that our strategy finds the location of an expected answer without depending on the performance of an NER-tool. An example of 'OTHER-typed' questions is (CLEF 2006 #8) : \"What is the Bavarian National Anthem?\". The expected answer for this question is: \"God be with you, land of Bavarians\". The answer chunk constituent in the gold standard snippet is a sequence of \"VP-NP-O\", which comes from the following snippet: \"They ended their demonstration by singing the Bavarian Anthem \"God be with you, land of Bavarians\". Then many of them moved on to support another Bavarian tradition -Oktoberfest.\". If we look deeper into the feature set which is used in the AR training in Table 1 , our trigram chunk features actually consist of two bigrams: (left+answer)chunk and (answer+right)-chunk. During the final answer selection we consider these left and right chunk-bigrams as part of the selection process, not only the trigram sequence. This strategy is to ensure that the answer could be covered in one of the possibilities: a chunk trigram, a chunk left-bigram, or a chunk right-bigram.", | |
| "cite_spans": [ | |
| { | |
| "start": 541, | |
| "end": 555, | |
| "text": "(CLEF 2006 #8)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 1062, | |
| "end": 1069, | |
| "text": "Table 1", | |
| "ref_id": "TABREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Gold Standard Snippets", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "In this first experiment, the most difficult questions to be answered are the TIME and MEASURE question-types. The answer of the TIME answertype can be in the form of: dd/mm/yy, dd-mmmyy, a single year number, or in the form of hh:mm a.m./p.m. These variations give rise to problems during the feature extraction process, because sometimes the chunker recognizes variations as numbers or as nouns. This problem also occurred in the MEASURE-typed questions. A measurement can be written as numbers (for example: \"40\") or as text (\"forty\"), and the chunker recognizes them differ-ently, even though they express the same thing. Figure 3(a) gives the number of expected \"AR trigram sequences\" in each answer-type which needs to be found in the snippets. We can see that for a factoid question answering task, the expected answers are mostly in the form of an NP (noun phrase).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 626, | |
| "end": 637, | |
| "text": "Figure 3(a)", | |
| "ref_id": "FIGREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Gold Standard Snippets", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "In our second experimental setting, we try to simulate our approach in a more realistic question answering system. In the real situation, we will not have any information about the semantic chunk of the final answer. We assume that the best pair (i.e. the top-1 pair after the re-ranking process) of the AR answer features will supply us with that information. We performed IR process by using Indri Search Engine to retrieve the top-5 documents and pass them on to Open Ephyra and our system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Indri Document Retrieval", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "In this experiment, we use the same AR feature set as in the first experiment during the training phase. However, unlike the first experiment, during the top-10 AR retrieval process, we only use the question feature set, i.e.: the question word and its chunk, and the question trigram of the semantic chunks. Due to the lack of the answer features, we need to adjust the way of the re-ranking process. We use a scoring function (sf ) which takes the AR and Indri retrieval results into consideration. We use the formula in equation 5. We adjust the weight of the parameters to fit the question features during the re-ranking process of the AR retrieval results. The result of this second experiment can be seen in Table 4 . Both the SAR approach and Open E- phyra have a lower accuracy compared to the first experiment. Once again, our approach achieves a higher accuracy. In the NER-based system, the errors are mainly caused by the model in the NER tools which cannot find the appropriate answer. For example for a person name \"Carl Lewis\", the NER tools can only recognize it either as Carl or Lewis, but not the whole name. We classify the error types of our approach in three groups, i.e.: (1) not covered by Indri retrieval, (2) decreasing rank of relevant document because of the AR re-ranking score function, and (3) irrelevant example from the best AR pair. The frequency of these error groups can be seen in Table 5 . In our opinion, the main drawback of our approach is that it suffers from the variations of sentence structures -those of the snippets in the training set and those of the retrieved documents. These variations influence the AR re-ranking and matching process of chunk sequences. For instance, if the AR best pair suggests that the answer should be located at the end of a sentence, while that chunk could not be found in the retrieved document, then we will have a negative result. An example of such case can be seen in Table 6 . The complete occurrences of the expected trigram sequences in this second experiment can be found in Figure 3 ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 714, | |
