Hot Papers | emotion - reason for extraction: a new topic text sentiment analysis

1. Summary

    Emotional reasons extract (Emotion cause extraction, ECE) is a text extraction task underlying causes some emotion behind aimed, due to its wide range of applications, recently has attracted widespread attention. But there are two deficiencies: 1) ECE must be carried out before the extraction cause of the emotional mark, which greatly limits its application in the real scene; Method 2) First, the emotional tagging, and then extract the reason they are ignored another indication of this fact. This paper presents a new task: emotion - reason for extraction (emod -cause pair extraction, ECPE), its purpose is to extract the document and the corresponding underlying emotional reasons.

2. Introduction

    Figure 1 shows the difference between traditional and new ECPE ECE task task. In FIG 1 an example is given of emotion marked: "happy", ECE's goal is to cause the corresponding track two clauses: "a policeman visit the old man with the lost money" and "and told him that the thief was caught ". In ECPE task, the goal is the direct extraction of all the feelings reasons (emotion cause pair), including ( "the old man very happy", "a police visit elderly loss") and ( "the old man very happy", "Tell him, the thief was caught live "), does not offer emotional comments" happiness. "

Figure 1: An example showing the difference between the ECE task and the ECPE task.

    To address this new ECPE task, the paper proposes a two-step framework. Step 1 in two multi-task learning network will be converted to the emotional reasons for the extraction task two separate sub-tasks are extracted emotion (emotion extraction) and cause extraction (cause extraction). Step 2 pairing and emotional reasons for screening. All papers will be combined elements of two sets of pairs, the last train a filter that excludes causality does not contain the emotional reasons right.

3. Methods

Step 1: extraction of individual emotion and reason

    Step 1 is to target a group of emotion were extracted from a set of clauses and the reasons for each document clause. To this end, we propose two multi-task learning network. Independent multi-task learning and interactive multi-task learning. The latter is an enhanced version further captures the correlation between emotion and reason on the basis of the former.

1. The independent multi-task learning

    A document comprising a plurality of clauses: D = [C _{. 1} , C ₂ , ...], each C _I also includes a plurality of word C _I = [W _I1 , W _I2 , ...]. In order to capture such a "word-clause-document" structure, a layered Bi-LSTM network, which contains two layers, as shown in FIG.

Figure 2: The Model for Independent Multitask Learning (Indep).

    Layer comprised of a group of words Bi-LSTM level modules, each module corresponds to a clause, and accumulate information for each word in context clause. An upper layer comprising two parts: one for extracting emotion, another reason for extraction. Each component is a clause level Bi-LSTM, which receives the input layer of the clause c _i after coding word vector S _i . Implicitly states of the two portions of the Bi-LSTM R & lt ^E _I and R & lt ^C _I can be seen from the clause C _I emotional reasons and extracted, and finally fed back to the emotion softmax layer prediction and prediction reasons.

2. Interactive multi-task learning

    So far, two Bi-LSTM the upper component are independent. However, these two sub-tasks (extraction and emotional reasons extract) are not independent. On the one hand, provide emotional help better identify the reasons; on the other hand, it may help to understand why more accurately extract emotion.

    Structure shown in Figure 3. It should be noted that the use of emotional reasons extract to improve the extraction method is called Inter-EC. To enhance the emotional cause extraction using the extraction method is called Inter-CE. Since the Inter-EC and Inter-CE is similar in structure, therefore only introduce Inter-EC (FIG. 3 (a) below).

Figure 3: Two Models for Interactive Multi-task Learning: (a) Inter-EC, which uses emotion extraction to improve cause extraction (b) Inter-CE, which uses cause extraction to enhance emotion extraction.

    Compared with the independent multi-task learning, the underlying Inter-EC is unchanged, while the upper layer consists of two parts, which are used to extract the emotional task and interactively predict reason extracting task. Each component is a clause level Bi-LSTM, then a softmax layer.

Step 2: matching and filtering on affective

    In step 1, the finally obtained a set and a set of emotional reasons set. The second step is to target these two sets of pairs.

    First, the E (emotional clause set) and C (clause set reasons) a Cartesian product application to obtain the set of all possible pairs: Pall = {?????, (C ^E _I , C ^C _J ),} ?????

    Next, using a feature vector composed of three feature represented in each of the Pall:

wherein, S ^E _I and S ^C _I represent C ^E _I , C ^C _J through word vector encoding the result, V ^D represents C ^E _I and C ^C _J distance between.

    Logistic regression model is then trained, each candidate pair is detected (c ^E _I , C ^C _J ), C ^E _I and C ^C _J whether there is a causal relationship between:

Y (c ^E _I and C ^C _J ) =. 1 shows that (c ^E _I , C ^C _J ) there is a causal relationship, Y (C ^E _I and C ^C _J ) = 0 indicates that (C ^E _I , C ^C _J ) no causal relationship, and δ (·) is a Sigmoid function from Pall the removing Y (C ^E _I and C ^C _J ) = 0 in (C ^E _I , C ^C _J) Pair, the emotion-cause to give the final pair.

4 Conclusion

    Papers emotional reasons based on the reference data set (emotion cause dataset) (Gui et al., 2016a) evaluate the method, without the use of emotional comments on the test data. The final emotional - achieved 61.28% of the F1 score for the extraction of reasons. The experimental results show the feasibility and effectiveness of this method. In addition to the emotional - the reason for the extraction of the evaluation, the paper also evaluated two separate tasks (extraction and emotional reasons extract) performance. Compared with the traditional method of ECE emotion marked dependent elimination, this method has great advantages.

More interesting information scan code concern BBIT