Intelligent Matching System of Clauses in International Investment Arbitration Cases Based on Big Data Statistical Model

International investment arbitration has the advantages of flexibility, autonomy, efficiency, neutrality, and other advantages in various methods of resolving investor and host country treaty disputes. The acceptance of international arbitration to resolve investment disputes in bilateral investment treaties concluded by China. has dramatically increased, which seems to help Chinese enterprises to go global and bring in foreign investment. Many arbitration institutions exist in mainland China, but none have ever accepted arbitration between foreign investors and host countries. This is due to legislative reasons and international cooperation at the government level. The rapid spread of information in the Internet age has led to the emergence of news. At the same time, discussions on case news have also been launched on Weibo, WeChat, Zhihu, and other major social platforms [1]. Although news about legal cases emerges endlessly in new media, such news information lacks professional legal interpretation. Therefore, more and more scholars have paid attention to the research necessity of information search in the legal field.

On October 1, 2016, the “Regulations of the Supreme People's Court on People's Courts Publishing Judgments on the Internet” came into effect. As of August 30, 2020, the relevant search websites for judgment documents include China Judgment Documents Network, Peking University Magic Treasure Website, etc. All central-local courts have corresponding official websites for document inquiries. On the China Judgment Documents website alone, the total number of uploaded judgment documents has exceeded 100 million. Since then, this kind of judgment document retrieval system has provided users with many legal research cases. This provides data protection for basic legal case searches. Judgment documents record the process and results of the trial of the people's court, and it is the carrier of the results of the litigation activities. The content includes the public prosecution, the defendant, the plaintiff, the defender, the judiciary, the evidence information, the court's judgment basis, and the referenced laws and regulations. The practical value of adjudication documents is reflected in the fact that the adjudication documents have been reviewed many times, the words are used carefully, the language is well standardized, and the case explanation is detailed. It can provide non-professional users with customized services for pushing related cases. Users can obtain relevant judgment documents according to unique case conditions.

In this paper, news-like factual texts are used as queries. In this paper, the structured and standardized judgment documents are used as the full-text corpus. This paper generalizes the professional retrieval system by using the structure and content features of the judgment documents [2]. This meets the retrieval needs of non-professional users who lack legal knowledge. This paper proposes the “news-documents” automatic recommendation system framework. The framework can be implemented to recommend relevant judgment documents for news-like authentic texts.

This paper integrates the structure information and content information of judgment documents into the information retrieval system of international investment judgment documents. This paper uses the BM25 model to calculate the similarity between feature words and documents [3]. And use the BM25 values of different feature words in factual texts as multi-dimensional features in the similarity algorithm. The overall similarity between the factual text and the international investment arbitration instrument is obtained.

The BM25 model is an algorithm for evaluating the relevance between search terms and documents. Its formula expression is (1) simBM 25(dj, q)~∑ki(dj,q)Bi,j×lg(N−ni+0.5ni+0.5) {sim}_{BM\,25}\left({{d_j},\,q} \right)\sim\sum\limits_{{k_i}\left({{d_j},q} \right)} {{B_{i,j}} \times \lg \left({{{N - {n_i} + 0.5} \over {{n_i} + 0.5}}} \right)}

Where Bi,j=(K1+1)fi,jK1[(1−b)+b×len(dj)avgdoclen]+fi,j {B_{i,j}} = {{\left({{K_1} + 1} \right){f_{i,j}}} \over {{K_1}\left[{\left({1 - b} \right) + b \times {{len\left({{d_j}} \right)} \over {{avg}_{doclen}}}} \right] + {f_{i,j}}}} ; K₁ and b are empirically determined constants. f_i,j represents the frequency of index item i in document j. It refers specifically to feature words in this paper. Feature words in factual texts contain a lot of semantic information. This paper extracts feature words in factual texts to calculate the correlation between feature words and international investment arbitration documents. This achieves the purpose of assessing the overall relevance between the factual text and the international investment arbitration instrument.

This paper uses the SVM Rank and Lambda MART algorithms to fit the BM25 values of multi-dimensional feature words [4]. In this way, the overall similarity calculation between news corpus and international investment arbitration documents is realized.

Based on SVM, some scholars have proposed a ranking learning algorithm, SVM Rank. They transformed the ranking problem into a binary classification problem. The basic idea is to give a dataset {x_i, y_i}. where y_i ∈ {1, ⋯, R}. There exists a function h(x) that satisfies h(x_i) > h(x_j) ⇔ y_i > y_j. Therefore, we set the relevant international investment arbitration document set of factual text as {x_i, y_i}. The relevant international investment arbitration document association pair {x_i, y_i} and its correlation label y constitute training data x_i, y_i meaning that the international investment arbitration document and the factual text document are opposite relationships. Assuming m = |ρ|, the optimization problem of SVM Rank can be transformed into the mathematical form: (2) min w, ζi,j≥012wTw+Cm∑(i,j)∈ρξi,j \mathop {\min\,}\limits_{w,\,{\zeta _{i,j}} \ge 0} {1 \over 2}{w^T}w + {C \over m}\sum\limits_{\left({i,j} \right) \in \rho} {{\xi _{i,j}}} (3) s.t.∀(i,j)∈ρ:(wT xi)≥(wTxi)+1−ξi,j s.t.\forall \left({i,j} \right) \in \rho :\left({{w^T}\,{x_i}} \right) \ge \left({{w^T}{x_i}} \right) + 1 - {\xi _{i,j}}

We find a linear function h(x) such that the training corpus has a corresponding order. This is an orderly return. The algorithm can be integrated into the multi-dimensional vector represented by the BM25 value of the news corpus feature word. It effectively improves the calculation effect of document similarity [5]. Therefore, we can calculate the overall similarity between the feature words in the news corpus and the international investment arbitration documents.

The Lambda MART algorithm has achieved good results in information retrieval. Its algorithmic nature can be widely used in sorting tasks. The content includes advertisement recommendations, automatic scoring, etc. The basic idea of the RankNet algorithm is to provide a scoring function s_i = f(x_i). x_i represents the feature word vector extracted from the factual text. Then we calculate the probability value that IIA document i ranks ahead of IIA document j. Its calculation formula is P(Ui⊳Uj)=11+e−σ(si−sj) P\left({{U_i} \triangleright {U_j}} \right) = {1 \over {1 + {e^{- \sigma \left({{s_i} - {s_j}} \right)}}}} . Its loss function is C=−Pi,j¯lg(Pi,j)−(1−Pi,j¯)lg(1−Pi,j) C = - \overline {{P_{i,j}}} \lg \left({{P_{i,j}}} \right) - \left({1 - \overline {{P_{i,j}}}} \right)\lg \left({1 - {P_{i,j}}} \right) . Where Pi,j¯ \overline {{P_{i,j}}} refers to the actual probability that the International Investment Arbitration Instrument Document i precedes the International Investment Arbitration Instrument Document j. P_i,j is the predicted probability that the International Investment Arbitration Instrument Document i precedes the International Investment Arbitration Instrument Document j.

This paper proposes an automatic recommendation framework for international investment arbitration documents based on the characteristics of structure and content by using the characteristic words and structural content characteristics of international investment arbitration documents. The algorithm improves the shortcomings of the traditional full-text retrieval model BM25 in retrieving factual texts using a news-like corpus. On this basis, this paper uses the SVM Rank algorithm and the Lambda MART algorithm. This algorithm improves the effect of information retrieval based on factual texts of news-like corpora. This better fulfills the legal information retrieval needs of non-professional users.