| "end": 721, | |
| "text": "Table 4", | |
| "ref_id": "TABREF6" | |
| }, | |
| { | |
| "start": 1418, | |
| "end": 1425, | |
| "text": "Table 5", | |
| "ref_id": "TABREF8" | |
| }, | |
| { | |
| "start": 1949, | |
| "end": 1956, | |
| "text": "Table 6", | |
| "ref_id": "TABREF9" | |
| }, | |
| { | |
| "start": 2060, | |
| "end": 2068, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Indri Document Retrieval", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "In this paper we have shown that by learning analogical linkages of question-answer pairs we can predict the location of factoid answers of a given snippet or document. Our approach achieves a very good accuracy in the OTHER answer-type (cf. Section 4.1). It shows the potential of our approach for dealing with an answer-type with no available corresponding NER tool.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion and Future Work", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Another finding in our experiments is that there is no trigram answer chunk sequence that really dominates in each answer-type. This suggests that each question depends on the sentence structure of a given snippet, and has a different way to be answered. This fact also suggests that our approach could suffer from the variations of the sentence structures.In our opinion, this is one of the reasons why the accuracy drops when the AR retrieval does not guarantee the occurrence of an answer (cf. Section 4.2). However, our approach has achieved a higher accuracy than a pure NER-based question answering system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion and Future Work", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "For our future work, we plan to develop a hybrid method of our approach with NER-based methods on larger and different datasets with more answertype variations. We also plan to conduct another research in which we consider the trained question answer pairs as a kind of rule set. In this sense we look forward to combining the statistical approach, i.e. the analogical framework, and the semantic approach, i.e. the knowledge (rule) acquisition from the trained question answer pairs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion and Future Work", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Question Answering at Cross Language Evaluation Forum (http://celct.fbk.eu/ResPubliQA/index.php)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://opennlp.apache.org 3 http://nlp.stanford.edu/ner/index.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Linguistic Kernels for Answer Re-ranking in Question Answering Systems. Information Processing and Management", | |
| "authors": [ | |
| { | |
| "first": "Alessandro", | |
| "middle": [], | |
| "last": "Moschitti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Silvia", | |
| "middle": [], | |
| "last": "Quarteroni", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "", | |
| "volume": "47", | |
| "issue": "", | |
| "pages": "825--842", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Alessandro Moschitti and Silvia Quarteroni. 2011. Lin- guistic Kernels for Answer Re-ranking in Question Answering Systems. Information Processing and Management, 47:825-842.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Overview of ResPubliQA 2010: Question Answering Evaluation over European Legislation. CLEF Notebook Papers", | |
| "authors": [ | |
| { | |
| "first": "Anselmo", | |
| "middle": [], | |
| "last": "Pe\u00f1as", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Pamela", | |
| "middle": [], | |
| "last": "Forner", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alvaro", | |
| "middle": [], | |
| "last": "Rodrigo", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Richard", | |
| "middle": [], | |
| "last": "Sutcliffe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Corina", | |
| "middle": [], | |
| "last": "Forascu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Christina", | |
| "middle": [], | |
| "last": "Mota", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Anselmo Pe\u00f1as, Pamela Forner, Alvaro Rodrigo, Richard Sutcliffe, Corina Forascu, and Christina Mota. 2010. Overview of ResPubliQA 2010: Question Answering Evaluation over European Legislation. CLEF Note- book Papers.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Overview of the CLEF 2006 Multilingual Question Answering Track. CLEF Question Answering Working Notes", | |
| "authors": [ | |
| { | |
| "first": "Bernardo", | |
| "middle": [], | |
| "last": "Magnini", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Danilo", | |
| "middle": [], | |
| "last": "Giampiccolo", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Pamela", | |
| "middle": [], | |
| "last": "Forner", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Christelle", | |
| "middle": [], | |
