Extracting Message Sequence Charts from Hindi Narrative Text

In this paper, we propose the use of Message Sequence Charts (MSC) as a representation for visualizing narrative text in Hindi. An MSC is a formal representation allowing the depiction of actors and interactions among these actors in a scenario, apart from supporting a rich framework for formal inference. We propose an approach to extract MSC actors and interactions from a Hindi narrative. As a part of the approach, we enrich an existing event annotation scheme where we provide guidelines for annotation of the mood of events (realis vs irrealis) and guidelines for annotation of event arguments. We report performance on multiple evaluation criteria by experimenting with Hindi narratives from Indian History. Though Hindi is the fourth most-spoken first language in the world, from the NLP perspective it has comparatively lesser resources than English. Moreover, there is relatively less work in the context of event processing in Hindi. Hence, we believe that this work is among the initial works for Hindi event processing.


Introduction
Narratives are used to communicate complex ideas, detailed accounts of complex events or arguments about one's beliefs (Valls Vargas, 2017). Moreover, a narrative is a powerful tool not just from entertainment perspective but is one of the core component of human memory, knowledge and intelligence (Schank and Morson, 1995). Narrative texts are common in History where they mostly describe events that have happened in the past. Narrative text can also be routinely seen in news articles reporting events that have happened (or about to happen) in various political, corporate and social walks of a country. For multiple applications in text analysis, it becomes important to * Work done during internship at TCS Pune.  understand and analyse such narratives. A narrative has two key aspects, plot also referred to as story and discourse, the way in which the plot is described (Chatman, 1975). In this paper we focus on visualization of the plot aspect of a Hindi narrative.
Hindi is an Indo-Aryan language spoken by around 300 million people in India. Additionally, Hindi is the fourth most-spoken first language in the world 1 . In comparison to English, Hindi has different linguistic characteristics leading to a different set of NLP challenges. First of all, Hindi is a Subject-Object-Verb (SOV) language with relatively free word order, as against the SVO order in English. Secondly, Hindi does not have high accuracy NLP toolkits such as Stanford CoreNLP (Manning et al., 2014). In this paper, we make one of the first attempts to facilitate event processing in Hindi by proposing annotation guidelines for events as well as their arguments.
We propose to represent a Hindi narrative using Message Sequence Charts (MSC) (Rudolph  Table 1 et al., 1996). MSC is a widely used notation to depict interactions among components of communicating devices. In an MSC, a vertical line represents a component and horizontal arrows indicate message from one component to another. We assume actors (or entities) in a narrative as components and message among the components as interactions among the corresponding actors. As an example, consider the text narrative from the Indian Independence Movement in Table 1 and the corresponding MSC in Figure 1.
We also address the problem of automatically extracting the MSC representation from a given narrative text. MSC extends the notion of a single event-timeline (Bedi et al., 2017) for a narrative by providing a timeline per actor (entity of interest). MSC representation captures all the actors and interactions in an easy to visualize manner and hence make the text more comprehensible. Further, the representation's support for inference mechanisms opens up possibilities of tackling natural language understanding problems like text comprehension and question answering. Our work is similar to (Palshikar et al., 2019) to represent a English narrative using MSC. However, due to intricacies of events in Hindi language, their approach cannot be used for construction of MSC of a Hindi narrative.
The key contributions of the paper are (i) we extend an existing scheme of annotation of events (Goud et al., 2019); we provide guidelines for annotation of mood of events (realis vs irrealis) and guidelines for annotation of event arguments, (ii) we propose an approach to identify event predicates and their arguments, (iii) MSC based visu-alization of actors and interactions, (iv) we report performance on multiple evaluation criteria by experimenting with four Hindi narratives from Indian History.

Key Annotation and Extraction Challenges for Events in Hindi
Annotation and extraction of events and their arguments from English texts is a challenging task (Mitamura et al., 2015). In case of Hindi there are more challenges as compared to English. Following are the key challenges we observed while processing Hindi narratives. I. Absence of annotation guidelines and labelled data similar to ACE: Only few attempts have been made to define comprehensive event annotation guidelines for Hindi. (Goud et al., 2019;Goel et al., 2020) propose a set of guidelines for annotation of event mentions in Hindi. However, they do not consider arguments of events, which are vital for narrative processing.
II. Annotation of events with Light Verb Constructs (LVCs): LVCs are formed from a commonly used verb and usually a noun phrase (NP) in its direct object position, such as have a look or take an action. For example, in the sentence º Ú ('Ram ne Kitab di'; Ram gave the book), Ú (di;give) is the event, however in the sentence º Ú ('Ram ne Kitab kho di'; Ram lost the book.), Ú ('kho di'; lost) is the event. Similarly, in the sentence º ('Ram se kitab gum ho gayee'; Ram lost the book), ('gum ho gayee'; lost) is the event. As compared to English, LVCs are more common in Hindi, have different characteristics and are used as a preferred method for introducing new predicates into the language (Vaidya et al., 2016). A state-of-the-art approach (Chen et al., 2015) propose a supervised approach to identify LVCs in English using resources like PropBank, the OntoNotes sense groupings, Word-Net and the British National Corpus. However, it is difficult to extend this approach for Hindi as it would require extensive efforts to create labelled data based on such resources for Hindi.
III. Annotation of nominal events: An event is a nominal event if it is described by a noun. Annotation of nominal events is a challenging task, as eventiveness of a noun depends on the context in which it is used. In this paper, we restrict our focus to events pivoted around verbs in the past tense as interactions. We plan to explore nominal events in the future.

