New Frontiers in Artificial Intelligence

Legal documents are typically hierarchically structured. This paper focuses on ordinances and rules (OR documents for short) in the local governments, which are designed for social lives under the governments. OR documents are composed of provisions for social lives in various aspects such as healthy development of youths and landscape preservation. OR documents in different local governments share common provisions but also include different provisions depending on their social situations. There is a large demand on helping governmental officers draft OR documents, especially searching “relevant parts” of OR documents. To help drafting OR documents, this paper designs the relevancy of OR documents with two basic measurements; matching ratio and provision commonality. Based on the relevancy, this paper develops a structured document search algorithm for OR documents. Experimental evaluation on real OR documents in Japan demonstrates that the proposed algorithm successfully discovers relevant parts of OR documents.

PatentTransformer is our codename for patent text generation based on Transformer-based models. Our long-term goal of patent claim generation is to realize “augmented inventing” for inventors by leveraging new Deep Learning techniques. We envision the possibility of building an “auto-complete” function for inventors to conceive better inventions in the era of artificial intelligence. In order to generate patent claims with reasonable quality, a fundamental question is how to measure the quality. In PatentTransformer-1.5, we tackle the problem from the perspective of claim span relevancy as a proof of concept. Patent claim language was rarely explored in the NLP field. In this work, we propose a span-based approach and a generic framework to measure patent claim generation quantitatively. In order to study the effectiveness of patent claim generation, we define a metric to measure whether two consecutive spans in a generated patent claims are relevant. We treat such relevancy measurement as a span-pair classification problem, following the concept of natural language inference. Technically, the span-pair classifier is implemented by fine-tuning a pre-trained language model. The patent claim generation is implemented by fine-tuning the other pre-trained model. Specifically, we fine-tune a pre-trained Google BERT model to measure the patent claim spans generated by a fine-tuned OpenAI GPT-2 model. In this way, we re-use two of the state-of-the-art pre-trained models in the NLP field. Our result shows the effectiveness of the span-pair classifier after fine-tuning the pre-trained model. It further validates the quantitative metric of span relevancy in patent claim generation. Particularly, we found that the span relevancy ratio measured by BERT becomes lower when the diversity in GPT-2 text generation becomes higher.

We summarize the evaluation of the 6th Competition on Legal Information Extraction/Entailment (COLIEE 2019). The competition consists of four tasks: two on case law and two on statute law. The case law component includes an information retrieval task (Task 1), and the confirmation of an entailment relation between an existing case and an unseen case (Task 2). The statute law component also includes an information retrieval task (Task 3) and an entailment/question answering task (Task 4), which attempts to confirm whether a particular statute applies to a yes/no question. Participation was open to any group in the world, based on any approach. Eleven different teams participated in the case law competition tasks, some of them in more than one task. We received results from 7 teams for Task 1 (15 runs) and 7 teams for Task 2 (18 runs). For the statute law tasks, 8 different teams participated, some in more than one task. Seven teams submitted a total of 13 runs for Task 3, and 7 teams submitted a total of 15 runs for Task 4. Here we summarize each team’s approaches, our official evaluation, and some analysis of the variety of methods that produced the evaluation results.

Extracting and formalising legal norms from legal documents is a time-consuming and complex procedure. Therefore, the automatic methods that can accelerate this process are in high demand. In this paper, we address two major questions related to this problem: (i) what are the challenges in formalising legal documents into a machine understandable formalism? (ii) to what extent can the data-driven state-of-the-art approaches developed in the Natural Language Processing (NLP) community be used to automate the normative mining process. The results of our experiments indicate that NLP technologies such as relation extraction and semantic parsing are promising research avenues to advance research in this area.

Many studies have been undertaken with machine learning techniques to predict stock returns in terms of time-series prediction. However, from the viewpoint of the cross-sectional prediction with machine learning techniques, there are no examples that verify its profitability in regional and global stock markets. This paper implements deep learning for multi-factor models to predict stock returns in the cross-section in these stock markets and investigates the performance of the method. Our results show that deep neural networks generally outperform representative machine learning models all over the world. These results indicate that deep learning shows promise as a skillful machine learning method to predict stock returns in the cross-section.

Although deep learning performs quite well, it has significant disadvantages such as a lack of transparency and limitations to the interpretability of the prediction. Then, we present the application of layer-wise relevance propagation (LRP) to decompose attributes of the predicted return. By applying LRP to each stock and averaging them in a portfolio, we can determine which factor contributes to prediction. We illustrate which factor contributes to prediction in regional and global stock markets.