| "last": "Ayache", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Valentin", | |
| "middle": [], | |
| "last": "Jijkoun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Petya", | |
| "middle": [], | |
| "last": "Osenova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anselmo", | |
| "middle": [], | |
| "last": "Pe\u00f1as", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Paulo", | |
| "middle": [], | |
| "last": "Rocha", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bogdan", | |
| "middle": [], | |
| "last": "Sacaleanu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Richard", | |
| "middle": [], | |
| "last": "Sutcliffe", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bernardo Magnini, Danilo Giampiccolo, Pamela Forner, Christelle Ayache, Valentin Jijkoun, Petya Osenova, Anselmo Pe\u00f1as, Paulo Rocha, Bogdan Sacaleanu, and Richard Sutcliffe. 2006. Overview of the CLEF 2006 Multilingual Question Answering Track. CLEF Ques- tion Answering Working Notes.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Introduction to Information Retrieval", | |
| "authors": [ | |
| { | |
| "first": "Christopher", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Manning", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Prabhakar", | |
| "middle": [], | |
| "last": "Raghavan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hinrich", | |
| "middle": [], | |
| "last": "Schtze", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Contextual Approach for Paragraph Selection in Question Answering Task. CLEF Notebook Papers", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Christopher D. Manning, Prabhakar Raghavan, and Hin- rich Schtze. 2008. Introduction to Information Re- trieval. Cambridge University Press, New York, USA. Hapnes Toba, Mirna Adriani, Ruli Manurung. 2010. Contextual Approach for Paragraph Selection in Ques- tion Answering Task. CLEF Notebook Papers.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Expected Answer Type Construction using Analogical Reasoning in a Question Answering Task", | |
| "authors": [ | |
| { | |
| "first": "Hapnes", | |
| "middle": [], | |
| "last": "Toba", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mirna", | |
| "middle": [], | |
| "last": "Adriani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ruli", | |
| "middle": [], | |
| "last": "Manurung", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proc. of ICACSIS", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Hapnes Toba, Mirna Adriani, Ruli Manurung. 2011. Ex- pected Answer Type Construction using Analogical Reasoning in a Question Answering Task. Proc. of ICACSIS.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Probabilistic Models for Answer-Ranking in Multilingual Question-Answering", | |
| "authors": [ | |
| { | |
| "first": "Jeongwoo", | |
| "middle": [], | |
| "last": "Ko", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Luo", | |
| "middle": [], | |
| "last": "Si", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Eric", | |
| "middle": [], | |
| "last": "Nyberg", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Teruko", | |
| "middle": [], | |
| "last": "Mitamura", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "ACM Trans. on Information Systems", | |
| "volume": "28", | |
| "issue": "3", | |
| "pages": "1--35", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jeongwoo Ko, Luo Si, Eric Nyberg, and Teruko Mitamu- ra. 2010. Probabilistic Models for Answer-Ranking in Multilingual Question-Answering. ACM Trans. on Information Systems, 28(3) article 16: 1-35.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "A Pattern Learning Approach to Question Answering within the Ephyra Framework", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Schlaefer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Gieselmann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Schaaf", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Waibel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "LNAI", | |
| "volume": "4188", | |
| "issue": "", | |
| "pages": "687--694", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Schlaefer, P. Gieselmann, T. Schaaf, and A. Waibel. 2006. A Pattern Learning Approach to Question An- swering within the Ephyra Framework. LNAI, 4188: 687-694. Springer, Heidelberg.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Combining Semantic Information in Question Answering System. Information Processing and Management", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Moreda", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Llorens", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Saquete", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Palomar", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1016/j.ipm.2010.03.008" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P. Moreda, H. Llorens, E. Saquete, and M. Palomar. 2010. Combining Semantic Information in Question Answering System. Information Processing and Man- agement, doi: 10.1016/j.ipm.2010.03.008.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Analogical Reasoning with Relational Bayesian-sets", | |