Proposed Annotation Guidelines and Approach
As highlighted earlier, an MSC consists of a set of actors and interactions among them. The proposed approach is accordingly divided into two major stages: (1) Extraction of actors and their coreference resolution; (2) Extraction of interactions with their sender and receiver arguments.

Actor Extraction and Coreference Resolution
Since we are focusing on narratives of historical events, we assume the actors to be entities of type PERSON, LOCATION and ORGANIZA-TION. Further, we are interested in identifying all types of actor mentions -proper nouns, common nouns and pronouns. To extract proper noun mentions, we perform Named Entity Recognition of the text and mark all noun phrases of type PER-SON, LOCATION and ORGANIZATION as actors. Since Hindi is a relatively resource-poor language and it is very costly to generate necessary labeled data to train a supervised model, we propose an unsupervised technique to identify actors based on Wordnet. Given a noun which is the head of a noun phrase, we query its hypernym hierarchy to check if specific senses of (sthaan; place), (kshetra; region), -(bhu-bhaag; geographical area), (vyakti; person), (maanav; person) and Ú (samudaay; group) are found. If so, we tag the type of the respective noun phrase corresponding to the hypernym found. As pronouns are a closed set of words, we use a manually prepared list of pronouns and corresponding types to identify all pronoun mentions of the actors. It is ensured that the list does not include any demonstrative pronouns (such as , , u , u ). Each pronoun in the text is checked against the list and marked as an actor, if found.
As the actor mentions are to be visualized on a MSC, it is not sufficient to depict only head words of actor phrases. For example, Ú (Sardar Patel is more informative as well as describes the complete entity than just the headword (Patel). Hence, we propose a simple dependency parse based approach to identify the complete noun phrase given the head word. We append to the head word all nouns which are dependent on it through the compound dependency. For example, the word (Mahatma) is dependent on the head word (Gandhi) through the compound dependency and hence, the complete phrase becomes (Mahatma Gandhi). We also append any adjectives/quantifiers which modify the head word using the amod dependency. For example, (pratham pradhanmantri; first prime minister) is formed as a complete phrase where (pratham; first) modifies (pradhanmantri; prime minister) using the amod dependency. A complex construct observed is a noun dependent on the head word with a nmod dependency with no verbs or prepositions occurring between the two. Such nouns are also considered part of the complete phrase and are handled through this approach. For example, in the phrase (swatantrata sangram senani; freedom fighter), the word (swatantrata; independence) is dependent on the head word (senani; soldier) through the nmod dependency with only a noun (sangram; struggle) appearing in between.
For coreference resolution, we employ the technique proposed in (Ramrakhiyani et al., 2018). The authors assume gold actor mentions as input and predict coreferences between the actor mentions. They apply a set of linguistically motivated rules coded in a Markov Logic Network (MLN) based framework to perform the coreference resolution. In this paper, we develop an MLN with the stated rules and input it the extracted actor mentions to perform the coreference resolution. We also annotate the first observed mention in each mention cluster as the canonical actor mention for that coreference group. The canonical actor mention is used for its depiction on the output MSC.

Interaction Extraction
An interaction in the MSC notation is composed of the central event and its sender and receiver actor arguments. To extract interactions it hence becomes important to identify the meaningful events and their actor arguments. In this paper, we focus on events which are primarily described by verbs. We define such verbal events based on linguistic properties of Hindi verbs (Vaidya, 2015). We also propose a set of annotation guidelines to mark such verbal events and their arguments.

Verbal Events
A verbal event can be of two types: 1) simple predicate where a verb triggers the event. For example, in the sentence, a (Ram ne aam khaya; Ram ate a mango.), the verb is a simple predicate, 2) complex predicate where an event is triggered by a verb and an additional element of type verb or adjective or adverb or borrowed English verb or noun (Vaidya, 2015). A complex predicate further has two important types: 1) conjunct predicate where the additional element is of type noun. For example, in the sentence, 2017 a (2017 me Ram ka janma hua; Ram was born in 2017.) the verbal predicate a triggers an event with the additional noun element , 2) compound predicate where the additional element is of type verb. For example, in the sentence, º Ú (Ram ne kitab kho di; Ram lost the book.), the verbal predicate Ú triggers an event with the additional verb element .

