1 Introduction

1The present work has been conducted within the Babylonian Talmud Translation Project (in Italian, Progetto Traduzione Talmud Babilonese - PTTB) which aims at the translation of the Babylonian Talmud (BT) into Italian.

2The translation is being carried out with the aid of tools for text and language processing integrated into an application, called Traduco (Bellandi et al., 2016), developed by the Institute of Computational Linguistics “Antonio Zampolli” of the CNR in collaboration with the PTTB team. Traduco is a collaborative computer-assisted translation (CAT) tool conceived to ease the translation, revision and editing of the BT.

3The research described here fits exactly in this context: we want to provide the system with additional informative elements as a further aid in the translation of the Talmud. In particular, we intend to linguistically analyze the Talmudic text starting from the automatic attribution of the Part-Of-Speech to words by adopting a stochastic POS tagging approach.

4The first difficulty that has emerged regards the text and the languages it contains. In this regard we can say, simplifying, that the Babylonian Talmud is essentially composed of two languages which, in turn, correspond to two distinct texts: the Mishna and the Gemara. The first is the oldest one written in mishnaic Hebrew, one of the most homogeneous and coherent languages appearing in the Talmud that, for this reason, has been chosen to start from in the POS tagging experiment.

5The main purpose of linguistic analysis in the context of our translation project is to improve the suggestions provided by the system through the so-called Translation Memory (TM).

6Moreover, on a linguistically annotated text it is possible to carry out linguistic-based searches, useful both for the scholar (in this case a talmudist), and, during the translation work, for the revisor and the curator, who have the possibility, for example, to make bulk editing of polysemous words by discarding out words with undesired POS.

7The rest of the paper is organized as follows: Section 2 summarizes the state of the art in NLP of Hebrew. The construction of the linguistically annotated corpus is described in Section 3. The training process and evaluation of the POS taggers used in the experiments is detailed in Section 4. Lastly, Section 5 outlines the next steps of the research.

2 State of the art

8The aforementioned linguistic richness and the intrinsic complexity of the Babylonian Talmud make automatic linguistic analysis of the BT particularly hard (Bellandi et al., 2015).

9However, some linguistic resources of ancient Hebrew and Aramaic have been (and are being) developed, among which we cite: i) the Hebrew Text Database (Van Peursen and Sikkel, 2014) (ETCBC) accessible by SHEBANQ¹ an online environment for the study of Biblical Hebrew (with emphasis on syntax), developed by the Eep Talstra Centre for Bible and Computer of the Vrije Universiteit in Amsterdam; ii) the Historical Dictionary² project of the Academy of the Hebrew Language of Israel; iii) the Comprehensive Aramaic Lexicon (CAL)³ developed by the Hebrew Union College of Cincinnati; iv) the Digital Mishna⁴ project, concerning the creation of a digital scholarly edition of the Mishna conducted by the Maryland Institute of Technology in the Humanities.

10Apart from the aforementioned resources, to date there are no available NLP tools suitable for the processing of ancient north-western Semitic languages, such as the different Aramaic idioms and the historical variants of Hebrew attested in the BT. The only existing projects and tools for the processing of Jewish languages (Kamir et al., 2002) (Cohen and Smith, 2007) have been developed for modern Hebrew, a language that has been artificially revitalized from the end of the XIX century and that does not correspond to the idioms recurring in the BT. Among them we cite HebTokenizer⁵ for tokenization, MILA (Barhaim et al., 2008), HebMorph⁶, MorphTagger ⁷ and NLPH⁸ for morphological analysis and lemmatization, yap⁹, hebdepparser¹⁰, UD_Hebrew¹¹ for syntactic analysis. We conducted some preliminary tests by starting with MILA’s (ambiguous) morphological analyzer applied to the three main languages of the Talmud:

1. Aramaic: Hebrew and Aramaic are different languages. There are even some cases in which the very same root has different semantics in the two languages. In addition, MILA did not recognize many aramaic roots, tagging the relative words, derived from them, as proper nouns.

2. Biblical Hebrew: MILA recognized most of the words, since Modern Hebrew preserved almost the entire biblical lexicon. However, syntax of Modern Hebrew is quite different from that of Biblical Hebrew, leading MILA to output wrong analyses.

3. Mishnaic Hebrew: this is the language where MILA performed better. Modern Hebrew inherits some of the morpho-syntactic features of mishnaic Hebrew, however, the two idioms differ substantially on the lexicon, since in modern Hebrew many archaic words have been lost (Skolnik and Berenbaum, 2007).

11In the light of the above, we decided to create a novel linguistically annotated resource to start developing our own tools for the processing of ancient Jewish languages. In the next section, we will describe how the resource was built.

3 Building the resource

12The linguistic annotation of Semitic languages poses several problems. Although we here discuss the analysis of Hebrew, many of the critical points that must be taken into account are common to other languages belonging to the same family. As already mentioned in the previous section, the first problem concerns the access to existing linguistic resources and analytical tools which, in the case of Hebrew, are available exclusively for the modern language.

13One of the major challenges posed by the morphological analysis of Semitic languages is the orthographic disambiguation of words. Since writing is almost exclusively consonantal, every word can have multiple readings. The problem of orthographic ambiguity, crucial in all studies on large corpora (typically in Hebrew and modern Arabic), does not prove to be so difficult when the text under examination is vocalized.

