When the disaster occurs, the social such as Twitter is increasingly being helping direct rescue operati.onTshis article describes the methods we used Firien201t7h.e We regarded tdhiestinction onfeed-tweets and availability-tweets as classicfaition tasks, and the logistic regressionand Support Vector Machine are used to decidetypethe of the tw.eeItns the need and availability matching, we regard information retrievaltask, using thretrieval model to complete the task.
2. Task 1 2.1 Problem Description 2.2 Classifier Models
For Task w1,e need to judge availability-tweets and othe.rs We classification problem.We use the distinguish the useful -tnweedets and the
th-etweentese,d deem it as classifiers availa-bility to
In this working note, two a is a binary model, which significant liner classifier in learning strategy is the interval groups adopt SVM, which is defined as the most the feature space. The maximization. Solving the problem into Convex dratQicua Programming Remove URL Problem. We chose the RBF kernel, which can map the samples no-nlinearly to a higher dimension. The RBFID core has fewer numerical complexity characteristics.
Ultimately two groups completed Task 1 with LibSVMR.emove stop word
The third group uses LRl. mEaocdhe data point of the LR model has an impact on the classification plane.
Its influence is far from its distance to the classificRaetimonove punctuation plane, and if the data dimension is high, the LR Rmemodoevle URL will match the parameters regularization method. In classification, the computation is minimal, and the Note: speed is breakneck, so the storage resources are low, so it is convenient to observe the probability scores o(1f, samples. 2.3 Feature Selection
For feature selection, we conduct experiments with words and -grnam (n = 2, 3, 4, 5, respectively. Th(e4, results show that we have the best effect of using as features. Table 1 shows the comparison of the results of each feature. For the sipnrgeprocfes tweet text, each team adopts different methods to preprocess the data. Table 2 specifically describes the differ3en.tTask 2 teams in the -treparetment differences exis(t√. indicates that this method is used, and × is not 3u.s1edP,roblem Description w(5o,rdsHLJIT201-7IRMIDIS_1_task2_2 test (6, HLJIT201-7IRMIDIS_1_task2_3
HLJIT201-7IRMIDIS_1_task2_1 √ 2_1 √ √ √ √ √ 2_2 √ √ √ √ = = = = = = √ 2_3 × √ √ √ 1_1 1_2 1_3 2_1 2_2 2_3
HLJIT201-7IRMIDIS_1_task1_1 (2,
HLJIT201-7IRMIDIS_1_task1_2 (3, HLJIT2017-IRMIDIS_1_task1_3
word
Pre 0.281 0.482 0.428 0.385 0.517
Re 0.245 0.254 0.245 0.2 0.281 1_1 √ √ √
Pre 0.477 0.671 0.641 0.604 0.709 1_2 √ √ √
Re 0.474 0.597 0.606 0.556 0.644 1_3 × √ √
HLJIT2017-IRMIDIS_1_task1_1 and HLJIT20-17 IRMIDIS_1_task1_2 is different in stop words list.
Task 2 requires that th-teweentesed match in Task 1 be searched by 1. TaNskee-dtweets is used as a query set Q. Availabi-ltiwtyeets can be used as a collection of documents D. We use statistical language models to solve the problem of Task 2. Language models are used to assess what kind of word sequences are more typical according to language usages, and inject the right bias accordingly into the system to prefer an output sequence of words with high probability according to the language model. If a document language model gives the query a high probability, the query words must have top opportunities according to the document language model, which further means that the query words frequently occur in the document.
The correlation calculation can be expressed as shown Fingure ,1 using the -NTeweedets (Na,s the query set Q, the Avai-ltawbeileitsy (A, as document set D, and then the correlation obtain the correlation R (Q, D,.
briefly the calculation to N
Q
D
A 3.3 Language Model 4. Experimental Result
At released tweets Later, a the and
start theof track, (training set,, along about with 20,000
tweets a samp-le of availabi-ltiwtyeets in these K20 tweets. new set of 50,000 tweets released (test set,. 4.2 Evaluation index number of extracted messages
= number of correct mesrsaacgteesd e/xt Recall = the number of correct messages extracted the number of messages in the sample − = ( 2 + 1) 2 + β is the parameteri,s the precision, ainsd the recall.
Map： ̂ = ̂ =
( | ) ( ) ∫⊖ ( | ′) ( ′) ′ probability
of the provide does extra
value not appear in the SVM LR word to Task precision
2 but other
the niened Task threshold.
/ ability ID
Map
Average MAP 1_1 0.550 1_2 0.100 1_3 0.760 4.4 Experimental results 2_1 2_2 2_3
0.021 0.217 0.147
Through the above experimerenstualts, that the classification and sorting some informal occasions are very the future experiments, we will machine learning and try to features, to filter andose chtoext occasions.
we found of text content for regarding words. In deepen the study of select more different content of informal Acknowledgments This work is supported by Philosophy and Social Science Planning Project of Heilongjiang Province, China (No. 16EDD05)
A.