News articles have great impacts on asset prices in the financial markets. Many attempts have been reported to ascertain how news influences stock prices. Stock price fluctuations of highly influential companies can have a major impact on the economy as a whole. In particular, the automobile industry is a colossal industry that leads the Japanese industry. However, the limitations in the number of available data sets usually become the hurdle for the model accuracy. In this study, we constructed a news evaluation model utilizing GPT-2. A news evaluation model is a model that evaluates news articles distributed to financial markets based on price fluctuation rates and predicts fluctuations in stock prices. We have added news articles generated by GPT-2 as data for analysis. Besides, we used a co-occurrence network analysis to review the overview of the news articles. News articles were classified through Long Short-Term Memory (LSTM). The results showed that the accuracy of the news evaluation model improved by generating news articles using a language generation model through GPT-2. More detailed analyses are planned for the future.

The purpose of this study is to investigate relationships between performances of start-ups and external supports to them. For the analysis, we used a questionnaire survey with responses from 2,897 start-ups (as of the 1st survey). As the analysis means, we adopted propensity score matching, which is one of the causal inference methods. First, after categorizing external supports and performances, and examining the effect of each support, only consulting information support was found to contribute to some performance improvement. Next, when we examined some external supports in combination with other supports, new causal effects such as significant growth in sales were confirmed by combining non-public and public funding supports. These results suggest the potential for contributing to the performances by the proper combination of these supports, considering the characteristics of each type of external support.

Certain attitude verbs in Korean such as al- and gieogha- (standardly translated as ‘know’ and ‘remember’, respectively) may give rise to veridicality inferences, i.e., inferences that their propositional complements are true. These inferences arise systematically, but selectively. In particular, they arise only under certain prosody. When they do arise, they project through various entailment-canceling operators and are understood to be backgrounded, suggesting that they are presuppositional in nature. I characterize these patterns as prosodically-conditioned factivity inferences. I propose an analysis that can capture this systematic variation in factivity, which crucially occurs below the level of projection (i.e., variation within ‘local contexts’). The analysis is in the vein of Abusch (2010) and Simons et al. (2017), in that it makes use of a general pragmatic reasoning process involving alternatives. I argue that asymmetries in meaning between the positive verbs (al- ‘know’, gieokha- ‘remember’) and their negative suppletive counterparts (moreu- ‘not know’, ggameok- ‘forget’) play an important role in deriving the prosodically-conditioned factivity inferences. In connection with this claim, I propose a new pragmatic principle that governs how alternatives come into contrast with each other. Via the activation of this principle, interpretations of verbs that are presuppositionally underspecified can obtain factive interpretations whenever their contrasting factive alternatives are activated.

In studies of linguistic meaning, it is often assumed that the relevant expressions exhibit many semantic types: <e> for entity denoters; <t> for truth-evaluable sentences; and the non-basic types <α, β> such that <α> and <β> are types. Expressions of a type <α, β>—e.g., <e, t> or <<e, t>, <<e, t>, t>—are said to signify functions, from things of the sort associated with expressions of type <α> to things of the sort associated with expressions of type <β>. On this view, children acquire languages that are importantly like the language that Frege invented to study the foundations of arithmetic. I think this conception of human linguistic meaning overgenerates wildly, even distinguishing—as we should—competence from performance. I sketch an alternative, defended elsewhere, to illustrate a broader point: when offering theories of natural languages, we shouldn’t be surprised if vocabulary designed for other purposes is inadequate, and attention to relevant phenomena motivates a spare semantic typology.

This paper compares two systems of functional type logic that have been applied to the analysis of meaning composition in natural language: Montague’s Intensional Logic IL and its extensional substratum Ty2 of two-sorted type theory. The two systems differ in their treatment of reference and quantification over indices (like possible worlds or times): whereas the denotations of IL-formulae (inter alia) depend on indices as parameters, their Ty2-counterparts contain explicit free and bound variables for them. Building on earlier results, it is argued that, appearances to the contrary, the two systems are largely equivalent; that any differences in expressivity are irrelevant to said applications; and that the equivalence also extends to variations of the systems that make use of multiple indices (as in mixed systems of modal and temporal interpretation) or additional dimensions (as in standard accounts of context dependence).