| "authors": [ | |
| { | |
| "first": "Ricardo", | |
| "middle": [], | |
| "last": "Silva", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Katherine", | |
| "middle": [], | |
| "last": "Heller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Zoubin", | |
| "middle": [], | |
| "last": "Ghahramani", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proc. of AISTATS", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ricardo Silva, Katherine Heller, and Zoubin Ghahra- mani. 2007. Analogical Reasoning with Relational Bayesian-sets. Proc. of AISTATS.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Ranking Relations Using Analogies in Biological and Information Networks", | |
| "authors": [ | |
| { | |
| "first": "Ricardo", | |
| "middle": [], | |
| "last": "Silva", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Katherine", | |
| "middle": [], | |
| "last": "Heller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Zoubin", | |
| "middle": [], | |
| "last": "Ghahramani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Eduardo", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Airoldi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "The Annals of Applied Statistics", | |
| "volume": "4", | |
| "issue": "2", | |
| "pages": "615--644", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ricardo Silva, Katherine Heller, Zoubin Ghahramani, and Eduardo M. Airoldi. 2010. Ranking Relations Us- ing Analogies in Biological and Information Network- s. The Annals of Applied Statistics, 4(2):615-644.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Dependency Relation Matching for Answer Selection", | |
| "authors": [ | |
| { | |
| "first": "Renxu", | |
| "middle": [], | |
| "last": "Sun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hang", | |
| "middle": [], | |
| "last": "Cui", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Keya", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Min-Yen", | |
| "middle": [], | |
| "last": "Kan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tat-Seng", | |
| "middle": [], | |
| "last": "Chua", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proc. of SIGIR", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Renxu Sun, Hang Cui, Keya Li, Min-Yen Kan, and Tat- Seng Chua. 2005. Dependency Relation Matching for Answer Selection. Proc. of SIGIR.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Ranking Community Answers by Modeling Question-Answer Relationship via Analogical Reasoning", | |
| "authors": [ | |
| { | |
| "first": "X-J", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Tu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Feng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proc. of SIGIR", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "X-J. Wang, X. Tu, D. Feng, and L. Zhang. 2009. Ranking Community Answers by Modeling Question- Answer Relationship via Analogical Reasoning. Proc. of SIGIR.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "Idea of Semantic Analogical Reasoning framework from" | |
| }, | |
| "FIGREF1": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "and syntactic chunk of the final answer. Example: NP-NNP Trigram of syntactic chunk sequence which appears in the question. Example: PP-NP-PP, VP-NP-VP Trigram of the final answer, [leftanswer -right chunk] (during training). Example: PP-NP-VP. Trigram chunk sequences of the whole answer passage (during testing)" | |
| }, | |
| "FIGREF2": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "sf = {\u03b1OV (ai, aj) + \u03b2OV (bi, bj)} * log(AR) * IR * overlap expected answer type (0.08) AR = score of the AR retrieval (see eq. 4) IR = the weight of the Indri top-5 retrieval rank (5 to 1) OV(x,y)= 1, if there is overlap between x and y in a question i and its analogy pair j, otherwise 0" | |
| }, | |
| "FIGREF3": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "Number of Chunk Trigram Sequences in each EAT: (a) Snippet Experiment (b) Indri Retrieval Experiment" | |
| }, | |
| "TABREF0": { | |
| "text": "Question Answer Semantic Features used in the Analogical Reasoning", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table/>" | |
| }, | |
| "TABREF2": { | |
| "text": "Weighting Factors of the Feature Sets. 'Expected Answer Type' is not part of the extracted and learnt feature set in the AR model. The information about EAT during the experiments is provided by the gold standard.", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table/>" | |
| }, | |
| "TABREF4": { | |
| "text": "Gold Standard Experiment Accuracy", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table/>" | |
| }, | |
| "TABREF6": { | |
| "text": "Indri Retrieval Experiment Accuracy", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table/>" | |
| }, | |
| "TABREF8": { | |
| "text": "Frequency of Error Classification of SAR Approach", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table/>" | |
| }, | |
| "TABREF9": { | |
| "text": "Influence of Sentence Structures", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table/>" | |
| } | |
| } | |
| } | |
| } |