Annotation of Events
The proposed annotation guidelines for marking verbal events are described as follows: I. In this paper, we restrict our scope to events represented using verbs. For example, in the following sentence (mahotsav; festival) is not annotated as an event: Ú a (film mahotsav ke dauran purane abhineta padhare; senior actors arrived during the film festival) II. The head verb of an event predicate is tagged as PIVOT. If an event predicate is a conjunct predicate, its noun element is tagged as P-CONJ. In case of a compound predicate, the verb element of the head verb is tagged as P-COMP. IV. Only punctual events are annotated as events. An event is punctual if it "does not have a transitional phase between its start and end point" (Kay and Aylett, 1996). This implies that a process in continuation is not punctual and hence not marked as a valid event.
• Punctual event: a (kisano ne angrejh sarkar se bhari kar me chut ki mang ki; The farmers demanded the British government, a waiver in the heavy taxes) • Non-punctual event: Ú Ú u Ú Ê (gujrat ka kheda khand un dino sukhe ki • Verbs appearing in the authors' opinions are not considered valid events. For example, the verbs in the quoted part e (jayenge;will go) and ( hai) of the following sentence are ignored: Ú i " e , Ú º " (sardar patel ne kewal itna kaha "kya hum goa jayenge, kewal do ghante ki baat hai"; Sardar Patel only said "Will we go to Goa, its a matter of only two hours".)

Annotation of Event Arguments
We consider an actor who initiates or triggers an event as the sender of the event. All other actors that participate in the event are considered as receivers of the event. For example, in the sentence Ú (ram ne raju ko london bheja; Ram sent Raju to London), is the sender of the event , while and Ú are its receivers.

Approach to Identify Verbal Events
In this paper, we propose an approach to identify the above described verbal events and their arguments. Our proposed approach makes use of POS tagging and dependency parsing. Algorithm 1 shows our approach to identify event predicates and their compound and conjunct predicates (if available) from a sentence. Figure 2 shows an illustration of our approach on an example sentence. The verb e (hue; became) is the root of the dependency tree of the sentence. We can determine the tense of e based on its child (thhe) with the dependency relation aux. As e is the root of the sentence with past tense, we identify it as the PIVOT of an event.
In the dependency tree of this sentence, the word (naraj; unhappy) is a noun child of e with dependency relation compound and hence, we identify as P-CONJ of e. Similarly, we identify as its P-COMP. In order to filter out an irrealis or non-punctual event, we check its respective P-COMP against a manually curated list of verbs such as , , , , Ê , Ê , , indicating either continuity or uncertainty of events. We also ignore events appearing in quotes as they are likely to be authors' opinions.

Approach to Identify Event Arguments
Algorithm 2 discusses our event argument identification approach leading to formation of the interactions to be shown on an MSC. Figure 3 shows an illustration of our approach on an example sentence. For the event pivot (likha; wrote), words (Gandhi) and (Patel) are its dependency children with relations nsubj and dobj respectively. They are also present in the list of extracted actors and hence are designated as sender and receiver of this event.
In Figure 4, the word (chhuut; waiver) is dependent on the P-CONJ -(maang; demand) through the dependency relation nmod, but is not a valid actor and hence not a valid receiver.

Datasets
We carry out our experiments on the four text narratives from Indian History, contributed by (Ramrakhiyani et al., 2018). We obtain the dataset text and gold actor annotations and we carry out the event annotations for these datasets with the help of three annotators. The statistics about the datasets are described in Table 2.

Experimental Setup
We used the Google SyntaxNet 2 for dependency parsing the Hindi sentences. SyntaxNet is a Ten-sorFlow toolkit for deep learning powered natural language understanding developed at Google. The Parsey Universal component of SyntaxNet supports NLP preprocessing tasks such as POS tagging, morphological analysis and dependency parsing for 40 different languages including Hindi. We employed two different NER approaches proposed for Hindi and consider a word as part of a named entity if either or both of them identify it as a named entity. One of the approaches is Polyglot, proposed in (Al-Rfou et al., 2013). It is based on using language agnostic techniques involving Wikipedia and Freebase and no human annotated NER training data. The second approach is proposed by (Murthy et al., 2019) and is based on a supervised deep learning architecture for NER in Hindi. To access the Hindi WordNet (Narayan et al., 2002), we use the pyiwn toolkit (Panjwani et al., 2018) which is a python API to access Indian language WordNets.
We use the Sardar dataset as a training set to iteratively revise and improve the extraction algorithms while keeping the other datasets unseen.