14The edition of the Talmud used in the project is actually vocalized and the text, consequently, is orthographically unambiguous. An additional critical aspect is represented by the definition of the tagset. Most of the computational studies on language analysis have been conducted on Indo-european languages (especially on English).

15As a result, it may be difficult to reuse tagsets created for these languages. Not surprisingly, there are still many discussions about how it is better to catalog some POS and each language has its own part under discussion. Each tagset must ultimately be created in the light of a specific purpose. For example, the tagging of the (Modern) Hebrew Treebank developed at the Technion (Sima’an et al., 2001) was syntax-oriented, while the work on participles of Hebrew described in (Adler et al., 2008) was more lexicon-oriented. We considered the idea of adopting the tagset used in the already cited Universal Dependency Corpus for Hebrew. However, its 16 tags appeared to be too “coarse grained" for our purposes.¹² In particular, the UD tagset lacks of all the prefix tags that we needed. For this reason we decided to define our own tagset.

16Once the tagset has been defined, it remains to decide which is the most suitable grammatical category to associate with each token. You can collect essentially two types of information, the problem is how and if you can keep both, in particular: i) the definition of the token from a syntagmatic perspective (i.e. what the token represents in context) and ii) the lexical information that the token gives by itself (without context). To give a couple of examples:

• Verb/noun: is “the one who makes a vow” or “the vowing”? (the one who consecrates his wife): should it be assigned to verb or noun category?

• Adjective/verb: is adjective or verb (given that most of the mishnaic language dictionaries provide both options)?

17We could discuss about which category would be the best for each and why, but, for now, we decided to keep both by introducing two parallel annotations, by “category” (without context) and by “function” (in context). The tagset we used for this work are the following: agg., avv., cong., interiez., nome pr., num. card., num. ord., pref. art., pref. cong., pref. prep., pref. pron. rel., prep., pron. dim., pron. indef., pron. interr., pron. pers., pron. suff., punt., sost., vb.

18One could also envisage the refining of the tagset by adding: interrogative, modal, negation, and quantifier (Adler, 2007) (Netzer and Elhadad, 1998) (Netzer et al., 2007).

19As anticipated, in order to build the morphologically annotated resource, all of the Mishna sentences were extracted from the Talmud and annotated using an ad hoc developed Web application (Fig. 1).

Figure 1: The interface for the linguistic annotation of the corpus to be used to train the POS tagger

20All the annotations have been made with the aim of training a stochastic POS tagger in charge of the automatic analysis of the entire Mishna: to obtain a good accuracy it was thus necessary to manually annotate as many sentences as possible. To date, 10442 tokens have been annotated.

21The software created for the annotation shows, in a tabular form, the information of the analysis carried out on a sentence by sentence basis.

22The system, once a sentence is selected for annotation, checks whether the tokens composing it have already been analyzed and, in case, calculates a possible subdivision into sub-tokens (i.e. the stems, prefixes and suffixes constituting each word) by exploiting previous annotations. If the system finds that a word is associated with multiple different annotations, it proposes the most frequent one.

23Regarding the linguistic annotation, the grammar of Pérez Fernández (Fernández and Elwolde, 1999) was adopted and, for lemmatization, the dictionary of M. Jastrow (Jastrow, 1971).

24The software allows to gather as much information as possible for each word by providing a double annotation: by “category” to represent the POS from a grammatical point of view, and by “function” to describe the function the word assumes in its context. For the POS tagging experiments, described below, we used the annotation made by “function”.

4 Training and testing of POS taggers

25Once the mishnaic corpus has been linguistically annotated three of the most used algorithms for POS tagging have been used and evaluated: HunPos (Halácsy et al., 2007), the Stanford Log-linear Part-Of-Speech Tagger (Toutanova et al., 2003), and TreeTagger (Schmid, 1994). The three algorithms implement supervised stochastic models and, consequently, they need to be trained with a manually annotated corpus.

26To evaluate the accuracy of the algorithms we adopted the strategy of k-fold cross validation (Brink et al., 2016), with k set to 10, and thus dividing the corpus in 10 partitions.

27Table 1 summarizes the results of the experiment by showing the tagging accuracy of the three tested algorithms. With a number of tokens slightly higher than ten thousands the Stanford POS tagger provided the best results over HunPos and Treetagger, with an accuracy of 87,9%.

5 Next steps

28In this work, the tagging experiments have been limited to the attribution of the Part-Of-Speech: the next, natural step, will be the addition of the lemma. Furthermore, we will try to modify the parameters affecting the behaviour of the three adopted POS taggers (left at their default values for the experiments) and see how they influence the results.

29Once the Mishna will be lemmatized, Traduco, the software used to translate the Talmud in Italian, will be able to exploit this additional information mainly to provide translators with translation suggestions based on lemmas, but also to allow users to query the mishnaic text by POS and lemma.

30As a further step we will also take into account the linguistic annotation of portions of the Babylonian Talmud written in other languages, starting from the Babylonian Aramaic, the language of the Gemara, which constitutes the earlier portion of the Talmud.

Acknowledgments

31This work was conducted in the context of the TALMUD project and the scientific cooperation between S.c.a r.l. PTTB and ILC-CNR.