This paper describes combining the parsing of Stanford CoreNLP with the transformations of Treebank Semantics to realise a system for taking raw text as input to reach logical representation output. The analysis converts tree content into the structures of a formal language which is then processed against a locally constrained global calculation. The calculation resolves the interpretation, including the accessibility of antecedents for pronouns and definites, but, most in particular, it is the system that decides predicate valency through self-regulation of the calculation to ensure results from a minimum of explicit input.

This paper proposes a new composition rule for discourse particles and prosodic morphemes that paratactically-associate with the main text. Furthermore, the data and analyses support the framework of inquisitive semantics since the morphemes at issue often embed both declarative and interrogative clauses.

This paper takes an initial step to boil down the notion of (anti-)locality (or Binding Conditions A and B) of anaphoric expressions into semantic and morphological formalisms. Inspired by [15], this paper proposes a substructure internal to NPs which mirrors the verb cartography. This mirroring tree constitutes a stacked continuation which takes all the verbal heads one by one from the bottom to the root of the sentence. The tree can then be broken down to a part of a covert reflexivizer and the remaining dummy parts, the latter being filled with overt anaphoric morphemes. This treatment of anaphoric NPs enables a compositional analysis of complex anaphors (which tend to be local) and is subsumed under an established semantics and anti-lexalistic morphology. This paper also discusses more complex cases of multiple anaphors, verbal syncretism and non-c-commanding antecedents.

The idea that wh-phrases can quantify over generalized quantifiers emerged following two main observations: (i) disjunctive answers to modalized questions lead to free choice inferences if the wh-phrases’s restrictor is plural and (ii) questions with collective predicates do not lead to uniqueness presuppositions. Such proposals, however, fail to derive the connection between (i-ii) and plurality. We propose a novel analysis in which (i-ii) are derived via the presence of an existential distributivity operator. By tying these phenomena to distributivity, our analysis is able to establish the desired connection to plurality.

In this paper, we give a groundwork for the foundations of the semantic concept of logical consequence. We first give an opinionated survey of recent discussions on the model-theoretic concept, in particular Etchemendy’s criticisms and responses, alluding to Kreisel’s squeezing argument. We then present a view that in a sense the semantic concept of logical consequence irreducibly depends on the meaning of logical expressions but in another sense the extensional adequacy of the semantic account of first-order logical consequence is also of fundamental importance. We further point out a connection with proof-theoretic semantics.

This paper gives a compositional semantics for attitude reports with nominal, gerund, and that-clause complements that captures the intuitive entailment relations between these reports (e.g. Ida sees/imagines a penguin diving \(\Rightarrow \) Ida sees/imagines a penguin). These relations are identified through the familiar diagnostic tests. We observe that entailments that are licensed by counterfactual attitude verbs (here: imagine) are largely different from the entailments between veridical vision reports that are described in (Barwise 1981). To capture this difference, we give a non-clausal syntax for gerund attitude reports and assign factive clausal complements a different semantics from non-factive and gerund complements. The resulting account captures the entailment patterns of imagination and vision reports without assuming special axioms in the lexical semantics of see or imagine. On our account, the ‘logic’ of the above reports thus falls directly out of their semantics.

In this paper, I propose an analysis that covers both the wide scope or reading of the either/or construction and the availability of Alternative Question and Yes/No Question readings, namely a hybrid of an ellipsis analysis and a choice function analysis of either. After presenting two sets of data, I introduce two hybrid analyses that combine an ellipsis analysis and a choice function analysis. The two differ from each other in terms of the item that introduces the choice function variable: in the first analysis, the disjunction particle or introduces the choice function variable while in the second analysis, either has that semantic role. It is demonstrated that the two analyses both account for the either/or construction data, whereas only the second hybrid analysis, in which either introduces the choice function variable, explains the Alternative Question and Yes/No Question data. Finally, I review another account proposed in previous research, namely the focus alternative semantics analysis, and point out its problems.

The Japanese modals hazu and beki respectively correspond to alleged weak epistemic and deontic necessity readings of English ought. I propose a novel analysis of weak necessity as generic as opposed to the individual modality of both strong necessity and possibility modals, using ingredients from extant analysis of modality in a possible-world framework. On my view, the modal flavor of hazu is anticipative, replacing the epistemic modal base with a circumstantial one, that of beki ethic, differing from deontic modality in that its replaces individuated with idealistic norms. Both share the absence of agent-variables in the conversational background, which distinguishes them from individual modals. I conclude that the view from Japanese with its articulated and unambiguous modal inventory explains the strong/weak necessity distinction better than extant analyses.