Evaluation and Results
The approach comprises of multiple facets like actor identification, actor coreference resolution, event extraction and event argument finding. To assess each of these facets, we carry out evaluation of the proposed approach at multiple levels.
As the first level, we check the performance of actor identification and coreference resolution. If an actor predicted by the approach is present in the gold standard, it is marked as a true positive. False positives and false negatives are computed accordingly. We report the F1 scores for actor mention identification for each dataset in Table 3. At this level, we also report the MUC (Vilain et al., 1995), the B 3 (Bagga and Baldwin, 1998)  CEAFe (Luo, 2005) scores to measure the coreference resolution performance.  At the next level, we check the correctness of events which are basis for interactions. If the predicted pivot is present in the gold standard, it is counted as a true positive and the false positives and false negatives are computed accordingly. In a similar way, true positives, false positives and false negatives are computed for P-CONJ and P-COMP parts. We report F1 scores for each of PIVOT, P-CONJ and P-COMP separately and the three combined i.e. the complete event, in Table 4.
At the final level, we check correctness of the extracted interactions which involves checking the correctness of both the event and its actor arguments. In Table 5, we report F1 scores for three settings namely Sender+Pivot (combination of both event pivot and corresponding sender should be correct), Receiver+Pivot (combination of both event pivot and corresponding receiver should be correct), Complete interaction (complete combination of event pivot, corresponding sender and corresponding receiver should be correct).

Analysis
From the low to moderate results, it can be observed that each stage of the approach is chal-lenging. It is important to note here that this is a pipeline based approach and errors propagate from one stage to the next. A considerable number of errors is also attributed to the performance of linguistic tools such as SyntaxNet and Hindi NER.
We now discuss some error cases for the actor extraction stage. In the sentence i Ê a Ú a Ú Ú Ú (iske chalte kai avsaro par dono ne hi apne pad ka tyaag karne ki dhamki de di thi; Meanwhile, on multiple occasions both had threatened to resign from their posts), the gold actor Ú (dono; both) is a noun but is not tagged as a valid actor. This is because none of the hypernyms of Ú carry a PERSON sense and hence, our WordNet based approach fails to identify Ú as a valid actor. This example highlights the need of a more richer knowledge resource that can disambiguate words like Ú as an actor from its quantifier sense. Another error case is regarding pronouns like i (isko; this/him/her) which can be used to refer to events or actors. Pronouns like these are present in our pronoun list because of their actor referring property, but at places in the text when it appears referring a non-actor it is still considered as an actor leading to a false positive.
We also discuss error cases for the interaction extraction stage. In the sentence Ú u º a a a Ú º Ú (landan jakar unhone baristri ki padhai ki aur wapas aakar ahmedabad me vakalat karne lage; He went to London, studied law, came back and started practising law in Ahmedabad.), the PIVOT a (aakar; came) does not find the right sender argument u (unhone; he) using the dependency   Table 5: F1 scores -Interaction Evaluation Figure 5: Predicted MSC of the example sentence " Ú u º a a a Ú º Ú " (He went to London, studied law, came back and started practising law in Ahmedabad.) parse based argument finding approach. This highlights the importance of more robust approach to identify arguments of events.
As part of the analysis, we show the visual MSC based depiction of the earlier sentence Ú u º a a a Ú º Ú in Figure 5. As pointed out, the event a does not form a valid interaction because of absence of correct arguments.

Related Work
Automatic extraction of MSCs from narratives is studied for English language. Recently, (Palshikar et al., 2019) proposed a semantic role labelling and dependency parsing based approach to extract messages discussed in a narrative. They further use the Document Level Time-anchors algorithm for temporal ordering of extracted messages.
Several researchers such as (Rao and Devi, 2017, 2018) have focused on extraction of events from Hindi. However, these works are focused on news and social media texts. (Kuila et al., 2018) propose a neural network based approach for event extraction. However, the approach is supervised and needs labelled data. On the other hand, our approach does not need labelled data.

Conclusions
In this paper we proposed the use of a knowledge representation known as Message Sequence Charts (MSC) for visualizing Hindi narratives by identifying and depicting the multiple actors and interactions involved. As per our knowledge this is the first attempt to visualize Hindi narratives and represent them formally. We extend a set of annotation guidelines for marking events and their arguments pivoted on verbs. We also propose linguistic knowledge based approach for actor and interaction identification. We report results on four state-of-the-art Hindi narrative datasets and present a brief analysis of the approaches. As future work, we would like to extend this work on lines of Semantic Role Labelling (SRL) for Hindi including nominal predicates and more arguments. We also aim to work on the temporal ordering aspect of Hindi narratives which would allow us to depict interactions in the true chronological order.