This paper investigates the ambiguity of tense in the Japanese mirative sentence with nante/towa. Unlike an English sentence exclamative (e.g., (Wow), John won the race!), a Japanese sentence with nante/towa has a property of ambiguity with regard to tense. When nante or towa is combined with a proposition that contains the so-called non-past form ru, the sentence can be ambiguous between a non-past (future/present) reading and a past reading. This fact is surprising because the non-past form ru can never be used for describing a past event. We argue that the ambiguous interpretation of nante/towa comes from the conventional implicature of nante/towa. Unlike an English sentence exclamation (Rett 2011), the Japanese nante/towa takes a “tenseless” proposition p (i.e., ru does not specify a tense) and conventionally implies that (i) p is settled (i.e., p is/was true or predicted to be true) and (ii) the speaker had not expected that p. We will also consider the case where p + nante/towa is embedded under a surprising predicate and claim that both the embedded and non-embedded nante/towa can be analyzed in a uniform way, suggesting that the embedded nante/towa clause is an instance of a main clause phenomenon (rather than a relative tense phenomenon).

This paper investigates the semantics and pragmatics of the Japanese equative marker hodo, which has the interesting property that it patterns as a negative polarity item on some but not all of its uses. We argue that the distributional patterns characterizing hodo derive from its weak existential semantics, which result in a trivial meaning in certain configurations. We further propose a pragmatic account of the presuppositional effects found with hodo, and discuss potential extensions to other data in Japanese and beyond. Overall, our findings add to other recent work demonstrating that the presence or absence of maximality represents an important dimension of cross-linguistic variation in the semantics of equative constructions.

The relation between a sentence type and an illocutionary force is ‘one-to-many’ but not ‘one-to-any.’ The goal of this paper is to provide a formal theory capable of describing this association. The primary data for this study comes from Japanese imperatives. In this language, the illocutionary force of an imperative sentence is determined by the interaction between high-applicatives and subject-honorific markers. Inheriting important insights from Portner et al. (2019), this paper develops the idea that all of these constructions are involved in the process of determining authority among the discourse participants. Integrating the Optimality Theory into Dynamic Pragmatics, I propose that there are pragmatic constraints which are relevant in determining (i) who is in authority and (ii) what illocutionary force is appropriate for a given sentence, before we update the structured discourse context.

This paper contributes to the understanding of Japanese yori-comparatives, focusing on the alleged lack of negative island effect in this language. The lack of the effect has been taken to be one piece of evidence for the negative setting of the Degree Abstraction Parameter [1]. We argue that this parameter setting does not explain the whole picture of the negative island effect in Japanese comparatives and advocate a more traditional analysis that utilizes the maximality operator.

The field of Neural Style Transfer (NST) has led to interesting applications that enables to transform the reality as human beings perceive. Particularly, NST for material translation aims to change the material (texture) of an object to a different material from a desired image. In order to generate more realistic results, in this paper, we propose a partial texture style transfer method by combining NST with semantic segmentation. The original NST algorithm changes the style of an entire image including the style of background even though the texture is contained only in object regions. Therefore, we segment target objects using a weakly supervised segmentation method, and transfer the material of the style image to only material-based segmented areas. As a result, we achieved partial style transfer for only specific object regions, which enables us to change materials of objects in a given image as we like. Furthermore, we analyze the material translation capability of state-of-the-art image-to-image (I2I) translation algorithms, including the conventional NST method of Gatys, WCT, StarGAN, MUNIT, and DRIT++. The analysis of our experimental results suggests that the conventional NST produces more realistic results than other I2I translation methods. Moreover, there are certain materials that are easier to synthesize than others.

Our impression can be effectively delivered by a color. In this paper we present a novel model for generating advertising copies using machine learning techniques. Unlike most of the previously reported advertising copy generators take specified keyword(s) which a user wants to embed in a copy, our proposed model takes one colored image as its input. We use the previously reported database that provides the potential color impression of a word for the purpose of selecting several words assumed to give a similar perceptual impression of the input image. We also use a deep neural network based binary classifier to extract appropriate words for advertising copies from an increased vocabulary. To output advertising copies of relatively natural expression out of the ones generated, we use a word embedding model of a shallow neural network called Skip-gram. The qualities of the advertising copies were evaluated by online survey and were compared with other copies generated by various models. As the result of the evaluation, our proposed model outperformed the other models.